IB/mlx4: Add support for IBoE

Add support for IBoE to mlx4_ib.  The bulk of the code is handling the
new address vector fields; mlx4 needs the MAC address of a remote node
to include it in a WQE (for datagrams) or in the QP context (for
connected QPs).  Address resolution is done by assuming all unicast
GIDs are either link-local IPv6 addresses.

Multicast group attach/detach needs to update the NIC's multicast
filters; but since attaching a QP to a multicast group can be done
before the QP is bound to a port, for IBoE we need to keep track of
all multicast groups that a QP is attached too before it transitions
from INIT to RTR (since it does not have a port in the INIT state).
Signed-off-by: Eli Cohen <eli@mellanox.co.il>

[ Many things cleaned up and otherwise monkeyed with; hope I didn't
  introduce too many bugs.  - Roland ]
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/hw/mlx4/ah.c

@@ -30,66 +30,153 @@
  * SOFTWARE.
  */
 
+#include <rdma/ib_addr.h>
+
 #include <linux/slab.h>
+#include <linux/inet.h>
+#include <linux/string.h>
 
 #include "mlx4_ib.h"
 
-struct ib_ah *mlx4_ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
+int mlx4_ib_resolve_grh(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah_attr,
+			u8 *mac, int *is_mcast, u8 port)
 {
-	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
-	struct mlx4_ib_ah *ah;
+	struct in6_addr in6;
 
-	ah = kmalloc(sizeof *ah, GFP_ATOMIC);
-	if (!ah)
-		return ERR_PTR(-ENOMEM);
+	*is_mcast = 0;
 
-	memset(&ah->av, 0, sizeof ah->av);
+	memcpy(&in6, ah_attr->grh.dgid.raw, sizeof in6);
+	if (rdma_link_local_addr(&in6))
+		rdma_get_ll_mac(&in6, mac);
+	else if (rdma_is_multicast_addr(&in6)) {
+		rdma_get_mcast_mac(&in6, mac);
+		*is_mcast = 1;
+	} else
+		return -EINVAL;
 
-	ah->av.port_pd = cpu_to_be32(to_mpd(pd)->pdn | (ah_attr->port_num << 24));
-	ah->av.g_slid  = ah_attr->src_path_bits;
-	ah->av.dlid    = cpu_to_be16(ah_attr->dlid);
-	if (ah_attr->static_rate) {
-		ah->av.stat_rate = ah_attr->static_rate + MLX4_STAT_RATE_OFFSET;
-		while (ah->av.stat_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
-		       !(1 << ah->av.stat_rate & dev->caps.stat_rate_support))
-			--ah->av.stat_rate;
-	}
-	ah->av.sl_tclass_flowlabel = cpu_to_be32(ah_attr->sl << 28);
+	return 0;
+}
+
+static struct ib_ah *create_ib_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr,
+				  struct mlx4_ib_ah *ah)
+{
+	struct mlx4_dev *dev = to_mdev(pd->device)->dev;
+
+	ah->av.ib.port_pd = cpu_to_be32(to_mpd(pd)->pdn | (ah_attr->port_num << 24));
+	ah->av.ib.g_slid  = ah_attr->src_path_bits;
 	if (ah_attr->ah_flags & IB_AH_GRH) {
-		ah->av.g_slid   |= 0x80;
-		ah->av.gid_index = ah_attr->grh.sgid_index;
-		ah->av.hop_limit = ah_attr->grh.hop_limit;
-		ah->av.sl_tclass_flowlabel |=
+		ah->av.ib.g_slid   |= 0x80;
+		ah->av.ib.gid_index = ah_attr->grh.sgid_index;
+		ah->av.ib.hop_limit = ah_attr->grh.hop_limit;
+		ah->av.ib.sl_tclass_flowlabel |=
 			cpu_to_be32((ah_attr->grh.traffic_class << 20) |
 				    ah_attr->grh.flow_label);
-		memcpy(ah->av.dgid, ah_attr->grh.dgid.raw, 16);
+		memcpy(ah->av.ib.dgid, ah_attr->grh.dgid.raw, 16);
+	}
+
+	ah->av.ib.dlid    = cpu_to_be16(ah_attr->dlid);
+	if (ah_attr->static_rate) {
+		ah->av.ib.stat_rate = ah_attr->static_rate + MLX4_STAT_RATE_OFFSET;
+		while (ah->av.ib.stat_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
+		       !(1 << ah->av.ib.stat_rate & dev->caps.stat_rate_support))
+			--ah->av.ib.stat_rate;
 	}
+	ah->av.ib.sl_tclass_flowlabel = cpu_to_be32(ah_attr->sl << 28);
 
