[ipoib] Expose Ethernet-compatible eIPoIB link-layer addresses and headers

Almost all clients of the raw-packet interfaces (UNDI and SNP) can
handle only Ethernet link layers.  Expose an Ethernet-compatible link
layer to local clients, while remaining compatible with IPoIB on the
wire.  This requires manipulation of ARP (but not DHCP) packets within
the IPoIB driver.

This is ugly, but it's the only viable way to allow IPoIB devices to
be driven via the raw-packet interfaces.

Signed-off-by: Michael Brown <mcb30@ipxe.org>

1 files changed, 375 insertions, 260 deletions

diff --git a/src/drivers/net/ipoib.c b/src/drivers/net/ipoib.c
index 41ba348..c1b8cad 100644
--- a/src/drivers/net/ipoib.c
+++ b/src/drivers/net/ipoib.c
@@ -20,18 +20,23 @@
 FILE_LICENCE ( GPL2_OR_LATER );
 
 #include <stdint.h>
+#include <stdlib.h>
 #include <stdio.h>
 #include <unistd.h>
 #include <string.h>
 #include <byteswap.h>
 #include <errno.h>
 #include <ipxe/errortab.h>
+#include <ipxe/malloc.h>
 #include <ipxe/if_arp.h>
+#include <ipxe/if_ether.h>
+#include <ipxe/ethernet.h>
 #include <ipxe/iobuf.h>
 #include <ipxe/netdevice.h>
 #include <ipxe/infiniband.h>
 #include <ipxe/ib_pathrec.h>
 #include <ipxe/ib_mcast.h>
+#include <ipxe/retry.h>
 #include <ipxe/ipoib.h>
 
 /** @file
@@ -58,6 +63,8 @@ struct ipoib_device {
 	struct ib_completion_queue *cq;
 	/** Queue pair */
 	struct ib_queue_pair *qp;
+	/** Local MAC */
+	struct ipoib_mac mac;
 	/** Broadcast MAC */
 	struct ipoib_mac broadcast;
 	/** Joined to IPv4 broadcast multicast group
@@ -68,6 +75,8 @@ struct ipoib_device {
 	int broadcast_joined;
 	/** IPv4 broadcast multicast group membership */
 	struct ib_mc_membership broadcast_membership;
+	/** REMAC cache */
+	struct list_head peers;
 };
 
 /** Broadcast IPoIB address */
@@ -89,184 +98,140 @@ struct errortab ipoib_errors[] __errortab = {
 
 /****************************************************************************
  *
- * IPoIB peer cache
+ * IPoIB REMAC cache
  *
  ****************************************************************************
  */
 
-/**
- * IPoIB peer address
- *
- * The IPoIB link-layer header is only four bytes long and so does not
- * have sufficient room to store IPoIB MAC address(es).  We therefore
- * maintain a cache of MAC addresses identified by a single-byte key,
- * and abuse the spare two bytes within the link-layer header to
- * communicate these MAC addresses between the link-layer code and the
- * netdevice driver.
- */
+/** An IPoIB REMAC cache entry */
 struct ipoib_peer {
-	/** Key */
-	uint8_t key;
+	/** List of REMAC cache entries */
+	struct list_head list;
+	/** Remote Ethermet MAC */
+	struct ipoib_remac remac;
 	/** MAC address */
 	struct ipoib_mac mac;
 };
 
-/** Number of IPoIB peer cache entries
- *
- * Must be a power of two.
- */
-#define IPOIB_NUM_CACHED_PEERS 4
-
-/** IPoIB peer address cache */
-static struct ipoib_peer ipoib_peer_cache[IPOIB_NUM_CACHED_PEERS];
-
-/** Oldest IPoIB peer cache entry index */
-static unsigned int ipoib_peer_cache_idx = 0;
-
-/** IPoIB peer cache entry validity flag */
-#define IPOIB_PEER_KEY_VALID 0x80
-
 /**
- * Look up cached peer by key
+ * Find IPoIB MAC from REMAC
  *
- * @v key		Peer cache key
- * @ret peer		Peer cache entry, or NULL
+ * @v ipoib		IPoIB device
+ * @v remac		Remote Ethernet MAC
+ * @ret mac		IPoIB MAC (or NULL if not found)
  */
-static struct ipoib_peer * ipoib_lookup_peer_by_key ( unsigned int key ) {
+static struct ipoib_mac * ipoib_find_remac ( struct ipoib_device *ipoib,
+					     const struct ipoib_remac *remac ) {
 	struct ipoib_peer *peer;
-	unsigned int i;
 
