chiark / gitweb /
clients: unify how we invoke getopt_long()
[elogind.git] / src / nspawn / nspawn.c
index a0809da7439cd730833137a2e0fee66a33cf3ecd..34a0dafa645803024956b62d016debfc17fe95e0 100644 (file)
@@ -33,8 +33,6 @@
 #include <sys/prctl.h>
 #include <sys/capability.h>
 #include <getopt.h>
-#include <sys/poll.h>
-#include <sys/epoll.h>
 #include <termios.h>
 #include <sys/signalfd.h>
 #include <grp.h>
@@ -43,9 +41,9 @@
 #include <sys/socket.h>
 #include <linux/netlink.h>
 
-#include <systemd/sd-daemon.h>
-#include <systemd/sd-bus.h>
-
+#include "sd-daemon.h"
+#include "sd-bus.h"
+#include "sd-id128.h"
 #include "log.h"
 #include "util.h"
 #include "mkdir.h"
 #include "strv.h"
 #include "path-util.h"
 #include "loopback-setup.h"
-#include "sd-id128.h"
 #include "dev-setup.h"
 #include "fdset.h"
 #include "build.h"
 #include "fileio.h"
-#include "bus-internal.h"
-#include "bus-message.h"
+#include "bus-util.h"
+#include "bus-error.h"
+#include "ptyfwd.h"
 
 #ifndef TTY_GID
 #define TTY_GID 5
@@ -167,7 +165,7 @@ static int parse_argv(int argc, char *argv[]) {
                 { "bind-ro",         required_argument, NULL, ARG_BIND_RO         },
                 { "machine",         required_argument, NULL, 'M'                 },
                 { "slice",           required_argument, NULL, 'S'                 },
-                { NULL,              0,                 NULL, 0                   }
+                {}
         };
 
         int c, r;
@@ -180,8 +178,7 @@ static int parse_argv(int argc, char *argv[]) {
                 switch (c) {
 
                 case 'h':
-                        help();
-                        return 0;
+                        return help();
 
                 case ARG_VERSION:
                         puts(PACKAGE_STRING);
@@ -224,6 +221,9 @@ static int parse_argv(int argc, char *argv[]) {
 
                 case 'S':
                         arg_slice = strdup(optarg);
+                        if (!arg_slice)
+                                return log_oom();
+
                         break;
 
                 case 'M':
@@ -315,11 +315,11 @@ static int parse_argv(int argc, char *argv[]) {
 
                         r = strv_extend(x, a);
                         if (r < 0)
-                                return r;
+                                return log_oom();
 
                         r = strv_extend(x, b);
                         if (r < 0)
-                                return r;
+                                return log_oom();
 
                         break;
                 }
@@ -328,8 +328,7 @@ static int parse_argv(int argc, char *argv[]) {
                         return -EINVAL;
 
                 default:
-                        log_error("Unknown option code %c", c);
-                        return -EINVAL;
+                        assert_not_reached("Unhandled option");
                 }
         }
 
@@ -424,7 +423,7 @@ static int mount_binds(const char *dest, char **l, unsigned long flags) {
 
                 if (stat(where, &dest_st) == 0) {
                         if ((source_st.st_mode & S_IFMT) != (dest_st.st_mode & S_IFMT)) {
-                                log_error("The file types of %s and %s do not matching. Refusing bind mount",
+                                log_error("The file types of %s and %s do not match. Refusing bind mount",
                                                 *x, where);
                                 return -EINVAL;
                         }
@@ -521,7 +520,6 @@ static int setup_timezone(const char *dest) {
 
 static int setup_resolv_conf(const char *dest) {
         char _cleanup_free_ *where = NULL;
-        _cleanup_close_ int fd = -1;
 
         assert(dest);
 
@@ -533,18 +531,9 @@ static int setup_resolv_conf(const char *dest) {
         if (!where)
                 return log_oom();
 