 	return &ah->ibah;
 }
 
+static struct ib_ah *create_iboe_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr,
+				    struct mlx4_ib_ah *ah)
+{
+	struct mlx4_ib_dev *ibdev = to_mdev(pd->device);
+	struct mlx4_dev *dev = ibdev->dev;
+	u8 mac[6];
+	int err;
+	int is_mcast;
+
+	err = mlx4_ib_resolve_grh(ibdev, ah_attr, mac, &is_mcast, ah_attr->port_num);
+	if (err)
+		return ERR_PTR(err);
+
+	memcpy(ah->av.eth.mac, mac, 6);
+	ah->av.eth.port_pd = cpu_to_be32(to_mpd(pd)->pdn | (ah_attr->port_num << 24));
+	ah->av.eth.gid_index = ah_attr->grh.sgid_index;
+	if (ah_attr->static_rate) {
+		ah->av.eth.stat_rate = ah_attr->static_rate + MLX4_STAT_RATE_OFFSET;
+		while (ah->av.eth.stat_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
+		       !(1 << ah->av.eth.stat_rate & dev->caps.stat_rate_support))
+			--ah->av.eth.stat_rate;
+	}
+
+	/*
+	 * HW requires multicast LID so we just choose one.
+	 */
+	if (is_mcast)
+		ah->av.ib.dlid = cpu_to_be16(0xc000);
+
+	memcpy(ah->av.eth.dgid, ah_attr->grh.dgid.raw, 16);
+	ah->av.eth.sl_tclass_flowlabel = cpu_to_be32(ah_attr->sl << 28);
+
+	return &ah->ibah;
+}
+
+struct ib_ah *mlx4_ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
+{
+	struct mlx4_ib_ah *ah;
+	struct ib_ah *ret;
+
+	ah = kzalloc(sizeof *ah, GFP_ATOMIC);
+	if (!ah)
+		return ERR_PTR(-ENOMEM);
+
+	if (rdma_port_get_link_layer(pd->device, ah_attr->port_num) == IB_LINK_LAYER_ETHERNET) {
+		if (!(ah_attr->ah_flags & IB_AH_GRH)) {
+			ret = ERR_PTR(-EINVAL);
+		} else {
+			/*
+			 * TBD: need to handle the case when we get
+			 * called in an atomic context and there we
+			 * might sleep.  We don't expect this
+			 * currently since we're working with link
+			 * local addresses which we can translate
+			 * without going to sleep.
+			 */
+			ret = create_iboe_ah(pd, ah_attr, ah);
+		}
+
+		if (IS_ERR(ret))
+			kfree(ah);
+
+		return ret;
+	} else
+		return create_ib_ah(pd, ah_attr, ah); /* never fails */
+}
+
 int mlx4_ib_query_ah(struct ib_ah *ibah, struct ib_ah_attr *ah_attr)
 {
 	struct mlx4_ib_ah *ah = to_mah(ibah);
+	enum rdma_link_layer ll;
 
