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;
}
/* 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;
}
/* 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;
}
static int ci_tcp_connect_ul_start(ci_netif *ni, ci_tcp_state* ts,
ci_uint32 dst_be32, unsigned dport_be16,
int* fail_rc)
{
ci_ip_pkt_fmt* pkt;
int rc = 0;
ci_assert(ts->s.pkt.mtu);
/* Now that we know the outgoing route, set the MTU related values.
* Note, even these values are speculative since the real MTU
* could change between now and passing the packet to the lower layers
*/
ts->amss = ts->s.pkt.mtu - sizeof(ci_tcp_hdr) - sizeof(ci_ip4_hdr);
#if CI_CFG_LIMIT_AMSS
ts->amss = ci_tcp_limit_mss(ts->amss, ni, __FUNCTION__);
#endif
/* Default smss until discovered by MSS option in SYN - RFC1122 4.2.2.6 */
ts->smss = CI_CFG_TCP_DEFAULT_MSS;
/* set pmtu, eff_mss, snd_buf and adjust windows */
ci_pmtu_set(ni, &ts->pmtus, ts->s.pkt.mtu);
ci_tcp_set_eff_mss(ni, ts);
ci_tcp_set_initialcwnd(ni, ts);
/* Send buffer adjusted by ci_tcp_set_eff_mss(), but we want it to stay
* zero until the connection is established.
*/
ts->so_sndbuf_pkts = 0;
/*
* 3. State and address are OK. It's address routed through our NIC.
* Do connect().
*/
ci_assert_nequal(ts->s.pkt.ip.ip_saddr_be32, INADDR_ANY);
if( ts->s.s_flags & CI_SOCK_FLAG_CONNECT_MUST_BIND ) {
ci_sock_cmn* s = &ts->s;
ci_uint16 source_be16 = 0;
if( s->s_flags & CI_SOCK_FLAG_ADDR_BOUND )
rc = __ci_bind(ni, &ts->s, ts->s.pkt.ip.ip_saddr_be32, &source_be16);
else
rc = __ci_bind(ni, &ts->s, INADDR_ANY, &source_be16);
if(CI_LIKELY( rc == 0 )) {
TS_TCP(ts)->tcp_source_be16 = source_be16;
ts->s.cp.lport_be16 = source_be16;
LOG_TC(log(LNT_FMT "connect: our bind returned %s:%u",
LNT_PRI_ARGS(ni, ts),
ip_addr_str(INADDR_ANY),
(unsigned) CI_BSWAP_BE16(TS_TCP(ts)->tcp_source_be16)));
}
else {
LOG_U(ci_log("__ci_bind returned %d at %s:%d", CI_GET_ERROR(rc),
__FILE__, __LINE__));
*fail_rc = rc;
return CI_CONNECT_UL_FAIL;
}
if(CI_UNLIKELY( ts->s.pkt.ip.ip_saddr_be32 == 0 )) {
CI_SET_ERROR(*fail_rc, EINVAL);
return CI_CONNECT_UL_FAIL;
}
}
ci_tcp_set_peer(ts, dst_be32, dport_be16);
/* Make sure we can get a buffer before we change state. */
pkt = ci_netif_pkt_tx_tcp_alloc(ni);
if( CI_UNLIKELY(! pkt) ) {
/* NB. We've already done a poll above. */
rc = ci_netif_pkt_wait(ni, &ts->s, CI_SLEEP_NETIF_LOCKED|CI_SLEEP_NETIF_RQ);
if( ci_netif_pkt_wait_was_interrupted(rc) ) {
CI_SET_ERROR(*fail_rc, -rc);
return CI_CONNECT_UL_LOCK_DROPPED;
}
/* OK, there are (probably) packets available - go try again. Note we
* jump back to the top of the function because someone may have
* connected this socket in the mean-time, so we need to check the
* state once more.
