nspawn: allocate a new pty instead of passing ours through to avoid terminal settings chaos

1 files changed, 399 insertions, 51 deletions

diff --git a/src/nspawn.c b/src/nspawn.c
index 94e03b58dc..f8a190097a 100644
--- a/src/nspawn.c
+++ b/src/nspawn.c
@@ -33,12 +33,16 @@
 #include <sys/prctl.h>
 #include <sys/capability.h>
 #include <getopt.h>
+#include <sys/epoll.h>
+#include <termios.h>
+#include <sys/signalfd.h>
 
 #include "log.h"
 #include "util.h"
 #include "missing.h"
 #include "cgroup-util.h"
 #include "sd-daemon.h"
+#include "strv.h"
 
 static char *arg_directory = NULL;
 
@@ -166,7 +170,7 @@ static int mount_all(const char *dest) {
         return r;
 }
 
-static int copy_devnodes(const char *dest) {
+static int copy_devnodes(const char *dest, const char *console) {
 
         static const char devnodes[] =
                 "null\0"
@@ -181,15 +185,17 @@ static int copy_devnodes(const char *dest) {
 
         const char *d;
         int r = 0, k;
-        char *tty = NULL;
-        dev_t tty_devnum;
         mode_t u;
+        struct stat st;
+        char *from = NULL, *to = NULL;
+
+        assert(dest);
+        assert(console);
 
         u = umask(0000);
 
         NULSTR_FOREACH(d, devnodes) {
-                char *from = NULL, *to = NULL;
-                struct stat st;
+                from = to = NULL;
 
                 asprintf(&from, "/dev/%s", d);
                 asprintf(&to, "%s/dev/%s", dest, d);
@@ -200,6 +206,8 @@ static int copy_devnodes(const char *dest) {
                         free(from);
                         free(to);
 
+                        from = to = NULL;
+
                         if (r == 0)
                                 r = -ENOMEM;
 
@@ -210,70 +218,80 @@ static int copy_devnodes(const char *dest) {
 
                         if (errno != ENOENT) {
                                 log_error("Failed to stat %s: %m", from);
-
                                 if (r == 0)
                                         r = -errno;
                         }
 
-                } else {
-                        if (mknod(to, st.st_mode, st.st_rdev) < 0) {
-                                log_error("mknod(%s) failed: %m", dest);
+                } else if (!S_ISCHR(st.st_mode) && !S_ISBLK(st.st_mode)) {
 
-                                if (r == 0)
-                                        r = -errno;
-                        }
+                        log_error("%s is not a char or block device, cannot copy.", from);
+                        if (r == 0)
+                                r = -EIO;
+
+                } else if (mknod(to, st.st_mode, st.st_rdev) < 0) {
+
+                        log_error("mknod(%s) failed: %m", dest);
+                        if (r == 0)
+                                r = -errno;
                 }
 
                 free(from);
                 free(to);
         }
 
-        if ((k = get_ctty(&tty, &tty_devnum)) < 0) {
-                log_error("Failed to determine controlling tty: %s", strerror(-k));
+        if (stat(console, &st) < 0) {
 
+                log_error("Failed to stat %s: %m", console);
                 if (r == 0)
-                        r = k;
-        } else {
-                char *from = NULL, *to = NULL;
+                        r = -errno;
 
-                asprintf(&from, "/dev/%s", tty);
-                asprintf(&to, "%s/dev/console", dest);
+                goto finish;
 
-                if (!from || !to) {
-                        log_error("Out of memory");
+        } else if (!S_ISCHR(st.st_mode)) {
 
-                        if (r == 0)
-                                r = k;
-                } else {
-                        /* We need to bind mount our own tty on
-                         * /dev/console, since ptys cannot be used
-                         * unless on a devpts file system. But to bind
-                         * mount it we first have to create a device
-                         * node where we can bind mount it on. This is
-                         * kinda ugly since the TTY will very likely
-                         * be owned by a user/group that does not
-                         * exist in the container. */
-
-                        if (mknod(to, S_IFCHR|0600, tty_devnum) < 0) {
-                                log_error("mknod for /dev/console failed: %m");
+                log_error("/dev/console is not a char device.");
+                if (r == 0)
+                        r = -EIO;
 
