gmlnd_comm.c

非常经典的一个分布式系统

                        spin_unlock(&gmni->gmni_tx_lock);

                        /* Unlocking here allows sends to get re-ordered,
                         * but we want to allow other CPUs to progress... */

                        tx->tx_ltxb = ltxb;

                        /* marshall message in tx_ltxb...
                         * 1. Copy what was marshalled so far (in tx_buf) */
                        memcpy(GMNAL_NETBUF_MSG(&ltxb->txb_buf),
                               GMNAL_NETBUF_MSG(&tx->tx_buf), tx->tx_msgnob);

                        /* 2. Copy the payload */
                        if (tx->tx_large_iskiov)
                                lnet_copy_kiov2kiov(
                                        gmni->gmni_large_pages,
                                        ltxb->txb_buf.nb_kiov,
                                        tx->tx_msgnob,
                                        tx->tx_large_niov,
                                        tx->tx_large_frags.kiov,
                                        tx->tx_large_offset,
                                        tx->tx_large_nob);
                        else
                                lnet_copy_iov2kiov(
                                        gmni->gmni_large_pages,
                                        ltxb->txb_buf.nb_kiov,
                                        tx->tx_msgnob,
                                        tx->tx_large_niov,
                                        tx->tx_large_frags.iov,
                                        tx->tx_large_offset,
                                        tx->tx_large_nob);

                        tx->tx_msgnob += tx->tx_large_nob;

                        spin_lock(&gmni->gmni_tx_lock);
                }

                list_add_tail(&tx->tx_list, &gmni->gmni_cred_txq);
        }

        if (!list_empty(&gmni->gmni_cred_txq) &&
            gmni->gmni_tx_credits != 0) {

                tx = list_entry(gmni->gmni_cred_txq.next, gmnal_tx_t, tx_list);

                /* consume tx and 1 credit */
                list_del(&tx->tx_list);
                gmni->gmni_tx_credits--;

                spin_unlock(&gmni->gmni_tx_lock);

                /* Unlocking here allows sends to get re-ordered, but we want
                 * to allow other CPUs to progress... */

                LASSERT(!tx->tx_credit);
                tx->tx_credit = 1;

                tx->tx_launchtime = jiffies;

                if (tx->tx_msgnob <= gmni->gmni_small_msgsize) {
                        LASSERT (tx->tx_ltxb == NULL);
                        netaddr = GMNAL_NETBUF_LOCAL_NETADDR(&tx->tx_buf);
                        gmsize = gmni->gmni_small_gmsize;
                        pri = GMNAL_SMALL_PRIORITY;
                } else {
                        LASSERT (tx->tx_ltxb != NULL);
                        netaddr = GMNAL_NETBUF_LOCAL_NETADDR(&tx->tx_ltxb->txb_buf);
                        gmsize = gmni->gmni_large_gmsize;
                        pri = GMNAL_LARGE_PRIORITY;
                }

                spin_lock(&gmni->gmni_gm_lock);

                gm_send_to_peer_with_callback(gmni->gmni_port, 
                                              netaddr, gmsize, 
                                              tx->tx_msgnob,
                                              pri, 
                                              tx->tx_gmlid,
                                              gmnal_tx_callback, 
                                              (void*)tx);

                spin_unlock(&gmni->gmni_gm_lock);
                spin_lock(&gmni->gmni_tx_lock);
        }
}

void
gmnal_post_rx(gmnal_ni_t *gmni, gmnal_rx_t *rx)
{
        int   gmsize = rx->rx_islarge ? gmni->gmni_large_gmsize :
                                        gmni->gmni_small_gmsize;
        int   pri    = rx->rx_islarge ? GMNAL_LARGE_PRIORITY :
                                        GMNAL_SMALL_PRIORITY;
        void *buffer = GMNAL_NETBUF_LOCAL_NETADDR(&rx->rx_buf);

	CDEBUG(D_NET, "posting rx %p buf %p\n", rx, buffer);

	spin_lock(&gmni->gmni_gm_lock);
        gm_provide_receive_buffer_with_tag(gmni->gmni_port, 
                                           buffer, gmsize, pri, 0);
	spin_unlock(&gmni->gmni_gm_lock);
}

void
gmnal_version_reply (gmnal_ni_t *gmni, gmnal_rx_t *rx)
{
        /* Future protocol version compatibility support!
         * The next gmlnd-specific protocol rev will first send a message to
         * check version; I reply with a stub message containing my current
         * magic+version... */
        gmnal_msg_t *msg;
        gmnal_tx_t  *tx = gmnal_get_tx(gmni);

        if (tx == NULL) {
                CERROR("Can't allocate tx to send version info to %u\n",
                       rx->rx_recv_gmid);
                return;
        }

        LASSERT (tx->tx_lntmsg == NULL);        /* no finalize */

        tx->tx_nid = LNET_NID_ANY;
        tx->tx_gmlid = rx->rx_recv_gmid;

        msg = GMNAL_NETBUF_MSG(&tx->tx_buf);
        msg->gmm_magic   = GMNAL_MSG_MAGIC;
        msg->gmm_version = GMNAL_MSG_VERSION;

