Example #1
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;
  }
Example #2
0
/* c_ni is assumed to be locked on enterance and is always unlocked on
 * exit. */
int ci_tcp_connect_lo_toconn(ci_netif *c_ni, oo_sp c_id, ci_uint32 dst,
                             ci_netif *l_ni, oo_sp l_id)
{
  ci_tcp_state *ts;
  ci_tcp_socket_listen *tls, *alien_tls;
  citp_waitable_obj *wo;
  citp_waitable *w;
  int rc;

  ci_assert(ci_netif_is_locked(c_ni));
  ci_assert(OO_SP_NOT_NULL(c_id));
  ci_assert(OO_SP_NOT_NULL(l_id));

  LOG_TC(log("%s: connect %d:%d to %d:%d", __FUNCTION__,
             c_ni->state->stack_id, OO_SP_TO_INT(c_id),
             l_ni->state->stack_id, OO_SP_TO_INT(l_id)));

  alien_tls = SP_TO_TCP_LISTEN(l_ni, l_id);
  if( (int)ci_tcp_acceptq_n(alien_tls) >= alien_tls->acceptq_max ) {
    ci_netif_unlock(c_ni);
    return -EBUSY;
  }

  /* In c_ni, create shadow listening socket tls (copy l_id) */
  ts = ci_tcp_get_state_buf(c_ni);
  if( ts == NULL ) {
    ci_netif_unlock(c_ni);
    LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__, NI_ID(c_ni)));
    return -ENOMEM;
  }

  /* init common tcp fields */
  ts->s.so = alien_tls->s.so;
  ts->s.cp.ip_ttl = alien_tls->s.cp.ip_ttl;
  S_TCP_HDR(&ts->s)->tcp_source_be16 =
      S_TCP_HDR(&alien_tls->s)->tcp_source_be16;
  ts->s.domain = alien_tls->s.domain;
  ts->c = alien_tls->c;
  ts->c.tcp_defer_accept = OO_TCP_DEFER_ACCEPT_OFF;

  /* make sure nobody will ever connect to our "shadow" socket
   * except us */
  ci_bit_set(&ts->s.b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);

  ci_tcp_set_slow_state(c_ni, ts, CI_TCP_LISTEN);
  tls = SOCK_TO_TCP_LISTEN(&ts->s);
  /* no timer: */
  tls->s.s_flags = alien_tls->s.s_flags | CI_SOCK_FLAG_BOUND_ALIEN;

  tls->acceptq_max = 1;
  rc = ci_tcp_listen_init(c_ni, tls);
  if( rc != 0 ) {
    citp_waitable_obj_free(c_ni, &tls->s.b);
    return rc;
  }

  /* Connect c_id to tls */
  ts = SP_TO_TCP(c_ni, c_id);
  rc = ci_tcp_connect_lo_samestack(c_ni, ts, tls->s.b.bufid);

  /* Accept as from tls */
  if( !ci_tcp_acceptq_not_empty(tls) ) {
    /* it is possible, for example, if ci_tcp_listenq_try_promote() failed
     * because there are no endpoints */
    ci_tcp_listenq_drop_all(c_ni, tls);
    citp_waitable_obj_free(c_ni, &tls->s.b);
    ci_netif_unlock(c_ni);
    return -EBUSY;
  }
  w = ci_tcp_acceptq_get(c_ni, tls);
  ci_assert(w);
  LOG_TV(ci_log("%s: %d:%d to %d:%d shadow %d:%d accepted %d:%d",
                __FUNCTION__,
                c_ni->state->stack_id, OO_SP_TO_INT(c_id),
                l_ni->state->stack_id, OO_SP_TO_INT(l_id),
                c_ni->state->stack_id, tls->s.b.bufid,
                c_ni->state->stack_id, w->bufid));

  ci_assert(w->state & CI_TCP_STATE_TCP);
  ci_assert(w->state != CI_TCP_LISTEN);

