src/nspawn/nspawn-stub-pid1.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <sys/reboot.h>
   4 #include <sys/wait.h>
   5 #include <sys/prctl.h>
   6 #include <unistd.h>
   7
   8 #include "fd-util.h"
   9 #include "log.h"
  10 #include "missing.h"
  11 #include "nspawn-stub-pid1.h"
  12 #include "process-util.h"
  13 #include "signal-util.h"
  14 #include "time-util.h"
  15 #include "def.h"
  16
  17 static int reset_environ(const char *new_environment, size_t length) {
  18         unsigned long start, end;
  19
  20         start = (unsigned long) new_environment;
  21         end = start + length;
  22
  23         if (prctl(PR_SET_MM, PR_SET_MM_ENV_START, start, 0, 0) < 0)
  24                 return -errno;
  25
  26         if (prctl(PR_SET_MM, PR_SET_MM_ENV_END, end, 0, 0) < 0)
  27                 return -errno;
  28
  29         return 0;
  30 }
  31
  32 int stub_pid1(sd_id128_t uuid) {
  33         enum {
  34                 STATE_RUNNING,
  35                 STATE_REBOOT,
  36                 STATE_POWEROFF,
  37         } state = STATE_RUNNING;
  38
  39         sigset_t fullmask, oldmask, waitmask;
  40         usec_t quit_usec = USEC_INFINITY;
  41         pid_t pid;
  42         int r;
  43
  44         /* The new environment we set up, on the stack. */
  45         char new_environment[] =
  46                 "container=systemd-nspawn\0"
  47                 "container_uuid=XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX";
  48
  49         /* Implements a stub PID 1, that reaps all processes and processes a couple of standard signals. This is useful
  50          * for allowing arbitrary processes run in a container, and still have all zombies reaped. */
  51
  52         assert_se(sigfillset(&fullmask) >= 0);
  53         assert_se(sigprocmask(SIG_BLOCK, &fullmask, &oldmask) >= 0);
  54
  55         pid = fork();
  56         if (pid < 0)
  57                 return log_error_errno(errno, "Failed to fork child pid: %m");
  58
  59         if (pid == 0) {
  60                 /* Return in the child */
  61                 assert_se(sigprocmask(SIG_SETMASK, &oldmask, NULL) >= 0);
  62                 setsid();
  63                 return 0;
  64         }
  65
  66         reset_all_signal_handlers();
  67
  68         log_close();
  69         close_all_fds(NULL, 0);
  70         log_open();
  71
  72         /* Flush out /proc/self/environ, so that we don't leak the environment from the host into the container. Also,
  73          * set $container= and $container_uuid= so that clients in the container that query it from /proc/1/environ
  74          * find them set. */
  75         sd_id128_to_string(uuid, new_environment + sizeof(new_environment) - SD_ID128_STRING_MAX);
  76         reset_environ(new_environment, sizeof(new_environment));
  77
  78         (void) rename_process("(sd-stubinit)");
  79
  80         assert_se(sigemptyset(&waitmask) >= 0);
  81         assert_se(sigset_add_many(&waitmask,
  82                                   SIGCHLD,          /* posix: process died */
  83                                   SIGINT,           /* sysv: ctrl-alt-del */
  84                                   SIGRTMIN+3,       /* systemd: halt */
  85                                   SIGRTMIN+4,       /* systemd: poweroff */
  86                                   SIGRTMIN+5,       /* systemd: reboot */
  87                                   SIGRTMIN+6,       /* systemd: kexec */
  88                                   SIGRTMIN+13,      /* systemd: halt */
  89                                   SIGRTMIN+14,      /* systemd: poweroff */
  90                                   SIGRTMIN+15,      /* systemd: reboot */
  91                                   SIGRTMIN+16,      /* systemd: kexec */
  92                                   -1) >= 0);
  93
  94         /* Note that we ignore SIGTERM (sysv's reexec), SIGHUP (reload), and all other signals here, since we don't
  95          * support reexec/reloading in this stub process. */
  96
  97         for (;;) {
  98                 siginfo_t si;
  99                 usec_t current_usec;
 100
 101                 si.si_pid = 0;
