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;
}
/* unpick the ci_ip_timer structure to actually do the callback */
static void ci_ip_timer_docallback(ci_netif *netif, ci_ip_timer* ts)
{
ci_assert( TIME_LE(ts->time, ci_ip_time_now(netif)) );
ci_assert( ts->time == IPTIMER_STATE(netif)->sched_ticks );
switch(ts->fn){
case CI_IP_TIMER_TCP_RTO:
CHECK_TS(netif, SP_TO_TCP(netif, ts->param1));
ci_tcp_timeout_rto(netif, SP_TO_TCP(netif, ts->param1));
break;
case CI_IP_TIMER_TCP_DELACK:
CHECK_TS(netif, SP_TO_TCP(netif, ts->param1));
ci_tcp_timeout_delack(netif, SP_TO_TCP(netif, ts->param1));
break;
case CI_IP_TIMER_TCP_ZWIN:
CHECK_TS(netif, SP_TO_TCP(netif, ts->param1));
ci_tcp_timeout_zwin(netif, SP_TO_TCP(netif, ts->param1));
break;
case CI_IP_TIMER_TCP_KALIVE:
CHECK_TS(netif, SP_TO_TCP(netif, ts->param1));
ci_tcp_timeout_kalive(netif, SP_TO_TCP(netif, ts->param1));
break;
case CI_IP_TIMER_TCP_LISTEN:
ci_tcp_timeout_listen(netif, SP_TO_TCP_LISTEN(netif, ts->param1));
break;
case CI_IP_TIMER_TCP_CORK:
ci_tcp_timeout_cork(netif, SP_TO_TCP(netif, ts->param1));
break;
case CI_IP_TIMER_NETIF_TIMEOUT:
ci_netif_timeout_state(netif);
break;
case CI_IP_TIMER_PMTU_DISCOVER:
ci_pmtu_timeout_pmtu(netif, SP_TO_TCP(netif, ts->param1));
break;
#if CI_CFG_TCP_SOCK_STATS
case CI_IP_TIMER_TCP_STATS:
ci_tcp_stats_action(netif, SP_TO_TCP(netif, ts->param1),
CI_IP_STATS_FLUSH,
CI_IP_STATS_OUTPUT_NONE, NULL, NULL );
break;
#endif
#if CI_CFG_SUPPORT_STATS_COLLECTION
case CI_IP_TIMER_NETIF_STATS:
ci_netif_stats_action(netif, CI_IP_STATS_FLUSH,
CI_IP_STATS_OUTPUT_NONE, NULL, NULL );
break;
#endif
#if CI_CFG_IP_TIMER_DEBUG
case CI_IP_TIMER_DEBUG_HOOK:
ci_ip_timer_debug_fn(netif, ts->link.addr, ts->param1);
break;
#endif
default:
LOG_U(log( LPF "unknown timer callback code:%x param1:%d",
ts->fn, OO_SP_FMT(ts->param1)));
CI_DEBUG(ci_fail_stop_fn());
}
}
/*!
* Tests for valid sockaddr & sockaddr length & AF_INET or AF_INET6.
*/
static int ci_tcp_validate_sa( sa_family_t domain,
const struct sockaddr* sa, socklen_t sa_len )
{
/*
* Linux deviates from documented behaviour here;
* On Linux we return EINVAL if sa and sa_len are NULL and 0 respectively,
* and we return EFAULT if sa is NULL and sa_len != 0....
*/
if( !sa ) {
LOG_U(ci_log(LPF "invalid sockaddr : sa = %lx, sa_len = %d",
(long) sa, sa_len));
if( sa_len == 0 )
RET_WITH_ERRNO( EINVAL );
else
RET_WITH_ERRNO( EFAULT );
}
if( sa_len < sizeof(struct sockaddr_in)
#if CI_CFG_FAKE_IPV6
|| (domain == AF_INET6 && sa_len < SIN6_LEN_RFC2133)
#endif
) {
LOG_U( ci_log(LPF "struct too short to be sockaddr_in(6)" ));
RET_WITH_ERRNO( EINVAL );
}
/* It should be sa->sa_family, but MS wdm does not understand it,
* so let's use CI_SIN(sa)->sin_family. */
if (CI_SIN(sa)->sin_family != domain &&
CI_SIN(sa)->sin_family != AF_UNSPEC) {
LOG_U(ci_log(LPF "address family %d does not match "
"with socket domain %d", CI_SIN(sa)->sin_family, domain));
RET_WITH_ERRNO(EAFNOSUPPORT);
}
#if CI_CFG_FAKE_IPV6
if (sa->sa_family == AF_INET6 && !ci_tcp_ipv6_is_ipv4(sa)) {
LOG_TC(ci_log(LPF "Pure IPv6 address is not supported"));
RET_WITH_ERRNO(EAFNOSUPPORT);
}
#endif
return 0;
}
/* Kill an orphan stack in the thc
*
* You must hold the thc_mutex before calling this function.
