LU-3322 lnet: make connect parameters persistent

Store map-on-demand and peertx credits in the peer, since the peer
is persistent. Also made sure that when assigning the parameters
received on the connection to the peer structure through create,
that if another peer is added before grabbing the lock we assign
these parameters to it as well.

Signed-off-by: Amir Shehata <amir.shehata@intel.com>
Change-Id: Ie68f1ba1349d15b0a31eff9a2ca454df8e408ea9
Reviewed-on: http://review.whamcloud.com/17074
Tested-by: Jenkins
Reviewed-by: Doug Oucharek <doug.s.oucharek@intel.com>
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c
index c7edbba..5576ee8 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/lnet/klnds/o2iblnd/o2iblnd.c
@@ -337,6 +337,8 @@ kiblnd_create_peer(lnet_ni_t *ni, kib_peer_t **peerp, lnet_nid_t nid)
        peer->ibp_nid = nid;
        peer->ibp_error = 0;
        peer->ibp_last_alive = 0;
+       peer->ibp_max_frags = IBLND_CFG_RDMA_FRAGS;
+       peer->ibp_queue_depth = *kiblnd_tunables.kib_peertxcredits;
        atomic_set(&peer->ibp_refcount, 1);     /* 1 ref for caller */
 
        INIT_LIST_HEAD(&peer->ibp_list);        /* not in the peer table yet */
@@ -704,7 +706,7 @@ kiblnd_get_completion_vector(kib_conn_t *conn, int cpt)
 
 kib_conn_t *
 kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
-                  int state, int version, kib_connparams_t *cp)
+                  int state, int version)
 {
        /* CAVEAT EMPTOR:
         * If the new conn is created successfully it takes over the caller's
@@ -758,14 +760,8 @@ kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
        conn->ibc_peer = peer;                  /* I take the caller's ref */
        cmid->context = conn;                   /* for future CM callbacks */
        conn->ibc_cmid = cmid;
-
-       if (cp == NULL) {
-               conn->ibc_max_frags = IBLND_CFG_RDMA_FRAGS;
-               conn->ibc_queue_depth = *kiblnd_tunables.kib_peertxcredits;
-       } else {
-               conn->ibc_max_frags = cp->ibcp_max_frags;
-               conn->ibc_queue_depth = cp->ibcp_queue_depth;
-       }
+       conn->ibc_max_frags = peer->ibp_max_frags;
+       conn->ibc_queue_depth = peer->ibp_queue_depth;
 
        INIT_LIST_HEAD(&conn->ibc_early_rxs);
        INIT_LIST_HEAD(&conn->ibc_tx_noops);

diff --git a/lnet/klnds/o2iblnd/o2iblnd.h b/lnet/klnds/o2iblnd/o2iblnd.h
index c03de37..bdb7ac5 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd.h
+++ b/lnet/klnds/o2iblnd/o2iblnd.h
@@ -736,6 +736,10 @@ typedef struct kib_peer
        int                     ibp_error;
        /* when (in jiffies) I was last alive */
        cfs_time_t              ibp_last_alive;
+       /* max map_on_demand */
+       __u16                   ibp_max_frags;
+       /* max_peer_credits */
+       __u16                   ibp_queue_depth;
 } kib_peer_t;
 
 extern kib_data_t      kiblnd_data;
@@ -1096,8 +1100,8 @@ int  kiblnd_close_stale_conns_locked (kib_peer_t *peer,
                                       int version, __u64 incarnation);
 int  kiblnd_close_peer_conns_locked (kib_peer_t *peer, int why);
 
-kib_conn_t *kiblnd_create_conn (kib_peer_t *peer, struct rdma_cm_id *cmid,
-                               int state, int version, kib_connparams_t *cp);
+kib_conn_t *kiblnd_create_conn(kib_peer_t *peer, struct rdma_cm_id *cmid,
+                              int state, int version);
 void kiblnd_destroy_conn (kib_conn_t *conn);
 void kiblnd_close_conn (kib_conn_t *conn, int error);
 void kiblnd_close_conn_locked (kib_conn_t *conn, int error);

diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c
index 1223906..ff64291 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd_cb.c
+++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c
@@ -2312,6 +2312,10 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
                goto failed;
        }
 
+       /* We have validated the peer's parameters so use those */
+       peer->ibp_max_frags = reqmsg->ibm_u.connparams.ibcp_max_frags;
+       peer->ibp_queue_depth = reqmsg->ibm_u.connparams.ibcp_queue_depth;
+
        write_lock_irqsave(g_lock, flags);
 
         peer2 = kiblnd_find_peer_locked(nid);
@@ -2350,6 +2354,12 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
                 peer2->ibp_accepting++;
                 kiblnd_peer_addref(peer2);
 