  /* Destroy tls.
   * NB: nobody could possibly connect to it, so no need to do proper
   * shutdown.
   */
  ci_assert_equal(ci_tcp_acceptq_n(tls), 0);
  ci_tcp_listenq_drop_all(c_ni, tls);
  citp_waitable_obj_free(c_ni, &tls->s.b);
  ci_netif_unlock(c_ni);

  /* Keep a port reference */
  {
    tcp_helper_endpoint_t *l_ep, *a_ep;
    struct oo_file_ref* os_sock_ref;
    ci_irqlock_state_t lock_flags;

    l_ep = ci_trs_ep_get(netif2tcp_helper_resource(l_ni), l_id);
    a_ep = ci_trs_ep_get(netif2tcp_helper_resource(c_ni), W_SP(w));
    ci_irqlock_lock(&l_ep->thr->lock, &lock_flags);
    os_sock_ref = l_ep->os_socket;
    ci_assert_equal(a_ep->os_port_keeper, NULL);
    if( os_sock_ref != NULL ) {
      os_sock_ref = oo_file_ref_add(os_sock_ref);
      os_sock_ref = oo_file_ref_xchg(&a_ep->os_port_keeper, os_sock_ref);
      ci_irqlock_unlock(&l_ep->thr->lock, &lock_flags);
      if( os_sock_ref != NULL )
        oo_file_ref_drop(os_sock_ref);
    }
    else {
      ci_irqlock_unlock(&l_ep->thr->lock, &lock_flags);
      goto cleanup;
    }
  }

  /* lock l_ni: Check that l_id is the same socket it used to be */
  /* create ref-sock in l_ni, put it into acc q */
  if( ci_netif_lock(l_ni) != 0 )
    goto cleanup;
  if( alien_tls->s.b.state != CI_TCP_LISTEN ||
      (alien_tls->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN) ||
      S_TCP_HDR(&alien_tls->s)->tcp_source_be16 != TS_TCP(ts)->tcp_dest_be16 ||
      (alien_tls->s.pkt.ip.ip_saddr_be32 != INADDR_ANY &&
       alien_tls->s.pkt.ip.ip_saddr_be32 != ts->s.pkt.ip.ip_daddr_be32) ) {
    ci_netif_unlock(l_ni);
    goto cleanup;
  }

  ci_bit_mask_set(&w->sb_aflags,
                  CI_SB_AFLAG_TCP_IN_ACCEPTQ | CI_SB_AFLAG_ORPHAN);

  wo = citp_waitable_obj_alloc(l_ni);
  if( wo == NULL ) {
    ci_netif_unlock(l_ni);
    goto cleanup;
  }
  wo->waitable.state = CI_TCP_CLOSED;
  wo->waitable.sb_aflags |= CI_SB_AFLAG_MOVED_AWAY;
  wo->waitable.moved_to_stack_id = c_ni->state->stack_id;
  wo->waitable.moved_to_sock_id = W_SP(w);
  LOG_TC(log("%s: put to acceptq %d:%d referencing %d:%d", __func__,
             l_ni->state->stack_id, OO_SP_TO_INT(W_SP(&wo->waitable)),
             c_ni->state->stack_id, OO_SP_TO_INT(W_SP(w))));

  ci_tcp_acceptq_put(l_ni, alien_tls, &wo->waitable);
  citp_waitable_wake_not_in_poll(l_ni, &alien_tls->s.b, CI_SB_FLAG_WAKE_RX);
  ci_netif_unlock(l_ni);

  return rc;

cleanup:
  ci_assert(w->sb_aflags & CI_SB_AFLAG_ORPHAN);
  ci_bit_mask_clear(&w->sb_aflags,
                    CI_SB_AFLAG_TCP_IN_ACCEPTQ | CI_SB_AFLAG_ORPHAN);
  efab_tcp_helper_close_endpoint(netif2tcp_helper_resource(c_ni), w->bufid);
  /* we can not guarantee c_ni lock, so we can' call
   * ci_tcp_drop(c_ni, ts).  So, we return error; UL will handover
   * and close ts endpoint. */
  return -EBUSY;
}