-                                if (r == 0)
-                                        r = -errno;
-                        }
+                goto finish;
+        }
 
-                        if (mount(from, to, "bind", MS_BIND|MS_RDONLY, NULL) < 0) {
-                                log_error("bind mount for /dev/console failed: %m");
+        if (asprintf(&to, "%s/dev/console", dest) < 0) {
 
-                                if (r == 0)
-                                        r = -errno;
-                        }
-                }
+                log_error("Out of memory");
+                if (r == 0)
+                        r = -ENOMEM;
 
-                free(from);
-                free(to);
+                 goto finish;
         }
 
-        free(tty);
+        /* We need to bind mount the right tty to /dev/console since
+         * ptys can only exist on pts file systems. To have something
+         * to bind mount things on we create a device node first, that
+         * has the right major/minor (note that the major minor
+         * doesn't actually matter here, since we mount it over
+         * anyway). */
+
+        if (mknod(to, (st.st_mode & ~07777) | 0600, st.st_rdev) < 0)
+                log_error("mknod for /dev/console failed: %m");
+
+        if (mount(console, to, "bind", MS_BIND, NULL) < 0) {
+                log_error("bind mount for /dev/console failed: %m");
+
+                if (r == 0)
+                        r = -errno;
+        }
+
+        free(to);
+
+        if ((k = chmod_and_chown(console, 0600, 0, 0)) < 0) {
+                log_error("Failed to correct access mode for TTY: %s", strerror(-k));
+
+                if (r == 0)
+                        r = k;
+        }
+
+finish:
 
         umask(u);
 
@@ -346,11 +364,259 @@ static int is_os_tree(const char *path) {
         return r < 0 ? 0 : 1;
 }
 