-	if ( ! key )
-		return NULL;
-
-	for ( i = 0 ; i < IPOIB_NUM_CACHED_PEERS ; i++ ) {
-		peer = &ipoib_peer_cache[i];
-		if ( peer->key == key )
-			return peer;
+	/* Check for broadcast REMAC */
+	if ( is_broadcast_ether_addr ( remac ) )
+		return &ipoib->broadcast;
+
+	/* Try to find via REMAC cache */
+	list_for_each_entry ( peer, &ipoib->peers, list ) {
+		if ( memcmp ( remac, &peer->remac,
+			      sizeof ( peer->remac ) ) == 0 ) {
+			/* Move peer to start of list */
+			list_del ( &peer->list );
+			list_add ( &peer->list, &ipoib->peers );
+			return &peer->mac;
+		}
 	}
 
-	DBG ( "IPoIB warning: peer cache lost track of key %x while still in "
-	      "use\n", key );
+	DBGC ( ipoib, "IPoIB %p unknown REMAC %s\n",
+	       ipoib, eth_ntoa ( remac ) );
 	return NULL;
 }
 
 /**
- * Store GID and QPN in peer cache
+ * Add IPoIB MAC to REMAC cache
  *
- * @v mac		Peer MAC address
- * @ret peer		Peer cache entry
+ * @v ipoib		IPoIB device
+ * @v remac		Remote Ethernet MAC
+ * @v mac		IPoIB MAC
+ * @ret rc		Return status code
  */
-static struct ipoib_peer * ipoib_cache_peer ( const struct ipoib_mac *mac ) {
+static int ipoib_map_remac ( struct ipoib_device *ipoib,
+			     const struct ipoib_remac *remac,
+			     const struct ipoib_mac *mac ) {
 	struct ipoib_peer *peer;
-	uint8_t key;
-	unsigned int i;
 
-	/* Look for existing cache entry */
-	for ( i = 0 ; i < IPOIB_NUM_CACHED_PEERS ; i++ ) {
-		peer = &ipoib_peer_cache[i];
-		if ( memcmp ( &peer->mac, mac, sizeof ( peer->mac ) ) == 0 )
-			return peer;
+	/* Check for existing entry in REMAC cache */
+	list_for_each_entry ( peer, &ipoib->peers, list ) {
+		if ( memcmp ( remac, &peer->remac,
+			      sizeof ( peer->remac ) ) == 0 ) {
+			/* Move peer to start of list */
+			list_del ( &peer->list );
+			list_add ( &peer->list, &ipoib->peers );
+			/* Update MAC */
+			memcpy ( &peer->mac, mac, sizeof ( peer->mac ) );
+			return 0;
+		}
 	}
 
-	/* No entry found: create a new one */
-	key = ( ipoib_peer_cache_idx++ | IPOIB_PEER_KEY_VALID );
-	peer = &ipoib_peer_cache[ key % IPOIB_NUM_CACHED_PEERS ];
-	if ( peer->key )
-		DBG ( "IPoIB peer %x evicted from cache\n", peer->key );
-
-	memset ( peer, 0, sizeof ( *peer ) );
-	peer->key = key;
+	/* Create new entry */
+	peer = malloc ( sizeof ( *peer ) );
+	if ( ! peer )
+		return -ENOMEM;
+	memcpy ( &peer->remac, remac, sizeof ( peer->remac ) );
 	memcpy ( &peer->mac, mac, sizeof ( peer->mac ) );
-	DBG ( "IPoIB peer %x has MAC %s\n",
-	      peer->key, ipoib_ntoa ( &peer->mac ) );
-	return peer;
-}
+	list_add ( &peer->list, &ipoib->peers );
 
-/****************************************************************************
- *
- * IPoIB link layer
- *
- ****************************************************************************
- */
+	return 0;
+}
 
 /**
- * Add IPoIB link-layer header
+ * Flush REMAC cache
  *
- * @v netdev		Network device
- * @v iobuf		I/O buffer
- * @v ll_dest		Link-layer destination address
- * @v ll_source		Source link-layer address
- * @v net_proto		Network-layer protocol, in network-byte order
- * @ret rc		Return status code
+ * @v ipoib		IPoIB device
  */
-static int ipoib_push ( struct net_device *netdev __unused,
-			struct io_buffer *iobuf, const void *ll_dest,
-			const void *ll_source __unused, uint16_t net_proto ) {
-	struct ipoib_hdr *ipoib_hdr =
-		iob_push ( iobuf, sizeof ( *ipoib_hdr ) );
-	const struct ipoib_mac *dest_mac = ll_dest;
-	const struct ipoib_mac *src_mac = ll_source;
-	struct ipoib_peer *dest;
-	struct ipoib_peer *src;
-
-	/* Add link-layer addresses to cache */
-	dest = ipoib_cache_peer ( dest_mac );
-	src = ipoib_cache_peer ( src_mac );
-
-	/* Build IPoIB header */
-	ipoib_hdr->proto = net_proto;
-	ipoib_hdr->u.peer.dest = dest->key;
-	ipoib_hdr->u.peer.src = src->key;
+static void ipoib_flush_remac ( struct ipoib_device *ipoib ) {
+	struct ipoib_peer *peer;
+	struct ipoib_peer *tmp;
 