*/
return CI_CONNECT_UL_START_AGAIN;
}
#ifdef ONLOAD_OFE
if( ni->ofe != NULL )
ts->s.ofe_code_start = ofe_socktbl_find(
ni->ofe, OFE_SOCKTYPE_TCP_ACTIVE,
tcp_laddr_be32(ts), tcp_raddr_be32(ts),
tcp_lport_be16(ts), tcp_rport_be16(ts));
#endif
rc = ci_tcp_ep_set_filters(ni, S_SP(ts), ts->s.cp.so_bindtodevice,
OO_SP_NULL);
if( rc < 0 ) {
/* Perhaps we've run out of filters? See if we can push a socket out
* of timewait and steal its filter.
*/
ci_assert_nequal(rc, -EFILTERSSOME);
if( rc != -EBUSY || ! ci_netif_timewait_try_to_free_filter(ni) ||
(rc = ci_tcp_ep_set_filters(ni, S_SP(ts),
ts->s.cp.so_bindtodevice,
OO_SP_NULL)) < 0 ) {
ci_assert_nequal(rc, -EFILTERSSOME);
/* Either a different error, or our efforts to free a filter did not
* work.
*/
if( ! (ts->s.s_flags & CI_SOCK_FLAG_ADDR_BOUND) ) {
ts->s.pkt.ip.ip_saddr_be32 = 0;
ts->s.cp.ip_laddr_be32 = 0;
}
ci_netif_pkt_release(ni, pkt);
CI_SET_ERROR(*fail_rc, -rc);
return CI_CONNECT_UL_FAIL;
}
}
LOG_TC(log(LNT_FMT "CONNECT %s:%u->%s:%u", LNT_PRI_ARGS(ni, ts),
ip_addr_str(ts->s.pkt.ip.ip_saddr_be32),
(unsigned) CI_BSWAP_BE16(TS_TCP(ts)->tcp_source_be16),
ip_addr_str(ts->s.pkt.ip.ip_daddr_be32),
(unsigned) CI_BSWAP_BE16(TS_TCP(ts)->tcp_dest_be16)));
/* We are going to send the SYN - set states appropriately */
tcp_snd_una(ts) = tcp_snd_nxt(ts) = tcp_enq_nxt(ts) = tcp_snd_up(ts) =
ci_tcp_initial_seqno(ni);
ts->snd_max = tcp_snd_nxt(ts) + 1;
/* Must be after initialising snd_una. */
ci_tcp_clear_rtt_timing(ts);
ci_tcp_set_flags(ts, CI_TCP_FLAG_SYN);
ts->tcpflags &=~ CI_TCPT_FLAG_OPT_MASK;
ts->tcpflags |= NI_OPTS(ni).syn_opts;
if( (ts->tcpflags & CI_TCPT_FLAG_WSCL) ) {
ts->rcv_wscl = ci_tcp_wscl_by_buff(ni, ci_tcp_rcvbuf_established(ni, &ts->s));
CI_IP_SOCK_STATS_VAL_RXWSCL(ts, ts->rcv_wscl);
}
else {
ts->rcv_wscl = 0;
CI_IP_SOCK_STATS_VAL_RXWSCL(ts, 0);
}
ci_tcp_set_rcvbuf(ni, ts);
ci_tcp_init_rcv_wnd(ts, "CONNECT");
/* outgoing_hdrs_len is initialised to include timestamp option. */
if( ! (ts->tcpflags & CI_TCPT_FLAG_TSO) )
ts->outgoing_hdrs_len = sizeof(ci_ip4_hdr)+sizeof(ci_tcp_hdr);
if( ci_tcp_can_stripe(ni, ts->s.pkt.ip.ip_saddr_be32,
ts->s.pkt.ip.ip_daddr_be32) )
ts->tcpflags |= CI_TCPT_FLAG_STRIPE;
ci_tcp_set_slow_state(ni, ts, CI_TCP_SYN_SENT);