-        fd = open(where, O_WRONLY|O_CREAT|O_EXCL|O_CLOEXEC|O_NOCTTY|O_NOFOLLOW, 0644);
-
         /* We don't really care for the results of this really. If it
          * fails, it fails, but meh... */
-        if (mount("/etc/resolv.conf", where, "bind", MS_BIND, NULL) < 0)
-                log_warning("Failed to bind mount /etc/resolv.conf: %m");
-        else
-                if (mount("/etc/resolv.conf", where, "bind",
-                          MS_BIND|MS_REMOUNT|MS_RDONLY, NULL) < 0) {
-                        log_error("Failed to remount /etc/resolv.conf readonly: %m");
-                        return -errno;
-                }
+        copy_file("/etc/resolv.conf", where, O_TRUNC|O_NOFOLLOW);
 
         return 0;
 }
@@ -934,248 +923,6 @@ static int drop_capabilities(void) {
         return capability_bounding_set_drop(~arg_retain, false);
 }
 
-static int process_pty(int master, pid_t pid, sigset_t *mask) {
-
-        char in_buffer[LINE_MAX], out_buffer[LINE_MAX];
-        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;
-        int ep = -1, signal_fd = -1, r;
-        bool tried_orderly_shutdown = false;
-
-        assert(master >= 0);
-        assert(pid > 0);
-        assert(mask);
-
-        fd_nonblock(STDIN_FILENO, 1);
-        fd_nonblock(STDOUT_FILENO, 1);
-        fd_nonblock(master, 1);
-
-        signal_fd = signalfd(-1, mask, SFD_NONBLOCK|SFD_CLOEXEC);
-        if (signal_fd < 0) {
-                log_error("signalfd(): %m");
-                r = -errno;
-                goto finish;
-        }
-
-        ep = epoll_create1(EPOLL_CLOEXEC);
-        if (ep < 0) {
-                log_error("Failed to create epoll: %m");
-                r = -errno;
-                goto finish;
-        }
-
-        /* We read from STDIN only if this is actually a TTY,
-         * otherwise we assume non-interactivity. */
-        if (isatty(STDIN_FILENO)) {
-                zero(stdin_ev);
-                stdin_ev.events = EPOLLIN|EPOLLET;
-                stdin_ev.data.fd = STDIN_FILENO;
-
-                if (epoll_ctl(ep, EPOLL_CTL_ADD, STDIN_FILENO, &stdin_ev) < 0) {
-                        log_error("Failed to register STDIN in epoll: %m");
-                        r = -errno;
-                        goto finish;
-                }
-        }
-
-        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, STDOUT_FILENO, &stdout_ev) < 0) {
-                if (errno != EPERM) {
-                        log_error("Failed to register stdout in epoll: %m");
-                        r = -errno;
-                        goto finish;
-                }
-                /* stdout without epoll support. Likely redirected to regular file. */
-                stdout_writable = true;
-        }
-
-        if (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 register fds in epoll: %m");
-                r = -errno;
-                goto finish;
-        }
-
-        for (;;) {
-                struct epoll_event ev[16];
-                ssize_t k;
-                int i, nfds;
-
-                nfds = epoll_wait(ep, ev, ELEMENTSOF(ev), -1);
-                if (nfds < 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 (ev[i].events & (EPOLLIN|EPOLLHUP))
-                                        stdin_readable = true;
-
-                        } else if (ev[i].data.fd == STDOUT_FILENO) {
-
-                                if (ev[i].events & (EPOLLOUT|EPOLLHUP))
-                                        stdout_writable = true;
-
-                        } else if (ev[i].data.fd == master) {
-
-                                if (ev[i].events & (EPOLLIN|EPOLLHUP))
-                                        master_readable = true;
-
-                                if (ev[i].events & (EPOLLOUT|EPOLLHUP))
-                                        master_writable = true;
-
-                        } else if (ev[i].data.fd == signal_fd) {
-                                struct signalfd_siginfo sfsi;
-                                ssize_t n;
-
-                                n = read(signal_fd, &sfsi, sizeof(sfsi));
-                                if (n != sizeof(sfsi)) {
-
-                                        if (n >= 0) {
-                                                log_error("Failed to read from signalfd: invalid block size");
-                                                r = -EIO;
-                                                goto finish;
-                                        }
-
-                                        if (errno != EINTR && errno != EAGAIN) {
-                                                log_error("Failed to read from signalfd: %m");
-                                                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 if (sfsi.ssi_signo == SIGTERM && arg_boot && !tried_orderly_shutdown) {
-
-                                                log_info("Trying to halt container. Send SIGTERM again to trigger immediate termination.");
