1st working ib integrated ctdb

TODO: bugfix of ctdb_bench (wants to sent via uninitialised connection - see bench_ring/dest)

(This used to be ctdb commit 61516461e9d45dc7ba87518d134894fed7d7b7cd)

6 files changed, 74 insertions, 33 deletions

diff --git a/ctdb/ib/README.txt b/ctdb/ib/README.txt
index 3e6150afe33..74fc129c352 100644
--- a/ctdb/ib/README.txt
+++ b/ctdb/ib/README.txt
@@ -9,3 +9,12 @@ After then:
 
 ./configure --enable-infiniband
 
+Example for testing
+===================
+bin/ctdb_test --transport ib --nlist ../2nodes_rm.txt --listen 10.0.0.1:9001
+bin/ctdb_test --transport ib --nlist ../2nodes_rm.txt --listen 10.0.0.2:9001
+
+where 2nodes_rm.txt:
+10.0.0.1:9001
+10.0.0.2:9001
+
diff --git a/ctdb/ib/ibw_ctdb.c b/ctdb/ib/ibw_ctdb.c
index 5152062d246..53293810ca8 100644
--- a/ctdb/ib/ibw_ctdb.c
+++ b/ctdb/ib/ibw_ctdb.c
@@ -29,6 +29,36 @@
 #include "ibwrapper.h"
 #include "ibw_ctdb.h"
 
+int ctdb_ibw_node_connect(struct ibw_ctx *ictx, struct ctdb_node *node)
+{
+	struct sockaddr_in sock_out;
+
+	memset(&sock_out, 0, sizeof(struct sockaddr_in));
+	inet_pton(AF_INET, node->address.address, &sock_out.sin_addr);
+	sock_out.sin_port = htons(node->address.port);
+	sock_out.sin_family = PF_INET;
+
+	if (ibw_connect(ictx, &sock_out, node)) {
+		DEBUG(0, ("ctdb_ibw_node_connect: ibw_connect failed - retrying in 1 sec...\n"));
+		/* try again once a second */
+		event_add_timed(node->ctdb->ev, node, timeval_current_ofs(1, 0), 
+			ctdb_ibw_node_connect_event, node);
+		return -1;
+	}
+
+	/* continues at ibw_ctdb.c/IBWC_CONNECTED in good case */
+	return 0;
+}
+
+void ctdb_ibw_node_connect_event(struct event_context *ev, struct timed_event *te, 
+	struct timeval t, void *private)
+{
+	struct ctdb_node *node = talloc_get_type(private, struct ctdb_node);
+	struct ibw_ctx *ictx = talloc_get_type(node->ctdb->private, struct ibw_ctx);
+
+	ctdb_ibw_node_connect(ictx, node);
+}
+
 int ctdb_ibw_connstate_handler(struct ibw_ctx *ctx, struct ibw_conn *conn)
 {
 	if (ctx!=NULL) {
@@ -76,11 +106,17 @@ int ctdb_ibw_connstate_handler(struct ibw_ctx *ctx, struct ibw_conn *conn)
 			if (node!=NULL)
 				node->ctdb->upcalls->node_dead(node);
 			talloc_free(conn);
+			/* normal + intended disconnect => not reconnecting in this layer */
 		} break;
 		case IBWC_ERROR: {
-/*			struct ctdb_node *node = talloc_get_type(conn->conn_userdata, struct ctdb_node);
+			struct ctdb_node *node = talloc_get_type(conn->conn_userdata, struct ctdb_node);
 			if (node!=NULL)
-				node->ctdb->upcalls->node_dead(node);*/
+				node->private = NULL; /* not to use again */
+
+			DEBUG(10, ("IBWC_ERROR, reconnecting immediately...\n"));
+			talloc_free(conn);
+			event_add_timed(node->ctdb->ev, node, timeval_current_ofs(1, 0),
+				ctdb_ibw_node_connect_event, node);
 		} break;
 		default:
 			assert(0);
@@ -94,17 +130,20 @@ int ctdb_ibw_connstate_handler(struct ibw_ctx *ctx, struct ibw_conn *conn)
 int ctdb_ibw_receive_handler(struct ibw_conn *conn, void *buf, int n)
 {
 	struct ctdb_context *ctdb = talloc_get_type(conn->ctx->ctx_userdata, struct ctdb_context);
+	void	*buf2; /* future TODO: a solution for removal of this */
 
 	assert(ctdb!=NULL);
+	assert(buf!=NULL);
+	assert(conn!=NULL);
 	assert(conn->state==IBWC_CONNECTED);
 
