Merge branch 'fsc' into pathbased

Resolved Conflicts:
	core/fs/ext2/ext2.c

Signed-off-by: H. Peter Anvin <hpa@zytor.com>

47 files changed, 10711 insertions, 373 deletions

diff --git a/gpxe/src/net/80211/net80211.c b/gpxe/src/net/80211/net80211.c
new file mode 100644
index 00000000..1c54597f
--- /dev/null
+++ b/gpxe/src/net/80211/net80211.c
@@ -0,0 +1,2829 @@
+/*
+ * The gPXE 802.11 MAC layer.
+ *
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <string.h>
+#include <byteswap.h>
+#include <stdlib.h>
+#include <gpxe/settings.h>
+#include <gpxe/if_arp.h>
+#include <gpxe/ethernet.h>
+#include <gpxe/ieee80211.h>
+#include <gpxe/netdevice.h>
+#include <gpxe/net80211.h>
+#include <gpxe/sec80211.h>
+#include <gpxe/timer.h>
+#include <gpxe/nap.h>
+#include <unistd.h>
+#include <errno.h>
+
+/** @file
+ *
+ * 802.11 device management
+ */
+
+/* Disambiguate the EINVAL's a bit */
+#define EINVAL_PKT_TOO_SHORT	( EINVAL | EUNIQ_01 )
+#define EINVAL_PKT_VERSION	( EINVAL | EUNIQ_02 )
+#define EINVAL_PKT_NOT_DATA	( EINVAL | EUNIQ_03 )
+#define EINVAL_PKT_NOT_FROMDS	( EINVAL | EUNIQ_04 )
+#define EINVAL_PKT_LLC_HEADER	( EINVAL | EUNIQ_05 )
+#define EINVAL_CRYPTO_REQUEST	( EINVAL | EUNIQ_06 )
+#define EINVAL_ACTIVE_SCAN	( EINVAL | EUNIQ_07 )
+
+/*
+ * 802.11 error codes: The AP can give us a status code explaining why
+ * authentication failed, or a reason code explaining why we were
+ * deauthenticated/disassociated. These codes range from 0-63 (the
+ * field is 16 bits wide, but only up to 45 or so are defined yet; we
+ * allow up to 63 for extensibility). This is encoded into an error
+ * code as such:
+ *
+ *                                      status & 0x1f goes here --vv--
+ *   Status code 0-31:  ECONNREFUSED | EUNIQ_(status & 0x1f) (0e1a6038)
+ *   Status code 32-63: EHOSTUNREACH | EUNIQ_(status & 0x1f) (171a6011)
+ *   Reason code 0-31:  ECONNRESET | EUNIQ_(reason & 0x1f)   (0f1a6039)
+ *   Reason code 32-63: ENETRESET | EUNIQ_(reason & 0x1f)    (271a6001)
+ *
+ * The POSIX error codes more or less convey the appropriate message
+ * (status codes occur when we can't associate at all, reason codes
+ * when we lose association unexpectedly) and let us extract the
+ * complete 802.11 error code from the rc value.
+ */
+
+/** Make return status code from 802.11 status code */
+#define E80211_STATUS( stat )  ( ((stat & 0x20)? EHOSTUNREACH : ECONNREFUSED) \
+					| ((stat & 0x1f) << 8) )
+
+/** Make return status code from 802.11 reason code */
+#define E80211_REASON( reas )  ( ((reas & 0x20)? ENETRESET : ECONNRESET) \
+					| ((reas & 0x1f) << 8) )
+
+
+/** List of 802.11 devices */
+static struct list_head net80211_devices = LIST_HEAD_INIT ( net80211_devices );
+
+/** Set of device operations that does nothing */
+static struct net80211_device_operations net80211_null_ops;
+
+/** Information associated with a received management packet
+ *
+ * This is used to keep beacon signal strengths in a parallel queue to
+ * the beacons themselves.
+ */
+struct net80211_rx_info {
+	int signal;
+	struct list_head list;
+};
+
+/** Context for a probe operation */
+struct net80211_probe_ctx {
+	/** 802.11 device to probe on */
+	struct net80211_device *dev;
+
+	/** Value of keep_mgmt before probe was started */
+	int old_keep_mgmt;
+
+	/** If scanning actively, pointer to probe packet to send */
+	struct io_buffer *probe;
+
+	/** If non-"", the ESSID to limit ourselves to */
+	const char *essid;
+
+	/** Time probe was started */
+	u32 ticks_start;
+
+	/** Time last useful beacon was received */
+	u32 ticks_beacon;
+
+	/** Time channel was last changed */
+	u32 ticks_channel;
+
+	/** Time to stay on each channel */
+	u32 hop_time;
+
+	/** Channels to hop by when changing channel */
+	int hop_step;
+
+	/** List of best beacons for each network found so far */
+	struct list_head *beacons;
+};
+
+/** Context for the association task */
+struct net80211_assoc_ctx {
+	/** Next authentication method to try using */
+	int method;
+
+	/** Time (in ticks) of the last sent association-related packet */
+	int last_packet;
+
+	/** Number of times we have tried sending it */
+	int times_tried;
+};
+
+/**
+ * @defgroup net80211_netdev Network device interface functions
+ * @{
+ */
+static int net80211_netdev_open ( struct net_device *netdev );
+static void net80211_netdev_close ( struct net_device *netdev );
+static int net80211_netdev_transmit ( struct net_device *netdev,
+				      struct io_buffer *iobuf );
+static void net80211_netdev_poll ( struct net_device *netdev );
+static void net80211_netdev_irq ( struct net_device *netdev, int enable );
+/** @} */
+
+/**
+ * @defgroup net80211_linklayer 802.11 link-layer protocol functions
+ * @{
+ */
+static int net80211_ll_push ( struct net_device *netdev,
+			      struct io_buffer *iobuf, const void *ll_dest,
+			      const void *ll_source, uint16_t net_proto );
+static int net80211_ll_pull ( struct net_device *netdev,
+			      struct io_buffer *iobuf, const void **ll_dest,
+			      const void **ll_source, uint16_t * net_proto );
+/** @} */
+
+/**
+ * @defgroup net80211_help 802.11 helper functions
+ * @{
+ */
+static void net80211_add_channels ( struct net80211_device *dev, int start,
+				    int len, int txpower );
+static void net80211_filter_hw_channels ( struct net80211_device *dev );
+static void net80211_set_rtscts_rate ( struct net80211_device *dev );
+static int net80211_process_capab ( struct net80211_device *dev,
+				    u16 capab );
+static int net80211_process_ie ( struct net80211_device *dev,
+				 union ieee80211_ie *ie, void *ie_end );
+static union ieee80211_ie *
+net80211_marshal_request_info ( struct net80211_device *dev,
+				union ieee80211_ie *ie );
+/** @} */
+
+/**
+ * @defgroup net80211_assoc_ll 802.11 association handling functions
+ * @{
+ */
+static void net80211_step_associate ( struct process *proc );
+static void net80211_handle_auth ( struct net80211_device *dev,
+				   struct io_buffer *iob );
+static void net80211_handle_assoc_reply ( struct net80211_device *dev,
+					  struct io_buffer *iob );
+static int net80211_send_disassoc ( struct net80211_device *dev, int reason,
+				    int deauth );
+static void net80211_handle_mgmt ( struct net80211_device *dev,
+				   struct io_buffer *iob, int signal );
+/** @} */
+
+/**
+ * @defgroup net80211_frag 802.11 fragment handling functions
+ * @{
+ */
+static void net80211_free_frags ( struct net80211_device *dev, int fcid );
+static struct io_buffer *net80211_accum_frags ( struct net80211_device *dev,
+						int fcid, int nfrags, int size );
+static void net80211_rx_frag ( struct net80211_device *dev,
+			       struct io_buffer *iob, int signal );
+/** @} */
+
+/**
+ * @defgroup net80211_settings 802.11 settings handlers
+ * @{
+ */
+static int net80211_check_settings_update ( void );
+
+/** 802.11 settings applicator
+ *
+ * When the SSID is changed, this will cause any open devices to
+ * re-associate; when the encryption key is changed, we similarly
+ * update their state.
+ */
+struct settings_applicator net80211_applicator __settings_applicator = {
+	.apply = net80211_check_settings_update,
+};
+
+/** The network name to associate with
+ *
+ * If this is blank, we scan for all networks and use the one with the
+ * greatest signal strength.
+ */
+struct setting net80211_ssid_setting __setting = {
+	.name = "ssid",
+	.description = "802.11 SSID (network name)",
+	.type = &setting_type_string,
+};
+
+/** Whether to use active scanning
+ *
+ * In order to associate with a hidden SSID, it's necessary to use an
+ * active scan (send probe packets). If this setting is nonzero, an
+ * active scan on the 2.4GHz band will be used to associate.
+ */
+struct setting net80211_active_setting __setting = {
+	.name = "active-scan",
+	.description = "Use an active scan during 802.11 association",
+	.type = &setting_type_int8,
+};
+
+/** The cryptographic key to use
+ *
+ * For hex WEP keys, as is common, this must be entered using the
+ * normal gPXE method for entering hex settings; an ASCII string of
+ * hex characters will not behave as expected.
