There is a window of opportunity to leak file descriptors between
their creation and the fcntl(FD_CLOEXEC) call. If another thread forks
during this window, the descriptors will not have a the close-on-exec
flag, and get leaked after exec(). This is a limitation of POSIX.
While we're using the Linux-specific SOCK_CLOEXEC, we might as well
use SOCK_NONBLOCK, and spare ourselves the three fcntl() calls.
int net_Socket (vlc_object_t *p_this, int family, int socktype,
int protocol)
{
int net_Socket (vlc_object_t *p_this, int family, int socktype,
int protocol)
{
- int fd = socket (family, socktype, protocol);
- if (fd == -1)
+ int fd;
+
+#ifdef SOCK_CLOEXEC
+ fd = socket (family, socktype | SOCK_NONBLOCK | SOCK_CLOEXEC, protocol);
+ if (fd == -1 && errno == EINVAL)
+#endif
- if (net_errno != EAFNOSUPPORT)
- msg_Err (p_this, "cannot create socket: %m");
- return -1;
+ fd = socket (family, socktype, protocol);
+ if (fd == -1)
+ {
+ if (net_errno != EAFNOSUPPORT)
+ msg_Err (p_this, "cannot create socket: %m");
+ return -1;
+ }
+#ifndef WIN32
+ fcntl (fd, F_SETFD, FD_CLOEXEC);
+ fcntl (fd, F_SETFL, fcntl (fd, F_GETFL, 0) | O_NONBLOCK);
+#else
+ ioctlsocket (fd, FIONBIO, &(unsigned long){ 1 });
+#endif
+ setsockopt (fd, SOL_SOCKET, SO_REUSEADDR, &(int){ 1 }, sizeof (int));