LU-10800 lnet: Revert "LU-10270 lnet: remove an early rx code"

This reverts commit c3894ff80fe4b48f2d62ea33ddc54fb5891e6484. Dropping
early receives caused pings to be ignored and interacted badly with
dynamic discovery.

Signed-off-by: John L. Hammond <john.hammond@intel.com>
Change-Id: I99a87a8f58ea67c59d5e85b964295472c2e15de4
Reviewed-on: https://review.whamcloud.com/31675
Tested-by: Jenkins
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Amir Shehata <amir.shehata@intel.com>
Reviewed-by: Doug Oucharek <dougso@me.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>

diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c
index d1ea6a2..dc7981d 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/lnet/klnds/o2iblnd/o2iblnd.c
@@ -604,6 +604,10 @@ kiblnd_debug_conn (kib_conn_t *conn)
               conn->ibc_outstanding_credits, conn->ibc_reserved_credits);
        CDEBUG(D_CONSOLE, "   comms_err %d\n", conn->ibc_comms_error);
 
+       CDEBUG(D_CONSOLE, "   early_rxs:\n");
+       list_for_each(tmp, &conn->ibc_early_rxs)
+               kiblnd_debug_rx(list_entry(tmp, kib_rx_t, rx_list));
+
        CDEBUG(D_CONSOLE, "   tx_noops:\n");
        list_for_each(tmp, &conn->ibc_tx_noops)
                kiblnd_debug_tx(list_entry(tmp, kib_tx_t, tx_list));
@@ -810,6 +814,7 @@ kiblnd_create_conn(kib_peer_ni_t *peer_ni, struct rdma_cm_id *cmid,
        conn->ibc_max_frags = peer_ni->ibp_max_frags;
        conn->ibc_queue_depth = peer_ni->ibp_queue_depth;
 
+       INIT_LIST_HEAD(&conn->ibc_early_rxs);
        INIT_LIST_HEAD(&conn->ibc_tx_noops);
        INIT_LIST_HEAD(&conn->ibc_tx_queue);
        INIT_LIST_HEAD(&conn->ibc_tx_queue_rsrvd);
@@ -999,6 +1004,7 @@ kiblnd_destroy_conn(kib_conn_t *conn)
 
        LASSERT (!in_interrupt());
        LASSERT (atomic_read(&conn->ibc_refcount) == 0);
+       LASSERT(list_empty(&conn->ibc_early_rxs));
        LASSERT(list_empty(&conn->ibc_tx_noops));
        LASSERT(list_empty(&conn->ibc_tx_queue));
        LASSERT(list_empty(&conn->ibc_tx_queue_rsrvd));

diff --git a/lnet/klnds/o2iblnd/o2iblnd.h b/lnet/klnds/o2iblnd/o2iblnd.h
index 7a465aa..8ac9c45 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd.h
+++ b/lnet/klnds/o2iblnd/o2iblnd.h
@@ -703,6 +703,8 @@ typedef struct kib_conn
        ktime_t                 ibc_last_send;
        /** link chain for kiblnd_check_conns only */
        struct list_head        ibc_connd_list;
+       /** rxs completed before ESTABLISHED */
+       struct list_head        ibc_early_rxs;
        /** IBLND_MSG_NOOPs for IBLND_MSG_VERSION_1 */
        struct list_head        ibc_tx_noops;
        /* sends that need a credit */

diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c
index 3323741..cf753dd 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd_cb.c
+++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c
@@ -516,10 +516,21 @@ kiblnd_rx_complete (kib_rx_t *rx, int status, int nob)
         /* set time last known alive */
         kiblnd_peer_alive(conn->ibc_peer);
 
-       /* racing with connection establishment/teardown! */
-       if (conn->ibc_state < IBLND_CONN_ESTABLISHED)
-               goto ignore;
+        /* racing with connection establishment/teardown! */
 
+        if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
+               rwlock_t  *g_lock = &kiblnd_data.kib_global_lock;
+               unsigned long  flags;
+
+               write_lock_irqsave(g_lock, flags);
+               /* must check holding global lock to eliminate race */
+               if (conn->ibc_state < IBLND_CONN_ESTABLISHED) {
+                       list_add_tail(&rx->rx_list, &conn->ibc_early_rxs);
+                       write_unlock_irqrestore(g_lock, flags);
+                       return;
+               }
+               write_unlock_irqrestore(g_lock, flags);
+        }
         kiblnd_handle_rx(rx);
         return;
 
@@ -2045,6 +2056,29 @@ kiblnd_close_conn(kib_conn_t *conn, int error)
 }
 
 static void
+kiblnd_handle_early_rxs(kib_conn_t *conn)
+{
+       unsigned long    flags;
+       kib_rx_t        *rx;
+
+       LASSERT(!in_interrupt());
+       LASSERT(conn->ibc_state >= IBLND_CONN_ESTABLISHED);
+
+       write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+       while (!list_empty(&conn->ibc_early_rxs)) {
+               rx = list_entry(conn->ibc_early_rxs.next,
+                                   kib_rx_t, rx_list);
+               list_del(&rx->rx_list);
+               write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+
+               kiblnd_handle_rx(rx);
+
+               write_lock_irqsave(&kiblnd_data.kib_global_lock, flags);
+       }
+       write_unlock_irqrestore(&kiblnd_data.kib_global_lock, flags);
+}
+
+static void
 kiblnd_abort_txs(kib_conn_t *conn, struct list_head *txs)
 {
        struct list_head         zombies = LIST_HEAD_INIT(zombies);
@@ -2101,6 +2135,8 @@ kiblnd_finalise_conn (kib_conn_t *conn)
        kiblnd_abort_txs(conn, &conn->ibc_tx_queue_rsrvd);
        kiblnd_abort_txs(conn, &conn->ibc_tx_queue_nocred);
        kiblnd_abort_txs(conn, &conn->ibc_active_txs);
+
+       kiblnd_handle_early_rxs(conn);
 }
 
 static void
@@ -2257,6 +2293,8 @@ kiblnd_connreq_done(kib_conn_t *conn, int status)
        kiblnd_check_sends_locked(conn);
        spin_unlock(&conn->ibc_lock);
 
+       /* schedule blocked rxs */
+       kiblnd_handle_early_rxs(conn);
        kiblnd_conn_decref(conn);
 }