+ */
+struct setting net80211_key_setting __setting = {
+	.name = "key",
+	.description = "Encryption key for protected 802.11 networks",
+	.type = &setting_type_string,
+};
+
+/** @} */
+
+
+/* ---------- net_device wrapper ---------- */
+
+/**
+ * Open 802.11 device and start association
+ *
+ * @v netdev	Wrapping network device
+ * @ret rc	Return status code
+ *
+ * This sets up a default conservative set of channels for probing,
+ * and starts the auto-association task unless the @c
+ * NET80211_NO_ASSOC flag is set in the wrapped 802.11 device's @c
+ * state field.
+ */
+static int net80211_netdev_open ( struct net_device *netdev )
+{
+	struct net80211_device *dev = netdev->priv;
+	int rc = 0;
+
+	if ( dev->op == &net80211_null_ops )
+		return -EFAULT;
+
+	if ( dev->op->open )
+		rc = dev->op->open ( dev );
+
+	if ( rc < 0 )
+		return rc;
+
+	if ( ! ( dev->state & NET80211_NO_ASSOC ) )
+		net80211_autoassociate ( dev );
+
+	return 0;
+}
+
+/**
+ * Close 802.11 device
+ *
+ * @v netdev	Wrapping network device.
+ *
+ * If the association task is running, this will stop it.
+ */
+static void net80211_netdev_close ( struct net_device *netdev )
+{
+	struct net80211_device *dev = netdev->priv;
+
+	if ( dev->state & NET80211_WORKING )
+		process_del ( &dev->proc_assoc );
+
+	/* Send disassociation frame to AP, to be polite */
+	if ( dev->state & NET80211_ASSOCIATED )
+		net80211_send_disassoc ( dev, IEEE80211_REASON_LEAVING, 0 );
+
+	if ( dev->handshaker && dev->handshaker->stop &&
+	     dev->handshaker->started )
+		dev->handshaker->stop ( dev );
+
+	free ( dev->crypto );
+	free ( dev->handshaker );
+	dev->crypto = NULL;
+	dev->handshaker = NULL;
+
+	netdev_link_down ( netdev );
+	dev->state = 0;
+
+	if ( dev->op->close )
+		dev->op->close ( dev );
+}
+
+/**
+ * Transmit packet on 802.11 device
+ *
+ * @v netdev	Wrapping network device
+ * @v iobuf	I/O buffer
+ * @ret rc	Return status code
+ *
+ * If encryption is enabled for the currently associated network, the
+ * packet will be encrypted prior to transmission.
+ */
+static int net80211_netdev_transmit ( struct net_device *netdev,
+				      struct io_buffer *iobuf )
+{
+	struct net80211_device *dev = netdev->priv;
+	struct ieee80211_frame *hdr = iobuf->data;
+	int rc = -ENOSYS;
+
+	if ( dev->crypto && ! ( hdr->fc & IEEE80211_FC_PROTECTED ) &&
+	     ( ( hdr->fc & IEEE80211_FC_TYPE ) == IEEE80211_TYPE_DATA ) ) {
+		struct io_buffer *niob = dev->crypto->encrypt ( dev->crypto,
+								iobuf );
+		if ( ! niob )
+			return -ENOMEM;	/* only reason encryption could fail */
+
+		/* Free the non-encrypted iob */
+		netdev_tx_complete ( netdev, iobuf );
+
+		/* Transmit the encrypted iob; the Protected flag is
+		   set, so we won't recurse into here again */
+		netdev_tx ( netdev, niob );
+
+		/* Don't transmit the freed packet */
+		return 0;
+	}
+
+	if ( dev->op->transmit )
+		rc = dev->op->transmit ( dev, iobuf );
+
+	return rc;
+}
+
+/**
+ * Poll 802.11 device for received packets and completed transmissions
+ *
+ * @v netdev	Wrapping network device
+ */
+static void net80211_netdev_poll ( struct net_device *netdev )
+{
+	struct net80211_device *dev = netdev->priv;
+
+	if ( dev->op->poll )
+		dev->op->poll ( dev );
+}
+
+/**
+ * Enable or disable interrupts for 802.11 device
+ *
+ * @v netdev	Wrapping network device
+ * @v enable	Whether to enable interrupts
+ */
+static void net80211_netdev_irq ( struct net_device *netdev, int enable )
+{
+	struct net80211_device *dev = netdev->priv;
+
+	if ( dev->op->irq )
+		dev->op->irq ( dev, enable );
+}
+
+/** Network device operations for a wrapped 802.11 device */
+static struct net_device_operations net80211_netdev_ops = {
+	.open = net80211_netdev_open,
+	.close = net80211_netdev_close,
+	.transmit = net80211_netdev_transmit,
+	.poll = net80211_netdev_poll,
+	.irq = net80211_netdev_irq,
+};
+
+
+/* ---------- 802.11 link-layer protocol ---------- */
+
+/** 802.11 broadcast MAC address */
+static u8 net80211_ll_broadcast[] = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff };
+
+/**
+ * Determine whether a transmission rate uses ERP/OFDM
+ *
+ * @v rate	Rate in 100 kbps units
+ * @ret is_erp	TRUE if the rate is an ERP/OFDM rate
+ *
+ * 802.11b supports rates of 1.0, 2.0, 5.5, and 11.0 Mbps; any other
+ * rate than these on the 2.4GHz spectrum is an ERP (802.11g) rate.
+ */
+static inline int net80211_rate_is_erp ( u16 rate )
+{
+	if ( rate == 10 || rate == 20 || rate == 55 || rate == 110 )
+		return 0;
+	return 1;
+}
+
+
+/**
+ * Calculate one frame's contribution to 802.11 duration field
+ *
+ * @v dev	802.11 device
+ * @v bytes	Amount of data to calculate duration for
+ * @ret dur	Duration field in microseconds
+ *
+ * To avoid multiple stations attempting to transmit at once, 802.11
+ * provides that every packet shall include a duration field
+ * specifying a length of time for which the wireless medium will be
+ * reserved after it is transmitted. The duration is measured in
+ * microseconds and is calculated with respect to the current
+ * physical-layer parameters of the 802.11 device.
+ *
+ * For an unfragmented data or management frame, or the last fragment
+ * of a fragmented frame, the duration captures only the 10 data bytes
+ * of one ACK; call once with bytes = 10.
+ *
+ * For a fragment of a data or management rame that will be followed
+ * by more fragments, the duration captures an ACK, the following
+ * fragment, and its ACK; add the results of three calls, two with
+ * bytes = 10 and one with bytes set to the next fragment's size.
+ *
+ * For an RTS control frame, the duration captures the responding CTS,
+ * the frame being sent, and its ACK; add the results of three calls,
+ * two with bytes = 10 and one with bytes set to the next frame's size
+ * (assuming unfragmented).
+ *
+ * For a CTS-to-self control frame, the duration captures the frame
+ * being protected and its ACK; add the results of two calls, one with
+ * bytes = 10 and one with bytes set to the next frame's size.
+ *
+ * No other frame types are currently supported by gPXE.
+ */
+u16 net80211_duration ( struct net80211_device *dev, int bytes, u16 rate )
+{
+	struct net80211_channel *chan = &dev->channels[dev->channel];
+	u32 kbps = rate * 100;
+
+	if ( chan->band == NET80211_BAND_5GHZ || net80211_rate_is_erp ( rate ) ) {
+		/* OFDM encoding (802.11a/g) */
+		int bits_per_symbol = ( kbps * 4 ) / 1000;	/* 4us/symbol */
+		int bits = 22 + ( bytes << 3 );	/* 22-bit PLCP */
+		int symbols = ( bits + bits_per_symbol - 1 ) / bits_per_symbol;
+
+		return 16 + 20 + ( symbols * 4 ); /* 16us SIFS, 20us preamble */
+	} else {
+		/* CCK encoding (802.11b) */
+		int phy_time = 144 + 48;	/* preamble + PLCP */
+		int bits = bytes << 3;
+		int data_time = ( bits * 1000 + kbps - 1 ) / kbps;
+
+		if ( dev->phy_flags & NET80211_PHY_USE_SHORT_PREAMBLE )
+			phy_time >>= 1;
+
+		return 10 + phy_time + data_time; /* 10us SIFS */
+	}
+}
+
+/**
+ * Add 802.11 link-layer header
+ *
+ * @v netdev		Wrapping network device
+ * @v iobuf		I/O buffer
+ * @v ll_dest		Link-layer destination address
+ * @v ll_source		Link-layer source address
+ * @v net_proto		Network-layer protocol, in network byte order
+ * @ret rc		Return status code
+ *
+ * This adds both the 802.11 frame header and the 802.2 LLC/SNAP
+ * header used on data packets.
+ *
+ * We also check here for state of the link that would make it invalid
+ * to send a data packet; every data packet must pass through here,
+ * and no non-data packet (e.g. management frame) should.
