src/process/posix_spawn.c

   1 #define _GNU_SOURCE
   2 #include <spawn.h>
   3 #include <sched.h>
   4 #include <unistd.h>
   5 #include <signal.h>
   6 #include <fcntl.h>
   7 #include <sys/wait.h>
   8 #include "syscall.h"
   9 #include "lock.h"
  10 #include "pthread_impl.h"
  11 #include "fdop.h"
  12
  13 struct args {
  14         int p[2];
  15         sigset_t oldmask;
  16         const char *path;
  17         const posix_spawn_file_actions_t *fa;
  18         const posix_spawnattr_t *restrict attr;
  19         char *const *argv, *const *envp;
  20 };
  21
  22 static int __sys_dup2(int old, int new)
  23 {
  24 #ifdef SYS_dup2
  25         return __syscall(SYS_dup2, old, new);
  26 #else
  27         return __syscall(SYS_dup3, old, new, 0);
  28 #endif
  29 }
  30
  31 static int child(void *args_vp)
  32 {
  33         int i, ret;
  34         struct sigaction sa = {0};
  35         struct args *args = args_vp;
  36         int p = args->p[1];
  37         const posix_spawn_file_actions_t *fa = args->fa;
  38         const posix_spawnattr_t *restrict attr = args->attr;
  39         sigset_t hset;
  40
  41         close(args->p[0]);
  42
  43         /* All signal dispositions must be either SIG_DFL or SIG_IGN
  44          * before signals are unblocked. Otherwise a signal handler
  45          * from the parent might get run in the child while sharing
  46          * memory, with unpredictable and dangerous results. To
  47          * reduce overhead, sigaction has tracked for us which signals
  48          * potentially have a signal handler. */
  49         __get_handler_set(&hset);
  50         for (i=1; i<_NSIG; i++) {
  51                 if ((attr->__flags & POSIX_SPAWN_SETSIGDEF)
  52                      && sigismember(&attr->__def, i)) {
  53                         sa.sa_handler = SIG_DFL;
  54                 } else if (sigismember(&hset, i)) {
  55                         if (i-32<3U) {
  56                                 sa.sa_handler = SIG_IGN;
  57                         } else {
  58                                 __libc_sigaction(i, 0, &sa);
  59                                 if (sa.sa_handler==SIG_IGN) continue;
  60                                 sa.sa_handler = SIG_DFL;
  61                         }
  62                 } else {
  63                         continue;
  64                 }
  65                 __libc_sigaction(i, &sa, 0);
  66         }
  67
  68         if (attr->__flags & POSIX_SPAWN_SETSID)
  69                 if ((ret=__syscall(SYS_setsid)) < 0)
  70                         goto fail;
  71
  72         if (attr->__flags & POSIX_SPAWN_SETPGROUP)
  73                 if ((ret=__syscall(SYS_setpgid, 0, attr->__pgrp)))
  74                         goto fail;
  75
  76         /* Use syscalls directly because the library functions attempt
  77          * to do a multi-threaded synchronized id-change, which would
  78          * trash the parent's state. */
  79         if (attr->__flags & POSIX_SPAWN_RESETIDS)
  80                 if ((ret=__syscall(SYS_setgid, __syscall(SYS_getgid))) ||
  81                     (ret=__syscall(SYS_setuid, __syscall(SYS_getuid))) )
  82                         goto fail;
  83
  84         if (fa && fa->__actions) {
  85                 struct fdop *op;
  86                 int fd;
  87                 for (op = fa->__actions; op->next; op = op->next);
  88                 for (; op; op = op->prev) {
  89                         /* It's possible that a file operation would clobber
  90                          * the pipe fd used for synchronizing with the
  91                          * parent. To avoid that, we dup the pipe onto
  92                          * an unoccupied fd. */
  93                         if (op->fd == p) {
  94                                 ret = __syscall(SYS_dup, p);
  95                                 if (ret < 0) goto fail;
  96                                 __syscall(SYS_close, p);
  97                                 p = ret;
  98                         }
  99                         switch(op->cmd) {
 100                         case FDOP_CLOSE:
 101                                 __syscall(SYS_close, op->fd);
 102                                 break;
 103                         case FDOP_DUP2:
 104                                 fd = op->srcfd;
 105                                 if (fd == p) {
 106                                         ret = -EBADF;
 107                                         goto fail;