*
* You cannot hold the THR_TABLE.lock when calling this function.
*/
static void thc_kill_an_orphan(tcp_helper_cluster_t* thc)
{
tcp_helper_resource_t* thr = NULL;
int rc;
rc = thc_get_an_orphan(thc, &thr);
ci_assert_equal(rc, 0);
/* This is generally called when the stack is being freed. But as
* we are holding the thc_mutex, we will deadlock if we took that
* path. So we remove thr from the thc now. */
thc_remove_thr(thc, thr);
LOG_U(ci_log("Clustering: Killing orphan stack %d", thr->id));
rc = tcp_helper_kill_stack_by_id(thr->id);
#ifndef NDEBUG
if( rc != 0 && rc != -EBUSY )
LOG_U(ci_log("%s: tcp_helper_kill_stack_by_id(%d): failed %d", __FUNCTION__,
thr->id, rc));
#endif
}
ci_fd_t ci_udp_ep_ctor(citp_socket* ep, ci_netif* netif, int domain, int type)
{
ci_udp_state* us;
ci_fd_t fd;
VERB( log(LPFIN "ctor( )" ) );
ci_assert(ep);
ci_assert(netif);
ci_netif_lock(netif);
us = ci_udp_get_state_buf(netif);
if (!us) {
ci_netif_unlock(netif);
LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__,NI_ID(netif)));
return -ENOMEM;
}
/* It's required to set protocol before ci_tcp_helper_sock_attach()
* since it's used to determine if TCP or UDP file operations should be
* attached to the file descriptor in kernel. */
sock_protocol(&us->s) = IPPROTO_UDP;
/* NB: this attach will close the os_sock_fd */
fd = ci_tcp_helper_sock_attach(ci_netif_get_driver_handle(netif),
SC_SP(&us->s), domain, type);
if( fd < 0 ) {
if( fd == -EAFNOSUPPORT )
LOG_U(ci_log("%s: ci_tcp_helper_sock_attach (domain=%d, type=%d) "
"failed %d", __FUNCTION__, domain, type, fd));
else
LOG_E(ci_log("%s: ci_tcp_helper_sock_attach (domain=%d, type=%d) "
"failed %d", __FUNCTION__, domain, type, fd));
ci_netif_unlock(netif);
return fd;
}
ci_assert(~us->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN);
us->s.rx_errno = 0;
us->s.tx_errno = 0;
us->s.so_error = 0;
us->s.cp.sock_cp_flags |= OO_SCP_UDP_WILD;
ep->s = &us->s;
ep->netif = netif;
CHECK_UEP(ep);
ci_netif_unlock(netif);
return fd;
}
/* 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;
}
ci_fd_t ci_tcp_ep_ctor(citp_socket* ep, ci_netif* netif, int domain, int type)
{
ci_tcp_state* ts;
ci_fd_t fd;
ci_assert(ep);
ci_assert(netif);
ci_netif_lock(netif);
ts = ci_tcp_get_state_buf(netif);
if( ts == NULL ) {
ci_netif_unlock(netif);
LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__,NI_ID(netif)));
return -ENOMEM;
}
fd = ci_tcp_helper_sock_attach(ci_netif_get_driver_handle(netif), S_SP(ts),
domain, type);
if( fd < 0 ) {
if( fd == -EAFNOSUPPORT )