+ */
+static int net80211_ll_push ( struct net_device *netdev,
+			      struct io_buffer *iobuf, const void *ll_dest,
+			      const void *ll_source, uint16_t net_proto )
+{
+	struct net80211_device *dev = netdev->priv;
+	struct ieee80211_frame *hdr = iob_push ( iobuf,
+						 IEEE80211_LLC_HEADER_LEN +
+						 IEEE80211_TYP_FRAME_HEADER_LEN );
+	struct ieee80211_llc_snap_header *lhdr =
+		( void * ) hdr + IEEE80211_TYP_FRAME_HEADER_LEN;
+
+	/* We can't send data packets if we're not associated. */
+	if ( ! ( dev->state & NET80211_ASSOCIATED ) ) {
+		if ( dev->assoc_rc )
+			return dev->assoc_rc;
+		return -ENETUNREACH;
+	}
+
+	hdr->fc = IEEE80211_THIS_VERSION | IEEE80211_TYPE_DATA |
+	    IEEE80211_STYPE_DATA | IEEE80211_FC_TODS;
+
+	/* We don't send fragmented frames, so duration is the time
+	   for an SIFS + 10-byte ACK. */
+	hdr->duration = net80211_duration ( dev, 10, dev->rates[dev->rate] );
+
+	memcpy ( hdr->addr1, dev->bssid, ETH_ALEN );
+	memcpy ( hdr->addr2, ll_source, ETH_ALEN );
+	memcpy ( hdr->addr3, ll_dest, ETH_ALEN );
+
+	hdr->seq = IEEE80211_MAKESEQ ( ++dev->last_tx_seqnr, 0 );
+
+	lhdr->dsap = IEEE80211_LLC_DSAP;
+	lhdr->ssap = IEEE80211_LLC_SSAP;
+	lhdr->ctrl = IEEE80211_LLC_CTRL;
+	memset ( lhdr->oui, 0x00, 3 );
+	lhdr->ethertype = net_proto;
+
+	return 0;
+}
+
+/**
+ * Remove 802.11 link-layer header
+ *
+ * @v netdev		Wrapping network device
+ * @v iobuf		I/O buffer
+ * @ret ll_dest		Link-layer destination address
+ * @ret ll_source	Link-layer source
+ * @ret net_proto	Network-layer protocol, in network byte order
+ * @ret rc		Return status code
+ *
+ * This expects and removes both the 802.11 frame header and the 802.2
+ * LLC/SNAP header that are used on data packets.
+ */
+static int net80211_ll_pull ( struct net_device *netdev __unused,
+			      struct io_buffer *iobuf,
+			      const void **ll_dest, const void **ll_source,
+			      uint16_t * net_proto )
+{
+	struct ieee80211_frame *hdr = iobuf->data;
+	struct ieee80211_llc_snap_header *lhdr =
+		( void * ) hdr + IEEE80211_TYP_FRAME_HEADER_LEN;
+
+	/* Bunch of sanity checks */
+	if ( iob_len ( iobuf ) < IEEE80211_TYP_FRAME_HEADER_LEN +
+	     IEEE80211_LLC_HEADER_LEN ) {
+		DBGC ( netdev->priv, "802.11 %p packet too short (%zd bytes)\n",
+		       netdev->priv, iob_len ( iobuf ) );
+		return -EINVAL_PKT_TOO_SHORT;
+	}
+
+	if ( ( hdr->fc & IEEE80211_FC_VERSION ) != IEEE80211_THIS_VERSION ) {
+		DBGC ( netdev->priv, "802.11 %p packet invalid version %04x\n",
+		       netdev->priv, hdr->fc & IEEE80211_FC_VERSION );
+		return -EINVAL_PKT_VERSION;
+	}
+
+	if ( ( hdr->fc & IEEE80211_FC_TYPE ) != IEEE80211_TYPE_DATA ||
+	     ( hdr->fc & IEEE80211_FC_SUBTYPE ) != IEEE80211_STYPE_DATA ) {
+		DBGC ( netdev->priv, "802.11 %p packet not data/data (fc=%04x)\n",
+		       netdev->priv, hdr->fc );
+		return -EINVAL_PKT_NOT_DATA;
+	}
+
+	if ( ( hdr->fc & ( IEEE80211_FC_TODS | IEEE80211_FC_FROMDS ) ) !=
+	     IEEE80211_FC_FROMDS ) {
+		DBGC ( netdev->priv, "802.11 %p packet not from DS (fc=%04x)\n",
+		       netdev->priv, hdr->fc );
+		return -EINVAL_PKT_NOT_FROMDS;
+	}
+
+	if ( lhdr->dsap != IEEE80211_LLC_DSAP || lhdr->ssap != IEEE80211_LLC_SSAP ||
+	     lhdr->ctrl != IEEE80211_LLC_CTRL || lhdr->oui[0] || lhdr->oui[1] ||
+	     lhdr->oui[2] ) {
+		DBGC ( netdev->priv, "802.11 %p LLC header is not plain EtherType "
+		       "encapsulator: %02x->%02x [%02x] %02x:%02x:%02x %04x\n",
+		       netdev->priv, lhdr->dsap, lhdr->ssap, lhdr->ctrl,
+		       lhdr->oui[0], lhdr->oui[1], lhdr->oui[2], lhdr->ethertype );
+		return -EINVAL_PKT_LLC_HEADER;
+	}
+
+	iob_pull ( iobuf, sizeof ( *hdr ) + sizeof ( *lhdr ) );
+
+	*ll_dest = hdr->addr1;
+	*ll_source = hdr->addr3;
+	*net_proto = lhdr->ethertype;
+	return 0;
+}
+
+/** 802.11 link-layer protocol */
+static struct ll_protocol net80211_ll_protocol __ll_protocol = {
+	.name = "802.11",
+	.push = net80211_ll_push,
+	.pull = net80211_ll_pull,
+	.init_addr = eth_init_addr,
+	.ntoa = eth_ntoa,
+	.mc_hash = eth_mc_hash,
+	.eth_addr = eth_eth_addr,
+	.ll_proto = htons ( ARPHRD_ETHER ),	/* "encapsulated Ethernet" */
+	.hw_addr_len = ETH_ALEN,
+	.ll_addr_len = ETH_ALEN,
+	.ll_header_len = IEEE80211_TYP_FRAME_HEADER_LEN +
+				IEEE80211_LLC_HEADER_LEN,
+};
+
+
+/* ---------- 802.11 network management API ---------- */
+
+/**
+ * Get 802.11 device from wrapping network device
+ *
+ * @v netdev	Wrapping network device
+ * @ret dev	802.11 device wrapped by network device, or NULL
+ *
+ * Returns NULL if the network device does not wrap an 802.11 device.
+ */
+struct net80211_device * net80211_get ( struct net_device *netdev )
+{
+	struct net80211_device *dev;
+
+	list_for_each_entry ( dev, &net80211_devices, list ) {
+		if ( netdev->priv == dev )
+			return netdev->priv;
+	}
+
+	return NULL;
+}
+
+/**
+ * Set state of 802.11 device keeping management frames
+ *
+ * @v dev	802.11 device
+ * @v enable	Whether to keep management frames
+ * @ret oldenab	Whether management frames were enabled before this call
+ *
+ * If enable is TRUE, beacon, probe, and action frames will be kept
+ * and may be retrieved by calling net80211_mgmt_dequeue().
+ */
+int net80211_keep_mgmt ( struct net80211_device *dev, int enable )
+{
+	int oldenab = dev->keep_mgmt;
+
+	dev->keep_mgmt = enable;
+	return oldenab;
+}
+
+/**
+ * Get 802.11 management frame
+ *
+ * @v dev	802.11 device
+ * @ret signal	Signal strength of returned management frame
+ * @ret iob	I/O buffer, or NULL if no management frame is queued
+ *
+ * Frames will only be returned by this function if
+ * net80211_keep_mgmt() has been previously called with enable set to
+ * TRUE.
+ *
+ * The calling function takes ownership of the returned I/O buffer.
+ */
+struct io_buffer * net80211_mgmt_dequeue ( struct net80211_device *dev,
+					   int *signal )
+{
+	struct io_buffer *iobuf;
+	struct net80211_rx_info *rxi;
+
+	list_for_each_entry ( rxi, &dev->mgmt_info_queue, list ) {
+		list_del ( &rxi->list );
+		if ( signal )
+			*signal = rxi->signal;
+		free ( rxi );
+
+		list_for_each_entry ( iobuf, &dev->mgmt_queue, list ) {
+			list_del ( &iobuf->list );
+			return iobuf;
+		}
+		assert ( 0 );
+	}
+
+	return NULL;
+}
+
+/**
+ * Transmit 802.11 management frame
+ *
+ * @v dev	802.11 device
+ * @v fc	Frame Control flags for management frame
+ * @v dest	Destination access point
+ * @v iob	I/O buffer
+ * @ret rc	Return status code
+ *
+ * The @a fc argument must contain at least an IEEE 802.11 management
+ * subtype number (e.g. IEEE80211_STYPE_PROBE_REQ). If it contains
+ * IEEE80211_FC_PROTECTED, the frame will be encrypted prior to
+ * transmission.
+ *
+ * It is required that @a iob have at least 24 bytes of headroom
+ * reserved before its data start.
+ */
+int net80211_tx_mgmt ( struct net80211_device *dev, u16 fc, u8 dest[6],
+		       struct io_buffer *iob )
+{
+	struct ieee80211_frame *hdr = iob_push ( iob,
+						 IEEE80211_TYP_FRAME_HEADER_LEN );
+
+	hdr->fc = IEEE80211_THIS_VERSION | IEEE80211_TYPE_MGMT |
+	    ( fc & ~IEEE80211_FC_PROTECTED );
+	hdr->duration = net80211_duration ( dev, 10, dev->rates[dev->rate] );
+	hdr->seq = IEEE80211_MAKESEQ ( ++dev->last_tx_seqnr, 0 );
+
+	memcpy ( hdr->addr1, dest, ETH_ALEN );	/* DA = RA */
+	memcpy ( hdr->addr2, dev->netdev->ll_addr, ETH_ALEN );	/* SA = TA */
+	memcpy ( hdr->addr3, dest, ETH_ALEN );	/* BSSID */
+
+	if ( fc & IEEE80211_FC_PROTECTED ) {
+		if ( ! dev->crypto )
+			return -EINVAL_CRYPTO_REQUEST;
+
+		struct io_buffer *eiob = dev->crypto->encrypt ( dev->crypto,
+								iob );
+		free_iob ( iob );
+		iob = eiob;
+	}
+
+	return netdev_tx ( dev->netdev, iob );
+}
+
+
+/* ---------- Driver API ---------- */
+
+/**
+ * Allocate 802.11 device
+ *
+ * @v priv_size		Size of driver-private allocation area
+ * @ret dev		Newly allocated 802.11 device
+ *
+ * This function allocates a net_device with space in its private area
+ * for both the net80211_device it will wrap and the driver-private
+ * data space requested. It initializes the link-layer-specific parts
+ * of the net_device, and links the net80211_device to the net_device
+ * appropriately.