-
-                                                /* This only works for systemd... */
-                                                tried_orderly_shutdown = true;
-                                                kill(pid, SIGRTMIN+3);
-
-                                        } else {
-                                                r = 0;
-                                                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 < LINE_MAX) {
-
-                                k = read(STDIN_FILENO, in_buffer + in_buffer_full, LINE_MAX - in_buffer_full);
-                                if (k < 0) {
-
-                                        if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
-                                                stdin_readable = false;
-                                        else {
-                                                log_error("read(): %m");
-                                                r = -errno;
-                                                goto finish;
-                                        }
-                                } else
-                                        in_buffer_full += (size_t) k;
-                        }
-
-                        if (master_writable && in_buffer_full > 0) {
-
-                                k = write(master, in_buffer, in_buffer_full);
-                                if (k < 0) {
-
-                                        if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
-                                                master_writable = false;
-                                        else {
-                                                log_error("write(): %m");
-                                                r = -errno;
-                                                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 < LINE_MAX) {
-
-                                k = read(master, out_buffer + out_buffer_full, LINE_MAX - out_buffer_full);
-                                if (k < 0) {
-
-                                        if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
-                                                master_readable = false;
-                                        else {
-                                                log_error("read(): %m");
-                                                r = -errno;
-                                                goto finish;
-                                        }
-                                }  else
-                                        out_buffer_full += (size_t) k;
-                        }
-
-                        if (stdout_writable && out_buffer_full > 0) {
-
-                                k = write(STDOUT_FILENO, out_buffer, out_buffer_full);
-                                if (k < 0) {
-
-                                        if (errno == EAGAIN || errno == EPIPE || errno == ECONNRESET || errno == EIO)
-                                                stdout_writable = false;
-                                        else {
-                                                log_error("write(): %m");
-                                                r = -errno;
-                                                goto finish;
-                                        }
-
-                                } else {
-                                        assert(out_buffer_full >= (size_t) k);
-                                        memmove(out_buffer, out_buffer + k, out_buffer_full - k);
-                                        out_buffer_full -= k;
-                                }
-                        }
-                }
-        }
-
-finish:
-        if (ep >= 0)
-                close_nointr_nofail(ep);
-
-        if (signal_fd >= 0)
-                close_nointr_nofail(signal_fd);
-
-        return r;
-}
-
 static int register_machine(void) {
         _cleanup_bus_error_free_ sd_bus_error error = SD_BUS_ERROR_NULL;
         _cleanup_bus_unref_ sd_bus *bus = NULL;
@@ -1197,20 +944,74 @@ static int register_machine(void) {
                         NULL,
                         "sayssusa(sv)",
                         arg_machine,
-                        SD_BUS_APPEND_ID128(arg_uuid),
+                        SD_BUS_MESSAGE_APPEND_ID128(arg_uuid),
                         "nspawn",
                         "container",
                         (uint32_t) 0,
                         strempty(arg_directory),
-                        1, "Slice", "s", strempty(arg_slice));
+                        !isempty(arg_slice), "Slice", "s", arg_slice);
         if (r < 0) {
-                log_error("Failed to register machine: %s", error.message ? error.message : strerror(-r));
+                log_error("Failed to register machine: %s", bus_error_message(&error, r));
                 return r;
         }
 
