int ci_udp_shutdown(citp_socket* ep, ci_fd_t fd, int how)
{
ci_fd_t os_sock;
int rc;
CHECK_UEP(ep);
LOG_UV(log(LPF "shutdown("SF_FMT", %d)", SF_PRI_ARGS(ep,fd), how));
os_sock = ci_get_os_sock_fd(fd);
if( CI_IS_VALID_SOCKET( os_sock ) ) {
rc = ci_sys_shutdown(os_sock, how);
ci_rel_os_sock_fd( os_sock );
if( rc < 0 )
return CI_SOCKET_ERROR;
}
rc = __ci_udp_shutdown(ep->netif, SOCK_TO_UDP(ep->s), how);
if( rc < 0 ) {
CI_SET_ERROR(rc, -rc);
return rc;
}
return 0;
}
void ci_udp_handle_force_reuseport(ci_fd_t fd, citp_socket* ep,
const struct sockaddr* sa, socklen_t sa_len)
{
int rc;
if( CITP_OPTS.udp_reuseports != 0 &&
((struct sockaddr_in*)sa)->sin_port != 0 ) {
struct ci_port_list *force_reuseport;
CI_DLLIST_FOR_EACH2(struct ci_port_list, force_reuseport, link,
(ci_dllist*)(ci_uintptr_t)CITP_OPTS.udp_reuseports) {
if( force_reuseport->port == ((struct sockaddr_in*)sa)->sin_port ) {
int one = 1;
ci_fd_t os_sock = ci_get_os_sock_fd(fd);
ci_assert(CI_IS_VALID_SOCKET(os_sock));
rc = ci_sys_setsockopt(os_sock, SOL_SOCKET, SO_REUSEPORT, &one,
sizeof(one));
ci_rel_os_sock_fd(os_sock);
/* Fixme: shouldn't we handle errors? */
if( rc != 0 ) {
log("%s: failed to set SO_REUSEPORT on OS socket: "
"rc=%d errno=%d", __func__, rc, errno);
}
ep->s->s_flags |= CI_SOCK_FLAG_REUSEPORT;
LOG_UC(log("%s "SF_FMT", applied legacy SO_REUSEPORT flag for port %u",
__FUNCTION__, SF_PRI_ARGS(ep, fd), force_reuseport->port));
}
}
}
}
/* To handle bind we just let the underlying OS socket make all
* of the decisions for us. If The bind leaves things such that
* the source address is not one of ours then we hand it over to the
* OS (by returning CI_SOCKET_HANDOVER) - in which case the OS socket
* will be bound as expected. */
int ci_udp_bind(citp_socket* ep, ci_fd_t fd, const struct sockaddr* addr,
socklen_t addrlen)
{
int rc;
ci_uint16 local_port;
CHECK_UEP(ep);
LOG_UC(log("%s("SF_FMT", addrlen=%d)", __FUNCTION__,
SF_PRI_ARGS(ep,fd), addrlen));
/* Make sure we have no filters.
*
* ?? TODO: Under what circumstances could we possibly have filters here?
* _WIN32 only perhaps?
*/
ci_udp_clr_filters(ep);
rc = ci_tcp_helper_bind_os_sock(fd, addr, addrlen, &local_port);
if( rc == CI_SOCKET_ERROR )
return rc;
return ci_udp_bind_conclude(ep, addr, local_port );
}
/* create a pt->pt association with a server
* This uses the OS to do all the work so that we don't have to emulate
* some of the more unpleasant "tricks" of Linux.
*
* When we're either handing-over OS-dest connects or when we're "no
* failing" connects we may return -2 (unhandled). In this case the
* OS socket _has_ been connected & we therefore are handing-over to
* a socket in the right state.
*
* NOTE: WINDOWS the WSPConnect() API is quite a lot more complex than
* the BSD one. Therefore, to stop polluting the core code with masses
* of Windows frippery, the backing socket connection is successfully
* established _before_ this function is called. This function will use
* the state of the backing socket to configure the Efab socket - so the
* end result is the same (right down to the race between the OS socket
* connection being established and our filters being inserted).
*/
int ci_udp_connect(citp_socket* ep, ci_fd_t fd,
const struct sockaddr* serv_addr, socklen_t addrlen )
{
int rc;
ci_fd_t os_sock;
CHECK_UEP(ep);
LOG_UC(log("%s("SF_FMT", addrlen=%d)", __FUNCTION__,
SF_PRI_ARGS(ep,fd), addrlen));
os_sock = ci_get_os_sock_fd(fd);
if( !CI_IS_VALID_SOCKET( os_sock ) ) {
LOG_U(ci_log("%s: no backing socket", __FUNCTION__));
return -1;
}
/* Because we have not handed over the fd to the OS all calls to bind()
* and connect() will have been seen by us - therefore our copies of
* the local/remote address & port will be accurate. */
/* Let the OS do the connection - it'll also do the data validation
* for free. On failure the OS changes nothing - therefore we
* need to leave the filters in place (if such they were).