LOG_U(ci_log("%s: ci_tcp_helper_sock_attach" \
"(domain=%d, type=%d) failed %d",
__FUNCTION__, domain, type, fd));
else
LOG_E(ci_log("%s: ci_tcp_helper_sock_attach" \
"(domain=%d, type=%d) failed %d",
__FUNCTION__, domain, type, fd));
}
else {
ci_assert(~ts->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN);
/* Apply default sockbuf sizes now we've updated them from the kernel
** defaults. */
ts->s.so.sndbuf = NI_OPTS(netif).tcp_sndbuf_def;
ts->s.so.rcvbuf = NI_OPTS(netif).tcp_rcvbuf_def;
ep->netif = netif;
ep->s = &ts->s;
CHECK_TEP(ep);
}
ci_netif_unlock(netif);
return fd;
}
int onload_zc_alloc_buffers(int fd, struct onload_zc_iovec* iovecs,
int iovecs_len,
enum onload_zc_buffer_type_flags flags)
{
int rc = 0, i;
citp_lib_context_t lib_context;
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ci_netif* ni;
ci_ip_pkt_fmt *pkt;
unsigned max_len;
Log_CALL(ci_log("%s(%d, %p, %d, %x)", __FUNCTION__, fd, iovecs,
iovecs_len, flags));
citp_enter_lib(&lib_context);
if( (fdi = citp_fdtable_lookup(fd)) != NULL ) {
switch( citp_fdinfo_get_type(fdi) ) {
case CITP_UDP_SOCKET:
case CITP_TCP_SOCKET:
epi = fdi_to_sock_fdi(fdi);
ni = epi->sock.netif;
ci_netif_lock(ni);
for( i = 0; i < iovecs_len; ++i ) {
max_len = CI_CFG_PKT_BUF_SIZE;
pkt = ci_netif_pkt_tx_tcp_alloc(ni);
if( pkt == NULL ) {
while( --i >= 0 )
ci_netif_pkt_release(ni, (ci_ip_pkt_fmt*)iovecs[i].buf);
rc = -ENOMEM;
ci_netif_unlock(ni);
goto out;
}
/* Make sure this is clear as it affects behaviour when freeing */
pkt->pf.udp.rx_flags = 0;
iovecs[i].buf = (struct oo_zc_buf *)pkt;
if( flags & ONLOAD_ZC_BUFFER_HDR_TCP ) {
if( (citp_fdinfo_get_type(fdi) == CITP_TCP_SOCKET) &&
(epi->sock.s->b.state & CI_TCP_STATE_TCP_CONN) ) {
ci_tcp_state* ts = SOCK_TO_TCP(epi->sock.s);
oo_tx_pkt_layout_init(pkt);
iovecs[i].iov_base = ((char *)oo_tx_ip_hdr(pkt)) +
ts->outgoing_hdrs_len;
max_len = tcp_eff_mss(ts);
}
else {
/* Best guess. We can fix it up later. Magic 12 leaves
* space for time stamp option (common case)
*/
oo_tx_pkt_layout_init(pkt);
iovecs[i].iov_base =
(uint8_t*) oo_tx_ip_data(pkt) + sizeof(ci_tcp_hdr) + 12;
}
}
else if( flags & ONLOAD_ZC_BUFFER_HDR_UDP ) {
oo_tx_pkt_layout_init(pkt);
iovecs[i].iov_base =
(uint8_t*) oo_tx_ip_data(pkt) + sizeof(ci_udp_hdr);
}
else
iovecs[i].iov_base = PKT_START(pkt);
iovecs[i].iov_len = CI_CFG_PKT_BUF_SIZE -
((char *)iovecs[i].iov_base - (char *)pkt);
if( iovecs[i].iov_len > max_len )
iovecs[i].iov_len = max_len;
}
ni->state->n_async_pkts += iovecs_len;
ci_netif_unlock(ni);
break;
