diff options
Diffstat (limited to 'networking/inetd.c')
| -rw-r--r-- | networking/inetd.c | 101 |
1 files changed, 60 insertions, 41 deletions
diff --git a/networking/inetd.c b/networking/inetd.c index 0620188d6..41824dbc8 100644 --- a/networking/inetd.c +++ b/networking/inetd.c @@ -130,8 +130,13 @@ * tening service socket, and must accept at least one connection request * before exiting. Such a server would normally accept and process incoming * connection requests until a timeout. - * - * In short: "stream" can be "wait" or "nowait"; "dgram" must be "wait". + */ + +/* Despite of above doc saying that dgram services must use "wait", + * "udp nowait" servers are implemented in busyboxed inetd. + * IPv6 addresses are also implemented. However, they may look ugly - + * ":::service..." means "address '::' (IPv6 wildcard addr)":"service"... + * You have to put "tcp6"/"udp6" in protocol field to select IPv6. */ /* Here's the scoop concerning the user[:group] feature: @@ -832,9 +837,6 @@ static NOINLINE servtab_t *parse_one_line(void) if (sep->se_socktype == SOCK_DGRAM) { if (sep->se_proto_no == IPPROTO_TCP) goto parse_err; - /* "udp nowait" is a small fork bomb :) */ - if (!sep->se_wait) - goto parse_err; } /* check if the hostname specifier is a comma separated list @@ -1195,7 +1197,7 @@ int inetd_main(int argc, char **argv) for (;;) { int ready_fd_cnt; - int ctrl, accepted_fd; + int ctrl, accepted_fd, new_udp_fd; fd_set readable; if (maxsock < 0) @@ -1220,12 +1222,43 @@ int inetd_main(int argc, char **argv) ready_fd_cnt--; ctrl = sep->se_fd; accepted_fd = -1; - if (!sep->se_wait && sep->se_socktype == SOCK_STREAM) { - ctrl = accepted_fd = accept(sep->se_fd, NULL, NULL); - if (ctrl < 0) { - if (errno != EINTR) - bb_perror_msg("accept (for %s)", sep->se_service); - continue; + new_udp_fd = -1; + if (!sep->se_wait) { + if (sep->se_socktype == SOCK_STREAM) { + ctrl = accepted_fd = accept(sep->se_fd, NULL, NULL); + if (ctrl < 0) { + if (errno != EINTR) + bb_perror_msg("accept (for %s)", sep->se_service); + continue; + } + } + /* "nowait" udp */ + if (sep->se_socktype == SOCK_DGRAM + && sep->se_family != AF_UNIX + ) { +/* How udp "nowait" works: + * child peeks at (received and buffered by kernel) UDP packet, + * performs connect() on the socket so that it is linked only + * to this peer. But this also affects parent, because descriptors + * are shared after fork() a-la dup(). When parent performs + * select(), it will see this descriptor connected to the peer (!) + * and still readable, will act on it and mess things up + * (can create many copies of same child, etc). + * Parent must create and use new socket instead. */ + new_udp_fd = socket(sep->se_family, SOCK_DGRAM, 0); + if (new_udp_fd < 0) { /* error: eat packet, forget about it */ + udp_err: + recv(sep->se_fd, line, LINE_SIZE, MSG_DONTWAIT); + continue; + } + setsockopt_reuseaddr(new_udp_fd); + /* TODO: better do bind after vfork in parent, + * so that we don't have two wildcard bound sockets + * even for a brief moment? */ + if (bind(new_udp_fd, &sep->se_lsa->u.sa, sep->se_lsa->len) < 0) { + close(new_udp_fd); + goto udp_err; + } } } @@ -1283,10 +1316,15 @@ int inetd_main(int argc, char **argv) if (pid > 0) { /* parent */ if (sep->se_wait) { + /* tcp wait: we passed listening socket to child, + * will wait for child to terminate */ sep->se_wait = pid; remove_fd_from_set(sep->se_fd); - /* we passed listening socket to child, - * will wait for child to terminate */ + } + if (new_udp_fd >= 0) { + /* udp nowait: child connected the socket, + * we created and will use new, unconnected one */ + xmove_fd(new_udp_fd, sep->se_fd); } restore_sigmask(&omask); maybe_close(accepted_fd); @@ -1313,39 +1351,20 @@ int inetd_main(int argc, char **argv) #endif /* child */ setsid(); -#if 0 -/* This does not work. - * Actually, it _almost_ works. The idea behind it is: child - * can peek at (already received and buffered by kernel) UDP packet, - * and perform connect() on the socket so that it is linked only - * to this peer. But this also affects parent, because descriptors - * are shared after fork() a-la dup(). When parent returns to - * select(), it will see this descriptor attached to the peer (!) - * and likely still readable, will act on it and mess things up - * (can create many copies of same child, etc). - * If child will create new socket instead, then bind() and - * connect() it to peer's address, descriptor aliasing problem - * is solved, but first packet cannot be "transferred" to the new - * socket. It is not a problem if child can account for this, - * but our child will exec - and exec'ed program does not know - * about this "lost packet" problem! Pity... */ - /* "nowait" udp[6]. Hmmm... */ - if (!sep->se_wait - && sep->se_socktype == SOCK_DGRAM - && sep->se_family != AF_UNIX - ) { + /* "nowait" udp */ + if (new_udp_fd >= 0) { len_and_sockaddr *lsa = xzalloc_lsa(sep->se_family); /* peek at the packet and remember peer addr */ int r = recvfrom(ctrl, NULL, 0, MSG_PEEK|MSG_DONTWAIT, &lsa->u.sa, &lsa->len); - if (r >= 0) - /* make this socket "connected" to peer addr: - * only packets from this peer will be recv'ed, - * and bare write()/send() will work on it */ - connect(ctrl, &lsa->u.sa, lsa->len); + if (r < 0) + goto do_exit1; + /* make this socket "connected" to peer addr: + * only packets from this peer will be recv'ed, + * and bare write()/send() will work on it */ + connect(ctrl, &lsa->u.sa, lsa->len); free(lsa); } -#endif /* prepare env and exec program */ pwd = getpwnam(sep->se_user); if (pwd == NULL) { |