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;
}
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;
}
