LU-10185 gnilnd: SLES 12 SP2 support and cleanup

[fs/lustre-release.git] / lnet / klnds / gnilnd / gnilnd.c

diff --git a/lnet/klnds/gnilnd/gnilnd.c b/lnet/klnds/gnilnd/gnilnd.c
index 2922f5e..4976a07 100644 (file)
--- a/lnet/klnds/gnilnd/gnilnd.c
+++ b/lnet/klnds/gnilnd/gnilnd.c
@@ -26,11 +26,7 @@
 
 /* Primary entry points from LNET.  There are no guarantees against reentrance. */
 lnd_t the_kgnilnd = {
-#ifdef CONFIG_CRAY_XT
        .lnd_type       = GNILND,
-#else
-       .lnd_type       = GNIIPLND,
-#endif
        .lnd_startup    = kgnilnd_startup,
        .lnd_shutdown   = kgnilnd_shutdown,
        .lnd_ctl        = kgnilnd_ctl,
@@ -266,6 +262,7 @@ kgnilnd_create_conn(kgn_conn_t **connp, kgn_device_t *dev)
        INIT_LIST_HEAD(&conn->gnc_schedlist);
        INIT_LIST_HEAD(&conn->gnc_fmaq);
        INIT_LIST_HEAD(&conn->gnc_mdd_list);
+       INIT_LIST_HEAD(&conn->gnc_delaylist);
        spin_lock_init(&conn->gnc_list_lock);
        spin_lock_init(&conn->gnc_tx_lock);
        conn->gnc_magic = GNILND_CONN_MAGIC;
@@ -451,8 +448,9 @@ kgnilnd_destroy_conn(kgn_conn_t *conn)
                list_empty(&conn->gnc_hashlist) &&
                list_empty(&conn->gnc_schedlist) &&
                list_empty(&conn->gnc_mdd_list) &&
+               list_empty(&conn->gnc_delaylist) &&
                conn->gnc_magic == GNILND_CONN_MAGIC,
-               "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d\n",
+               "conn 0x%p->%s IRQ %d sched %d purg %d ep 0x%p Mg %d lists %d/%d/%d/%d/%d\n",
                conn, conn->gnc_peer ? libcfs_nid2str(conn->gnc_peer->gnp_nid)
                                     : "<?>",
                !!in_interrupt(), conn->gnc_scheduled,
@@ -462,7 +460,8 @@ kgnilnd_destroy_conn(kgn_conn_t *conn)
                list_empty(&conn->gnc_list),
                list_empty(&conn->gnc_hashlist),
                list_empty(&conn->gnc_schedlist),
-               list_empty(&conn->gnc_mdd_list));
+               list_empty(&conn->gnc_mdd_list),
+               list_empty(&conn->gnc_delaylist));
 
        /* Tripping these is especially bad, as it means we have items on the
         *  lists that didn't keep their refcount on the connection - or
@@ -738,6 +737,11 @@ kgnilnd_complete_closed_conn(kgn_conn_t *conn)
                 kgnilnd_conn_state2str(conn));
 
        LASSERT(list_empty(&conn->gnc_hashlist));
+       /* We shouldnt be on the delay list, the conn can 
+        * get added to this list during a retransmit, and retransmits
+        * only occur within scheduler threads.
+        */
+       LASSERT(list_empty(&conn->gnc_delaylist));
 
        /* we've sent the close, start nuking */
        if (CFS_FAIL_CHECK(CFS_FAIL_GNI_SCHEDULE_COMPLETE))
@@ -2046,12 +2050,6 @@ kgnilnd_dev_init(kgn_device_t *dev)
                }
        }
 