 108                                 }
 109                                 if (fd != op->fd) {
 110                                         if ((ret=__sys_dup2(fd, op->fd))<0)
 111                                                 goto fail;
 112                                 } else {
 113                                         ret = __syscall(SYS_fcntl, fd, F_GETFD);
 114                                         ret = __syscall(SYS_fcntl, fd, F_SETFD,
 115                                                         ret & ~FD_CLOEXEC);
 116                                         if (ret<0)
 117                                                 goto fail;
 118                                 }
 119                                 break;
 120                         case FDOP_OPEN:
 121                                 fd = __sys_open(op->path, op->oflag, op->mode);
 122                                 if ((ret=fd) < 0) goto fail;
 123                                 if (fd != op->fd) {
 124                                         if ((ret=__sys_dup2(fd, op->fd))<0)
 125                                                 goto fail;
 126                                         __syscall(SYS_close, fd);
 127                                 }
 128                                 break;
 129                         case FDOP_CHDIR:
 130                                 ret = __syscall(SYS_chdir, op->path);
 131                                 if (ret<0) goto fail;
 132                                 break;
 133                         case FDOP_FCHDIR:
 134                                 ret = __syscall(SYS_fchdir, op->fd);
 135                                 if (ret<0) goto fail;
 136                                 break;
 137                         }
 138                 }
 139         }
 140
 141         /* Close-on-exec flag may have been lost if we moved the pipe
 142          * to a different fd. We don't use F_DUPFD_CLOEXEC above because
 143          * it would fail on older kernels and atomicity is not needed --
 144          * in this process there are no threads or signal handlers. */
 145         __syscall(SYS_fcntl, p, F_SETFD, FD_CLOEXEC);
 146
 147         pthread_sigmask(SIG_SETMASK, (attr->__flags & POSIX_SPAWN_SETSIGMASK)
 148                 ? &attr->__mask : &args->oldmask, 0);
 149
 150         int (*exec)(const char *, char *const *, char *const *) =
 151                 attr->__fn ? (int (*)())attr->__fn : execve;
 152
 153         exec(args->path, args->argv, args->envp);
 154         ret = -errno;
 155
 156 fail:
 157         /* Since sizeof errno < PIPE_BUF, the write is atomic. */
 158         ret = -ret;
 159         if (ret) while (__syscall(SYS_write, p, &ret, sizeof ret) < 0);
 160         _exit(127);
 161 }
 162
 163
 164 int posix_spawn(pid_t *restrict res, const char *restrict path,
 165         const posix_spawn_file_actions_t *fa,
 166         const posix_spawnattr_t *restrict attr,
 167         char *const argv[restrict], char *const envp[restrict])
 168 {
 169         pid_t pid;
 170         char stack[1024+PATH_MAX];
 171         int ec=0, cs;
 172         struct args args;
 173
 174         pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &cs);
 175
 176         args.path = path;
 177         args.fa = fa;
 178         args.attr = attr ? attr : &(const posix_spawnattr_t){0};
 179         args.argv = argv;
 180         args.envp = envp;
 181         pthread_sigmask(SIG_BLOCK, SIGALL_SET, &args.oldmask);
 182
 183         /* The lock guards both against seeing a SIGABRT disposition change
 184          * by abort and against leaking the pipe fd to fork-without-exec. */
 185         LOCK(__abort_lock);
 186
 187         if (pipe2(args.p, O_CLOEXEC)) {
 188                 UNLOCK(__abort_lock);
 189                 ec = errno;
 190                 goto fail;
 191         }
 192
 193         pid = __clone(child, stack+sizeof stack,
 194                 CLONE_VM|CLONE_VFORK|SIGCHLD, &args);
 195         close(args.p[1]);
 196         UNLOCK(__abort_lock);
 197
 198         if (pid > 0) {
 199                 if (read(args.p[0], &ec, sizeof ec) != sizeof ec) ec = 0;
 200                 else waitpid(pid, &(int){0}, 0);
 201         } else {
 202                 ec = -pid;
 203         }
 204
 205         close(args.p[0]);
 206
 207         if (!ec && res) *res = pid;
 208
 209 fail:
 210         pthread_sigmask(SIG_SETMASK, &args.oldmask, 0);
 211         pthread_setcancelstate(cs, 0);
 212
 213         return ec;
 214 }