+#define BUFFER_SIZE 1024
+
+static int process_pty(int master, sigset_t *mask) {
+        char in_buffer[BUFFER_SIZE], out_buffer[BUFFER_SIZE];
+        size_t in_buffer_full = 0, out_buffer_full = 0;
+        struct epoll_event stdin_ev, stdout_ev, master_ev, signal_ev;
+        bool stdin_readable = false, stdout_writable = false, master_readable = false, master_writable = false;
+        bool stdin_rhup = false, stdout_whup = false, master_rhup = false, master_whup = false;
+        int ep = -1, signal_fd = -1, r;
+
+        fd_nonblock(STDIN_FILENO, 1);
+        fd_nonblock(STDOUT_FILENO, 1);
+        fd_nonblock(master, 1);
+
+        if ((signal_fd = signalfd(-1, mask, SFD_NONBLOCK|SFD_CLOEXEC)) < 0) {
+                log_error("signalfd(): %m");
+                r = -errno;
+                goto finish;
+        }
+
+        if ((ep = epoll_create1(EPOLL_CLOEXEC)) < 0) {
+                log_error("Failed to create epoll: %m");
+                r = -errno;
+                goto finish;
+        }
+
+        zero(stdin_ev);
+        stdin_ev.events = EPOLLIN|EPOLLET;
+        stdin_ev.data.fd = STDIN_FILENO;
+
+        zero(stdout_ev);
+        stdout_ev.events = EPOLLOUT|EPOLLET;
+        stdout_ev.data.fd = STDOUT_FILENO;
+
+        zero(master_ev);
+        master_ev.events = EPOLLIN|EPOLLOUT|EPOLLET;
+        master_ev.data.fd = master;
+
+        zero(signal_ev);
+        signal_ev.events = EPOLLIN;
+        signal_ev.data.fd = signal_fd;
+
+        if (epoll_ctl(ep, EPOLL_CTL_ADD, STDIN_FILENO, &stdin_ev) < 0 ||
+            epoll_ctl(ep, EPOLL_CTL_ADD, STDOUT_FILENO, &stdout_ev) < 0 ||
+            epoll_ctl(ep, EPOLL_CTL_ADD, master, &master_ev) < 0 ||
+            epoll_ctl(ep, EPOLL_CTL_ADD, signal_fd, &signal_ev) < 0) {
+                log_error("Failed to regiser fds in epoll: %m");
+                r = -errno;
+                goto finish;
+        }
+
+        do {
+                struct epoll_event ev[16];
+                ssize_t k;
+                int i, nfds;
+
+                if ((nfds = epoll_wait(ep, ev, ELEMENTSOF(ev), -1)) < 0) {
+
+                        if (errno == EINTR || errno == EAGAIN)
+                                continue;
+
+                        log_error("epoll_wait(): %m");
+                        r = -errno;
+                        goto finish;
+                }
+
+                assert(nfds >= 1);
+
+                for (i = 0; i < nfds; i++) {
+                        if (ev[i].data.fd == STDIN_FILENO) {
+
+                                if (!stdin_rhup && (ev[i].events & (EPOLLHUP|EPOLLIN)))
+                                        stdin_readable = true;
+
+                        } else if (ev[i].data.fd == STDOUT_FILENO) {
+
+                                if (ev[i].events & EPOLLHUP) {
+                                        stdout_writable = false;
+                                        stdout_whup = true;
+                                }
+
+                                if (!stdout_whup && (ev[i].events & EPOLLOUT))
+                                        stdout_writable = true;
+
+                        } else if (ev[i].data.fd == master) {
+
+                                /* We don't connect EPOLLHUP to
+                                 * master_whup here, since EPOLLHUP
+                                 * can happen when noone has opened
+                                 * the other side */
+
+                                if (!master_rhup && (ev[i].events & (EPOLLHUP|EPOLLIN)))
+                                        master_readable = true;
+
+                                if (!master_whup && (ev[i].events & EPOLLOUT))
+                                        master_writable = true;
+
+                        } else if (ev[i].data.fd == signal_fd) {
+                                struct signalfd_siginfo sfsi;
+                                ssize_t n;
+
+                                if ((n = read(signal_fd, &sfsi, sizeof(sfsi))) != sizeof(sfsi)) {
+
+                                        if (n >= 0) {
+                                                r = -EIO;
+                                                goto finish;
+                                        }
+
+                                        if (errno != EINTR && errno != EAGAIN) {
+                                                r = -errno;
+                                                goto finish;
+                                        }
+                                } else {
+
+                                        if (sfsi.ssi_signo == SIGWINCH) {
+                                                struct winsize ws;
+
+                                                /* The window size changed, let's forward that. */
+
+                                                if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
+                                                        ioctl(master, TIOCSWINSZ, &ws);
+                                        } else {
+                                                r = -EINTR;
+                                                goto finish;
+                                        }
+                                }
+                        }
+                }
+
+                while ((stdin_readable && in_buffer_full <= 0) ||