#if CI_CFG_USERSPACE_EPOLL
case CITP_EPOLL_FD:
rc = -ENOTSOCK;
break;
#endif
#if CI_CFG_USERSPACE_PIPE
case CITP_PIPE_FD:
rc = -ENOTSOCK;
break;
#endif
case CITP_PASSTHROUGH_FD:
rc = -ESOCKTNOSUPPORT;
break;
default:
LOG_U(log("%s: unknown fdinfo type %d", __FUNCTION__,
citp_fdinfo_get_type(fdi)));
rc = -EINVAL;
}
citp_fdinfo_release_ref(fdi, 0);
}
else {
/* Not onload socket */
rc = -ESOCKTNOSUPPORT;
}
out:
citp_exit_lib(&lib_context, TRUE);
Log_CALL_RESULT(rc);
return rc;
}
int onload_zc_release_buffers(int fd, onload_zc_handle* bufs, int bufs_len)
{
int rc = 0, i;
citp_lib_context_t lib_context;
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ci_netif* ni;
Log_CALL(ci_log("%s(%d, %p, %d)", __FUNCTION__, fd, bufs, bufs_len));
citp_enter_lib(&lib_context);
if( (fdi = citp_fdtable_lookup(fd)) != NULL ) {
switch( citp_fdinfo_get_type(fdi) ) {
case CITP_UDP_SOCKET:
case CITP_TCP_SOCKET:
epi = fdi_to_sock_fdi(fdi);
ni = epi->sock.netif;
ci_netif_lock(ni);
for( i = 0; i < bufs_len; ++i ) {
ci_ip_pkt_fmt* pkt = (ci_ip_pkt_fmt*)bufs[i];
if( pkt->stack_id != ni->state->stack_id ) {
LOG_U(log("%s: attempt to free buffer from stack %d to stack %d",
__FUNCTION__, pkt->stack_id, ni->state->stack_id));
rc = -EINVAL;
break;
}
}
if( rc == 0 ) {
for( i = 0; i < bufs_len; ++i )
ci_netif_pkt_release_check_keep(ni, (ci_ip_pkt_fmt*)bufs[i]);
}
ci_netif_unlock(ni);
break;
#if CI_CFG_USERSPACE_EPOLL
case CITP_EPOLL_FD:
rc = -ENOTSOCK;
break;
#endif
#if CI_CFG_USERSPACE_PIPE
case CITP_PIPE_FD:
rc = -ENOTSOCK;
break;
#endif
default:
LOG_U(log("%s: unknown fdinfo type %d", __FUNCTION__,
citp_fdinfo_get_type(fdi)));
rc = -EINVAL;
}
citp_fdinfo_release_ref(fdi, 0);
}
else {
/* Not onload socket */
rc = -ESOCKTNOSUPPORT;
}
citp_exit_lib(&lib_context, TRUE);
Log_CALL_RESULT(rc);
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;
}
static int ci_udp_filter_kernel_pkt(ci_netif* ni, ci_udp_state* us,
struct msghdr* msg, int *bytes)
{
enum onload_zc_callback_rc rc;
struct onload_zc_msg zc_msg;
struct onload_zc_iovec zc_iovec[CI_UDP_ZC_IOVEC_MAX];
unsigned cb_flags = 0;
int i = 0, bytes_remaining = *bytes;
if( msg->msg_iovlen > CI_UDP_ZC_IOVEC_MAX ) {
LOG_U(log("%s: too many fragments (%d), passing packet unfiltered",
__FUNCTION__, (int)msg->msg_iovlen));
return 1;
}
zc_msg.iov = zc_iovec;
zc_msg.msghdr = *msg;
zc_msg.msghdr.msg_iov = NULL;
ci_assert_gt(msg->msg_iovlen, 0);
do {
zc_msg.iov[i].iov_base = msg->msg_iov[i].iov_base;
zc_msg.iov[i].iov_len = msg->msg_iov[i].iov_len > bytes_remaining ?