+ */
+struct net80211_device * net80211_alloc ( size_t priv_size )
+{
+	struct net80211_device *dev;
+	struct net_device *netdev =
+		alloc_netdev ( sizeof ( *dev ) + priv_size );
+
+	if ( ! netdev )
+		return NULL;
+
+	netdev->ll_protocol = &net80211_ll_protocol;
+	netdev->ll_broadcast = net80211_ll_broadcast;
+	netdev->max_pkt_len = IEEE80211_MAX_DATA_LEN;
+	netdev_init ( netdev, &net80211_netdev_ops );
+
+	dev = netdev->priv;
+	dev->netdev = netdev;
+	dev->priv = ( u8 * ) dev + sizeof ( *dev );
+	dev->op = &net80211_null_ops;
+
+	process_init_stopped ( &dev->proc_assoc, net80211_step_associate,
+			       &netdev->refcnt );
+	INIT_LIST_HEAD ( &dev->mgmt_queue );
+	INIT_LIST_HEAD ( &dev->mgmt_info_queue );
+
+	return dev;
+}
+
+/**
+ * Register 802.11 device with network stack
+ *
+ * @v dev	802.11 device
+ * @v ops	802.11 device operations
+ * @v hw	802.11 hardware information
+ *
+ * This also registers the wrapping net_device with the higher network
+ * layers.
+ */
+int net80211_register ( struct net80211_device *dev,
+			struct net80211_device_operations *ops,
+			struct net80211_hw_info *hw )
+{
+	dev->op = ops;
+	dev->hw = malloc ( sizeof ( *hw ) );
+	if ( ! dev->hw )
+		return -ENOMEM;
+
+	memcpy ( dev->hw, hw, sizeof ( *hw ) );
+	memcpy ( dev->netdev->hw_addr, hw->hwaddr, ETH_ALEN );
+
+	/* Set some sensible channel defaults for driver's open() function */
+	memcpy ( dev->channels, dev->hw->channels,
+		 NET80211_MAX_CHANNELS * sizeof ( dev->channels[0] ) );
+	dev->channel = 0;
+
+	list_add_tail ( &dev->list, &net80211_devices );
+	return register_netdev ( dev->netdev );
+}
+
+/**
+ * Unregister 802.11 device from network stack
+ *
+ * @v dev	802.11 device
+ *
+ * After this call, the device operations are cleared so that they
+ * will not be called.
+ */
+void net80211_unregister ( struct net80211_device *dev )
+{
+	unregister_netdev ( dev->netdev );
+	list_del ( &dev->list );
+	dev->op = &net80211_null_ops;
+}
+
+/**
+ * Free 802.11 device
+ *
+ * @v dev	802.11 device
+ *
+ * The device should be unregistered before this function is called.
+ */
+void net80211_free ( struct net80211_device *dev )
+{
+	free ( dev->hw );
+	rc80211_free ( dev->rctl );
+	netdev_nullify ( dev->netdev );
+	netdev_put ( dev->netdev );
+}
+
+
+/* ---------- 802.11 network management workhorse code ---------- */
+
+/**
+ * Set state of 802.11 device
+ *
+ * @v dev	802.11 device
+ * @v clear	Bitmask of flags to clear
+ * @v set	Bitmask of flags to set
+ * @v status	Status or reason code for most recent operation
+ *
+ * If @a status represents a reason code, it should be OR'ed with
+ * NET80211_IS_REASON.
+ *
+ * Clearing authentication also clears association; clearing
+ * association also clears security handshaking state. Clearing
+ * association removes the link-up flag from the wrapping net_device,
+ * but setting it does not automatically set the flag; that is left to
+ * the judgment of higher-level code.
+ */
+static inline void net80211_set_state ( struct net80211_device *dev,
+					short clear, short set,
+					u16 status )
+{
+	/* The conditions in this function are deliberately formulated
+	   to be decidable at compile-time in most cases. Since clear
+	   and set are generally passed as constants, the body of this
+	   function can be reduced down to a few statements by the
+	   compiler. */
+
+	const int statmsk = NET80211_STATUS_MASK | NET80211_IS_REASON;
+
+	if ( clear & NET80211_PROBED )
+		clear |= NET80211_AUTHENTICATED;
+
+	if ( clear & NET80211_AUTHENTICATED )
+		clear |= NET80211_ASSOCIATED;
+
+	if ( clear & NET80211_ASSOCIATED )
+		clear |= NET80211_CRYPTO_SYNCED;
+
+	dev->state = ( dev->state & ~clear ) | set;
+	dev->state = ( dev->state & ~statmsk ) | ( status & statmsk );
+
+	if ( clear & NET80211_ASSOCIATED )
+		netdev_link_down ( dev->netdev );
+
+	if ( ( clear | set ) & NET80211_ASSOCIATED )
+		dev->op->config ( dev, NET80211_CFG_ASSOC );
+
+	if ( status != 0 ) {
+		if ( status & NET80211_IS_REASON )
+			dev->assoc_rc = -E80211_REASON ( status );
+		else
+			dev->assoc_rc = -E80211_STATUS ( status );
+		netdev_link_err ( dev->netdev, dev->assoc_rc );
+	}
+}
+
+/**
+ * Add channels to 802.11 device
+ *
+ * @v dev	802.11 device
+ * @v start	First channel number to add
+ * @v len	Number of channels to add
+ * @v txpower	TX power (dBm) to allow on added channels
+ *
+ * To replace the current list of channels instead of adding to it,
+ * set the nr_channels field of the 802.11 device to 0 before calling
+ * this function.
+ */
+static void net80211_add_channels ( struct net80211_device *dev, int start,
+				    int len, int txpower )
+{
+	int i, chan = start;
+
+	for ( i = dev->nr_channels; len-- && i < NET80211_MAX_CHANNELS; i++ ) {
+		dev->channels[i].channel_nr = chan;
+		dev->channels[i].maxpower = txpower;
+		dev->channels[i].hw_value = 0;
+
+		if ( chan >= 1 && chan <= 14 ) {
+			dev->channels[i].band = NET80211_BAND_2GHZ;
+			if ( chan == 14 )
+				dev->channels[i].center_freq = 2484;
+			else
+				dev->channels[i].center_freq = 2407 + 5 * chan;
+			chan++;
+		} else {
+			dev->channels[i].band = NET80211_BAND_5GHZ;
+			dev->channels[i].center_freq = 5000 + 5 * chan;
+			chan += 4;
+		}
+	}
+
+	dev->nr_channels = i;
+}
+
+/**
+ * Filter 802.11 device channels for hardware capabilities
+ *
+ * @v dev	802.11 device
+ *
+ * Hardware may support fewer channels than regulatory restrictions
+ * allow; this function filters out channels in dev->channels that are
+ * not supported by the hardware list in dev->hwinfo. It also copies
+ * over the net80211_channel::hw_value and limits maximum TX power
+ * appropriately.
+ *
+ * Channels are matched based on center frequency, ignoring band and
+ * channel number.
+ *
+ * If the driver specifies no supported channels, the effect will be
+ * as though all were supported.