+                       (master_writable && in_buffer_full > 0) ||
+                       (master_readable && out_buffer_full <= 0) ||
+                       (stdout_writable && out_buffer_full > 0)) {
+
+                        if (stdin_readable && in_buffer_full < BUFFER_SIZE) {
+
+                                if ((k = read(STDIN_FILENO, in_buffer + in_buffer_full, BUFFER_SIZE - in_buffer_full)) < 0) {
+
+                                        if (errno == EAGAIN)
+                                                stdin_readable = false;
+                                        else if (errno == EPIPE || errno == ECONNRESET || errno == EIO)
+                                                k = 0;
+                                        else {
+                                                log_error("read(): %m");
+                                                goto finish;
+                                        }
+                                } else
+                                        in_buffer_full += (size_t) k;
+
+                                if (k == 0) {
+                                        stdin_rhup = true;
+                                        stdin_readable = false;
+                                        shutdown(STDIN_FILENO, SHUT_RD);
+                                }
+                        }
+
+                        if (master_writable && in_buffer_full > 0) {
+
+                                if ((k = write(master, in_buffer, in_buffer_full)) < 0) {
+
+                                        if (errno == EAGAIN)
+                                                master_writable = false;
+                                        else if (errno == EPIPE || errno == ECONNRESET || errno == EIO) {
+                                                master_whup = true;
+                                                master_writable = false;
+                                        } else {
+                                                log_error("write(): %m");
+                                                goto finish;
+                                        }
+
+                                } else {
+                                        assert(in_buffer_full >= (size_t) k);
+                                        memmove(in_buffer, in_buffer + k, in_buffer_full - k);
+                                        in_buffer_full -= k;
+                                }
+                        }
+
+                        if (master_readable && out_buffer_full < BUFFER_SIZE) {
+
+                                if ((k = read(master, out_buffer + out_buffer_full, BUFFER_SIZE - out_buffer_full)) < 0) {
+
+                                        if (errno == EAGAIN)
+                                                master_readable = false;
+                                        else if (errno == EPIPE || errno == ECONNRESET || errno == EIO)
+                                                k = 0;
+                                        else {
+                                                log_error("read(): %m");
+                                                goto finish;
+                                        }
+                                }  else
+                                        out_buffer_full += (size_t) k;
+
+                                if (k == 0) {
+                                        master_rhup = true;
+                                        master_readable = false;
+                                }
+                        }
+
+                        if (stdout_writable && out_buffer_full > 0) {
+
+                                if ((k = write(STDOUT_FILENO, out_buffer, out_buffer_full)) < 0) {
+
+                                        if (errno == EAGAIN)
+                                                stdout_writable = false;
+                                        else if (errno == EPIPE || errno == ECONNRESET || errno == EIO) {
+                                                stdout_whup = true;
+                                                stdout_writable = false;
+                                        } else {
+                                                log_error("write(): %m");
+                                                goto finish;
+                                        }
+
+                                } else {
+                                        assert(out_buffer_full >= (size_t) k);
+                                        memmove(out_buffer, out_buffer + k, out_buffer_full - k);
+                                        out_buffer_full -= k;
+                                }
+                        }
+                }
+
+                if (stdin_rhup && in_buffer_full <= 0 && !master_whup) {
+                        master_whup = true;
+                        master_writable = false;
+                }
+
+                if (master_rhup && out_buffer_full <= 0 && !stdout_whup) {
+                        stdout_whup = true;
+                        stdout_writable = false;
+                        shutdown(STDOUT_FILENO, SHUT_WR);
+                }
+
+        } while (!stdout_whup || !master_whup);
+
+finish:
+        if (ep >= 0)
+                close_nointr_nofail(ep);
+
+        if (signal_fd >= 0)
+                close_nointr_nofail(signal_fd);
+
+        return r;
+}
 