-	return 0;
+	list_for_each_entry_safe ( peer, tmp, &ipoib->peers, list ) {
+		list_del ( &peer->list );
+		free ( peer );
+	}
 }
 
 /**
- * Remove IPoIB link-layer header
+ * Discard some entries from the REMAC cache
  *
- * @v netdev		Network device
- * @v iobuf		I/O buffer
- * @ret ll_dest		Link-layer destination address
- * @ret ll_source	Source link-layer address
- * @ret net_proto	Network-layer protocol, in network-byte order
- * @ret flags		Packet flags
- * @ret rc		Return status code
+ * @ret discarded	Number of cached items discarded
  */
-static int ipoib_pull ( struct net_device *netdev,
-			struct io_buffer *iobuf, const void **ll_dest,
-			const void **ll_source, uint16_t *net_proto,
-			unsigned int *flags ) {
-	struct ipoib_device *ipoib = netdev->priv;
-	struct ipoib_hdr *ipoib_hdr = iobuf->data;
-	struct ipoib_peer *dest;
-	struct ipoib_peer *source;
-
-	/* Sanity check */
-	if ( iob_len ( iobuf ) < sizeof ( *ipoib_hdr ) ) {
-		DBG ( "IPoIB packet too short for link-layer header\n" );
-		DBG_HD ( iobuf->data, iob_len ( iobuf ) );
-		return -EINVAL;
+static unsigned int ipoib_discard_remac ( void ) {
+	struct ib_device *ibdev;
+	struct ipoib_device *ipoib;
+	struct ipoib_peer *peer;
+	unsigned int discarded = 0;
+
+	/* Try to discard one cache entry for each IPoIB device */
+	for_each_ibdev ( ibdev ) {
+		ipoib = ib_get_ownerdata ( ibdev );
+		list_for_each_entry_reverse ( peer, &ipoib->peers, list ) {
+			list_del ( &peer->list );
+			free ( peer );
+			discarded++;
+			break;
+		}
 	}
 
-	/* Strip off IPoIB header */
-	iob_pull ( iobuf, sizeof ( *ipoib_hdr ) );
-
-	/* Identify source and destination addresses, and clear
-	 * reserved word in IPoIB header
-	 */
-	dest = ipoib_lookup_peer_by_key ( ipoib_hdr->u.peer.dest );
-	source = ipoib_lookup_peer_by_key ( ipoib_hdr->u.peer.src );
-	ipoib_hdr->u.reserved = 0;
+	return discarded;
+}
 
-	/* Fill in required fields */
-	*ll_dest = ( dest ? &dest->mac : &ipoib->broadcast );
-	*ll_source = ( source ? &source->mac : &ipoib->broadcast );
-	*net_proto = ipoib_hdr->proto;
-	*flags = ( ( *ll_dest == &ipoib->broadcast ) ?
-		   ( LL_MULTICAST | LL_BROADCAST ) : 0 );
+/** IPoIB cache discarder */
+struct cache_discarder ipoib_discarder __cache_discarder ( CACHE_NORMAL ) = {
+	.discard = ipoib_discard_remac,
+};
 
-	return 0;
-}
+/****************************************************************************
+ *
+ * IPoIB link layer
+ *
+ ****************************************************************************
+ */
 
 /**
  * Initialise IPoIB link-layer address
@@ -275,133 +240,205 @@ static int ipoib_pull ( struct net_device *netdev,
  * @v ll_addr		Link-layer address
  */
 static void ipoib_init_addr ( const void *hw_addr, void *ll_addr ) {
-	const union ib_guid *guid = hw_addr;
-	struct ipoib_mac *mac = ll_addr;
+	const uint8_t *guid = hw_addr;
+	uint8_t *eth_addr = ll_addr;
+	uint8_t guid_mask = IPOIB_GUID_MASK;
+	unsigned int i;
 
-	memset ( mac, 0, sizeof ( *mac ) );
-	memcpy ( &mac->gid.s.guid, guid, sizeof ( mac->gid.s.guid ) );
+	/* Extract bytes from GUID according to mask */
+	for ( i = 0 ; i < 8 ; i++, guid++, guid_mask <<= 1 ) {
+		if ( guid_mask & 0x80 )
+			*(eth_addr++) = *guid;
+	}
 }
 