/* If the app trys to send data on a socket in SYN_SENT state
** then the data is queued for send until the SYN gets ACKed.
** (rfc793 p56)
**
** Receive calls on the socket should block until data arrives
** (rfc793 p58)
**
** Clearing tx_errno and rx_errno acheive this. The transmit window
** is set to 1 byte which ensures that only the SYN packet gets
** sent until the ACK is received with more window.
*/
ci_assert(ts->snd_max == tcp_snd_nxt(ts) + 1);
ts->s.rx_errno = 0;
ts->s.tx_errno = 0;
ci_tcp_enqueue_no_data(ts, ni, pkt);
ci_tcp_set_flags(ts, CI_TCP_FLAG_ACK);
if( ts->s.b.sb_aflags & (CI_SB_AFLAG_O_NONBLOCK | CI_SB_AFLAG_O_NDELAY) ) {
ts->tcpflags |= CI_TCPT_FLAG_NONBLOCK_CONNECT;
LOG_TC(log( LNT_FMT "Non-blocking connect - return EINPROGRESS",
LNT_PRI_ARGS(ni, ts)));
CI_SET_ERROR(*fail_rc, EINPROGRESS);
return CI_CONNECT_UL_FAIL;
}
return CI_CONNECT_UL_OK;
}
int ci_tcp_listen(citp_socket* ep, ci_fd_t fd, int backlog)
{
/*
** ?? error handling on possible fails not handled robustly...
** ?? Need to check port number is valid TODO
*/
/*! \todo If not bound then we have to be listening on all interfaces.
* It's likely that we won't be coming through here as we have to
* listen on the OS socket too! */
ci_tcp_state* ts;
ci_tcp_socket_listen* tls;
ci_netif* netif = ep->netif;
ci_sock_cmn* s = ep->s;
unsigned ul_backlog = backlog;
int rc;
oo_p sp;
LOG_TC(log("%s "SK_FMT" listen backlog=%d", __FUNCTION__, SK_PRI_ARGS(ep),
backlog));
CHECK_TEP(ep);
if( NI_OPTS(netif).tcp_listen_handover )
return CI_SOCKET_HANDOVER;
if( !NI_OPTS(netif).tcp_server_loopback) {
/* We should handover if the socket is bound to alien address. */
if( s->s_flags & CI_SOCK_FLAG_BOUND_ALIEN )
return CI_SOCKET_HANDOVER;
}
if( ul_backlog < 0 )
ul_backlog = NI_OPTS(netif).max_ep_bufs;
else if( ul_backlog < NI_OPTS(netif).acceptq_min_backlog )
ul_backlog = NI_OPTS(netif).acceptq_min_backlog;
if( s->b.state == CI_TCP_LISTEN ) {
tls = SOCK_TO_TCP_LISTEN(s);
tls->acceptq_max = ul_backlog;
ci_tcp_helper_listen_os_sock(fd, ul_backlog);
return 0;
}
if( s->b.state != CI_TCP_CLOSED ) {
CI_SET_ERROR(rc, EINVAL);
return rc;
}
ts = SOCK_TO_TCP(s);
/* Bug 3376: if socket used for a previous, failed, connect then the error
* numbers will not be as expected. Only seen when not using listening
* netifs (as moving the EP to the new netif resets them).