-	/* TODO: shall I short-circuit this in ibwrapper? */
-	/* maybe when everything go fine... */
+	/* so far "buf" is an ib-registered memory area
+	 * and being reused for next receive
+	 * noticed that HL requires talloc-ed memory to be stolen */
+	buf2 = talloc_zero_size(conn, n);
+	memcpy(buf2, buf, n);
 
-	/* TODO2: !!! here I can provide conn->conn_userdata (with no perf. penalty) -
-	 * as struct ctdb_node in case the connection
-	 * has been built up by ibw_connect !!! */
-	ctdb->upcalls->recv_pkt(ctdb, (uint8_t *)buf, (uint32_t)n);
+	ctdb->upcalls->recv_pkt(ctdb, (uint8_t *)buf2, (uint32_t)n);
 
 	return 0;
 }
diff --git a/ctdb/ib/ibw_ctdb.h b/ctdb/ib/ibw_ctdb.h
index c43aca7df15..14308682b21 100644
--- a/ctdb/ib/ibw_ctdb.h
+++ b/ctdb/ib/ibw_ctdb.h
@@ -23,3 +23,8 @@
 
 int ctdb_ibw_connstate_handler(struct ibw_ctx *ctx, struct ibw_conn *conn);
 int ctdb_ibw_receive_handler(struct ibw_conn *conn, void *buf, int n);
+
+int ctdb_ibw_node_connect(struct ibw_ctx *ictx, struct ctdb_node *node);
+void ctdb_ibw_node_connect_event(struct event_context *ev, struct timed_event *te, 
+	struct timeval t, void *private);
+
diff --git a/ctdb/ib/ibw_ctdb_init.c b/ctdb/ib/ibw_ctdb_init.c
index a1eaca9202b..78924632292 100644
--- a/ctdb/ib/ibw_ctdb_init.c
+++ b/ctdb/ib/ibw_ctdb_init.c
@@ -53,24 +53,6 @@ static int ctdb_ibw_listen(struct ctdb_context *ctdb, int backlog)
 	return 0;
 }
 
-int ctdb_ibw_node_connect(struct ibw_ctx *ictx, struct ctdb_node *node)
-{
-        struct sockaddr_in sock_out;
-
-	memset(&sock_out, 0, sizeof(struct sockaddr_in));
-	inet_pton(AF_INET, node->address.address, &sock_out.sin_addr);
-	sock_out.sin_port = htons(node->address.port);
-	sock_out.sin_family = PF_INET;
-
-	if (ibw_connect(ictx, &sock_out, node)) {
-		DEBUG(0, ("ctdb_ibw_node_connect: ibw_connect failed\n"));
-		return -1;
-	}
-
-	/* continues at ibw_ctdb.c/IBWC_CONNECTED in good case */
-	return 0;
-}
-
 /*
  * Start infiniband
  */
diff --git a/ctdb/ib/ibwrapper.c b/ctdb/ib/ibwrapper.c
index fef312678a5..f3ef0c4c5cd 100644
--- a/ctdb/ib/ibwrapper.c
+++ b/ctdb/ib/ibwrapper.c
@@ -775,7 +775,7 @@ static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc)
 			if (part->len<=sizeof(uint32_t) && part->to_read==0) {
 				assert(part->len==sizeof(uint32_t));
 				/* set it again now... */
-				part->to_read = ntohl(*((uint32_t *)(part->buf)));
+				part->to_read = *((uint32_t *)(part->buf)); /* TODO: ntohl */
 				if (part->to_read<sizeof(uint32_t)) {
 					sprintf(ibw_lasterr, "got msglen=%u #2\n", part->to_read);
 					goto error;
@@ -791,7 +791,7 @@ static int ibw_wc_recv(struct ibw_conn *conn, struct ibv_wc *wc)
 			}
 		} else {
 			if (remain>=sizeof(uint32_t)) {
-				uint32_t msglen = ntohl(*(uint32_t *)p);
+				uint32_t msglen = *(uint32_t *)p; /* TODO: ntohl */
 				if (msglen<sizeof(uint32_t)) {
 					sprintf(ibw_lasterr, "got msglen=%u\n", msglen);
 					goto error;
@@ -1028,7 +1028,8 @@ int ibw_connect(struct ibw_ctx *ctx, struct sockaddr_in *serv_addr, void *conn_u
 	if (rc) {
 		rc = errno;
 		sprintf(ibw_lasterr, "ibw_connect/rdma_create_id error %d\n", rc);
-		return rc;
+		talloc_free(conn);
+		return -1;
 	}
 	DEBUG(10, ("ibw_connect: rdma_create_id succeeded, cm_id=%p\n", pconn->cm_id));
 