        /* just send magic + version */
        tx->tx_msgnob = offsetof(gmnal_msg_t, gmm_type);
        tx->tx_large_nob = 0;

        spin_lock(&gmni->gmni_tx_lock);

        list_add_tail(&tx->tx_list, &gmni->gmni_buf_txq);
        gmnal_check_txqueues_locked(gmni);

        spin_unlock(&gmni->gmni_tx_lock);
}

int
gmnal_rx_thread(void *arg)
{
	gmnal_ni_t	*gmni = arg;
	gm_recv_event_t	*rxevent = NULL;
	gm_recv_t	*recv = NULL;
        gmnal_rx_t      *rx;
        int              rc;

	cfs_daemonize("gmnal_rxd");

        down(&gmni->gmni_rx_mutex);

	while (!gmni->gmni_shutdown) {

                spin_lock(&gmni->gmni_gm_lock);
		rxevent = gm_blocking_receive_no_spin(gmni->gmni_port);
                spin_unlock(&gmni->gmni_gm_lock);

                switch (GM_RECV_EVENT_TYPE(rxevent)) {
                default:
                        gm_unknown(gmni->gmni_port, rxevent);
                        continue;

                case GM_FAST_RECV_EVENT:
                case GM_FAST_PEER_RECV_EVENT:
                case GM_PEER_RECV_EVENT:
                case GM_FAST_HIGH_RECV_EVENT:
                case GM_FAST_HIGH_PEER_RECV_EVENT:
                case GM_HIGH_PEER_RECV_EVENT:
                case GM_RECV_EVENT:
                case GM_HIGH_RECV_EVENT:
                        break;
                }
                
                recv = &rxevent->recv;
                rx = gm_hash_find(gmni->gmni_rx_hash, 
                                  gm_ntohp(recv->buffer));
                LASSERT (rx != NULL);

                rx->rx_recv_nob  = gm_ntoh_u32(recv->length);
                rx->rx_recv_gmid = gm_ntoh_u16(recv->sender_node_id);
                rx->rx_recv_port = gm_ntoh_u8(recv->sender_port_id);
                rx->rx_recv_type = gm_ntoh_u8(recv->type);

                switch (GM_RECV_EVENT_TYPE(rxevent)) {
                case GM_FAST_RECV_EVENT:
                case GM_FAST_PEER_RECV_EVENT:
                case GM_FAST_HIGH_RECV_EVENT:
                case GM_FAST_HIGH_PEER_RECV_EVENT:
                        LASSERT (rx->rx_recv_nob <= PAGE_SIZE);

                        memcpy(GMNAL_NETBUF_MSG(&rx->rx_buf),
                               gm_ntohp(recv->message), rx->rx_recv_nob);
                        break;
                }

                up(&gmni->gmni_rx_mutex);

                CDEBUG (D_NET, "rx %p: buf %p(%p) nob %d\n", rx, 
                        GMNAL_NETBUF_LOCAL_NETADDR(&rx->rx_buf),
                        gm_ntohp(recv->buffer), rx->rx_recv_nob);

                /* We're connectionless: simply drop packets with
                 * errors */
                rc = gmnal_unpack_msg(gmni, rx);

                if (rc == 0) {
                        gmnal_msg_t *msg = GMNAL_NETBUF_MSG(&rx->rx_buf);
                        
                        LASSERT (msg->gmm_type == GMNAL_MSG_IMMEDIATE);
                        rc =  lnet_parse(gmni->gmni_ni, 
                                         &msg->gmm_u.immediate.gmim_hdr,
                                         msg->gmm_srcnid,
                                         rx, 0);
                } else if (rc > 0) {
                        gmnal_version_reply(gmni, rx);
                        rc = -EPROTO;           /* repost rx */
                }

                if (rc < 0)                     /* parse failure */
                        gmnal_post_rx(gmni, rx);

                down(&gmni->gmni_rx_mutex);
	}

        up(&gmni->gmni_rx_mutex);

	CDEBUG(D_NET, "exiting\n");
        atomic_dec(&gmni->gmni_nthreads);
	return 0;
}

void
gmnal_stop_threads(gmnal_ni_t *gmni)
{
        int count = 2;

        gmni->gmni_shutdown = 1;
        mb();
        
        /* wake rxthread owning gmni_rx_mutex with an alarm. */
	spin_lock(&gmni->gmni_gm_lock);
	gm_set_alarm(gmni->gmni_port, &gmni->gmni_alarm, 0, NULL, NULL);
	spin_unlock(&gmni->gmni_gm_lock);

	while (atomic_read(&gmni->gmni_nthreads) != 0) {
                count++;
                if ((count & (count - 1)) == 0)
                        CWARN("Waiting for %d threads to stop\n",
                              atomic_read(&gmni->gmni_nthreads));
                gmnal_yield(1);
	}
}

int
gmnal_start_threads(gmnal_ni_t *gmni)
{
        int     i;
        int     pid;

        LASSERT (!gmni->gmni_shutdown);
        LASSERT (atomic_read(&gmni->gmni_nthreads) == 0);

	gm_initialize_alarm(&gmni->gmni_alarm);

	for (i = 0; i < num_online_cpus(); i++) {

                pid = kernel_thread(gmnal_rx_thread, (void*)gmni, 0);
                if (pid < 0) {
                        CERROR("rx thread failed to start: %d\n", pid);
                        gmnal_stop_threads(gmni);
                        return pid;
                }

                atomic_inc(&gmni->gmni_nthreads);
	}

	return 0;
}