ci_inline ci_tcp_state*
get_ts_from_cache(ci_netif *netif,
ci_tcp_state_synrecv* tsr,
ci_tcp_socket_listen* tls)
{
ci_tcp_state *ts = NULL;
#if CI_CFG_FD_CACHING
if( ci_ni_dllist_not_empty(netif, &tls->epcache.cache) ) {
/* Take the entry from the cache */
ci_ni_dllist_link *link = ci_ni_dllist_pop(netif, &tls->epcache.cache);
ts = CI_CONTAINER (ci_tcp_state, epcache_link, link);
ci_assert (ts);
ci_ni_dllist_self_link(netif, &ts->epcache_link);
LOG_EP(ci_log("Taking cached fd %d off cached list, (onto acceptq)",
ts->cached_on_fd));
if( tcp_laddr_be32(ts) == tsr->l_addr ) {
ci_tcp_state_init(netif, ts, 1);
/* Shouldn't have touched these bits of state */
ci_assert(!(ts->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN));
ci_assert(ci_tcp_is_cached(ts));
CITP_STATS_NETIF(++netif->state->stats.sockcache_hit);
CITP_STATS_TCP_LISTEN(++tls->stats.n_sockcache_hit);
}
else {
/* Oh dear -- the tcp-state we cached was using a different local IP
* address. This means we've accepted a connection from a different
* interface as we did for the thing we've cached. Which means we
* can't share the hardware filter after all. For now, just bung it
* back on the list.
*/
LOG_EP(ci_log("changed interface of cached EP, re-queueing"));
ci_ni_dllist_push_tail(netif, &tls->epcache.cache, &ts->epcache_link);
ts = NULL;
CITP_STATS_NETIF(++netif->state->stats.sockcache_miss_intmismatch);
}
}
#endif
return ts;
}
static int ci_tcp_listen_init(ci_netif *ni, ci_tcp_socket_listen *tls)
{
int i;
oo_p sp;
tls->acceptq_n_in = tls->acceptq_n_out = 0;
tls->acceptq_put = CI_ILL_END;
tls->acceptq_get = OO_SP_NULL;
tls->n_listenq = 0;
tls->n_listenq_new = 0;
/* Allocate and initialise the listen bucket */
if( OO_P_IS_NULL(ni->state->free_aux_mem) )
return -ENOBUFS;
tls->bucket = ni->state->free_aux_mem;
ci_tcp_bucket_alloc(ni);
tls->n_buckets = 1;
/* Initialise the listenQ. */
for( i = 0; i <= CI_CFG_TCP_SYNACK_RETRANS_MAX; ++i ) {
sp = TS_OFF(ni, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, listenq[i]));
ci_ni_dllist_init(ni, &tls->listenq[i], sp, "lstq");
}
/* Initialize the cache and pending lists for the EP-cache.
* See comment at definition for details
*/
LOG_EP (log ("Initialise cache and pending list for id %d",
S_FMT(tls)));
#if CI_CFG_FD_CACHING
sp = TS_OFF(ni, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, epcache_cache));
ci_ni_dllist_init(ni, &tls->epcache_cache, sp, "epch");
sp = TS_OFF(ni, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, epcache_pending));
ci_ni_dllist_init(ni, &tls->epcache_pending, sp, "eppd");
sp = TS_OFF(ni, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, epcache_connected));
ci_ni_dllist_init(ni, &tls->epcache_connected, sp, "epco");
sp = TS_OFF(ni, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, epcache_fd_states));
ci_ni_dllist_init(ni, &tls->epcache_fd_states, sp, "ecfd");
tls->cache_avail_sock = ni->state->opts.per_sock_cache_max;
#endif
return 0;
}
/*
** 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;
}