*/
ts->s.tx_errno = EPIPE;
ts->s.rx_errno = ENOTCONN;
/* fill in address/ports and all TCP state */
if( !(ts->s.s_flags & CI_SOCK_FLAG_BOUND) ) {
ci_uint16 source_be16;
/* They haven't previously done a bind, so we need to choose
* a port. As we haven't been given a hint we let the OS choose. */
source_be16 = 0;
rc = __ci_bind(ep->netif, ep->s, ts->s.pkt.ip.ip_saddr_be32, &source_be16);
if (CI_LIKELY( rc==0 )) {
TS_TCP(ts)->tcp_source_be16 = source_be16;
ts->s.cp.lport_be16 = source_be16;
LOG_TC(log(LNT_FMT "listen: our bind returned %s:%u",
LNT_PRI_ARGS(ep->netif, ts),
ip_addr_str(ts->s.pkt.ip.ip_saddr_be32),
(unsigned) CI_BSWAP_BE16(TS_TCP(ts)->tcp_source_be16)));
} else {
LOG_U(ci_log("__ci_bind returned %d at %s:%d", CI_GET_ERROR(rc),
__FILE__, __LINE__));
return rc;
}
}
ci_sock_lock(netif, &ts->s.b);
ci_tcp_set_slow_state(netif, ts, CI_TCP_LISTEN);
tls = SOCK_TO_TCP_LISTEN(&ts->s);
tcp_raddr_be32(tls) = 0u;
tcp_rport_be16(tls) = 0u;
ci_assert_equal(tls->s.tx_errno, EPIPE);
ci_assert_equal(tls->s.rx_errno, ENOTCONN);
/* setup listen timer - do it before the first return statement,
* because __ci_tcp_listen_to_normal() will be called on error path. */
if( ~tls->s.s_flags & CI_SOCK_FLAG_BOUND_ALIEN ) {
sp = TS_OFF(netif, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, listenq_tid));
ci_ip_timer_init(netif, &tls->listenq_tid, sp, "lstq");
tls->listenq_tid.param1 = S_SP(tls);
tls->listenq_tid.fn = CI_IP_TIMER_TCP_LISTEN;
}
rc = ci_tcp_listen_init(netif, tls);
ci_sock_unlock(netif, &ts->s.b);
if( rc != 0 ) {
CI_SET_ERROR(rc, -rc);
goto listen_fail;
}
tls->acceptq_max = ul_backlog;
CITP_STATS_TCP_LISTEN(CI_ZERO(&tls->stats));
/* install all the filters needed for this connection
* - tcp_laddr_be32(ts) = 0 for IPADDR_ANY
*
* TODO: handle BINDTODEVICE by setting phys_port paramter to correct
* physical L5 port index
* TODO: handle REUSEADDR by setting last paramter to TRUE
*/
if( ~s->s_flags & CI_SOCK_FLAG_BOUND_ALIEN ) {
#ifdef ONLOAD_OFE
if( netif->ofe != NULL ) {
tls->s.ofe_code_start = ofe_socktbl_find(
netif->ofe, OFE_SOCKTYPE_TCP_LISTEN,
tcp_laddr_be32(tls), INADDR_ANY,
tcp_lport_be16(ts), 0);
tls->ofe_promote = ofe_socktbl_find(
netif->ofe, OFE_SOCKTYPE_TCP_PASSIVE,
tcp_laddr_be32(tls), INADDR_ANY,
tcp_lport_be16(ts), 0);
}
#endif
rc = ci_tcp_ep_set_filters(netif, S_SP(tls), tls->s.cp.so_bindtodevice,
OO_SP_NULL);
if( rc == -EFILTERSSOME ) {
if( CITP_OPTS.no_fail )
rc = 0;
else {
ci_tcp_ep_clear_filters(netif, S_SP(tls), 0);
rc = -ENOBUFS;
}
}
ci_assert_nequal(rc, -EFILTERSSOME);
VERB(ci_log("%s: set_filters returned %d", __FUNCTION__, rc));
if (rc < 0) {
CI_SET_ERROR(rc, -rc);
goto post_listen_fail;
}
}
/*
* Call of system listen() is required for listen any, local host
* communications server and multi-homed server (to accept connections
* to L5 assigned address(es), but incoming from other interfaces).
*/
#ifdef __ci_driver__
{
rc = efab_tcp_helper_listen_os_sock( netif2tcp_helper_resource(netif),
S_SP(tls), backlog);
}
#else
rc = ci_tcp_helper_listen_os_sock(fd, backlog);
#endif
if ( rc < 0 ) {
/* clear the filter we've just set */
ci_tcp_ep_clear_filters(netif, S_SP(tls), 0);
goto post_listen_fail;
}
return 0;
post_listen_fail:
ci_tcp_listenq_drop_all(netif, tls);
listen_fail:
/* revert TCP state to a non-listening socket format */
__ci_tcp_listen_to_normal(netif, tls);
/* Above function sets orphan flag but we are attached to an FD. */
ci_bit_clear(&tls->s.b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
#ifdef __ci_driver__
return rc;
#else
return CI_SOCKET_ERROR;
#endif
}