 	memset(ah_attr, 0, sizeof *ah_attr);
-	ah_attr->dlid	       = be16_to_cpu(ah->av.dlid);
-	ah_attr->sl	       = be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
-	ah_attr->port_num      = be32_to_cpu(ah->av.port_pd) >> 24;
-	if (ah->av.stat_rate)
-		ah_attr->static_rate = ah->av.stat_rate - MLX4_STAT_RATE_OFFSET;
-	ah_attr->src_path_bits = ah->av.g_slid & 0x7F;
+	ah_attr->sl = be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
+	ah_attr->port_num = be32_to_cpu(ah->av.ib.port_pd) >> 24;
+	ll = rdma_port_get_link_layer(ibah->device, ah_attr->port_num);
+	ah_attr->dlid = ll == IB_LINK_LAYER_INFINIBAND ? be16_to_cpu(ah->av.ib.dlid) : 0;
+	if (ah->av.ib.stat_rate)
+		ah_attr->static_rate = ah->av.ib.stat_rate - MLX4_STAT_RATE_OFFSET;
+	ah_attr->src_path_bits = ah->av.ib.g_slid & 0x7F;
 
 	if (mlx4_ib_ah_grh_present(ah)) {
 		ah_attr->ah_flags = IB_AH_GRH;
 
 		ah_attr->grh.traffic_class =
-			be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20;
+			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20;
 		ah_attr->grh.flow_label =
-			be32_to_cpu(ah->av.sl_tclass_flowlabel) & 0xfffff;
-		ah_attr->grh.hop_limit  = ah->av.hop_limit;
-		ah_attr->grh.sgid_index = ah->av.gid_index;
-		memcpy(ah_attr->grh.dgid.raw, ah->av.dgid, 16);
+			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) & 0xfffff;
+		ah_attr->grh.hop_limit  = ah->av.ib.hop_limit;
+		ah_attr->grh.sgid_index = ah->av.ib.gid_index;
+		memcpy(ah_attr->grh.dgid.raw, ah->av.ib.dgid, 16);
 	}
 
 	return 0;

drivers/infiniband/hw/mlx4/mad.c

@@ -311,19 +311,25 @@ int mlx4_ib_mad_init(struct mlx4_ib_dev *dev)
 	struct ib_mad_agent *agent;
 	int p, q;
 	int ret;
+	enum rdma_link_layer ll;
 
-	for (p = 0; p < dev->num_ports; ++p)
+	for (p = 0; p < dev->num_ports; ++p) {
+		ll = rdma_port_get_link_layer(&dev->ib_dev, p + 1);
 		for (q = 0; q <= 1; ++q) {
-			agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
-						      q ? IB_QPT_GSI : IB_QPT_SMI,
-						      NULL, 0, send_handler,
-						      NULL, NULL);
-			if (IS_ERR(agent)) {
-				ret = PTR_ERR(agent);
-				goto err;
-			}
-			dev->send_agent[p][q] = agent;
+			if (ll == IB_LINK_LAYER_INFINIBAND) {
+				agent = ib_register_mad_agent(&dev->ib_dev, p + 1,
+							      q ? IB_QPT_GSI : IB_QPT_SMI,
+							      NULL, 0, send_handler,
+							      NULL, NULL);
+				if (IS_ERR(agent)) {
+					ret = PTR_ERR(agent);
+					goto err;
+				}
+				dev->send_agent[p][q] = agent;
+			} else
+				dev->send_agent[p][q] = NULL;
 		}
+	}
 
 	return 0;
 
@@ -344,8 +350,10 @@ void mlx4_ib_mad_cleanup(struct mlx4_ib_dev *dev)
 	for (p = 0; p < dev->num_ports; ++p) {
 		for (q = 0; q <= 1; ++q) {
 			agent = dev->send_agent[p][q];
-			dev->send_agent[p][q] = NULL;
-			ib_unregister_mad_agent(agent);
+			if (agent) {
+				dev->send_agent[p][q] = NULL;
+				ib_unregister_mad_agent(agent);
+			}
 		}
 
 		if (dev->sm_ah[p])

drivers/infiniband/hw/mlx4/mlx4_ib.h

@@ -112,6 +112,13 @@ enum mlx4_ib_qp_flags {
 	MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK	= 1 << 1,
 };
 