-       rrc = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_IP, &kgnilnd_data.kgn_sock);
-       if (rrc < 0) {
-               CERROR("sock_create returned %d\n", rrc);
-               GOTO(failed, rrc);
-       }
-
        rc = kgnilnd_nicaddr_to_nid(dev->gnd_host_id, &dev->gnd_nid);
        if (rc < 0) {
                /* log messages during startup */
@@ -2114,9 +2112,12 @@ kgnilnd_dev_fini(kgn_device_t *dev)
        /* At quiesce or rest time, need to loop through and clear gnd_ready_conns ?*/
        LASSERTF(list_empty(&dev->gnd_ready_conns) &&
                 list_empty(&dev->gnd_map_tx) &&
-                list_empty(&dev->gnd_rdmaq),
-                "dev 0x%p ready_conns %d@0x%p map_tx %d@0x%p rdmaq %d@0x%p\n",
+                list_empty(&dev->gnd_rdmaq) &&
+                list_empty(&dev->gnd_delay_conns),
+                "dev 0x%p ready_conns %d@0x%p delay_conns %d@0x%p" 
+                "map_tx %d@0x%p rdmaq %d@0x%p\n",
                 dev, kgnilnd_count_list(&dev->gnd_ready_conns), &dev->gnd_ready_conns,
+                kgnilnd_count_list(&dev->gnd_delay_conns), &dev->gnd_delay_conns,
                 kgnilnd_count_list(&dev->gnd_map_tx), &dev->gnd_map_tx,
                 kgnilnd_count_list(&dev->gnd_rdmaq), &dev->gnd_rdmaq);
 
@@ -2190,9 +2191,6 @@ kgnilnd_dev_fini(kgn_device_t *dev)
                dev->gnd_domain = NULL;
        }
 
-       if (kgnilnd_data.kgn_sock)
-               sock_release(kgnilnd_data.kgn_sock);
-
        EXIT;
 }
 
@@ -2240,6 +2238,7 @@ int kgnilnd_base_startup(void)
 
                dev->gnd_id = i;
                INIT_LIST_HEAD(&dev->gnd_ready_conns);
+               INIT_LIST_HEAD(&dev->gnd_delay_conns);
                INIT_LIST_HEAD(&dev->gnd_map_tx);
                INIT_LIST_HEAD(&dev->gnd_fma_buffs);
                mutex_init(&dev->gnd_cq_mutex);
@@ -2706,8 +2705,12 @@ kgnilnd_startup(lnet_ni_t *ni)
        INIT_LIST_HEAD(&net->gnn_list);
        ni->ni_data = net;
        net->gnn_ni = ni;
-       ni->ni_maxtxcredits = *kgnilnd_tunables.kgn_credits;
-       ni->ni_peertxcredits = *kgnilnd_tunables.kgn_peer_credits;
+       if (!ni->ni_net->net_tunables_set) {
+               ni->ni_net->net_tunables.lct_max_tx_credits =
+                       *kgnilnd_tunables.kgn_credits;
+               ni->ni_net->net_tunables.lct_peer_tx_credits =
+                       *kgnilnd_tunables.kgn_peer_credits;
+       }
 
        if (*kgnilnd_tunables.kgn_peer_health) {
                int     fudge;
@@ -2717,9 +2720,10 @@ kgnilnd_startup(lnet_ni_t *ni)
                fudge = (GNILND_TO2KA(*kgnilnd_tunables.kgn_timeout) / GNILND_REAPER_NCHECKS);
                timeout = *kgnilnd_tunables.kgn_timeout + fudge;
 
-               if (*kgnilnd_tunables.kgn_peer_timeout >= timeout)
-                       ni->ni_peertimeout = *kgnilnd_tunables.kgn_peer_timeout;
-               else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
+               if (*kgnilnd_tunables.kgn_peer_timeout >= timeout) {
+                       ni->ni_net->net_tunables.lct_peer_timeout =
+                                *kgnilnd_tunables.kgn_peer_timeout;
+               } else if (*kgnilnd_tunables.kgn_peer_timeout > -1) {
                        LCONSOLE_ERROR("Peer_timeout is set to %d but needs to be >= %d\n",
                                        *kgnilnd_tunables.kgn_peer_timeout,
                                        timeout);
@@ -2727,10 +2731,10 @@ kgnilnd_startup(lnet_ni_t *ni)
                        LIBCFS_FREE(net, sizeof(*net));
                        GOTO(failed, rc = -EINVAL);
                } else
-                       ni->ni_peertimeout = timeout;
+                       ni->ni_net->net_tunables.lct_peer_timeout = timeout;
 
                LCONSOLE_INFO("Enabling LNet peer health for gnilnd, timeout %ds\n",
-                             ni->ni_peertimeout);
+                             ni->ni_net->net_tunables.lct_peer_timeout);
        }
 
        atomic_set(&net->gnn_refcount, 1);