static int
ci_udp_disconnect(citp_socket* ep, ci_udp_state* us, ci_fd_t os_sock)
{
int rc;
if( (rc = ci_udp_sys_getsockname(os_sock, ep)) != 0 ) {
LOG_E(log(FNS_FMT "ERROR: sys_getsockname failed (%d)",
FNS_PRI_ARGS(ep->netif, ep->s), errno));
return rc;
}
ci_udp_set_raddr(us, 0, 0);
/* TODO: We shouldn't really clear then set here; instead we should
* insert wildcard filters before removing the full-match ones. ie. The
* reverse of what we do in connect(). But probably not worth worrying
* about in this case.
*/
ci_udp_clr_filters(ep);
#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 )
/* Not too bad -- should still get packets via OS socket. */
LOG_U(log(FNS_FMT "ERROR: ci_udp_set_filters failed (%d)",
FNS_PRI_ARGS(ep->netif, ep->s), errno));
us->s.cp.sock_cp_flags &= ~OO_SCP_CONNECTED;
return 0;
}
static void ci_udp_set_raddr(ci_udp_state* us, unsigned raddr_be32,
int rport_be16)
{
ci_ip_cache_invalidate(&us->s.pkt);
udp_raddr_be32(us) = raddr_be32;
udp_rport_be16(us) = (ci_uint16) rport_be16;
us->s.pkt.dport_be16 = (ci_uint16) rport_be16;
}
static char * ci_udp_addr_str( ci_udp_state* us )
{
static char buf[128];
ci_assert(us);
sprintf( buf, "L[%s:%d] R[%s:%d]",
ip_addr_str( udp_laddr_be32(us)),
CI_BSWAP_BE16(udp_lport_be16(us)),
ip_addr_str( udp_raddr_be32(us)),
CI_BSWAP_BE16(udp_rport_be16(us)) );
return buf;
}
/* initialise all the fields that we can in the UDP state structure.
** There are no IP options, no destination addresses, no ports */
static void ci_udp_state_init(ci_netif* netif, ci_udp_state* us)
{
ci_sock_cmn_init(netif, &us->s, 1);
/* IP_MULTICAST_LOOP is 1 by default, so we should not send multicast
* unless specially permitted */
if( ! NI_OPTS(netif).force_send_multicast )
us->s.cp.sock_cp_flags |= OO_SCP_NO_MULTICAST;
/* Poison. */
CI_DEBUG(memset(&us->s + 1, 0xf0, (char*) (us + 1) - (char*) (&us->s + 1)));
/*! \todo This should be part of sock_cmn reinit, but the comment to that
* function suggests that it's possibly not a good plan to move it there */
#if CI_CFG_TIMESTAMPING
ci_udp_recv_q_init(&us->timestamp_q);
#endif
/*! \todo These two should really be handled in ci_sock_cmn_init() */
/* Make sure we don't hit any state assertions. Can use
* UDP_STATE_FROM_SOCKET_EPINFO() after this. */
us->s.b.state = CI_TCP_STATE_UDP;
us->s.so.sndbuf = NI_OPTS(netif).udp_sndbuf_def;
us->s.so.rcvbuf = NI_OPTS(netif).udp_rcvbuf_def;
/* Init the ip-caches (packet header templates). */
ci_udp_hdrs_init(&us->s.pkt);
ci_ip_cache_init(&us->ephemeral_pkt);
ci_udp_hdrs_init(&us->ephemeral_pkt);
udp_lport_be16(us) = 0;
udp_rport_be16(us) = 0;
#if CI_CFG_ZC_RECV_FILTER
us->recv_q_filter = 0;
us->recv_q_filter_arg = 0;
#endif
ci_udp_recv_q_init(&us->recv_q);
us->zc_kernel_datagram = OO_PP_NULL;
us->zc_kernel_datagram_count = 0;
us->tx_async_q = CI_ILL_END;
oo_atomic_set(&us->tx_async_q_level, 0);
us->tx_count = 0;
us->udpflags = CI_UDPF_MCAST_LOOP;
us->ip_pktinfo_cache.intf_i = -1;
us->stamp = 0;
memset(&us->stats, 0, sizeof(us->stats));
}
/* Conclude the EP's binding. This function is abstracted from the
* main bind code to allow implicit binds that occur when sendto() is
* called on an OS socket. [lport] and CI_SIN(addr)->sin_port do not
* have to be the same value. */
static int ci_udp_bind_conclude(citp_socket* ep, const struct sockaddr* addr,
ci_uint16 lport )
{
ci_udp_state* us;
ci_uint32 addr_be32;
int rc;
CHECK_UEP(ep);
ci_assert(addr != NULL);
if( ci_udp_should_handover(ep, addr, lport) )
goto handover;
addr_be32 = ci_get_ip4_addr(ep->s->domain, addr);
ci_udp_set_laddr(ep, addr_be32, lport);
us = SOCK_TO_UDP(ep->s);
if( addr_be32 != 0 )
us->s.cp.sock_cp_flags |= OO_SCP_LADDR_BOUND;
/* reset any rx/tx that have taken place already */
UDP_CLR_FLAG(us, CI_UDPF_EF_SEND);
#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
/* OS source addrs have already been handed-over, so this must be one of
* our src addresses.
*/
rc = ci_udp_set_filters( ep, us);
ci_assert( !UDP_GET_FLAG(us, CI_UDPF_EF_BIND) );
/*! \todo FIXME isn't the port the thing to be testing here? */
if( udp_laddr_be32(us) != INADDR_ANY_BE32 )
UDP_SET_FLAG(us, CI_UDPF_EF_BIND);
CI_UDPSTATE_SHOW_EP( ep );
if( rc == CI_SOCKET_ERROR && CITP_OPTS.no_fail) {
CITP_STATS_NETIF(++ep->netif->state->stats.udp_bind_no_filter);
goto handover;
}
return rc;
handover:
LOG_UV(log("%s: "SK_FMT" HANDOVER", __FUNCTION__, SK_PRI_ARGS(ep)));
return CI_SOCKET_HANDOVER;
}
/*! \todo we can simplify this a lot by letting the kernel have it! */
int ci_udp_getpeername(citp_socket*ep, struct sockaddr* name, socklen_t* namelen)
{
ci_udp_state* us;
CHECK_UEP(ep);
us = SOCK_TO_UDP(ep->s);
/*
* At first, it's necessary to check whether socket is connected or
* not, since we can return ENOTCONN even if name and/or namelen are
* not valid.
*/
if( udp_raddr_be32(us) == 0 ) {
RET_WITH_ERRNO(ENOTCONN);
} else if( name == NULL || namelen == NULL ) {
RET_WITH_ERRNO(EFAULT);
} else {
ci_addr_to_user(name, namelen, ep->s->domain,
udp_rport_be16(us), udp_raddr_be32(us));
return 0;
}
}
/* 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;
}