* Because the OS socket and our socket are socket-options-synchronized,
* the following call will also check the supplied address according to
* the SO_BROADCAST socket option settings. */
rc = ci_sys_connect(os_sock, serv_addr, addrlen);
if( rc != 0 ) {
LOG_U(log("%s: sys_connect failed errno:%d", __FUNCTION__, errno));
ci_rel_os_sock_fd(os_sock);
return -1;
}
rc = ci_udp_connect_conclude( ep, fd, serv_addr, addrlen, os_sock);
ci_rel_os_sock_fd(os_sock);
return rc;
}
/* Complete a UDP U/L connect. The sys connect() call must have been made
* (and succeeded) before calling this function. So if anything goes wrong
* in here, then it can be consider an internal error or failing of onload.
*/
int ci_udp_connect_conclude(citp_socket* ep, ci_fd_t fd,
const struct sockaddr* serv_addr,
socklen_t addrlen, ci_fd_t os_sock)
{
const struct sockaddr_in* serv_sin = (const struct sockaddr_in*) serv_addr;
ci_uint32 dst_be32;
ci_udp_state* us = SOCK_TO_UDP(ep->s);
int onloadable;
int rc = 0;
CHECK_UEP(ep);
UDP_CLR_FLAG(us, CI_UDPF_EF_SEND);
us->s.rx_errno = 0;
us->s.tx_errno = 0;
if( IS_DISCONNECTING(serv_sin) ) {
rc = ci_udp_disconnect(ep, us, os_sock);
goto out;
}
#if CI_CFG_FAKE_IPV6
if( us->s.domain == PF_INET6 && !ci_tcp_ipv6_is_ipv4(serv_addr) ) {
LOG_UC(log(FNT_FMT "HANDOVER not IPv4", FNT_PRI_ARGS(ep->netif, us)));
goto handover;
}
#endif
dst_be32 = ci_get_ip4_addr(serv_sin->sin_family, serv_addr);
if( (rc = ci_udp_sys_getsockname(os_sock, ep)) != 0 ) {
LOG_E(log(FNT_FMT "ERROR: (%s:%d) sys_getsockname failed (%d)",
FNT_PRI_ARGS(ep->netif, us), ip_addr_str(dst_be32),
CI_BSWAP_BE16(serv_sin->sin_port), errno));
goto out;
}
us->s.cp.sock_cp_flags |= OO_SCP_CONNECTED;
ci_udp_set_raddr(us, dst_be32, serv_sin->sin_port);
cicp_user_retrieve(ep->netif, &us->s.pkt, &us->s.cp);
switch( us->s.pkt.status ) {
case retrrc_success:
case retrrc_nomac:
onloadable = 1;
break;
default:
onloadable = 0;
if( NI_OPTS(ep->netif).udp_connect_handover ) {
LOG_UC(log(FNT_FMT "HANDOVER %s:%d", FNT_PRI_ARGS(ep->netif, us),
ip_addr_str(dst_be32), CI_BSWAP_BE16(serv_sin->sin_port)));
goto handover;
}
break;
}
if( dst_be32 == INADDR_ANY_BE32 || serv_sin->sin_port == 0 ) {
LOG_UC(log(FNT_FMT "%s:%d - route via OS socket",
FNT_PRI_ARGS(ep->netif, us), ip_addr_str(dst_be32),
CI_BSWAP_BE16(serv_sin->sin_port)));
ci_udp_clr_filters(ep);
return 0;
}
if( CI_IP_IS_LOOPBACK(dst_be32) ) {
/* After connecting via loopback it is not possible to connect anywhere
* else.
*/
LOG_UC(log(FNT_FMT "HANDOVER %s:%d", FNT_PRI_ARGS(ep->netif, us),
ip_addr_str(dst_be32), CI_BSWAP_BE16(serv_sin->sin_port)));
goto handover;
}
if( onloadable ) {
#ifdef ONLOAD_OFE
if( ep->netif->ofe != NULL )
us->s.ofe_code_start = ofe_socktbl_find(
ep->netif->ofe, OFE_SOCKTYPE_UDP,
udp_laddr_be32(us), udp_raddr_be32(us),
udp_lport_be16(us), udp_rport_be16(us));
#endif
if( (rc = ci_udp_set_filters(ep, us)) != 0 ) {
/* Failed to set filters. Most likely we've run out of h/w filters.
* Handover to O/S to avoid breaking the app.
*
* TODO: Actually we probably won't break the app if we don't
* handover, as packets will still get delivered via the kernel
* stack. Might be worth having a runtime option to choose whether
* or not to handover in such cases.
