void ci_sock_cmn_timestamp_q_enqueue(ci_netif* ni, ci_sock_cmn* s,
ci_ip_pkt_fmt* pkt)
{
ci_ip_pkt_queue* qu = &s->timestamp_q;
oo_pkt_p prev_head = qu->head;
/* This part is effectively ci_ip_queue_enqueue(ni, &s->timestamp_q, p);
* but inlined to allow using tsq_next field
*/
pkt->tsq_next = OO_PP_NULL;
if( ci_ip_queue_is_empty(qu) ) {
ci_assert(OO_PP_IS_NULL(qu->head));
qu->head = OO_PKT_P(pkt);
}
else {
ci_assert(OO_PP_NOT_NULL(qu->head));
/* This assumes the netif lock is held, so use
ci_ip_queue_enqueue_nnl() if it's not */
PKT(ni, qu->tail)->tsq_next = OO_PKT_P(pkt);
}
qu->tail = OO_PKT_P(pkt);
qu->num++;
if( OO_PP_IS_NULL(prev_head) ) {
ci_assert(OO_PP_IS_NULL(s->timestamp_q_extract));
s->timestamp_q_extract = qu->head;
}
else {
ci_sock_cmn_timestamp_q_reap(ni, s);
}
/* Tells post-poll loop to put socket on the [reap_list]. */
s->b.sb_flags |= CI_SB_FLAG_RX_DELIVERED;
}
int ci_ip_send_pkt_send(ci_netif* ni, ci_ip_pkt_fmt* pkt,
const ci_ip_cached_hdrs* ipcache)
{
int os_rc = 0;
switch( ipcache->status ) {
case retrrc_success:
ci_ip_set_mac_and_port(ni, ipcache, pkt);
ci_netif_send(ni, pkt);
return 0;
case retrrc_nomac:
cicp_user_defer_send(ni, retrrc_nomac, &os_rc, OO_PKT_P(pkt),
ipcache->ifindex);
return 0;
case retrrc_noroute:
return -EHOSTUNREACH;
case retrrc_alienroute:
return -ENETUNREACH;
case retrrc_localroute:
if( ipcache->flags & CI_IP_CACHE_IS_LOCALROUTE )
ci_assert(0);
/*passthrough*/
default:
if( ipcache->status < 0 )
return ipcache->status;
else
/* belt and braces... */
return 0;
}
}
/**
* Put an IP_PKTINFO control message into msg ancillary data buffer.
*/
static void ip_cmsg_recv_pktinfo(ci_netif* netif, const ci_ip_pkt_fmt* pkt,
struct cmsg_state *cmsg_state)
{
/* TODO: This is horribly inefficient -- two system calls. Could be made
* cheap with a user-level llap table.
*/
struct in_pktinfo info;
ci_uint32 addr;
int hwport;
addr = oo_ip_hdr_const(pkt)->ip_daddr_be32;
info.ipi_addr.s_addr = addr;
/* Set the ifindex the pkt was received at. */
{
ci_ifid_t ifindex = 0;
int rc = 0;
hwport = netif->state->intf_i_to_hwport[pkt->intf_i];
rc = cicp_llap_find(CICP_HANDLE(netif), &ifindex,
CI_HWPORT_ID(hwport), pkt->vlan);
if( rc != 0 )
LOG_E(ci_log("%s: cicp_llap_find(intf_i=%d, hwport=%d) failed rc=%d",
__FUNCTION__, pkt->intf_i, hwport, rc));
info.ipi_ifindex = ifindex;
}
/* RFC1122: The specific-destination address is defined to be the
* destination address in the IP header unless the header contains a
* broadcast or multicast address, in which case the specific-destination
* is an IP address assigned to the physical interface on which the
* datagram arrived. */
/*\ FIXME: we should drop the packet if this call fails */
cicp_ipif_pktinfo_query(CICP_HANDLE(netif), netif, OO_PKT_P(pkt),
info.ipi_ifindex,
&info.ipi_spec_dst.s_addr
);
ci_put_cmsg(cmsg_state, IPPROTO_IP, IP_PKTINFO, sizeof(info), &info);
}
void ci_ip_send_tcp_slow(ci_netif* ni, ci_tcp_state* ts, ci_ip_pkt_fmt* pkt)
{
/* We're here because the ipcache is not valid. */
int rc, prev_mtu = ts->s.pkt.mtu;
cicp_user_retrieve(ni, &ts->s.pkt, &ts->s.cp);
if( ts->s.pkt.status == retrrc_success ) {
if( ts->s.pkt.mtu != prev_mtu )
CI_PMTU_TIMER_NOW(ni, &ts->pmtus);
ci_ip_set_mac_and_port(ni, &ts->s.pkt, pkt);
ci_netif_send(ni, pkt);
return;
}
else if( ts->s.pkt.status == retrrc_localroute &&
(ts->s.pkt.flags & CI_IP_CACHE_IS_LOCALROUTE) )
ci_ip_local_send(ni, pkt, &ts->s, OO_SP_NULL);
/* For TCP, we want the ipcache to only be valid when onloadable. */
ci_ip_cache_invalidate(&ts->s.pkt);
switch( ts->s.pkt.status ) {
case retrrc_nomac:
rc = 0;
/* If we resend SYN, and there is no MAC - it means ARP failed.
* Connect() should return with EHOSTUNREACH.
* We verify twice - on the first and the second retransmit.
* Very hackish.
*/
if( ts->s.b.state == CI_TCP_SYN_SENT ) {
if( ts->retransmits == 1 )
ts->tcpflags |= CI_TCPT_FLAG_NO_ARP;
else if( (ts->tcpflags & CI_TCPT_FLAG_NO_ARP) &&
ts->retransmits == 2 ) {
ci_tcp_drop(ni, ts, EHOSTUNREACH);
return;
}
}
cicp_user_defer_send(ni, retrrc_nomac, &rc, OO_PKT_P(pkt),
ts->s.pkt.ifindex);
++ts->stats.tx_nomac_defer;
return;
case retrrc_noroute:
rc = -EHOSTUNREACH;
break;
case retrrc_alienroute:
case retrrc_localroute:
/* ?? TODO: inc some stat */
return;
default:
ci_assert_lt(ts->s.pkt.status, 0);
if( ts->s.pkt.status < 0 )
rc = ts->s.pkt.status;
else
/* belt and braces... */
rc = 0;
}
ci_assert_le(rc, 0);
/* In most cases, we should ignore return code; the packet will be resend
* later, because of RTO. However, in SYN-SENT we should pass errors to
* user. At the same time, we should not pass ENOBUFS to user - it is
* pretty internal problem of cplane, so we should try again. Possibly,
* there may be other internal problems, such as ENOMEM.
*
* Also, do not break connection when the first SYN fails:
* - Linux does not do it;
* - cplane has some latency, so we have false positives here;
* - ci_tcp_connect() does not expect it.
*/
if( ts->s.b.state == CI_TCP_SYN_SENT && rc < 0 && ts->retransmits > 0 &&
(rc == -EHOSTUNREACH || rc == -ENETUNREACH || rc == -ENETDOWN) )
ci_tcp_drop(ni, ts, -rc);
}
