LU-7734 socklnd: propagate errors on send failure

When an attempt to send a message fails, for example because no
connection could be established with the remote address, socklnd
drops the message. For a PUT or REPLY message with non-zero
payload, ksocknal_tx_done() calls lnet_finalize() with -EIO
as the error code. But for an ACK or GET message there is no
payload, and lnet_finalize() is called with 0 (no error) as the
error code. This leaves upper layers to rely on other means to
determine that sending the message did actually fail, and that
(for example) no REPLY will ever answer a failed GET.

Add an error code parameter to ksocknal_tx_done().

In ksocknal_txlist_done() change the 0/1 'error' indicator to be
an actual error code that is passed on the ksocknal_tx_done().
Update the callers of ksocknal_txlist_done() to pass in the error
code if they have encountered an error.

Test-Parameters: trivial
Signed-off-by: Olaf Weber <olaf@sgi.com>
Change-Id: Idf6dc7a3a2c95cecee6b272c35a66ef57a704b52

diff --git a/lnet/klnds/socklnd/socklnd.c b/lnet/klnds/socklnd/socklnd.c
index ad32457..0b1a69f 100644 (file)
--- a/lnet/klnds/socklnd/socklnd.c
+++ b/lnet/klnds/socklnd/socklnd.c
@@ -616,7 +616,7 @@ ksocknal_del_peer (lnet_ni_t *ni, lnet_process_id_t id, __u32 ip)
 
        write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 
-       ksocknal_txlist_done(ni, &zombies, 1);
+       ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
 
        return rc;
 }
@@ -1028,6 +1028,7 @@ ksocknal_create_conn(lnet_ni_t *ni, ksock_route_t *route,
         ksock_tx_t        *tx;
         ksock_tx_t        *txtmp;
         int                rc;
+       int                rc2;
         int                active;
         char              *warn = NULL;
 
@@ -1382,7 +1383,13 @@ failed_2:
                write_unlock_bh(global_lock);
         }
 
-        ksocknal_txlist_done(ni, &zombies, 1);
+       /*
+        * If we get here without an error code, just use -EALREADY.
+        * Depending on how we got here, the error may be positive
+        * or negative. Normalize the value for ksocknal_txlist_done().
+        */
+       rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
+       ksocknal_txlist_done(ni, &zombies, rc2);
         ksocknal_peer_decref(peer_ni);
 
 failed_1:

diff --git a/lnet/klnds/socklnd/socklnd.h b/lnet/klnds/socklnd/socklnd.h
index fd24314..d17dfec 100644 (file)
--- a/lnet/klnds/socklnd/socklnd.h
+++ b/lnet/klnds/socklnd/socklnd.h
@@ -564,14 +564,14 @@ ksocknal_tx_addref (ksock_tx_t *tx)
 }
 
 extern void ksocknal_tx_prep (ksock_conn_t *, ksock_tx_t *tx);
-extern void ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx);
+extern void ksocknal_tx_done(lnet_ni_t *ni, ksock_tx_t *tx, int error);
 
 static inline void
 ksocknal_tx_decref (ksock_tx_t *tx)
 {
        LASSERT (atomic_read(&tx->tx_refcount) > 0);
        if (atomic_dec_and_test(&tx->tx_refcount))
-               ksocknal_tx_done(NULL, tx);
+               ksocknal_tx_done(NULL, tx, 0);
 }
 
 static inline void

diff --git a/lnet/klnds/socklnd/socklnd_cb.c b/lnet/klnds/socklnd/socklnd_cb.c
index b69599b..9441366 100644 (file)
--- a/lnet/klnds/socklnd/socklnd_cb.c
+++ b/lnet/klnds/socklnd/socklnd_cb.c
@@ -389,50 +389,52 @@ ksocknal_receive (ksock_conn_t *conn)
 }
 
 void
-ksocknal_tx_done (lnet_ni_t *ni, ksock_tx_t *tx)
+ksocknal_tx_done(lnet_ni_t *ni, ksock_tx_t *tx, int rc)
 {
-        lnet_msg_t  *lnetmsg = tx->tx_lnetmsg;
-        int          rc = (tx->tx_resid == 0 && !tx->tx_zc_aborted) ? 0 : -EIO;
-        ENTRY;
+       lnet_msg_t  *lnetmsg = tx->tx_lnetmsg;
+       ENTRY;
 
-        LASSERT(ni != NULL || tx->tx_conn != NULL);
+       LASSERT(ni != NULL || tx->tx_conn != NULL);
 
