rds: ib: Revert "net/rds: Avoid stalled connection due to CM REQ retries"

This reverts commit 637d078.

Commit 637d078 ("net/rds: Avoid stalled connection due to CM REQ
retries") introduced a generation scheme allegedly to detect duplicate
RDMA CM REQs.

However, duplicate CM packets are filtered by the CM layer and should
not reach the client (RDS). Further, this commit introduced a bug when
a node being the active connector rebooted. Once rebooted, the peer
would classify the rebooted node's connection requests as duplicates.

The signature of this bug is:

rds_ib_cm_handle_connect: 1151: duplicated REQ in UP state. cm_id x conn y reject! ic->i_last_rej_seq 1 cm_req_seq 1

Orabug: 31648138

Tested-by: Aiman Al-Khammash <[email protected]>
Tested-by: Gerald Gibson <[email protected]>
Reviewed-by: Sharath Srinivasan <[email protected]>
cherry-pick-repo=uek/uek6/master

Signed-off-by: Håkon Bugge <[email protected]>

Conflicts in:
	net/rds/ib.h
	net/rds/ib_cm.c

	Due to context differences

(cherry picked from commit b6e8f987e83198acca42cb8a79fda3c7d739ba80)
cherry-pick-repo=uek/uek6/master

Signed-off-by: Håkon Bugge <[email protected]>

Author: Hakon-Bugge

net/rds/ib.h

@@ -56,8 +56,6 @@
 #define	RDS_IB_CQ_ERR		2
 
 #define RDS_IB_DEFAULT_FREG_PORT_NUM	1
-#define RDS_CM_RETRY_SEQ_EN	BIT(7)
-#define RDS_CM_REQ_SEQ_SZ	(RDS_CM_RETRY_SEQ_EN - 1)
 
 #define RDS_RDMA_RESOLVE_ADDR_TIMEOUT_MS(c) ((c)->c_loopback ? 1000 : 4000)
 
@@ -103,7 +101,7 @@ struct rds_ib_conn_priv_cmn {
 	u8			ricpc_protocol_minor;
 	__be16			ricpc_protocol_minor_mask;	/* bitmask */
 	u8			ricpc_tos;
-	u8			ricpc_cm_seq;
+	u8			ricpc_reserved1;
 	__be16			ricpc_frag_sz;
 	__be64			ricpc_ack_seq;
 	__be32			ricpc_credit;	/* non-zero enables flow ctl */
@@ -127,7 +125,7 @@ struct rds6_ib_connect_private {
 #define dp_protocol_minor	dp_cmn.ricpc_protocol_minor
 #define dp_protocol_minor_mask	dp_cmn.ricpc_protocol_minor_mask
 #define dp_tos			dp_cmn.ricpc_tos
-#define dp_cm_seq		dp_cmn.ricpc_cm_seq
+#define dp_reserved1		dp_cmn.ricpc_reserved1
 #define dp_frag_sz		dp_cmn.ricpc_frag_sz
 #define dp_ack_seq		dp_cmn.ricpc_ack_seq
 #define dp_credit		dp_cmn.ricpc_credit
@@ -292,9 +290,6 @@ struct rds_ib_connection {
 	unsigned int            i_rx_wait_for_handler;
 	atomic_t                i_worker_has_rx;
 	atomic_t		i_cq_quiesce;
-	u8			i_req_sequence;
-	u8			i_prev_seq;
-	u8			i_last_rej_seq;
 	uint			i_irq_local_cpu;
 };
 
@@ -303,7 +298,6 @@ struct rds_ib_connection {
 #define IB_GET_POST_CREDITS(v)	((v) >> 16)
 #define IB_SET_SEND_CREDITS(v)	((v) & 0xffff)
 #define IB_SET_POST_CREDITS(v)	((v) << 16)
-#define IB_GET_CM_SEQ_NUM(v)	((v) & RDS_CM_REQ_SEQ_SZ)
 
 struct rds_ib_ipaddr {
 	struct list_head	list;

net/rds/ib_cm.c

@@ -384,7 +384,7 @@ static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 				      u32 protocol_version,
 				      u32 max_responder_resources,
 				      u32 max_initiator_depth, u16 frag,
-				      bool isv6, u8 seq)
+				      bool isv6)
 {
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct rds_ib_device *rds_ibdev = ic->rds_ibdev;
@@ -414,7 +414,6 @@ static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 			    cpu_to_be64(rds_ib_piggyb_ack(ic));
 			dp->ricp_v6.dp_tos = conn->c_tos;
 			dp->ricp_v6.dp_frag_sz = cpu_to_be16(frag);
-			dp->ricp_v6.dp_cm_seq = seq;
 