 102                 r = waitid(P_ALL, 0, &si, WEXITED|WNOHANG);
 103                 if (r < 0) {
 104                         r = log_error_errno(errno, "Failed to reap children: %m");
 105                         goto finish;
 106                 }
 107
 108                 current_usec = now(CLOCK_MONOTONIC);
 109
 110                 if (si.si_pid == pid || current_usec >= quit_usec) {
 111
 112                         /* The child we started ourselves died or we reached a timeout. */
 113
 114                         if (state == STATE_REBOOT) { /* dispatch a queued reboot */
 115                                 (void) reboot(RB_AUTOBOOT);
 116                                 r = log_error_errno(errno, "Failed to reboot: %m");
 117                                 goto finish;
 118
 119                         } else if (state == STATE_POWEROFF)
 120                                 (void) reboot(RB_POWER_OFF); /* if this fails, fall back to normal exit. */
 121
 122                         if (si.si_pid == pid && si.si_code == CLD_EXITED)
 123                                 r = si.si_status; /* pass on exit code */
 124                         else
 125                                 r = 255; /* signal, coredump, timeout, … */
 126
 127                         goto finish;
 128                 }
 129                 if (si.si_pid != 0)
 130                         /* We reaped something. Retry until there's nothing more to reap. */
 131                         continue;
 132
 133                 if (quit_usec == USEC_INFINITY)
 134                         r = sigwaitinfo(&waitmask, &si);
 135                 else {
 136                         struct timespec ts;
 137                         r = sigtimedwait(&waitmask, &si, timespec_store(&ts, quit_usec - current_usec));
 138                 }
 139                 if (r < 0) {
 140                         if (errno == EINTR) /* strace -p attach can result in EINTR, let's handle this nicely. */
 141                                 continue;
 142                         if (errno == EAGAIN) /* timeout reached */
 143                                 continue;
 144
 145                         r = log_error_errno(errno, "Failed to wait for signal: %m");
 146                         goto finish;
 147                 }
 148
 149                 if (si.si_signo == SIGCHLD)
 150                         continue; /* Let's reap this */
 151
 152                 if (state != STATE_RUNNING)
 153                         continue;
 154
 155                 /* Would love to use a switch() statement here, but SIGRTMIN is actually a function call, not a
 156                  * constant… */
 157
 158                 if (si.si_signo == SIGRTMIN+3 ||
 159                     si.si_signo == SIGRTMIN+4 ||
 160                     si.si_signo == SIGRTMIN+13 ||
 161                     si.si_signo == SIGRTMIN+14)
 162
 163                         state = STATE_POWEROFF;
 164
 165                 else if (si.si_signo == SIGINT ||
 166                          si.si_signo == SIGRTMIN+5 ||
 167                          si.si_signo == SIGRTMIN+6 ||
 168                          si.si_signo == SIGRTMIN+15 ||
 169                          si.si_signo == SIGRTMIN+16)
 170
 171                         state = STATE_REBOOT;
 172                 else
 173                         assert_not_reached("Got unexpected signal");
 174
 175                 r = kill_and_sigcont(pid, SIGTERM);
 176
 177                 /* Let's send a SIGHUP after the SIGTERM, as shells tend to ignore SIGTERM but do react to SIGHUP. We
 178                  * do it strictly in this order, so that the SIGTERM is dispatched first, and SIGHUP second for those
 179                  * processes which handle both. That's because services tend to bind configuration reload or something
 180                  * else to SIGHUP. */
 181
 182                 if (r != -ESRCH)
 183                         (void) kill(pid, SIGHUP);
 184
 185                 quit_usec = now(CLOCK_MONOTONIC) + DEFAULT_TIMEOUT_USEC;
 186         }
 187
 188 finish:
 189         _exit(r < 0 ? EXIT_FAILURE : r);
 190 }