-        if (tx->tx_conn != NULL)
-                ksocknal_conn_decref(tx->tx_conn);
+       if (!rc && (tx->tx_resid != 0 || tx->tx_zc_aborted))
+               rc = -EIO;
 
-        if (ni == NULL && tx->tx_conn != NULL)
-                ni = tx->tx_conn->ksnc_peer->ksnp_ni;
+       if (tx->tx_conn != NULL)
+               ksocknal_conn_decref(tx->tx_conn);
 
-        ksocknal_free_tx (tx);
-        if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
-                lnet_finalize (ni, lnetmsg, rc);
+       if (ni == NULL && tx->tx_conn != NULL)
+               ni = tx->tx_conn->ksnc_peer->ksnp_ni;
+
+       ksocknal_free_tx(tx);
+       if (lnetmsg != NULL) /* KSOCK_MSG_NOOP go without lnetmsg */
+               lnet_finalize(ni, lnetmsg, rc);
 
-        EXIT;
+       EXIT;
 }
 
 void
 ksocknal_txlist_done(lnet_ni_t *ni, struct list_head *txlist, int error)
 {
-        ksock_tx_t *tx;
+       ksock_tx_t *tx;
 
        while (!list_empty(txlist)) {
                tx = list_entry(txlist->next, ksock_tx_t, tx_list);
 
-                if (error && tx->tx_lnetmsg != NULL) {
-                        CNETERR("Deleting packet type %d len %d %s->%s\n",
-                                le32_to_cpu (tx->tx_lnetmsg->msg_hdr.type),
-                                le32_to_cpu (tx->tx_lnetmsg->msg_hdr.payload_length),
-                                libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
-                                libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
-                } else if (error) {
-                        CNETERR("Deleting noop packet\n");
-                }
+               if (error && tx->tx_lnetmsg != NULL) {
+                       CNETERR("Deleting packet type %d len %d %s->%s\n",
+                               le32_to_cpu(tx->tx_lnetmsg->msg_hdr.type),
+                               le32_to_cpu(tx->tx_lnetmsg->msg_hdr.payload_length),
+                               libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.src_nid)),
+                               libcfs_nid2str(le64_to_cpu(tx->tx_lnetmsg->msg_hdr.dest_nid)));
+               } else if (error) {
+                       CNETERR("Deleting noop packet\n");
+               }
 
                list_del(&tx->tx_list);
 
-               LASSERT (atomic_read(&tx->tx_refcount) == 1);
-                ksocknal_tx_done (ni, tx);
-        }
+               LASSERT(atomic_read(&tx->tx_refcount) == 1);
+               ksocknal_tx_done(ni, tx, error);
+       }
 }
 
 static void
@@ -1994,9 +1996,9 @@ ksocknal_connect (ksock_route_t *route)
 
        write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 
-        ksocknal_peer_failed(peer_ni);
-        ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, 1);
-        return 0;
+       ksocknal_peer_failed(peer_ni);
+       ksocknal_txlist_done(peer_ni->ksnp_ni, &zombies, rc);
+       return 0;
 }
 
 /*
@@ -2326,26 +2328,26 @@ ksocknal_find_timed_out_conn (ksock_peer_ni_t *peer_ni)
 static inline void
 ksocknal_flush_stale_txs(ksock_peer_ni_t *peer_ni)
 {
-        ksock_tx_t        *tx;
-       struct list_head        stale_txs = LIST_HEAD_INIT(stale_txs);
+       ksock_tx_t        *tx;
+       struct list_head        stale_txs = LIST_HEAD_INIT(stale_txs);
 
        write_lock_bh(&ksocknal_data.ksnd_global_lock);
 
        while (!list_empty(&peer_ni->ksnp_tx_queue)) {
                tx = list_entry(peer_ni->ksnp_tx_queue.next,
-                                     ksock_tx_t, tx_list);
+                                    ksock_tx_t, tx_list);
 
-                if (!cfs_time_aftereq(cfs_time_current(),
-                                      tx->tx_deadline))
-                        break;
+               if (!cfs_time_aftereq(cfs_time_current(),
+                                     tx->tx_deadline))
+                       break;
 
                list_del(&tx->tx_list);
                list_add_tail(&tx->tx_list, &stale_txs);
-        }
+       }
 
        write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 
-        ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, 1);
+       ksocknal_txlist_done(peer_ni->ksnp_ni, &stale_txs, -ETIMEDOUT);
 }
 
 static int