 			conn_param->private_data = &dp->ricp_v6;
 			conn_param->private_data_len = sizeof(dp->ricp_v6);
@@ -431,7 +430,6 @@ static void rds_ib_cm_fill_conn_param(struct rds_connection *conn,
 			    cpu_to_be64(rds_ib_piggyb_ack(ic));
 			dp->ricp_v4.dp_tos = conn->c_tos;
 			dp->ricp_v4.dp_frag_sz = cpu_to_be16(frag);
-			dp->ricp_v4.dp_cm_seq = seq;
 
 			conn_param->private_data = &dp->ricp_v4;
 			conn_param->private_data_len = sizeof(dp->ricp_v4);
@@ -1177,8 +1175,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 	u32 version;
 	int err = 1;
 	u16 frag;
-	u8 cm_req_seq = 0;
-	bool cm_seq_check_enable = false;
 
 	/* Check whether the remote protocol version matches ours. */
 	version = rds_ib_protocol_compatible(event, isv6);
@@ -1208,8 +1204,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 				goto out;
 			}
 		}
-		cm_seq_check_enable = dp->ricp_v6.dp_cm_seq & RDS_CM_RETRY_SEQ_EN;
-		cm_req_seq = IB_GET_CM_SEQ_NUM(dp->ricp_v6.dp_cm_seq);
 #else
 		err = -EOPNOTSUPP;
 		goto out;
@@ -1220,8 +1214,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 		ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr);
 		saddr6 = &s_mapped_addr;
 		daddr6 = &d_mapped_addr;
-		cm_seq_check_enable = dp->ricp_v4.dp_cm_seq & RDS_CM_RETRY_SEQ_EN;
-		cm_req_seq = IB_GET_CM_SEQ_NUM(dp->ricp_v4.dp_cm_seq);
 	}
 
 	rds_rtd_ptr(RDS_RTD_CM,
@@ -1273,22 +1265,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 	 */
 	mutex_lock(&conn->c_cm_lock);
 	ic = conn->c_transport_data;
-
-	if (ic && cm_seq_check_enable) {
-		if (cm_req_seq != ic->i_prev_seq) {
-			rds_rtd(RDS_RTD_CM_EXT_P,
-				"cm_id %p conn %p updating ic->i_prev_seq %d cm_req_seq %d\n",
-				cm_id, conn, ic->i_prev_seq, cm_req_seq);
-			ic->i_prev_seq = cm_req_seq;
-		} else if (cm_req_seq == ic->i_prev_seq && ic->i_last_rej_seq == cm_req_seq &&
-			   rds_conn_state(conn) == RDS_CONN_UP) {
-			rds_rtd(RDS_RTD_CM_EXT_P,
-				"duplicated REQ in UP state. cm_id %p conn %p reject! ic->i_last_rej_seq %d cm_req_seq %d\n",
-				cm_id, conn, ic->i_last_rej_seq, cm_req_seq);
-			goto out;
-		}
-	}
-
 	if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) {
 		/*
 		 * in both of the cases below, the conn is half setup.
@@ -1339,8 +1315,6 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 				rds_ib_stats_inc(s_ib_connect_raced);
 			}
 		}
-		if (ic && cm_seq_check_enable)
-			ic->i_last_rej_seq = cm_req_seq;
 		goto out;
 	} else {
 		/* Cancel any pending reconnect */
@@ -1395,7 +1369,7 @@ int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id,
 	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version,
 				  event->param.conn.responder_resources,
 				  event->param.conn.initiator_depth,
-				  frag, isv6, cm_req_seq);
+				  frag, isv6);
 
 	/* rdma_accept() calls rdma_reject() internally if it fails */
 	if (rds_ib_sysctl_local_ack_timeout &&
@@ -1450,7 +1424,6 @@ int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
 	union rds_ib_conn_priv dp;
 	u16 frag;
 	int ret;
-	u8 seq;
 
 #ifdef CONFIG_RDS_ACL
 
@@ -1495,11 +1468,9 @@ int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6)
 		goto out;
 	}
 	frag = rds_ib_set_frag_size(conn, ib_init_frag_size);
-	ic->i_req_sequence = IB_GET_CM_SEQ_NUM(ic->i_req_sequence + 1);
-	seq = RDS_CM_RETRY_SEQ_EN | ic->i_req_sequence;
 	rds_ib_cm_fill_conn_param(conn, &conn_param, &dp,
 				  conn->c_proposed_version, UINT_MAX, UINT_MAX,
-				  frag, isv6, seq);
+				  frag, isv6);
 	ret = rdma_connect(cm_id, &conn_param);
 	if (ret) {
 		rds_rtd(RDS_RTD_CM, "RDS/IB: rdma_connect failed (%d)\n", ret);