bytes_remaining : msg->msg_iov[i].iov_len;
zc_msg.iov[i].buf = ONLOAD_ZC_HANDLE_NONZC;
zc_msg.iov[i].iov_flags = 0;
bytes_remaining -= zc_msg.iov[i].iov_len;
} while(++i < msg->msg_iovlen && bytes_remaining);
zc_msg.msghdr.msg_iovlen = i;
rc = (*(onload_zc_recv_filter_callback)((ci_uintptr_t)us->recv_q_filter))
(&zc_msg, (void *)((ci_uintptr_t)us->recv_q_filter_arg), cb_flags);
ci_assert(!(rc & ONLOAD_ZC_KEEP));
if( rc & ONLOAD_ZC_TERMINATE )
return 0;
else {
if( rc & ONLOAD_ZC_MODIFIED ) {
int new_len = 0;
#ifndef NDEBUG
int found_shortened_iov = 0;
#endif
for( i = 0; i < zc_msg.msghdr.msg_iovlen; ++i ) {
new_len += zc_msg.iov[i].iov_len;
#ifndef NDEBUG
if( found_shortened_iov )
ci_assert_equal(zc_msg.iov[i].iov_len, 0);
ci_assert_equal(zc_msg.iov[i].iov_base, msg->msg_iov[i].iov_base);
if( zc_msg.iov[i].iov_len != msg->msg_iov[i].iov_len ) {
ci_assert_lt(zc_msg.iov[i].iov_len, msg->msg_iov[i].iov_len);
found_shortened_iov = 1;
}
#endif
}
#ifndef NDEBUG
if( found_shortened_iov )
ci_assert_lt(new_len, *bytes);
else
ci_assert_equal(new_len, *bytes);
#endif
*bytes = new_len;
}
}
return 1;
}
int onload_zc_release_buffers(int fd, onload_zc_handle* bufs, int bufs_len)
{
int rc = 0, i, rx_pkt, released;
citp_lib_context_t lib_context;
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ci_netif* ni;
ci_ip_pkt_fmt* pkt;
Log_CALL(ci_log("%s(%d, %p, %d)", __FUNCTION__, fd, bufs, bufs_len));
citp_enter_lib(&lib_context);
if( (fdi = citp_fdtable_lookup(fd)) != NULL ) {
switch( citp_fdinfo_get_type(fdi) ) {
case CITP_UDP_SOCKET:
case CITP_TCP_SOCKET:
epi = fdi_to_sock_fdi(fdi);
ni = epi->sock.netif;
ci_netif_lock(ni);
for( i = 0; i < bufs_len; ++i ) {
pkt = (ci_ip_pkt_fmt*)bufs[i];
if( pkt->stack_id != ni->state->stack_id ) {
LOG_U(log("%s: attempt to free buffer from stack %d to stack %d",
__FUNCTION__, pkt->stack_id, ni->state->stack_id));
rc = -EINVAL;
break;
}
}
if( rc == 0 ) {
for( i = 0; i < bufs_len; ++i ) {
pkt = (ci_ip_pkt_fmt*)bufs[i];
/* If we are releasing a packet without the RX_FLAG then the user
* allocated and then freed the packet (without using it).
* We detect this to decrement n_asyn_pkts.
* RX packets (kept via ONLOAD_ZC_KEEP) are counted differently
* so don't decrement here. (But may release)
*/
rx_pkt = pkt->flags & CI_PKT_FLAG_RX;
released = ci_netif_pkt_release_check_keep(ni, pkt);
if ( ! rx_pkt ) {
ci_assert(released == 1);
(void) released;
--ni->state->n_async_pkts;
}
}
}
ci_netif_unlock(ni);
break;
#if CI_CFG_USERSPACE_EPOLL
case CITP_EPOLL_FD:
rc = -ENOTSOCK;
break;
#endif
#if CI_CFG_USERSPACE_PIPE
case CITP_PIPE_FD:
rc = -ENOTSOCK;
break;
#endif
default:
LOG_U(log("%s: unknown fdinfo type %d", __FUNCTION__,
citp_fdinfo_get_type(fdi)));
rc = -EINVAL;
}
citp_fdinfo_release_ref(fdi, 0);
}
else {
/* Not onload socket */
rc = -ESOCKTNOSUPPORT;
}
citp_exit_lib(&lib_context, TRUE);
Log_CALL_RESULT(rc);
return rc;
}
/*
** promote a synrecv structure to an established socket
**
** Assumes that the caller will handle a fail if we can't allocate a new
** tcp_state structure due to memory pressure or the like
*/
int ci_tcp_listenq_try_promote(ci_netif* netif, ci_tcp_socket_listen* tls,
ci_tcp_state_synrecv* tsr,