+/** IPoIB protocol */
+struct ll_protocol ipoib_protocol __ll_protocol = {
+	.name		= "IPoIB",
+	.ll_proto	= htons ( ARPHRD_ETHER ),
+	.hw_addr_len	= sizeof ( union ib_guid ),
+	.ll_addr_len	= ETH_ALEN,
+	.ll_header_len	= ETH_HLEN,
+	.push		= eth_push,
+	.pull		= eth_pull,
+	.init_addr	= ipoib_init_addr,
+	.ntoa		= eth_ntoa,
+	.mc_hash	= eth_mc_hash,
+	.eth_addr	= eth_eth_addr,
+	.flags		= LL_NAME_ONLY,
+};
+
 /**
- * Transcribe IPoIB link-layer address
+ * Allocate IPoIB device
  *
- * @v ll_addr	Link-layer address
- * @ret string	Link-layer address in human-readable format
+ * @v priv_size		Size of driver private data
+ * @ret netdev		Network device, or NULL
  */
-const char * ipoib_ntoa ( const void *ll_addr ) {
-	static char buf[45];
-	const struct ipoib_mac *mac = ll_addr;
-
-	snprintf ( buf, sizeof ( buf ), "%08x:%08x:%08x:%08x:%08x",
-		   htonl ( mac->flags__qpn ), htonl ( mac->gid.dwords[0] ),
-		   htonl ( mac->gid.dwords[1] ),
-		   htonl ( mac->gid.dwords[2] ),
-		   htonl ( mac->gid.dwords[3] ) );
-	return buf;
+struct net_device * alloc_ipoibdev ( size_t priv_size ) {
+	struct net_device *netdev;
+
+	netdev = alloc_netdev ( priv_size );
+	if ( netdev ) {
+		netdev->ll_protocol = &ipoib_protocol;
+		netdev->ll_broadcast = eth_broadcast;
+		netdev->max_pkt_len = IB_MAX_PAYLOAD_SIZE;
+	}
+	return netdev;
 }
 
+/****************************************************************************
+ *
+ * IPoIB translation layer
+ *
+ ****************************************************************************
+ */
+
 /**
- * Hash multicast address
+ * Translate transmitted ARP packet
  *
- * @v af		Address family
- * @v net_addr		Network-layer address
- * @v ll_addr		Link-layer address to fill in
+ * @v netdev		Network device
+ * @v iobuf		Packet to be transmitted (with no link-layer headers)
  * @ret rc		Return status code
  */
-static int ipoib_mc_hash ( unsigned int af __unused,
-			   const void *net_addr __unused,
-			   void *ll_addr __unused ) {
+static int ipoib_translate_tx_arp ( struct net_device *netdev,
+				    struct io_buffer *iobuf ) {
+	struct ipoib_device *ipoib = netdev->priv;
+	struct arphdr *arphdr = iobuf->data;
+	struct ipoib_mac *target_ha = NULL;
+	void *sender_pa;
+	void *target_pa;
+
+	/* Do nothing unless ARP contains eIPoIB link-layer addresses */
+	if ( arphdr->ar_hln != ETH_ALEN )
+		return 0;
+
+	/* Fail unless we have room to expand packet */
+	if ( iob_tailroom ( iobuf ) < ( 2 * ( sizeof ( ipoib->mac ) -
+					      ETH_ALEN ) ) ) {
+		DBGC ( ipoib, "IPoIB %p insufficient space in TX ARP\n",
+		       ipoib );
+		return -ENOBUFS;
+	}
 
-	return -ENOTSUP;
+	/* Look up REMAC, if applicable */
+	if ( arphdr->ar_op == ARPOP_REPLY ) {
+		target_ha = ipoib_find_remac ( ipoib, arp_target_pa ( arphdr ));
+		if ( ! target_ha )
+			return -ENXIO;
+	}
+
+	/* Construct new packet */
+	iob_put ( iobuf, ( 2 * ( sizeof ( ipoib->mac ) - ETH_ALEN ) ) );
+	sender_pa = arp_sender_pa ( arphdr );
+	target_pa = arp_target_pa ( arphdr );
+	arphdr->ar_hrd = htons ( ARPHRD_INFINIBAND );
+	arphdr->ar_hln = sizeof ( ipoib->mac );
+	memcpy ( arp_target_pa ( arphdr ), target_pa, arphdr->ar_pln );
+	memcpy ( arp_sender_pa ( arphdr ), sender_pa, arphdr->ar_pln );
+	memcpy ( arp_sender_ha ( arphdr ), &ipoib->mac, sizeof ( ipoib->mac ) );
+	memset ( arp_target_ha ( arphdr ), 0, sizeof ( ipoib->mac ) );
+	if ( target_ha ) {
+		memcpy ( arp_target_ha ( arphdr ), target_ha,
+			 sizeof ( *target_ha ) );
+	}
+
+	return 0;
 }
 