+               /* Race with kiblnd_launch_tx (active connect) to create peer
+                * so copy validated parameters since we now know what the
+                * peer's limits are */
+               peer2->ibp_max_frags = peer->ibp_max_frags;
+               peer2->ibp_queue_depth = peer->ibp_queue_depth;
+
                write_unlock_irqrestore(g_lock, flags);
                 kiblnd_peer_decref(peer);
                 peer = peer2;
@@ -2372,8 +2382,7 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
                write_unlock_irqrestore(g_lock, flags);
         }
 
-       conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version,
-                                 &reqmsg->ibm_u.connparams);
+       conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_PASSIVE_WAIT, version);
         if (conn == NULL) {
                 kiblnd_peer_connect_failed(peer, 0, -ENOMEM);
                 kiblnd_peer_decref(peer);
@@ -2383,10 +2392,9 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
 
         /* conn now "owns" cmid, so I return success from here on to ensure the
          * CM callback doesn't destroy cmid. */
-
        conn->ibc_incarnation      = reqmsg->ibm_srcstamp;
-       conn->ibc_credits          = reqmsg->ibm_u.connparams.ibcp_queue_depth;
-       conn->ibc_reserved_credits = reqmsg->ibm_u.connparams.ibcp_queue_depth;
+       conn->ibc_credits          = conn->ibc_queue_depth;
+       conn->ibc_reserved_credits = conn->ibc_queue_depth;
        LASSERT(conn->ibc_credits + conn->ibc_reserved_credits +
                IBLND_OOB_MSGS(version) <= IBLND_RX_MSGS(conn));
 
@@ -2395,10 +2403,8 @@ kiblnd_passive_connect(struct rdma_cm_id *cmid, void *priv, int priv_nob)
 
         kiblnd_init_msg(ackmsg, IBLND_MSG_CONNACK,
                         sizeof(ackmsg->ibm_u.connparams));
-       ackmsg->ibm_u.connparams.ibcp_queue_depth  =
-               reqmsg->ibm_u.connparams.ibcp_queue_depth;
-       ackmsg->ibm_u.connparams.ibcp_max_frags    =
-               reqmsg->ibm_u.connparams.ibcp_max_frags;
+       ackmsg->ibm_u.connparams.ibcp_queue_depth  = conn->ibc_queue_depth;
+       ackmsg->ibm_u.connparams.ibcp_max_frags    = conn->ibc_max_frags;
        ackmsg->ibm_u.connparams.ibcp_max_msg_size = IBLND_MSG_SIZE;
 
         kiblnd_pack_msg(ni, ackmsg, version, 0, nid, reqmsg->ibm_srcstamp);
@@ -2488,6 +2494,9 @@ kiblnd_reconnect (kib_conn_t *conn, int version,
                 break;
 
        case IBLND_REJECT_RDMA_FRAGS:
+               if (!cp)
+                       goto failed;
+
                if (conn->ibc_max_frags <= cp->ibcp_max_frags) {
                        CNETERR("Unsupported max frags, peer supports %d\n",
                                cp->ibcp_max_frags);
@@ -2498,18 +2507,21 @@ kiblnd_reconnect (kib_conn_t *conn, int version,
                        goto failed;
                }
 
-               conn->ibc_max_frags = cp->ibcp_max_frags;
+               peer->ibp_max_frags = cp->ibcp_max_frags;
                reason = "rdma fragments";
                break;
 
        case IBLND_REJECT_MSG_QUEUE_SIZE:
+               if (!cp)
+                       goto failed;
+
                if (conn->ibc_queue_depth <= cp->ibcp_queue_depth) {
                        CNETERR("Unsupported queue depth, peer supports %d\n",
                                cp->ibcp_queue_depth);
                        goto failed;
                }
 
-               conn->ibc_queue_depth = cp->ibcp_queue_depth;
+               peer->ibp_queue_depth = cp->ibcp_queue_depth;
                reason = "queue depth";
                break;
 
@@ -2787,7 +2799,7 @@ kiblnd_active_connect (struct rdma_cm_id *cmid)
        read_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
 
        conn = kiblnd_create_conn(peer, cmid, IBLND_CONN_ACTIVE_CONNECT,
-                                 version, NULL);
+                                 version);
         if (conn == NULL) {
                 kiblnd_peer_connect_failed(peer, 1, -ENOMEM);
                 kiblnd_peer_decref(peer); /* lose cmid's ref */