 int main(int argc, char *argv[]) {
         pid_t pid = 0;
         int r = EXIT_FAILURE, k;
         char *oldcg = NULL, *newcg = NULL;
+        int master = -1;
+        const char *console = NULL;
+        struct termios saved_attr, raw_attr;
+        sigset_t mask;
+        bool saved_attr_valid = false;
+        struct winsize ws;
 
         log_parse_environment();
         log_open();
@@ -394,8 +660,6 @@ int main(int argc, char *argv[]) {
                 goto finish;
         }
 
-        log_info("Spawning namespace container on %s.", arg_directory);
-
         if ((k = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 0, &oldcg)) < 0) {
                 log_error("Failed to determine current cgroup: %s", strerror(-k));
                 goto finish;
@@ -411,31 +675,99 @@ int main(int argc, char *argv[]) {
                 goto finish;
         }
 
+        if ((master = posix_openpt(O_RDWR|O_NOCTTY|O_CLOEXEC|O_NDELAY)) < 0) {
+                log_error("Failed to acquire pseudo tty: %m");
+                goto finish;
+        }
+
+        if (!(console = ptsname(master))) {
+                log_error("Failed to determine tty name: %m");
+                goto finish;
+        }
+
+        log_info("Spawning namespace container on %s (console is %s).", arg_directory, console);
+
+        if (ioctl(STDIN_FILENO, TIOCGWINSZ, &ws) >= 0)
+                ioctl(master, TIOCSWINSZ, &ws);
+
+        if (unlockpt(master) < 0) {
+                log_error("Failed to unlock tty: %m");
+                goto finish;
+        }
+
+        if (tcgetattr(STDIN_FILENO, &saved_attr) < 0) {
+                log_error("Failed to get terminal attributes: %m");
+                goto finish;
+        }
+
+        saved_attr_valid = true;
+
+        raw_attr = saved_attr;
+        cfmakeraw(&raw_attr);
+        raw_attr.c_lflag &= ~ECHO;
+
+        if (tcsetattr(STDIN_FILENO, TCSANOW, &raw_attr) < 0) {
+                log_error("Failed to set terminal attributes: %m");
+                goto finish;
+        }
+
+        assert_se(sigemptyset(&mask) == 0);
+        sigset_add_many(&mask, SIGCHLD, SIGWINCH, SIGTERM, SIGINT, -1);
+        assert_se(sigprocmask(SIG_BLOCK, &mask, NULL) == 0);
+
         if ((pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS, NULL)) < 0) {
                 log_error("clone() failed: %m");
                 goto finish;
         }
 
         if (pid == 0) {
+                /* child */
+
                 const char *hn;
                 const char *envp[] = {
                         "HOME=/root",
                         "PATH=/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin",
+                        NULL,
                         NULL
                 };
 
-                /* child */
+                envp[2] = strv_find_prefix(environ, "TERM=");
+
+                close_nointr_nofail(master);
+
+                close_nointr(STDIN_FILENO);
+                close_nointr(STDOUT_FILENO);
+                close_nointr(STDERR_FILENO);
+
+                close_all_fds(NULL, 0);
+
+                reset_all_signal_handlers();
+
+                assert_se(sigemptyset(&mask) == 0);
+                assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
+
+                if (setsid() < 0)
+                        goto child_fail;
+
+                if (prctl(PR_SET_PDEATHSIG, SIGKILL) < 0)
+                        goto child_fail;
 
                 if (mount_all(arg_directory) < 0)
                         goto child_fail;
 
-                if (copy_devnodes(arg_directory) < 0)
+                if (copy_devnodes(arg_directory, console) < 0)
                         goto child_fail;
 
                 if (chdir(arg_directory) < 0) {
                         log_error("chdir(%s) failed: %m", arg_directory);
                         goto child_fail;
                 }
+
+                if (open_terminal("dev/console", O_RDWR) != STDIN_FILENO ||
+                    dup2(STDIN_FILENO, STDOUT_FILENO) != STDOUT_FILENO ||
+                    dup2(STDIN_FILENO, STDERR_FILENO) != STDERR_FILENO)
+                        goto child_fail;
+
                 if (mount(arg_directory, "/", "bind", MS_BIND|MS_MOVE, NULL) < 0) {
                         log_error("mount(MS_MOVE) failed: %m");
                         goto child_fail;
@@ -451,6 +783,8 @@ int main(int argc, char *argv[]) {
                         goto child_fail;
                 }
 
+                umask(0002);
+
                 if (drop_capabilities() < 0)
                         goto child_fail;
 
@@ -470,12 +804,26 @@ int main(int argc, char *argv[]) {
                 _exit(EXIT_FAILURE);
         }
 
+        if (process_pty(master, &mask) < 0)
+                goto finish;
+
+        if (saved_attr_valid) {
+                tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
+                saved_attr_valid = false;
+        }
+
         r = wait_for_terminate_and_warn(argc > optind ? argv[optind] : "bash", pid);
 
         if (r < 0)
                 r = EXIT_FAILURE;
 
 finish:
+        if (saved_attr_valid)
+                tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
+
+        if (master >= 0)
+                close_nointr_nofail(master);
+
         if (oldcg)
                 cg_attach(SYSTEMD_CGROUP_CONTROLLER, oldcg, 0);