ci_ip_cached_hdrs* ipcache,
ci_tcp_state** ts_out)
{
int rc = 0;
ci_assert(netif);
ci_assert(tls);
ci_assert(tls->s.b.state == CI_TCP_LISTEN);
ci_assert(tsr);
if( (int) ci_tcp_acceptq_n(tls) < tls->acceptq_max ) {
ci_tcp_state* ts;
/* grab a tcp_state structure that will go onto the accept queue. We take
* from the cache of EPs if any are available
*/
ts = get_ts_from_cache (netif, tsr, tls);
if( !ts ) {
/* None on cache; try allocating a new ts */
ts = ci_tcp_get_state_buf(netif);
#if CI_CFG_FD_CACHING
if( ts == NULL ) {
/* We've reaped. Did this result in any being cached */
ts = get_ts_from_cache(netif, tsr, tls);
if (ts == NULL ) {
/* No -- try again to allocate. */
ts = ci_tcp_get_state_buf(netif);
}
else {
CITP_STATS_NETIF(++netif->state->stats.sockcache_hit_reap);
}
}
#endif
if( ts == NULL ) {
LOG_TV(ci_log("%s: [%d] out of socket buffers",
__FUNCTION__, NI_ID(netif)));
CITP_STATS_TCP_LISTEN(++tls->stats.n_acceptq_no_sock);
CI_SET_SO_ERROR(&tls->s, ENOMEM);
citp_waitable_wake(netif, &tls->s.b, CI_SB_FLAG_WAKE_RX);
return -ENOMEM;
}
ci_assert(ci_tcp_is_cached(ts) ||
(ts->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN));
}
#ifdef ONLOAD_OFE
ts->s.ofe_code_start = tls->ofe_promote;
#endif
if( ! ci_tcp_is_cached(ts) ) {
/* Need to initialise address information for use when setting filters */
ci_tcp_set_addr_on_promote(netif, ts, tsr, tls);
/* "borrow" filter from listening socket. For loopback socket, we
* do not need filters, but we have to take a reference of the OS
* socket. */
rc = ci_tcp_ep_set_filters(netif, S_SP(ts), ts->s.cp.so_bindtodevice,
S_SP(tls));
if( rc < 0 ) {
LOG_U(ci_log("%s: Unable to set filters %d", __FUNCTION__, rc));
/* Either put this back on the list (at the head) or free it */
ci_tcp_state_free(netif, ts);
return rc;
}
}
#if CI_CFG_FD_CACHING
else {
/* Now set the s/w filter. We leave the hw filter in place for cached
* EPS. This will probably not have the correct raddr and rport, but as
* it's sharing the listening socket's filter that's not a problem. It
* will be updated if this is still around when the listener is closed.
*/
rc = ci_netif_filter_insert(netif, S_SP(ts), tsr->l_addr,
sock_lport_be16(&tls->s), tsr->r_addr,
tsr->r_port, tcp_protocol(ts));
if (rc < 0) {
/* Bung it back on the cache list */
LOG_EP(ci_log("Unable to create s/w filter!"));
ci_ni_dllist_push(netif, &tls->epcache.cache, &ts->epcache_link);
return rc;
}
/* Need to initialise address information. We do this after trying to
* insert the sw filter, so we can push the tcp state back onto the
* cache queue with as few changes as possible if we fail to add the
* sw filter.
*/
ci_tcp_set_addr_on_promote(netif, ts, tsr, tls);
LOG_EP(ci_log("Cached fd %d from cached to connected", ts->cached_on_fd));
ci_ni_dllist_push(netif, &tls->epcache_connected, &ts->epcache_link);
}
#endif
ci_assert(IS_VALID_SOCK_P(netif, S_SP(ts)));
ci_assert(ts->s.b.state == CI_TCP_CLOSED);
ts->s.domain = tls->s.domain;
cicp_ip_cache_update_from(netif, &ts->s.pkt, ipcache);
ci_pmtu_state_init(netif, &ts->s, &ts->pmtus,
CI_IP_TIMER_PMTU_DISCOVER);
ci_pmtu_set(netif, &ts->pmtus,
CI_MIN(ts->s.pkt.mtu,
tsr->tcpopts.smss + sizeof(ci_tcp_hdr)