+ */
+static void net80211_filter_hw_channels ( struct net80211_device *dev )
+{
+	int delta = 0, i = 0;
+	int old_freq = dev->channels[dev->channel].center_freq;
+	struct net80211_channel *chan, *hwchan;
+
+	if ( ! dev->hw->nr_channels )
+		return;
+
+	dev->channel = 0;
+	for ( chan = dev->channels; chan < dev->channels + dev->nr_channels;
+	      chan++, i++ ) {
+		int ok = 0;
+		for ( hwchan = dev->hw->channels;
+		      hwchan < dev->hw->channels + dev->hw->nr_channels;
+		      hwchan++ ) {
+			if ( hwchan->center_freq == chan->center_freq ) {
+				ok = 1;
+				break;
+			}
+		}
+
+		if ( ! ok )
+			delta++;
+		else {
+			chan->hw_value = hwchan->hw_value;
+			if ( hwchan->maxpower != 0 &&
+			     chan->maxpower > hwchan->maxpower )
+				chan->maxpower = hwchan->maxpower;
+			if ( old_freq == chan->center_freq )
+				dev->channel = i - delta;
+			if ( delta )
+				chan[-delta] = *chan;
+		}
+	}
+
+	dev->nr_channels -= delta;
+
+	if ( dev->channels[dev->channel].center_freq != old_freq )
+		dev->op->config ( dev, NET80211_CFG_CHANNEL );
+}
+
+/**
+ * Update 802.11 device state to reflect received capabilities field
+ *
+ * @v dev	802.11 device
+ * @v capab	Capabilities field in beacon, probe, or association frame
+ * @ret rc	Return status code
+ */
+static int net80211_process_capab ( struct net80211_device *dev,
+				    u16 capab )
+{
+	u16 old_phy = dev->phy_flags;
+
+	if ( ( capab & ( IEEE80211_CAPAB_MANAGED | IEEE80211_CAPAB_ADHOC ) ) !=
+	     IEEE80211_CAPAB_MANAGED ) {
+		DBGC ( dev, "802.11 %p cannot handle IBSS network\n", dev );
+		return -ENOSYS;
+	}
+
+	dev->phy_flags &= ~( NET80211_PHY_USE_SHORT_PREAMBLE |
+			     NET80211_PHY_USE_SHORT_SLOT );
+
+	if ( capab & IEEE80211_CAPAB_SHORT_PMBL )
+		dev->phy_flags |= NET80211_PHY_USE_SHORT_PREAMBLE;
+
+	if ( capab & IEEE80211_CAPAB_SHORT_SLOT )
+		dev->phy_flags |= NET80211_PHY_USE_SHORT_SLOT;
+
+	if ( old_phy != dev->phy_flags )
+		dev->op->config ( dev, NET80211_CFG_PHY_PARAMS );
+
+	return 0;
+}
+
+/**
+ * Update 802.11 device state to reflect received information elements
+ *
+ * @v dev	802.11 device
+ * @v ie	Pointer to first information element
+ * @v ie_end	Pointer to tail of packet I/O buffer
+ * @ret rc	Return status code
+ */
+static int net80211_process_ie ( struct net80211_device *dev,
+				 union ieee80211_ie *ie, void *ie_end )
+{
+	u16 old_rate = dev->rates[dev->rate];
+	u16 old_phy = dev->phy_flags;
+	int have_rates = 0, i;
+	int ds_channel = 0;
+	int changed = 0;
+	int band = dev->channels[dev->channel].band;
+
+	if ( ! ieee80211_ie_bound ( ie, ie_end ) )
+		return 0;
+
+	for ( ; ie; ie = ieee80211_next_ie ( ie, ie_end ) ) {
+		switch ( ie->id ) {
+		case IEEE80211_IE_SSID:
+			if ( ie->len <= 32 ) {
+				memcpy ( dev->essid, ie->ssid, ie->len );
+				dev->essid[ie->len] = 0;
+			}
+			break;
+
+		case IEEE80211_IE_RATES:
+		case IEEE80211_IE_EXT_RATES:
+			if ( ! have_rates ) {
+				dev->nr_rates = 0;
+				dev->basic_rates = 0;
+				have_rates = 1;
+			}
+			for ( i = 0; i < ie->len &&
+			      dev->nr_rates < NET80211_MAX_RATES; i++ ) {
+				u8 rid = ie->rates[i];
+				u16 rate = ( rid & 0x7f ) * 5;
+
+				if ( rid & 0x80 )
+					dev->basic_rates |=
+						( 1 << dev->nr_rates );
+
+				dev->rates[dev->nr_rates++] = rate;
+			}
+
+			break;
+
+		case IEEE80211_IE_DS_PARAM:
+			if ( dev->channel < dev->nr_channels && ds_channel ==
+			     dev->channels[dev->channel].channel_nr )
+				break;
+			ds_channel = ie->ds_param.current_channel;
+			net80211_change_channel ( dev, ds_channel );
+			break;
+
+		case IEEE80211_IE_COUNTRY:
+			dev->nr_channels = 0;
+
+			DBGC ( dev, "802.11 %p setting country regulations "
+			       "for %c%c\n", dev, ie->country.name[0],
+			       ie->country.name[1] );
+			for ( i = 0; i < ( ie->len - 3 ) / 3; i++ ) {
+				union ieee80211_ie_country_triplet *t =
+					&ie->country.triplet[i];
+				if ( t->first > 200 ) {
+					DBGC ( dev, "802.11 %p ignoring regulatory "
+					       "extension information\n", dev );
+				} else {
+					net80211_add_channels ( dev,
+							t->band.first_channel,
+							t->band.nr_channels,
+							t->band.max_txpower );
+				}
+			}
+			net80211_filter_hw_channels ( dev );
+			break;
+
+		case IEEE80211_IE_ERP_INFO:
+			dev->phy_flags &= ~( NET80211_PHY_USE_PROTECTION |
+					     NET80211_PHY_USE_SHORT_PREAMBLE );
+			if ( ie->erp_info & IEEE80211_ERP_USE_PROTECTION )
+				dev->phy_flags |= NET80211_PHY_USE_PROTECTION;
+			if ( ! ( ie->erp_info & IEEE80211_ERP_BARKER_LONG ) )
+				dev->phy_flags |= NET80211_PHY_USE_SHORT_PREAMBLE;
+			break;
+		}
+	}
+
+	if ( have_rates ) {
+		/* Allow only those rates that are also supported by
+		   the hardware. */
+		int delta = 0, j;
+
+		dev->rate = 0;
+		for ( i = 0; i < dev->nr_rates; i++ ) {
+			int ok = 0;
+			for ( j = 0; j < dev->hw->nr_rates[band]; j++ ) {
+				if ( dev->hw->rates[band][j] == dev->rates[i] ){
+					ok = 1;
+					break;
+				}
+			}
+
+			if ( ! ok )
+				delta++;
+			else {
+				dev->rates[i - delta] = dev->rates[i];
+				if ( old_rate == dev->rates[i] )
+					dev->rate = i - delta;
+			}
+		}
+
+		dev->nr_rates -= delta;
+
+		/* Sort available rates - sorted subclumps tend to already
+		   exist, so insertion sort works well. */
+		for ( i = 1; i < dev->nr_rates; i++ ) {
+			u16 rate = dev->rates[i];
+			u32 tmp, br, mask;
+
+			for ( j = i - 1; j >= 0 && dev->rates[j] >= rate; j-- )
+				dev->rates[j + 1] = dev->rates[j];
+			dev->rates[j + 1] = rate;
+
+			/* Adjust basic_rates to match by rotating the
+			   bits from bit j+1 to bit i left one position. */
+			mask = ( ( 1 << i ) - 1 ) & ~( ( 1 << ( j + 1 ) ) - 1 );
+			br = dev->basic_rates;
+			tmp = br & ( 1 << i );
+			br = ( br & ~( mask | tmp ) ) | ( ( br & mask ) << 1 );
+			br |= ( tmp >> ( i - j - 1 ) );
+			dev->basic_rates = br;
+		}
+
+		net80211_set_rtscts_rate ( dev );
+
+		if ( dev->rates[dev->rate] != old_rate )
+			changed |= NET80211_CFG_RATE;
+	}
+
+	if ( dev->hw->flags & NET80211_HW_NO_SHORT_PREAMBLE )
+		dev->phy_flags &= ~NET80211_PHY_USE_SHORT_PREAMBLE;
+	if ( dev->hw->flags & NET80211_HW_NO_SHORT_SLOT )
+		dev->phy_flags &= ~NET80211_PHY_USE_SHORT_SLOT;
+
+	if ( old_phy != dev->phy_flags )
+		changed |= NET80211_CFG_PHY_PARAMS;
+
+	if ( changed )
+		dev->op->config ( dev, changed );
+
+	return 0;
+}
+
+/**
+ * Create information elements for outgoing probe or association packet
+ *
+ * @v dev		802.11 device
+ * @v ie		Pointer to start of information element area
+ * @ret next_ie		Pointer to first byte after added information elements
+ */
+static union ieee80211_ie *
+net80211_marshal_request_info ( struct net80211_device *dev,
+				union ieee80211_ie *ie )
+{
+	int i;
+
+	ie->id = IEEE80211_IE_SSID;
+	ie->len = strlen ( dev->essid );
+	memcpy ( ie->ssid, dev->essid, ie->len );
+
+	ie = ieee80211_next_ie ( ie, NULL );
+
+	ie->id = IEEE80211_IE_RATES;
+	ie->len = dev->nr_rates;
+	if ( ie->len > 8 )
+		ie->len = 8;
+
+	for ( i = 0; i < ie->len; i++ ) {
+		ie->rates[i] = dev->rates[i] / 5;
+		if ( dev->basic_rates & ( 1 << i ) )
+			ie->rates[i] |= 0x80;
+	}
+
+	ie = ieee80211_next_ie ( ie, NULL );
+
+	if ( dev->rsn_ie && dev->rsn_ie->id == IEEE80211_IE_RSN ) {
+		memcpy ( ie, dev->rsn_ie, dev->rsn_ie->len + 2 );
+		ie = ieee80211_next_ie ( ie, NULL );
+	}
+
+	if ( dev->nr_rates > 8 ) {
+		/* 802.11 requires we use an Extended Basic Rates IE
+		   for the rates beyond the eighth. */
+
+		ie->id = IEEE80211_IE_EXT_RATES;
+		ie->len = dev->nr_rates - 8;
+
+		for ( ; i < dev->nr_rates; i++ ) {
+			ie->rates[i - 8] = dev->rates[i] / 5;
+			if ( dev->basic_rates & ( 1 << i ) )
+				ie->rates[i - 8] |= 0x80;
+		}
+
+		ie = ieee80211_next_ie ( ie, NULL );
+	}
+
+	if ( dev->rsn_ie && dev->rsn_ie->id == IEEE80211_IE_VENDOR ) {
+		memcpy ( ie, dev->rsn_ie, dev->rsn_ie->len + 2 );
+		ie = ieee80211_next_ie ( ie, NULL );
+	}
+
+	return ie;
+}
+
+/** Seconds to wait after finding a network, to possibly find better APs for it
+ *
+ * This is used when a specific SSID to scan for is specified.