 /**
- * Generate Mellanox Ethernet-compatible compressed link-layer address
+ * Translate transmitted packet
  *
- * @v ll_addr		Link-layer address
- * @v eth_addr		Ethernet-compatible address to fill in
+ * @v netdev		Network device
+ * @v iobuf		Packet to be transmitted (with no link-layer headers)
+ * @v net_proto		Network-layer protocol (in network byte order)
+ * @ret rc		Return status code
  */
-static int ipoib_mlx_eth_addr ( const union ib_guid *guid,
-				uint8_t *eth_addr ) {
-	eth_addr[0] = ( ( guid->bytes[3] == 2 ) ? 0x00 : 0x02 );
-	eth_addr[1] = guid->bytes[1];
-	eth_addr[2] = guid->bytes[2];
-	eth_addr[3] = guid->bytes[5];
-	eth_addr[4] = guid->bytes[6];
-	eth_addr[5] = guid->bytes[7];
-	return 0;
+static int ipoib_translate_tx ( struct net_device *netdev,
+				struct io_buffer *iobuf, uint16_t net_proto ) {
+
+	switch ( net_proto ) {
+	case htons ( ETH_P_ARP ) :
+		return ipoib_translate_tx_arp ( netdev, iobuf );
+	case htons ( ETH_P_IP ) :
+		/* No translation needed */
+		return 0;
+	default:
+		/* Cannot handle other traffic via eIPoIB */
+		return -ENOTSUP;
+	}
 }
 
-/** An IPoIB Ethernet-compatible compressed link-layer address generator */
-struct ipoib_eth_addr_handler {
-	/** GUID byte 1 */
-	uint8_t byte1;
-	/** GUID byte 2 */
-	uint8_t byte2;
-	/** Handler */
-	int ( * eth_addr ) ( const union ib_guid *guid,
-			     uint8_t *eth_addr );
-};
-
-/** IPoIB Ethernet-compatible compressed link-layer address generators */
-static struct ipoib_eth_addr_handler ipoib_eth_addr_handlers[] = {
-	{ 0x02, 0xc9, ipoib_mlx_eth_addr },
-};
-
 /**
- * Generate Ethernet-compatible compressed link-layer address
+ * Translate received ARP packet
  *
- * @v ll_addr		Link-layer address
- * @v eth_addr		Ethernet-compatible address to fill in
+ * @v netdev		Network device
+ * @v iobuf		Received packet (with no link-layer headers)
+ * @v remac		Constructed Remote Ethernet MAC
+ * @ret rc		Return status code
  */
-static int ipoib_eth_addr ( const void *ll_addr, void *eth_addr ) {
-	const struct ipoib_mac *ipoib_addr = ll_addr;
-	const union ib_guid *guid = &ipoib_addr->gid.s.guid;
-	struct ipoib_eth_addr_handler *handler;
-	unsigned int i;
+static int ipoib_translate_rx_arp ( struct net_device *netdev,
+				    struct io_buffer *iobuf,
+				    struct ipoib_remac *remac ) {
+	struct ipoib_device *ipoib = netdev->priv;
+	struct arphdr *arphdr = iobuf->data;
+	void *sender_pa;
+	void *target_pa;
+	int rc;
 
-	for ( i = 0 ; i < ( sizeof ( ipoib_eth_addr_handlers ) /
-			    sizeof ( ipoib_eth_addr_handlers[0] ) ) ; i++ ) {
-		handler = &ipoib_eth_addr_handlers[i];
-		if ( ( handler->byte1 == guid->bytes[1] ) &&
-		     ( handler->byte2 == guid->bytes[2] ) ) {
-			return handler->eth_addr ( guid, eth_addr );
-		}
+	/* Do nothing unless ARP contains IPoIB link-layer addresses */
+	if ( arphdr->ar_hln != sizeof ( ipoib->mac ) )
+		return 0;
+
+	/* Create REMAC cache entry */
+	if ( ( rc = ipoib_map_remac ( ipoib, remac,
+				      arp_sender_ha ( arphdr ) ) ) != 0 ) {
+		DBGC ( ipoib, "IPoIB %p could not map REMAC: %s\n",
+		       ipoib, strerror ( rc ) );
+		return rc;
 	}
-	return -ENOTSUP;
-}
 
-/** IPoIB protocol */
-struct ll_protocol ipoib_protocol __ll_protocol = {
-	.name		= "IPoIB",
-	.ll_proto	= htons ( ARPHRD_INFINIBAND ),
-	.hw_addr_len	= sizeof ( union ib_guid ),
-	.ll_addr_len	= IPOIB_ALEN,
-	.ll_header_len	= IPOIB_HLEN,
-	.push		= ipoib_push,
-	.pull		= ipoib_pull,
-	.init_addr	= ipoib_init_addr,
-	.ntoa		= ipoib_ntoa,
-	.mc_hash	= ipoib_mc_hash,
-	.eth_addr	= ipoib_eth_addr,
-};
+	/* Construct new packet */
+	sender_pa = arp_sender_pa ( arphdr );
+	target_pa = arp_target_pa ( arphdr );
+	arphdr->ar_hrd = htons ( ARPHRD_ETHER );
+	arphdr->ar_hln = ETH_ALEN;
+	memcpy ( arp_sender_pa ( arphdr ), sender_pa, arphdr->ar_pln );
+	memcpy ( arp_target_pa ( arphdr ), target_pa, arphdr->ar_pln );
+	memcpy ( arp_sender_ha ( arphdr ), remac, ETH_ALEN );
+	memset ( arp_target_ha ( arphdr ), 0, ETH_ALEN );
+	if ( arphdr->ar_op == ARPOP_REPLY ) {
+		/* Assume received replies were directed to us */
+		memcpy ( arp_target_ha ( arphdr ), netdev->ll_addr, ETH_ALEN );
+	}
+	iob_unput ( iobuf, ( 2 * ( sizeof ( ipoib->mac ) - ETH_ALEN ) ) );
+
+	return 0;
+}
 