*/
LOG_U(log(FNT_FMT "ERROR: (%s:%d) ci_udp_set_filters failed (%d)",
FNT_PRI_ARGS(ep->netif, us), ip_addr_str(dst_be32),
CI_BSWAP_BE16(serv_sin->sin_port), rc));
CITP_STATS_NETIF(++ep->netif->state->stats.udp_connect_no_filter);
goto out;
}
}
else {
ci_udp_clr_filters(ep);
}
LOG_UC(log(LPF "connect: "SF_FMT" %sCONNECTED L:%s:%u R:%s:%u (err:%d)",
SF_PRI_ARGS(ep,fd), udp_raddr_be32(us) ? "" : "DIS",
ip_addr_str(udp_laddr_be32(us)),
(unsigned) CI_BSWAP_BE16(udp_lport_be16(us)),
ip_addr_str(udp_raddr_be32(us)),
(unsigned) CI_BSWAP_BE16(udp_rport_be16(us)), errno));
return 0;
out:
if( rc < 0 && CITP_OPTS.no_fail )
goto handover;
return rc;
handover:
ci_udp_clr_filters(ep);
return CI_SOCKET_HANDOVER;
}
/* In this bind handler we just check that the address to which
* are binding is either "any" or one of ours.
* In the Linux kernel version [fd] is unused.
*/
int ci_tcp_bind(citp_socket* ep, const struct sockaddr* my_addr,
socklen_t addrlen, ci_fd_t fd )
{
struct sockaddr_in* my_addr_in;
ci_uint16 new_port;
ci_uint32 addr_be32;
ci_sock_cmn* s = ep->s;
ci_tcp_state* c = &SOCK_TO_WAITABLE_OBJ(s)->tcp;
int rc;
CHECK_TEP(ep);
my_addr_in = (struct sockaddr_in*) my_addr;
/* Check if state of the socket is OK for bind operation. */
/* \todo Earlier (TS_TCP( epi->tcpep.state )->tcp_source_be16) is used.
* What is better? */
if (my_addr == NULL)
RET_WITH_ERRNO( EINVAL );
if (s->b.state != CI_TCP_CLOSED)
RET_WITH_ERRNO( EINVAL );
if (c->tcpflags & CI_TCPT_FLAG_WAS_ESTAB)
RET_WITH_ERRNO( EINVAL );
if( my_addr->sa_family != s->domain )
RET_WITH_ERRNO( s->domain == PF_INET ? EAFNOSUPPORT : EINVAL );
/* Bug 4884: Windows regularly uses addrlen > sizeof(struct sockaddr_in)
* Linux is also relaxed about overlength data areas. */
if (s->domain == PF_INET && addrlen < sizeof(struct sockaddr_in))
RET_WITH_ERRNO( EINVAL );
#if CI_CFG_FAKE_IPV6
if (s->domain == PF_INET6 && addrlen < SIN6_LEN_RFC2133)
RET_WITH_ERRNO( EINVAL );
if( s->domain == PF_INET6 && !ci_tcp_ipv6_is_ipv4(my_addr) )
return CI_SOCKET_HANDOVER;
#endif
addr_be32 = ci_get_ip4_addr(s->domain, my_addr);
/* Using the port number provided, see if we can do this bind */
new_port = my_addr_in->sin_port;
if( CITP_OPTS.tcp_reuseports != 0 && new_port != 0 ) {
struct ci_port_list *force_reuseport;
CI_DLLIST_FOR_EACH2(struct ci_port_list, force_reuseport, link,
(ci_dllist*)(ci_uintptr_t)CITP_OPTS.tcp_reuseports) {
if( force_reuseport->port == new_port ) {
int one = 1;
ci_fd_t os_sock = ci_get_os_sock_fd(ep, fd);
ci_assert(CI_IS_VALID_SOCKET(os_sock));
rc = ci_sys_setsockopt(os_sock, SOL_SOCKET, SO_REUSEPORT, &one,
sizeof(one));
ci_rel_os_sock_fd(os_sock);
if( rc != 0 && errno == ENOPROTOOPT )
ep->s->s_flags |= CI_SOCK_FLAG_REUSEPORT_LEGACY;
ep->s->s_flags |= CI_SOCK_FLAG_REUSEPORT;
LOG_TC(log("%s "SF_FMT", applied legacy SO_REUSEPORT flag for port %u",
__FUNCTION__, SF_PRI_ARGS(ep, fd), new_port));
}
}
}
if( !(ep->s->s_flags & CI_SOCK_FLAG_REUSEPORT_LEGACY) )
CI_LOGLEVEL_TRY_RET(LOG_TV,
__ci_bind(ep->netif, ep->s, addr_be32, &new_port));
ep->s->s_flags |= CI_SOCK_FLAG_BOUND;
sock_lport_be16(s) = new_port;
sock_laddr_be32(s) = addr_be32;
if( CI_IP_IS_MULTICAST(addr_be32) )
s->cp.ip_laddr_be32 = 0;
else
s->cp.ip_laddr_be32 = addr_be32;
s->cp.lport_be16 = new_port;
sock_rport_be16(s) = sock_raddr_be32(s) = 0;
LOG_TC(log(LPF "bind to %s:%u n_p:%u lp:%u", ip_addr_str(addr_be32),
(unsigned) CI_BSWAP_BE16(my_addr_in->sin_port),
CI_BSWAP_BE16(new_port), CI_BSWAP_BE16(sock_lport_be16(s))));
return 0;
}