+ */
+#define NET80211_PROBE_GATHER    1
+
+/** Seconds to wait after finding a network, to possibly find other networks
+ *
+ * This is used when an empty SSID is specified, to scan for all
+ * networks.
+ */
+#define NET80211_PROBE_GATHER_ALL 2
+
+/** Seconds to allow a probe to take if no network has been found */
+#define NET80211_PROBE_TIMEOUT   6
+
+/**
+ * Begin probe of 802.11 networks
+ *
+ * @v dev	802.11 device
+ * @v essid	SSID to probe for, or "" to accept any (may not be NULL)
+ * @v active	Whether to use active scanning
+ * @ret ctx	Probe context
+ *
+ * Active scanning may only be used on channels 1-11 in the 2.4GHz
+ * band, due to gPXE's lack of a complete regulatory database. If
+ * active scanning is used, probe packets will be sent on each
+ * channel; this can allow association with hidden-SSID networks if
+ * the SSID is properly specified.
+ *
+ * A @c NULL return indicates an out-of-memory condition.
+ *
+ * The returned context must be periodically passed to
+ * net80211_probe_step() until that function returns zero.
+ */
+struct net80211_probe_ctx * net80211_probe_start ( struct net80211_device *dev,
+						   const char *essid,
+						   int active )
+{
+	struct net80211_probe_ctx *ctx = zalloc ( sizeof ( *ctx ) );
+
+	if ( ! ctx )
+		return NULL;
+
+	assert ( dev->netdev->state & NETDEV_OPEN );
+
+	ctx->dev = dev;
+	ctx->old_keep_mgmt = net80211_keep_mgmt ( dev, 1 );
+	ctx->essid = essid;
+	if ( dev->essid != ctx->essid )
+		strcpy ( dev->essid, ctx->essid );
+
+	if ( active ) {
+		struct ieee80211_probe_req *probe_req;
+		union ieee80211_ie *ie;
+
+		ctx->probe = alloc_iob ( 128 );
+		iob_reserve ( ctx->probe, IEEE80211_TYP_FRAME_HEADER_LEN );
+		probe_req = ctx->probe->data;
+
+		ie = net80211_marshal_request_info ( dev,
+						     probe_req->info_element );
+
+		iob_put ( ctx->probe, ( void * ) ie - ctx->probe->data );
+	}
+
+	ctx->ticks_start = currticks();
+	ctx->ticks_beacon = 0;
+	ctx->ticks_channel = currticks();
+	ctx->hop_time = ticks_per_sec() / ( active ? 2 : 6 );
+
+	/*
+	 * Channels on 2.4GHz overlap, and the most commonly used
+	 * are 1, 6, and 11. We'll get a result faster if we check
+	 * every 5 channels, but in order to hit all of them the
+	 * number of channels must be relatively prime to 5. If it's
+	 * not, tweak the hop.
+	 */
+	ctx->hop_step = 5;
+	while ( dev->nr_channels % ctx->hop_step == 0 && ctx->hop_step > 1 )
+		ctx->hop_step--;
+
+	ctx->beacons = malloc ( sizeof ( *ctx->beacons ) );
+	INIT_LIST_HEAD ( ctx->beacons );
+
+	dev->channel = 0;
+	dev->op->config ( dev, NET80211_CFG_CHANNEL );
+
+	return ctx;
+}
+
+/**
+ * Continue probe of 802.11 networks
+ *
+ * @v ctx	Probe context returned by net80211_probe_start()
+ * @ret rc	Probe status
+ *
+ * The return code will be 0 if the probe is still going on (and this
+ * function should be called again), a positive number if the probe
+ * completed successfully, or a negative error code if the probe
+ * failed for that reason.
+ *
+ * Whether the probe succeeded or failed, you must call
+ * net80211_probe_finish_all() or net80211_probe_finish_best()
+ * (depending on whether you want information on all networks or just
+ * the best-signal one) in order to release the probe context. A
+ * failed probe may still have acquired some valid data.
+ */
+int net80211_probe_step ( struct net80211_probe_ctx *ctx )
+{
+	struct net80211_device *dev = ctx->dev;
+	u32 start_timeout = NET80211_PROBE_TIMEOUT * ticks_per_sec();
+	u32 gather_timeout = ticks_per_sec();
+	u32 now = currticks();
+	struct io_buffer *iob;
+	int signal;
+	int rc;
+	char ssid[IEEE80211_MAX_SSID_LEN + 1];
+
+	gather_timeout *= ( ctx->essid[0] ? NET80211_PROBE_GATHER :
+			    NET80211_PROBE_GATHER_ALL );
+
+	/* Time out if necessary */
+	if ( now >= ctx->ticks_start + start_timeout )
+		return list_empty ( ctx->beacons ) ? -ETIMEDOUT : +1;
+
+	if ( ctx->ticks_beacon > 0 && now >= ctx->ticks_start + gather_timeout )
+		return +1;
+
+	/* Change channels if necessary */
+	if ( now >= ctx->ticks_channel + ctx->hop_time ) {
+		dev->channel = ( dev->channel + ctx->hop_step )
+			% dev->nr_channels;
+		dev->op->config ( dev, NET80211_CFG_CHANNEL );
+		udelay ( dev->hw->channel_change_time );
+
+		ctx->ticks_channel = now;
+
+		if ( ctx->probe ) {
+			struct io_buffer *siob = ctx->probe; /* to send */
+
+			/* make a copy for future use */
+			iob = alloc_iob ( siob->tail - siob->head );
+			iob_reserve ( iob, iob_headroom ( siob ) );
+			memcpy ( iob_put ( iob, iob_len ( siob ) ),
+				 siob->data, iob_len ( siob ) );
+
+			ctx->probe = iob;
+			rc = net80211_tx_mgmt ( dev, IEEE80211_STYPE_PROBE_REQ,
+						net80211_ll_broadcast,
+						iob_disown ( siob ) );
+			if ( rc ) {
+				DBGC ( dev, "802.11 %p send probe failed: "
+				       "%s\n", dev, strerror ( rc ) );
+				return rc;
+			}
+		}
+	}
+
+	/* Check for new management packets */
+	while ( ( iob = net80211_mgmt_dequeue ( dev, &signal ) ) != NULL ) {
+		struct ieee80211_frame *hdr;
+		struct ieee80211_beacon *beacon;
+		union ieee80211_ie *ie;
+		struct net80211_wlan *wlan;
+		u16 type;
+
+		hdr = iob->data;
+		type = hdr->fc & IEEE80211_FC_SUBTYPE;
+		beacon = ( struct ieee80211_beacon * ) hdr->data;
+
+		if ( type != IEEE80211_STYPE_BEACON &&
+		     type != IEEE80211_STYPE_PROBE_RESP ) {
+			DBGC2 ( dev, "802.11 %p probe: non-beacon\n", dev );
+			goto drop;
+		}
+
+		if ( ( void * ) beacon->info_element >= iob->tail ) {
+			DBGC ( dev, "802.11 %p probe: beacon with no IEs\n",
+			       dev );
+			goto drop;
+		}
+
+		ie = beacon->info_element;
+
+		if ( ! ieee80211_ie_bound ( ie, iob->tail ) )
+			ie = NULL;
+
+		while ( ie && ie->id != IEEE80211_IE_SSID )
+			ie = ieee80211_next_ie ( ie, iob->tail );
+
+		if ( ! ie ) {
+			DBGC ( dev, "802.11 %p probe: beacon with no SSID\n",
+			       dev );
+			goto drop;
+		}
+
+		memcpy ( ssid, ie->ssid, ie->len );
+		ssid[ie->len] = 0;
+
+		if ( ctx->essid[0] && strcmp ( ctx->essid, ssid ) != 0 ) {
+			DBGC2 ( dev, "802.11 %p probe: beacon with wrong SSID "
+				"(%s)\n", dev, ssid );
+			goto drop;
+		}
+
+		/* See if we've got an entry for this network */
+		list_for_each_entry ( wlan, ctx->beacons, list ) {
+			if ( strcmp ( wlan->essid, ssid ) != 0 )
+				continue;
+
+			if ( signal < wlan->signal ) {
+				DBGC2 ( dev, "802.11 %p probe: beacon for %s "
+					"(%s) with weaker signal %d\n", dev,
+					ssid, eth_ntoa ( hdr->addr3 ), signal );
+				goto drop;
+			}
+
+			goto fill;
+		}
+
+		/* No entry yet - make one */
+		wlan = zalloc ( sizeof ( *wlan ) );
+		strcpy ( wlan->essid, ssid );
+		list_add_tail ( &wlan->list, ctx->beacons );
+
+		/* Whether we're using an old entry or a new one, fill
+		   it with new data. */
+	fill:
+		memcpy ( wlan->bssid, hdr->addr3, ETH_ALEN );
+		wlan->signal = signal;
+		wlan->channel = dev->channels[dev->channel].channel_nr;
+
+		/* Copy this I/O buffer into a new wlan->beacon; the
+		 * iob we've got probably came from the device driver
+		 * and may have the full 2.4k allocation, which we
+		 * don't want to keep around wasting memory.