+ sizeof(ci_ip4_hdr)));
/* If we've got SYN via local route, we can handle it */
ci_assert_equiv(ts->s.pkt.status == retrrc_localroute,
OO_SP_NOT_NULL(tsr->local_peer));
if( ts->s.pkt.status == retrrc_localroute )
ts->s.pkt.flags |= CI_IP_CACHE_IS_LOCALROUTE;
ts->amss = tsr->amss;
/* options and flags */
ts->tcpflags = 0;
ts->tcpflags |= tsr->tcpopts.flags;
ts->tcpflags |= CI_TCPT_FLAG_PASSIVE_OPENED;
ts->outgoing_hdrs_len = sizeof(ci_ip4_hdr) + sizeof(ci_tcp_hdr);
if( ts->tcpflags & CI_TCPT_FLAG_WSCL ) {
ts->snd_wscl = tsr->tcpopts.wscl_shft;
ts->rcv_wscl = tsr->rcv_wscl;
} else {
ts->snd_wscl = ts->rcv_wscl = 0u;
}
CI_IP_SOCK_STATS_VAL_TXWSCL( ts, ts->snd_wscl);
CI_IP_SOCK_STATS_VAL_RXWSCL( ts, ts->rcv_wscl);
/* Send and receive sequence numbers */
tcp_snd_una(ts) = tcp_snd_nxt(ts) = tcp_enq_nxt(ts) = tcp_snd_up(ts) =
tsr->snd_isn + 1;
ci_tcp_set_snd_max(ts, tsr->rcv_nxt, tcp_snd_una(ts), 0);
ci_tcp_rx_set_isn(ts, tsr->rcv_nxt);
tcp_rcv_up(ts) = SEQ_SUB(tcp_rcv_nxt(ts), 1);
if( ts->tcpflags & CI_TCPT_FLAG_TSO ) {
ts->incoming_tcp_hdr_len += 12;
ts->outgoing_hdrs_len += 12;
ts->tspaws = ci_tcp_time_now(netif);
ts->tsrecent = tsr->tspeer;
ts->tslastack = tsr->rcv_nxt;
}
else {
/* Must be after initialising snd_una. */
ci_tcp_clear_rtt_timing(ts);
ts->timed_ts = tsr->timest;
}
/* SACK has nothing to be done. */
/* ?? ECN */
ci_tcp_set_hdr_len(ts, (ts->outgoing_hdrs_len - sizeof(ci_ip4_hdr)));
ts->smss = tsr->tcpopts.smss;
ts->c.user_mss = tls->c.user_mss;
if (ts->c.user_mss && ts->c.user_mss < ts->smss)
ts->smss = ts->c.user_mss;
#if CI_CFG_LIMIT_SMSS
ts->smss = ci_tcp_limit_mss(ts->smss, netif, __FUNCTION__);
#endif
ci_assert(ts->smss>0);
ci_tcp_set_eff_mss(netif, ts);
ci_tcp_set_initialcwnd(netif, ts);
/* Copy socket options & related fields that should be inherited.
* Note: Windows does not inherit rcvbuf until the call to accept
* completes. The assumption here is that all options can be
* inherited at the same time (most won't have an effect until there
* is a socket available for use by the app.).
*/
ci_tcp_inherit_accept_options(netif, tls, ts, "SYN RECV (LISTENQ PROMOTE)");
/* NB. Must have already set peer (which we have). */
ci_tcp_set_established_state(netif, ts);
CITP_STATS_NETIF(++netif->state->stats.synrecv2established);
ci_assert(ts->ka_probes == 0);
ci_tcp_kalive_restart(netif, ts, ci_tcp_kalive_idle_get(ts));
ci_tcp_set_flags(ts, CI_TCP_FLAG_ACK);
/* Remove the synrecv structure from the listen queue, and free the
** buffer. */
if( tsr->tcpopts.flags & CI_TCPT_FLAG_SYNCOOKIE )
ci_free(tsr);
else {
ci_tcp_listenq_remove(netif, tls, tsr);
ci_tcp_synrecv_free(netif, tsr);
}
ci_bit_set(&ts->s.b.sb_aflags, CI_SB_AFLAG_TCP_IN_ACCEPTQ_BIT);
ci_tcp_acceptq_put(netif, tls, &ts->s.b);
LOG_TC(log(LNT_FMT "new ts=%d SYN-RECV->ESTABLISHED flags=0x%x",
LNT_PRI_ARGS(netif, tls), S_FMT(ts), ts->tcpflags);
log(LNTS_FMT RCV_WND_FMT " snd=%08x-%08x-%08x enq=%08x",
LNTS_PRI_ARGS(netif, ts), RCV_WND_ARGS(ts),
tcp_snd_una(ts),
tcp_snd_nxt(ts), ts->snd_max, tcp_enq_nxt(ts)));
citp_waitable_wake(netif, &tls->s.b, CI_SB_FLAG_WAKE_RX);
*ts_out = ts;
return 0;
}
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
}