         return 0;
 }
 
+static int terminate_machine(pid_t pid) {
+        _cleanup_bus_error_free_ sd_bus_error error = SD_BUS_ERROR_NULL;
+        _cleanup_bus_message_unref_ sd_bus_message *reply = NULL;
+        _cleanup_bus_unref_ sd_bus *bus = NULL;
+        const char *path;
+        int r;
+
+        r = sd_bus_open_system(&bus);
+        if (r < 0) {
+                log_error("Failed to open system bus: %s", strerror(-r));
+                return r;
+        }
+
+        r = sd_bus_call_method(
+                        bus,
+                        "org.freedesktop.machine1",
+                        "/org/freedesktop/machine1",
+                        "org.freedesktop.machine1.Manager",
+                        "GetMachineByPID",
+                        &error,
+                        &reply,
+                        "u",
+                        (uint32_t) pid);
+        if (r < 0) {
+                /* Note that the machine might already have been
+                 * cleaned up automatically, hence don't consider it a
+                 * failure if we cannot get the machine object. */
+                log_debug("Failed to get machine: %s", bus_error_message(&error, r));
+                return 0;
+        }
+
+        r = sd_bus_message_read(reply, "o", &path);
+        if (r < 0) {
+                log_error("Failed to parse GetMachineByPID() reply: %s", bus_error_message(&error, r));
+                return r;
+        }
+
+        r = sd_bus_call_method(
+                        bus,
+                        "org.freedesktop.machine1",
+                        path,
+                        "org.freedesktop.machine1.Machine",
+                        "Terminate",
+                        &error,
+                        NULL,
+                        NULL);
+        if (r < 0) {
+                log_debug("Failed to terminate machine: %s", bus_error_message(&error, r));
+                return 0;
+        }
+
+        return 0;
+}
+
 static bool audit_enabled(void) {
         int fd;
 
@@ -1228,12 +1029,9 @@ int main(int argc, char *argv[]) {
         _cleanup_close_ int master = -1;
         int n_fd_passed;
         const char *console = NULL;
-        struct termios saved_attr, raw_attr;
         sigset_t mask;
-        bool saved_attr_valid = false;
-        struct winsize ws;
-        int kmsg_socket_pair[2] = { -1, -1 };
-        FDSet *fds = NULL;
+        _cleanup_close_pipe_ int kmsg_socket_pair[2] = { -1, -1 };
+        _cleanup_fdset_free_ FDSet *fds = NULL;
 
         log_parse_environment();
         log_open();
@@ -1327,24 +1125,13 @@ int main(int argc, char *argv[]) {
                 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);
+        log_info("Spawning container %s on %s. Press ^] three times within 1s to abort execution.", arg_machine, arg_directory);
 
         if (unlockpt(master) < 0) {
                 log_error("Failed to unlock tty: %m");
                 goto finish;
         }
 
-        if (tcgetattr(STDIN_FILENO, &saved_attr) >= 0) {
-                saved_attr_valid = true;
-
-                raw_attr = saved_attr;
-                cfmakeraw(&raw_attr);
-                raw_attr.c_lflag &= ~ECHO;
-        }
-
         if (socketpair(AF_UNIX, SOCK_DGRAM|SOCK_NONBLOCK|SOCK_CLOEXEC, 0, kmsg_socket_pair) < 0) {
                 log_error("Failed to create kmsg socket pair.");
                 goto finish;
@@ -1358,18 +1145,6 @@ int main(int argc, char *argv[]) {
 
         for (;;) {
                 siginfo_t status;
-                int pipefd[2], pipefd2[2];
-
-                if (pipe2(pipefd, O_NONBLOCK|O_CLOEXEC) < 0) {
-                        log_error("pipe2(): %m");
-                        goto finish;
-                }
-
-                if (pipe2(pipefd2, O_NONBLOCK|O_CLOEXEC) < 0) {
-                        log_error("pipe2(): %m");
-                        close_pipe(pipefd);
-                        goto finish;
-                }
 
                 pid = syscall(__NR_clone, SIGCHLD|CLONE_NEWIPC|CLONE_NEWNS|CLONE_NEWPID|CLONE_NEWUTS|(arg_private_network ? CLONE_NEWNET : 0), NULL);
                 if (pid < 0) {
@@ -1404,21 +1179,9 @@ int main(int argc, char *argv[]) {
                         if (envp[n_env])
                                 n_env ++;
 
-                        /* Wait for the parent process to log our PID */
-                        close_nointr_nofail(pipefd[1]);
-                        fd_wait_for_event(pipefd[0], POLLHUP, -1);
-                        close_nointr_nofail(pipefd[0]);
-
                         close_nointr_nofail(master);
                         master = -1;
 