+		 */
+		free_iob ( wlan->beacon );
+		wlan->beacon = alloc_iob ( iob_len ( iob ) );
+		memcpy ( iob_put ( wlan->beacon, iob_len ( iob ) ),
+			 iob->data, iob_len ( iob ) );
+
+		if ( ( rc = sec80211_detect ( wlan->beacon, &wlan->handshaking,
+					      &wlan->crypto ) ) == -ENOTSUP ) {
+			struct ieee80211_beacon *beacon =
+				( struct ieee80211_beacon * ) hdr->data;
+
+			if ( beacon->capability & IEEE80211_CAPAB_PRIVACY ) {
+				DBG ( "802.11 %p probe: secured network %s but "
+				      "encryption support not compiled in\n",
+				      dev, wlan->essid );
+				wlan->handshaking = NET80211_SECPROT_UNKNOWN;
+				wlan->crypto = NET80211_CRYPT_UNKNOWN;
+			} else {
+				wlan->handshaking = NET80211_SECPROT_NONE;
+				wlan->crypto = NET80211_CRYPT_NONE;
+			}
+		} else if ( rc != 0 ) {
+			DBGC ( dev, "802.11 %p probe warning: network "
+			       "%s with unidentifiable security "
+			       "settings: %s\n", dev, wlan->essid,
+			       strerror ( rc ) );
+		}
+
+		ctx->ticks_beacon = now;
+
+		DBGC2 ( dev, "802.11 %p probe: good beacon for %s (%s)\n",
+			dev, wlan->essid, eth_ntoa ( wlan->bssid ) );
+
+	drop:
+		free_iob ( iob );
+	}
+
+	return 0;
+}
+
+
+/**
+ * Finish probe of 802.11 networks, returning best-signal network found
+ *
+ * @v ctx	Probe context
+ * @ret wlan	Best-signal network found, or @c NULL if none were found
+ *
+ * If net80211_probe_start() was called with a particular SSID
+ * parameter as filter, only a network with that SSID (matching
+ * case-sensitively) can be returned from this function.
+ */
+struct net80211_wlan *
+net80211_probe_finish_best ( struct net80211_probe_ctx *ctx )
+{
+	struct net80211_wlan *best = NULL, *wlan;
+
+	if ( ! ctx )
+		return NULL;
+
+	list_for_each_entry ( wlan, ctx->beacons, list ) {
+		if ( ! best || best->signal < wlan->signal )
+			best = wlan;
+	}
+
+	if ( best )
+		list_del ( &best->list );
+	else
+		DBGC ( ctx->dev, "802.11 %p probe: found nothing for '%s'\n",
+		       ctx->dev, ctx->essid );
+
+	net80211_free_wlanlist ( ctx->beacons );
+
+	net80211_keep_mgmt ( ctx->dev, ctx->old_keep_mgmt );
+
+	if ( ctx->probe )
+		free_iob ( ctx->probe );
+
+	free ( ctx );
+
+	return best;
+}
+
+
+/**
+ * Finish probe of 802.11 networks, returning all networks found
+ *
+ * @v ctx	Probe context
+ * @ret list	List of net80211_wlan detailing networks found
+ *
+ * If net80211_probe_start() was called with a particular SSID
+ * parameter as filter, this will always return either an empty or a
+ * one-element list.
+ */
+struct list_head *net80211_probe_finish_all ( struct net80211_probe_ctx *ctx )
+{
+	struct list_head *beacons = ctx->beacons;
+
+	if ( ! ctx )
+		return NULL;
+
+	net80211_keep_mgmt ( ctx->dev, ctx->old_keep_mgmt );
+
+	if ( ctx->probe )
+		free_iob ( ctx->probe );
+
+	free ( ctx );
+
+	return beacons;
+}
+
+
+/**
+ * Free WLAN structure
+ *
+ * @v wlan	WLAN structure to free
+ */
+void net80211_free_wlan ( struct net80211_wlan *wlan )
+{
+	if ( wlan ) {
+		free_iob ( wlan->beacon );
+		free ( wlan );
+	}
+}
+
+
+/**
+ * Free list of WLAN structures
+ *
+ * @v list	List of WLAN structures to free
+ */
+void net80211_free_wlanlist ( struct list_head *list )
+{
+	struct net80211_wlan *wlan, *tmp;
+
+	if ( ! list )
+		return;
+
+	list_for_each_entry_safe ( wlan, tmp, list, list ) {
+		list_del ( &wlan->list );
+		net80211_free_wlan ( wlan );
+	}
+
+	free ( list );
+}
+
+
+/** Number of ticks to wait for replies to association management frames */
+#define ASSOC_TIMEOUT	TICKS_PER_SEC
+
+/** Number of times to try sending a particular association management frame */
+#define ASSOC_RETRIES	2
+
+/**
+ * Step 802.11 association process
+ *
+ * @v proc	Association process
+ */
+static void net80211_step_associate ( struct process *proc )
+{
+	struct net80211_device *dev =
+	    container_of ( proc, struct net80211_device, proc_assoc );
+	int rc = 0;
+	int status = dev->state & NET80211_STATUS_MASK;
+
+	/*
+	 * We use a sort of state machine implemented using bits in
+	 * the dev->state variable. At each call, we take the
+	 * logically first step that has not yet succeeded; either it
+	 * has not been tried yet, it's being retried, or it failed.
+	 * If it failed, we return an error indication; otherwise we
+	 * perform the step. If it succeeds, RX handling code will set
+	 * the appropriate status bit for us.
+	 *
+	 * Probe works a bit differently, since we have to step it
+	 * on every call instead of waiting for a packet to arrive
+	 * that will set the completion bit for us.
+	 */
+
+	/* If we're waiting for a reply, check for timeout condition */
+	if ( dev->state & NET80211_WAITING ) {
+		/* Sanity check */
+		if ( ! dev->associating )
+			return;
+
+		if ( currticks() - dev->ctx.assoc->last_packet > ASSOC_TIMEOUT ) {
+			/* Timed out - fail if too many retries, or retry */
+			dev->ctx.assoc->times_tried++;
+			if ( ++dev->ctx.assoc->times_tried > ASSOC_RETRIES ) {
+				rc = -ETIMEDOUT;
+				goto fail;
+			}
+		} else {
+			/* Didn't time out - let it keep going */
+			return;
+		}
+	} else {
+		if ( dev->state & NET80211_PROBED )
+			dev->ctx.assoc->times_tried = 0;
+	}
+
+	if ( ! ( dev->state & NET80211_PROBED ) ) {
+		/* state: probe */
+
+		if ( ! dev->ctx.probe ) {
+			/* start probe */
+			int active = fetch_intz_setting ( NULL,
+						&net80211_active_setting );
+			int band = dev->hw->bands;
+
+			if ( active )
+				band &= ~NET80211_BAND_BIT_5GHZ;
+
+			rc = net80211_prepare_probe ( dev, band, active );
+			if ( rc )
+				goto fail;
+
+			dev->ctx.probe = net80211_probe_start ( dev, dev->essid,
+								active );
+			if ( ! dev->ctx.probe ) {
+				dev->assoc_rc = -ENOMEM;
+				goto fail;
+			}
+		}
+
+		rc = net80211_probe_step ( dev->ctx.probe );
+		if ( ! rc ) {
+			return;	/* still going */
+		}
+
+		dev->associating = net80211_probe_finish_best ( dev->ctx.probe );
+		dev->ctx.probe = NULL;
+		if ( ! dev->associating ) {
+			if ( rc > 0 ) /* "successful" probe found nothing */
+				rc = -ETIMEDOUT;
+			goto fail;
+		}
+
+		/* If we probed using a broadcast SSID, record that
+		   fact for the settings applicator before we clobber
+		   it with the specific SSID we've chosen. */
+		if ( ! dev->essid[0] )
+			dev->state |= NET80211_AUTO_SSID;
+
+		DBGC ( dev, "802.11 %p found network %s (%s)\n", dev,
+		       dev->associating->essid,
+		       eth_ntoa ( dev->associating->bssid ) );
+
+		dev->ctx.assoc = zalloc ( sizeof ( *dev->ctx.assoc ) );
+		if ( ! dev->ctx.assoc ) {
+			rc = -ENOMEM;
+			goto fail;
+		}
+
+		dev->state |= NET80211_PROBED;
+		dev->ctx.assoc->method = IEEE80211_AUTH_OPEN_SYSTEM;
+
+		return;
+	}
+
+	/* Record time of sending the packet we're about to send, for timeout */
+	dev->ctx.assoc->last_packet = currticks();
+
+	if ( ! ( dev->state & NET80211_AUTHENTICATED ) ) {
+		/* state: prepare and authenticate */
+
+		if ( status != IEEE80211_STATUS_SUCCESS ) {
+			/* we tried authenticating already, but failed */
+			int method = dev->ctx.assoc->method;
+
+			if ( method == IEEE80211_AUTH_OPEN_SYSTEM &&
+			     ( status == IEEE80211_STATUS_AUTH_CHALL_INVALID ||
+			       status == IEEE80211_STATUS_AUTH_ALGO_UNSUPP ) ) {
+				/* Maybe this network uses Shared Key? */
+				dev->ctx.assoc->method =
+					IEEE80211_AUTH_SHARED_KEY;
+			} else {
+				goto fail;
+			}
+		}
+
+		DBGC ( dev, "802.11 %p authenticating with method %d\n", dev,
+		       dev->ctx.assoc->method );
+