 /**
- * Allocate IPoIB device
+ * Translate received packet
  *
- * @v priv_size		Size of driver private data
- * @ret netdev		Network device, or NULL
+ * @v netdev		Network device
+ * @v iobuf		Received packet (with no link-layer headers)
+ * @v remac		Constructed Remote Ethernet MAC
+ * @v net_proto		Network-layer protocol (in network byte order)
+ * @ret rc		Return status code
  */
-struct net_device * alloc_ipoibdev ( size_t priv_size ) {
-	struct net_device *netdev;
-
-	netdev = alloc_netdev ( priv_size );
-	if ( netdev ) {
-		netdev->ll_protocol = &ipoib_protocol;
-		netdev->ll_broadcast = ( uint8_t * ) &ipoib_broadcast;
-		netdev->max_pkt_len = IB_MAX_PAYLOAD_SIZE;
+static int ipoib_translate_rx ( struct net_device *netdev,
+				struct io_buffer *iobuf,
+				struct ipoib_remac *remac,
+				uint16_t net_proto ) {
+
+	switch ( net_proto ) {
+	case htons ( ETH_P_ARP ) :
+		return ipoib_translate_rx_arp ( netdev, iobuf, remac );
+	case htons ( ETH_P_IP ) :
+		/* No translation needed */
+		return 0;
+	default:
+		/* Cannot handle other traffic via eIPoIB */
+		return -ENOTSUP;
 	}
-	return netdev;
 }
 
 /****************************************************************************
@@ -422,17 +459,18 @@ static int ipoib_transmit ( struct net_device *netdev,
 			    struct io_buffer *iobuf ) {
 	struct ipoib_device *ipoib = netdev->priv;
 	struct ib_device *ibdev = ipoib->ibdev;
+	struct ethhdr *ethhdr;
 	struct ipoib_hdr *ipoib_hdr;
-	struct ipoib_peer *peer;
+	struct ipoib_mac *mac;
 	struct ib_address_vector dest;
+	uint16_t net_proto;
 	int rc;
 
 	/* Sanity check */
-	if ( iob_len ( iobuf ) < sizeof ( *ipoib_hdr ) ) {
+	if ( iob_len ( iobuf ) < sizeof ( *ethhdr ) ) {
 		DBGC ( ipoib, "IPoIB %p buffer too short\n", ipoib );
 		return -EINVAL;
 	}
-	ipoib_hdr = iobuf->data;
 
 	/* Attempting transmission while link is down will put the
 	 * queue pair into an error state, so don't try it.
@@ -440,17 +478,30 @@ static int ipoib_transmit ( struct net_device *netdev,
 	if ( ! ib_link_ok ( ibdev ) )
 		return -ENETUNREACH;
 
+	/* Strip eIPoIB header */
+	ethhdr = iobuf->data;
+	net_proto = ethhdr->h_protocol;
+	iob_pull ( iobuf, sizeof ( *ethhdr ) );
+
 	/* Identify destination address */
-	peer = ipoib_lookup_peer_by_key ( ipoib_hdr->u.peer.dest );
-	if ( ! peer )
+	mac = ipoib_find_remac ( ipoib, ( ( void *) ethhdr->h_dest ) );
+	if ( ! mac )
 		return -ENXIO;
-	ipoib_hdr->u.reserved = 0;
+
+	/* Translate packet if applicable */
+	if ( ( rc = ipoib_translate_tx ( netdev, iobuf, net_proto ) ) != 0 )
+		return rc;
+
+	/* Prepend real IPoIB header */
+	ipoib_hdr = iob_push ( iobuf, sizeof ( *ipoib_hdr ) );
+	ipoib_hdr->proto = net_proto;
+	ipoib_hdr->reserved = 0;
 