+struct mlx4_ib_gid_entry {
+	struct list_head	list;
+	union ib_gid		gid;
+	int			added;
+	u8			port;
+};
+
 struct mlx4_ib_qp {
 	struct ib_qp		ibqp;
 	struct mlx4_qp		mqp;
@@ -138,6 +145,8 @@ struct mlx4_ib_qp {
 	u8			resp_depth;
 	u8			sq_no_prefetch;
 	u8			state;
+	int			mlx_type;
+	struct list_head	gid_list;
 };
 
 struct mlx4_ib_srq {
@@ -157,7 +166,14 @@ struct mlx4_ib_srq {
 
 struct mlx4_ib_ah {
 	struct ib_ah		ibah;
-	struct mlx4_av		av;
+	union mlx4_ext_av       av;
+};
+
+struct mlx4_ib_iboe {
+	spinlock_t		lock;
+	struct net_device      *netdevs[MLX4_MAX_PORTS];
+	struct notifier_block 	nb;
+	union ib_gid		gid_table[MLX4_MAX_PORTS][128];
 };
 
 struct mlx4_ib_dev {
@@ -176,6 +192,7 @@ struct mlx4_ib_dev {
 
 	struct mutex		cap_mask_mutex;
 	bool			ib_active;
+	struct mlx4_ib_iboe	iboe;
 };
 
 static inline struct mlx4_ib_dev *to_mdev(struct ib_device *ibdev)
@@ -314,9 +331,20 @@ int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, int npages,
 int mlx4_ib_unmap_fmr(struct list_head *fmr_list);
 int mlx4_ib_fmr_dealloc(struct ib_fmr *fmr);
 
+int mlx4_ib_resolve_grh(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah_attr,
+			u8 *mac, int *is_mcast, u8 port);
+
 static inline int mlx4_ib_ah_grh_present(struct mlx4_ib_ah *ah)
 {
-	return !!(ah->av.g_slid & 0x80);
+	u8 port = be32_to_cpu(ah->av.ib.port_pd) >> 24 & 3;
+
+	if (rdma_port_get_link_layer(ah->ibah.device, port) == IB_LINK_LAYER_ETHERNET)
+		return 1;
+
+	return !!(ah->av.ib.g_slid & 0x80);
 }
 
+int mlx4_ib_add_mc(struct mlx4_ib_dev *mdev, struct mlx4_ib_qp *mqp,
+		   union ib_gid *gid);
+
 #endif /* MLX4_IB_H */

drivers/infiniband/hw/mlx4/qp.c

@@ -33,6 +33,7 @@
 
 #include <linux/log2.h>
 #include <linux/slab.h>
+#include <linux/netdevice.h>
 
 #include <rdma/ib_cache.h>
 #include <rdma/ib_pack.h>
@@ -48,17 +49,25 @@ enum {
 
 enum {
 	MLX4_IB_DEFAULT_SCHED_QUEUE	= 0x83,
-	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f
+	MLX4_IB_DEFAULT_QP0_SCHED_QUEUE	= 0x3f,
+	MLX4_IB_LINK_TYPE_IB		= 0,
+	MLX4_IB_LINK_TYPE_ETH		= 1
 };
 
 enum {
 	/*
-	 * Largest possible UD header: send with GRH and immediate data.
+	 * Largest possible UD header: send with GRH and immediate
+	 * data plus 14 bytes for an Ethernet header.  (LRH would only
+	 * use 8 bytes, so Ethernet is the biggest case)
 	 */
-	MLX4_IB_UD_HEADER_SIZE		= 72,
+	MLX4_IB_UD_HEADER_SIZE		= 78,
 	MLX4_IB_LSO_HEADER_SPARE	= 128,
 };
 
+enum {
+	MLX4_IB_IBOE_ETHERTYPE		= 0x8915
+};
+
 struct mlx4_ib_sqp {
 	struct mlx4_ib_qp	qp;
 	int			pkey_index;
@@ -462,6 +471,7 @@ static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 	mutex_init(&qp->mutex);
 	spin_lock_init(&qp->sq.lock);
 	spin_lock_init(&qp->rq.lock);
+	INIT_LIST_HEAD(&qp->gid_list);
 
 	qp->state	 = IB_QPS_RESET;
 	if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
@@ -649,6 +659,16 @@ static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *re
 	}
 }
 