-                        if (saved_attr_valid) {
-                                if (tcsetattr(STDIN_FILENO, TCSANOW, &raw_attr) < 0) {
-                                        log_error("Failed to set terminal attributes: %m");
-                                        goto child_fail;
-                                }
-                        }
-
                         close_nointr(STDIN_FILENO);
                         close_nointr(STDOUT_FILENO);
                         close_nointr(STDERR_FILENO);
@@ -1458,8 +1221,6 @@ int main(int argc, char *argv[]) {
                                 goto child_fail;
                         }
 
-                        close_pipe(pipefd2);
-
                         r = register_machine();
                         if (r < 0)
                                 goto finish;
@@ -1668,25 +1429,26 @@ int main(int argc, char *argv[]) {
                         _exit(EXIT_FAILURE);
                 }
 
-                log_info("Init process in the container running as PID %lu.", (unsigned long) pid);
-                close_nointr_nofail(pipefd[0]);
-                close_nointr_nofail(pipefd[1]);
-
-                /* Wait for the child process to establish cgroup hierarchy */
-                close_nointr_nofail(pipefd2[1]);
-                fd_wait_for_event(pipefd2[0], POLLHUP, -1);
-                close_nointr_nofail(pipefd2[0]);
-
                 fdset_free(fds);
                 fds = NULL;
 
-                if (process_pty(master, pid, &mask) < 0)
-                        goto finish;
+                k = process_pty(master, &mask, arg_boot ? pid : 0, SIGRTMIN+3);
+                if (k < 0) {
+                        r = EXIT_FAILURE;
+                        break;
+                }
+
+                putc('\n', stdout);
+
+                /* Kill if it is not dead yet anyway */
+                terminate_machine(pid);
 
-                if (saved_attr_valid)
-                        tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
+                /* Redundant, but better safe than sorry */
+                kill(pid, SIGKILL);
 
                 k = wait_for_terminate(pid, &status);
+                pid = 0;
+
                 if (k < 0) {
                         r = EXIT_FAILURE;
                         break;
@@ -1695,47 +1457,40 @@ int main(int argc, char *argv[]) {
                 if (status.si_code == CLD_EXITED) {
                         r = status.si_status;
                         if (status.si_status != 0) {
-                                log_error("Container failed with error code %i.", status.si_status);
+                                log_error("Container %s failed with error code %i.", arg_machine, status.si_status);
                                 break;
                         }
 
-                        log_debug("Container exited successfully.");
+                        log_debug("Container %s exited successfully.", arg_machine);
                         break;
                 } else if (status.si_code == CLD_KILLED &&
                            status.si_status == SIGINT) {
-                        log_info("Container has been shut down.");
+                        log_info("Container %s has been shut down.", arg_machine);
                         r = 0;
                         break;
                 } else if (status.si_code == CLD_KILLED &&
                            status.si_status == SIGHUP) {
-                        log_info("Container is being rebooted.");
+                        log_info("Container %s is being rebooted.", arg_machine);
                         continue;
                 } else if (status.si_code == CLD_KILLED ||
                            status.si_code == CLD_DUMPED) {
 
-                        log_error("Container terminated by signal %s.", signal_to_string(status.si_status));
+                        log_error("Container %s terminated by signal %s.", arg_machine,  signal_to_string(status.si_status));
                         r = EXIT_FAILURE;
                         break;
                 } else {
-                        log_error("Container failed due to unknown reason.");
+                        log_error("Container %s failed due to unknown reason.", arg_machine);
                         r = EXIT_FAILURE;
                         break;
                 }
         }
 
 finish:
-        if (saved_attr_valid)
-                tcsetattr(STDIN_FILENO, TCSANOW, &saved_attr);
-
-        close_pipe(kmsg_socket_pair);
-
         if (pid > 0)
                 kill(pid, SIGKILL);
 
         free(arg_directory);
         free(arg_machine);
 
-        fdset_free(fds);
-
         return r;
 }