@@ -1196,7 +1197,7 @@ int ibw_send(struct ibw_conn *conn, void *buf, void *key, uint32_t len)
 	int	rc;
 
 	assert(len>=sizeof(uint32_t));
-	*((uint32_t *)buf) = htonl(len);
+	assert((*((uint32_t *)buf)==len)); /* TODO: htonl */
 
 	if (len > pctx->opts.recv_bufsize) {
 		struct ibw_conn_priv *pconn = talloc_get_type(conn->internal, struct ibw_conn_priv);
diff --git a/ctdb/ib/ibwrapper_test.c b/ctdb/ib/ibwrapper_test.c
index dc07a1c313b..2fa590588ce 100644
--- a/ctdb/ib/ibwrapper_test.c
+++ b/ctdb/ib/ibwrapper_test.c
@@ -110,6 +110,7 @@ int ibwtest_send_id(struct ibw_conn *conn)
 	}
 
 	/* first sizeof(uint32_t) size bytes are for length */
+	*((uint32_t *)buf) = len;
 	buf[sizeof(uint32_t)] = (char)TESTOP_SEND_ID;
 	strcpy(buf+sizeof(uint32_t)+1, tcx->id);
 
@@ -137,6 +138,7 @@ int ibwtest_send_test_msg(struct ibwtest_ctx *tcx, struct ibw_conn *conn, const
 		return -1;
 	}
 
+	*((uint32_t *)buf) = len;
 	p = buf;
 	p += sizeof(uint32_t);
 	p[0] = (char)TESTOP_SEND_TEXT;
@@ -190,6 +192,7 @@ int ibwtest_do_varsize_scenario_conn_size(struct ibwtest_ctx *tcx, struct ibw_co
 		DEBUG(0, ("varsize/ibw_alloc_send_buf failed\n"));
 		return -1;
 	}
+	*((uint32_t *)buf) = len;
 	buf[sizeof(uint32_t)] = TESTOP_SEND_RND;
 	sum = ibwtest_fill_random(buf + sizeof(uint32_t) + 1, size);
 	buf[sizeof(uint32_t) + 1 + size] = sum;
@@ -329,7 +332,7 @@ int ibwtest_receive_handler(struct ibw_conn *conn, void *buf, int n)
 				DEBUG(0, ("ERROR: checksum mismatch %u!=%u\n",
 					(uint32_t)sum, (uint32_t)((unsigned char *)buf)[n-1]));
 				ibw_stop(tcx->ibwctx);
-				return -3;
+				goto error;
 			}
 		} else {
 			char *buf2;
@@ -338,12 +341,12 @@ int ibwtest_receive_handler(struct ibw_conn *conn, void *buf, int n)
 			/* bounce message regardless what it is */
 			if (ibw_alloc_send_buf(conn, (void **)&buf2, &key2, n)) {
 				fprintf(stderr, "ibw_alloc_send_buf error #2\n");
-				return -1;
+				goto error;
 			}
 			memcpy(buf2, buf, n);
 			if (ibw_send(conn, buf2, key2, n)) {
 				fprintf(stderr, "ibw_send error #2\n");
-				return -2;
+				goto error;
 			}
 			tcx->nsent++;
 		}
@@ -368,6 +371,8 @@ int ibwtest_receive_handler(struct ibw_conn *conn, void *buf, int n)
 		tcx->error = rc;
 
 	return rc;
+error:
+	return -1;
 }
 
 void ibwtest_timeout_handler(struct event_context *ev, struct timed_event *te,