+static void del_gid_entries(struct mlx4_ib_qp *qp)
+{
+	struct mlx4_ib_gid_entry *ge, *tmp;
+
+	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
+		list_del(&ge->list);
+		kfree(ge);
+	}
+}
+
 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
 			      int is_user)
 {
@@ -695,6 +715,8 @@ static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
 		if (!qp->ibqp.srq)
 			mlx4_db_free(dev->dev, &qp->db);
 	}
+
+	del_gid_entries(qp);
 }
 
 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
@@ -852,6 +874,12 @@ static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
 			 struct mlx4_qp_path *path, u8 port)
 {
+	int err;
+	int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
+		IB_LINK_LAYER_ETHERNET;
+	u8 mac[6];
+	int is_mcast;
+
 	path->grh_mylmc     = ah->src_path_bits & 0x7f;
 	path->rlid	    = cpu_to_be16(ah->dlid);
 	if (ah->static_rate) {
@@ -882,9 +910,35 @@ static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
 	path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
 		((port - 1) << 6) | ((ah->sl & 0xf) << 2);
 
+	if (is_eth) {
+		if (!(ah->ah_flags & IB_AH_GRH))
+			return -1;
+
+		err = mlx4_ib_resolve_grh(dev, ah, mac, &is_mcast, port);
+		if (err)
+			return err;
+
+		memcpy(path->dmac, mac, 6);
+		path->ackto = MLX4_IB_LINK_TYPE_ETH;
+		/* use index 0 into MAC table for IBoE */
+		path->grh_mylmc &= 0x80;
+	}
+
 	return 0;
 }
 
+static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
+{
+	struct mlx4_ib_gid_entry *ge, *tmp;
+
+	list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
+		if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
+			ge->added = 1;
+			ge->port = qp->port;
+		}
+	}
+}
+
 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
 			       const struct ib_qp_attr *attr, int attr_mask,
 			       enum ib_qp_state cur_state, enum ib_qp_state new_state)
@@ -980,7 +1034,7 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
 	}
 
 	if (attr_mask & IB_QP_TIMEOUT) {
-		context->pri_path.ackto = attr->timeout << 3;
+		context->pri_path.ackto |= attr->timeout << 3;
 		optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
 	}
 
@@ -1118,8 +1172,10 @@ static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
 		qp->atomic_rd_en = attr->qp_access_flags;
 	if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
 		qp->resp_depth = attr->max_dest_rd_atomic;
-	if (attr_mask & IB_QP_PORT)
+	if (attr_mask & IB_QP_PORT) {
 		qp->port = attr->port_num;
+		update_mcg_macs(dev, qp);
+	}
 	if (attr_mask & IB_QP_ALT_PATH)
 		qp->alt_port = attr->alt_port_num;
 
@@ -1226,35 +1282,45 @@ static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
 	int header_size;
 	int spc;
 	int i;
+	int is_eth;
+	int is_grh;
 
 	send_size = 0;
 	for (i = 0; i < wr->num_sge; ++i)
 		send_size += wr->sg_list[i].length;
 
-	ib_ud_header_init(send_size, 1, 0, mlx4_ib_ah_grh_present(ah), 0, &sqp->ud_header);
+	is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
+	is_grh = mlx4_ib_ah_grh_present(ah);
+	ib_ud_header_init(send_size, !is_eth, is_eth, is_grh, 0, &sqp->ud_header);
+
+	if (!is_eth) {
+		sqp->ud_header.lrh.service_level =
+			be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
+		sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
+		sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
+	}
 