 	/* Construct address vector */
 	memset ( &dest, 0, sizeof ( dest ) );
-	dest.qpn = ( ntohl ( peer->mac.flags__qpn ) & IB_QPN_MASK );
+	dest.qpn = ( ntohl ( mac->flags__qpn ) & IB_QPN_MASK );
 	dest.gid_present = 1;
-	memcpy ( &dest.gid, &peer->mac.gid, sizeof ( dest.gid ) );
+	memcpy ( &dest.gid, &mac->gid, sizeof ( dest.gid ) );
 	if ( ( rc = ib_resolve_path ( ibdev, &dest ) ) != 0 ) {
 		/* Path not resolved yet */
 		return rc;
@@ -487,14 +538,15 @@ static void ipoib_complete_send ( struct ib_device *ibdev __unused,
  */
 static void ipoib_complete_recv ( struct ib_device *ibdev __unused,
 				  struct ib_queue_pair *qp,
-				  struct ib_address_vector *dest __unused,
+				  struct ib_address_vector *dest,
 				  struct ib_address_vector *source,
 				  struct io_buffer *iobuf, int rc ) {
 	struct ipoib_device *ipoib = ib_qp_get_ownerdata ( qp );
 	struct net_device *netdev = ipoib->netdev;
 	struct ipoib_hdr *ipoib_hdr;
-	struct ipoib_mac ll_src;
-	struct ipoib_peer *src;
+	struct ethhdr *ethhdr;
+	struct ipoib_remac remac;
+	uint16_t net_proto;
 
 	/* Record errors */
 	if ( rc != 0 ) {
@@ -510,7 +562,6 @@ static void ipoib_complete_recv ( struct ib_device *ibdev __unused,
 		netdev_rx_err ( netdev, iobuf, -EIO );
 		return;
 	}
-	ipoib_hdr = iobuf->data;
 	if ( ! source ) {
 		DBGC ( ipoib, "IPoIB %p received packet without address "
 		       "vector\n", ipoib );
@@ -518,12 +569,37 @@ static void ipoib_complete_recv ( struct ib_device *ibdev __unused,
 		return;
 	}
 
-	/* Parse source address */
-	if ( source->gid_present ) {
-		ll_src.flags__qpn = htonl ( source->qpn );
-		memcpy ( &ll_src.gid, &source->gid, sizeof ( ll_src.gid ) );
-		src = ipoib_cache_peer ( &ll_src );
-		ipoib_hdr->u.peer.src = src->key;
+	/* Strip real IPoIB header */
+	ipoib_hdr = iobuf->data;
+	net_proto = ipoib_hdr->proto;
+	iob_pull ( iobuf, sizeof ( *ipoib_hdr ) );
+
+	/* Construct source address from remote QPN and LID */
+	remac.qpn = htonl ( source->qpn | EIPOIB_QPN_LA );
+	remac.lid = htons ( source->lid );
+
+	/* Translate packet if applicable */
+	if ( ( rc = ipoib_translate_rx ( netdev, iobuf, &remac,
+					 net_proto ) ) != 0 ) {
+		netdev_rx_err ( netdev, iobuf, rc );
+		return;
+	}
+
+	/* Prepend eIPoIB header */
+	ethhdr = iob_push ( iobuf, sizeof ( *ethhdr ) );
+	memcpy ( &ethhdr->h_source, &remac, sizeof ( ethhdr->h_source ) );
+	ethhdr->h_protocol = net_proto;
+
+	/* Construct destination address */
+	if ( dest->gid_present && ( memcmp ( &dest->gid, &ipoib->broadcast.gid,
+					     sizeof ( dest->gid ) ) == 0 ) ) {
+		/* Broadcast GID; use the Ethernet broadcast address */
+		memcpy ( &ethhdr->h_dest, eth_broadcast,
+			 sizeof ( ethhdr->h_dest ) );
+	} else {
+		/* Assume destination address is local Ethernet MAC */
+		memcpy ( &ethhdr->h_dest, netdev->ll_addr,
+			 sizeof ( ethhdr->h_dest ) );
 	}
 
 	/* Hand off to network layer */
@@ -536,9 +612,40 @@ static struct ib_completion_queue_operations ipoib_cq_op = {
 	.complete_recv = ipoib_complete_recv,
 };
 
+/**
+ * Allocate IPoIB receive I/O buffer
+ *
+ * @v len		Length of buffer
+ * @ret iobuf		I/O buffer, or NULL
+ *
+ * Some Infiniband hardware requires 2kB alignment of receive buffers
+ * and provides no way to disable header separation.  The result is
+ * that there are only four bytes of link-layer header (the real IPoIB
+ * header) before the payload.  This is not sufficient space to insert
+ * an eIPoIB link-layer pseudo-header.
+ *
+ * We therefore allocate I/O buffers offset to start slightly before
+ * the natural alignment boundary, in order to allow sufficient space.
+ */
+static struct io_buffer * ipoib_alloc_iob ( size_t len ) {
+	struct io_buffer *iobuf;
+	size_t reserve_len;
+
+	/* Calculate additional length required at start of buffer */
+	reserve_len = ( sizeof ( struct ethhdr ) -
+			sizeof ( struct ipoib_hdr ) );
+
+	/* Allocate buffer */
+	iobuf = alloc_iob_raw ( ( len + reserve_len ), len, -reserve_len );
+	if ( iobuf ) {
+		iob_reserve ( iobuf, reserve_len );
+	}
+	return iobuf;
+}
+
 /** IPoIB queue pair operations */
 static struct ib_queue_pair_operations ipoib_qp_op = {
-	.alloc_iob = alloc_iob,
+	.alloc_iob = ipoib_alloc_iob,
 };
 