-	sqp->ud_header.lrh.service_level   =
-		be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 28;
-	sqp->ud_header.lrh.destination_lid = ah->av.dlid;
-	sqp->ud_header.lrh.source_lid      = cpu_to_be16(ah->av.g_slid & 0x7f);
-	if (mlx4_ib_ah_grh_present(ah)) {
+	if (is_grh) {
 		sqp->ud_header.grh.traffic_class =
-			(be32_to_cpu(ah->av.sl_tclass_flowlabel) >> 20) & 0xff;
+			(be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
 		sqp->ud_header.grh.flow_label    =
-			ah->av.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
-		sqp->ud_header.grh.hop_limit     = ah->av.hop_limit;
-		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.port_pd) >> 24,
-				  ah->av.gid_index, &sqp->ud_header.grh.source_gid);
+			ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
+		sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
+		ib_get_cached_gid(ib_dev, be32_to_cpu(ah->av.ib.port_pd) >> 24,
+				  ah->av.ib.gid_index, &sqp->ud_header.grh.source_gid);
 		memcpy(sqp->ud_header.grh.destination_gid.raw,
-		       ah->av.dgid, 16);
+		       ah->av.ib.dgid, 16);
 	}
 
 	mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
-	mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
-				  (sqp->ud_header.lrh.destination_lid ==
-				   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
-				  (sqp->ud_header.lrh.service_level << 8));
-	mlx->rlid   = sqp->ud_header.lrh.destination_lid;
+
+	if (!is_eth) {
+		mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
+					  (sqp->ud_header.lrh.destination_lid ==
+					   IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
+					  (sqp->ud_header.lrh.service_level << 8));
+		mlx->rlid = sqp->ud_header.lrh.destination_lid;
+	}
 
 	switch (wr->opcode) {
 	case IB_WR_SEND:
@@ -1270,9 +1336,21 @@ static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
 		return -EINVAL;
 	}
 
-	sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
-	if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
-		sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
+	if (is_eth) {
+		u8 *smac;
+
+		memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
+		/* FIXME: cache smac value? */
+		smac = to_mdev(sqp->qp.ibqp.device)->iboe.netdevs[sqp->qp.port - 1]->dev_addr;
+		memcpy(sqp->ud_header.eth.smac_h, smac, 6);
+		if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
+			mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
+		sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
+	} else {
+		sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
+		if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
+			sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
+	}
 	sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
 	if (!sqp->qp.ibqp.qp_num)
 		ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
@@ -1434,6 +1512,8 @@ static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
 	memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
 	dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
 	dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
+	dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
+	memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
 }
 
 static void set_mlx_icrc_seg(void *dseg)

include/linux/mlx4/device.h

@@ -374,6 +374,27 @@ struct mlx4_av {
 	u8			dgid[16];
 };
 
+struct mlx4_eth_av {
+	__be32		port_pd;
+	u8		reserved1;
+	u8		smac_idx;
+	u16		reserved2;
+	u8		reserved3;
+	u8		gid_index;
+	u8		stat_rate;
+	u8		hop_limit;
+	__be32		sl_tclass_flowlabel;
+	u8		dgid[16];
+	u32		reserved4[2];
+	__be16		vlan;
+	u8		mac[6];
+};
+
+union mlx4_ext_av {
+	struct mlx4_av		ib;
+	struct mlx4_eth_av	eth;
+};
+
 struct mlx4_dev {
 	struct pci_dev	       *pdev;
 	unsigned long		flags;
@@ -402,6 +423,12 @@ struct mlx4_init_port_param {
 		if (((type) == MLX4_PORT_TYPE_IB ? (dev)->caps.port_mask : \
 		     ~(dev)->caps.port_mask) & 1 << ((port) - 1))
 
+#define mlx4_foreach_ib_transport_port(port, dev)			\
+	for ((port) = 1; (port) <= (dev)->caps.num_ports; (port)++)	\
+		if (((dev)->caps.port_mask & 1 << ((port) - 1)) ||	\
+		    ((dev)->caps.flags & MLX4_DEV_CAP_FLAG_IBOE))
+
+
 int mlx4_buf_alloc(struct mlx4_dev *dev, int size, int max_direct,
 		   struct mlx4_buf *buf);
 void mlx4_buf_free(struct mlx4_dev *dev, int size, struct mlx4_buf *buf);