 /**
@@ -550,7 +657,11 @@ static void ipoib_poll ( struct net_device *netdev ) {
 	struct ipoib_device *ipoib = netdev->priv;
 	struct ib_device *ibdev = ipoib->ibdev;
 
+	/* Poll Infiniband device */
 	ib_poll_eq ( ibdev );
+
+	/* Poll the retry timers (required for IPoIB multicast join) */
+	retry_poll();
 }
 
 /**
@@ -617,15 +728,14 @@ static void ipoib_leave_broadcast_group ( struct ipoib_device *ipoib ) {
 static void ipoib_link_state_changed ( struct ib_device *ibdev ) {
 	struct net_device *netdev = ib_get_ownerdata ( ibdev );
 	struct ipoib_device *ipoib = netdev->priv;
-	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
 	int rc;
 
 	/* Leave existing broadcast group */
 	ipoib_leave_broadcast_group ( ipoib );
 
 	/* Update MAC address based on potentially-new GID prefix */
-	memcpy ( &mac->gid.s.prefix, &ibdev->gid.s.prefix,
-		 sizeof ( mac->gid.s.prefix ) );
+	memcpy ( &ipoib->mac.gid.s.prefix, &ibdev->gid.s.prefix,
+		 sizeof ( ipoib->mac.gid.s.prefix ) );
 
 	/* Update broadcast GID based on potentially-new partition key */
 	ipoib->broadcast.gid.words[2] =
@@ -654,7 +764,6 @@ static void ipoib_link_state_changed ( struct ib_device *ibdev ) {
 static int ipoib_open ( struct net_device *netdev ) {
 	struct ipoib_device *ipoib = netdev->priv;
 	struct ib_device *ibdev = ipoib->ibdev;
-	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
 	int rc;
 
 	/* Open IB device */
@@ -686,7 +795,7 @@ static int ipoib_open ( struct net_device *netdev ) {
 	ib_qp_set_ownerdata ( ipoib->qp, ipoib );
 
 	/* Update MAC address with QPN */
-	mac->flags__qpn = htonl ( ipoib->qp->qpn );
+	ipoib->mac.flags__qpn = htonl ( ipoib->qp->qpn );
 
 	/* Fill receive rings */
 	ib_refill_recv ( ibdev, ipoib->qp );
@@ -713,13 +822,15 @@ static int ipoib_open ( struct net_device *netdev ) {
 static void ipoib_close ( struct net_device *netdev ) {
 	struct ipoib_device *ipoib = netdev->priv;
 	struct ib_device *ibdev = ipoib->ibdev;
-	struct ipoib_mac *mac = ( ( struct ipoib_mac * ) netdev->ll_addr );
+
+	/* Flush REMAC cache */
+	ipoib_flush_remac ( ipoib );
 
 	/* Leave broadcast group */
 	ipoib_leave_broadcast_group ( ipoib );
 
 	/* Remove QPN from MAC address */
-	mac->flags__qpn = 0;
+	ipoib->mac.flags__qpn = 0;
 
 	/* Tear down the queues */
 	ib_destroy_qp ( ibdev, ipoib->qp );
@@ -759,15 +870,19 @@ static int ipoib_probe ( struct ib_device *ibdev ) {
 	memset ( ipoib, 0, sizeof ( *ipoib ) );
 	ipoib->netdev = netdev;
 	ipoib->ibdev = ibdev;
+	INIT_LIST_HEAD ( &ipoib->peers );
 
 	/* Extract hardware address */
 	memcpy ( netdev->hw_addr, &ibdev->gid.s.guid,
 		 sizeof ( ibdev->gid.s.guid ) );
 
-	/* Set default broadcast address */
+	/* Set local MAC address */
+	memcpy ( &ipoib->mac.gid.s.guid, &ibdev->gid.s.guid,
+		 sizeof ( ipoib->mac.gid.s.guid ) );
+
+	/* Set default broadcast MAC address */
 	memcpy ( &ipoib->broadcast, &ipoib_broadcast,
 		 sizeof ( ipoib->broadcast ) );
-	netdev->ll_broadcast = ( ( uint8_t * ) &ipoib->broadcast );
 
 	/* Register network device */
 	if ( ( rc = register_netdev ( netdev ) ) != 0 )