+		rc = net80211_prepare_assoc ( dev, dev->associating );
+		if ( rc )
+			goto fail;
+
+		rc = net80211_send_auth ( dev, dev->associating,
+					  dev->ctx.assoc->method );
+		if ( rc )
+			goto fail;
+
+		return;
+	}
+
+	if ( ! ( dev->state & NET80211_ASSOCIATED ) ) {
+		/* state: associate */
+
+		if ( status != IEEE80211_STATUS_SUCCESS )
+			goto fail;
+
+		DBGC ( dev, "802.11 %p associating\n", dev );
+
+		if ( dev->handshaker && dev->handshaker->start &&
+		     ! dev->handshaker->started ) {
+			rc = dev->handshaker->start ( dev );
+			if ( rc < 0 )
+				goto fail;
+			dev->handshaker->started = 1;
+		}
+
+		rc = net80211_send_assoc ( dev, dev->associating );
+		if ( rc )
+			goto fail;
+
+		return;
+	}
+
+	if ( ! ( dev->state & NET80211_CRYPTO_SYNCED ) ) {
+		/* state: crypto sync */
+		DBGC ( dev, "802.11 %p security handshaking\n", dev );
+
+		if ( ! dev->handshaker || ! dev->handshaker->step ) {
+			dev->state |= NET80211_CRYPTO_SYNCED;
+			return;
+		}
+
+		rc = dev->handshaker->step ( dev );
+
+		if ( rc < 0 ) {
+			/* Only record the returned error if we're
+			   still marked as associated, because an
+			   asynchronous error will have already been
+			   reported to net80211_deauthenticate() and
+			   assoc_rc thereby set. */
+			if ( dev->state & NET80211_ASSOCIATED )
+				dev->assoc_rc = rc;
+			rc = 0;
+			goto fail;
+		}
+
+		if ( rc > 0 ) {
+			dev->assoc_rc = 0;
+			dev->state |= NET80211_CRYPTO_SYNCED;
+		}
+		return;
+	}
+
+	/* state: done! */
+	netdev_link_up ( dev->netdev );
+	dev->assoc_rc = 0;
+	dev->state &= ~NET80211_WORKING;
+
+	free ( dev->ctx.assoc );
+	dev->ctx.assoc = NULL;
+
+	net80211_free_wlan ( dev->associating );
+	dev->associating = NULL;
+
+	dev->rctl = rc80211_init ( dev );
+
+	process_del ( proc );
+
+	DBGC ( dev, "802.11 %p associated with %s (%s)\n", dev,
+	       dev->essid, eth_ntoa ( dev->bssid ) );
+
+	return;
+
+ fail:
+	dev->state &= ~( NET80211_WORKING | NET80211_WAITING );
+	if ( rc )
+		dev->assoc_rc = rc;
+
+	netdev_link_err ( dev->netdev, dev->assoc_rc );
+
+	/* We never reach here from the middle of a probe, so we don't
+	   need to worry about freeing dev->ctx.probe. */
+
+	if ( dev->state & NET80211_PROBED ) {
+		free ( dev->ctx.assoc );
+		dev->ctx.assoc = NULL;
+	}
+
+	net80211_free_wlan ( dev->associating );
+	dev->associating = NULL;
+
+	process_del ( proc );
+
+	DBGC ( dev, "802.11 %p association failed (state=%04x): "
+	       "%s\n", dev, dev->state, strerror ( dev->assoc_rc ) );
+
+	/* Try it again: */
+	net80211_autoassociate ( dev );
+}
+
+/**
+ * Check for 802.11 SSID or key updates
+ *
+ * This acts as a settings applicator; if the user changes netX/ssid,
+ * and netX is currently open, the association task will be invoked
+ * again. If the user changes the encryption key, the current security
+ * handshaker will be asked to update its state to match; if that is
+ * impossible without reassociation, we reassociate.
+ */
+static int net80211_check_settings_update ( void )
+{
+	struct net80211_device *dev;
+	char ssid[IEEE80211_MAX_SSID_LEN + 1];
+	int key_reassoc;
+
+	list_for_each_entry ( dev, &net80211_devices, list ) {
+		if ( ! ( dev->netdev->state & NETDEV_OPEN ) )
+			continue;
+
+		key_reassoc = 0;
+		if ( dev->handshaker && dev->handshaker->change_key &&
+		     dev->handshaker->change_key ( dev ) < 0 )
+			key_reassoc = 1;
+
+		fetch_string_setting ( netdev_settings ( dev->netdev ),
+				       &net80211_ssid_setting, ssid,
+				       IEEE80211_MAX_SSID_LEN + 1 );
+
+		if ( key_reassoc ||
+		     ( ! ( ! ssid[0] && ( dev->state & NET80211_AUTO_SSID ) ) &&
+		       strcmp ( ssid, dev->essid ) != 0 ) ) {
+			DBGC ( dev, "802.11 %p updating association: "
+			       "%s -> %s\n", dev, dev->essid, ssid );
+			net80211_autoassociate ( dev );
+		}
+	}
+
+	return 0;
+}
+
+/**
+ * Start 802.11 association process
+ *
+ * @v dev	802.11 device
+ *
+ * If the association process is running, it will be restarted.
+ */
+void net80211_autoassociate ( struct net80211_device *dev )
+{
+	if ( ! ( dev->state & NET80211_WORKING ) ) {
+		DBGC2 ( dev, "802.11 %p spawning association process\n", dev );
+		process_add ( &dev->proc_assoc );
+	} else {
+		DBGC2 ( dev, "802.11 %p restarting association\n", dev );
+	}
+
+	/* Clean up everything an earlier association process might
+	   have been in the middle of using */
+	if ( dev->associating )
+		net80211_free_wlan ( dev->associating );
+
+	if ( ! ( dev->state & NET80211_PROBED ) )
+		net80211_free_wlan (
+			net80211_probe_finish_best ( dev->ctx.probe ) );
+	else
+		free ( dev->ctx.assoc );
+
+	/* Reset to a clean state */
+	fetch_string_setting ( netdev_settings ( dev->netdev ),
+			       &net80211_ssid_setting, dev->essid,
+			       IEEE80211_MAX_SSID_LEN + 1 );
+	dev->ctx.probe = NULL;
+	dev->associating = NULL;
+	dev->assoc_rc = 0;
+	net80211_set_state ( dev, NET80211_PROBED, NET80211_WORKING, 0 );
+}
+
+/**
+ * Pick TX rate for RTS/CTS packets based on data rate
+ *
+ * @v dev	802.11 device
+ *
+ * The RTS/CTS rate is the fastest TX rate marked as "basic" that is
+ * not faster than the data rate.
+ */
+static void net80211_set_rtscts_rate ( struct net80211_device *dev )
+{
+	u16 datarate = dev->rates[dev->rate];
+	u16 rtsrate = 0;
+	int rts_idx = -1;
+	int i;
+
+	for ( i = 0; i < dev->nr_rates; i++ ) {
+		u16 rate = dev->rates[i];
+
+		if ( ! ( dev->basic_rates & ( 1 << i ) ) || rate > datarate )
+			continue;
+
+		if ( rate > rtsrate ) {
+			rtsrate = rate;
+			rts_idx = i;
+		}
+	}
+
+	/* If this is in initialization, we might not have any basic
+	   rates; just use the first data rate in that case. */
+	if ( rts_idx < 0 )
+		rts_idx = 0;
+
+	dev->rtscts_rate = rts_idx;
+}
+
+/**
+ * Set data transmission rate for 802.11 device
+ *
+ * @v dev	802.11 device
+ * @v rate	Rate to set, as index into @c dev->rates array
+ */
+void net80211_set_rate_idx ( struct net80211_device *dev, int rate )
+{
+	assert ( dev->netdev->state & NETDEV_OPEN );
+
+	if ( rate >= 0 && rate < dev->nr_rates && rate != dev->rate ) {
+		DBGC2 ( dev, "802.11 %p changing rate from %d->%d Mbps\n",
+			dev, dev->rates[dev->rate] / 10,
+			dev->rates[rate] / 10 );
+
+		dev->rate = rate;
+		net80211_set_rtscts_rate ( dev );
+		dev->op->config ( dev, NET80211_CFG_RATE );
+	}
+}
+
+/**
+ * Configure 802.11 device to transmit on a certain channel
+ *
+ * @v dev	802.11 device
+ * @v channel	Channel number (1-11 for 2.4GHz) to transmit on
+ */
+int net80211_change_channel ( struct net80211_device *dev, int channel )
+{
+	int i, oldchan = dev->channel;
+
+	assert ( dev->netdev->state & NETDEV_OPEN );
+
+	for ( i = 0; i < dev->nr_channels; i++ ) {
+		if ( dev->channels[i].channel_nr == channel ) {
+			dev->channel = i;
+			break;
+		}
+	}
+
+	if ( i == dev->nr_channels )
+		return -ENOENT;
+
+	if ( i != oldchan )
+		return dev->op->config ( dev, NET80211_CFG_CHANNEL );
+
+	return 0;
+}
+
+/**
+ * Prepare 802.11 device channel and rate set for scanning
+ *
+ * @v dev	802.11 device
+ * @v band	RF band(s) on which to prepare for scanning
+ * @v active	Whether the scanning will be active
+ * @ret rc	Return status code
+ */
+int net80211_prepare_probe ( struct net80211_device *dev, int band,
+			     int active )
+{
+	assert ( dev->netdev->state & NETDEV_OPEN );
+
+	if ( active && ( band & NET80211_BAND_BIT_5GHZ ) ) {
+		DBGC ( dev, "802.11 %p cannot perform active scanning on "
+		       "5GHz band\n", dev );
+		return -EINVAL_ACTIVE_SCAN;
+	}
+
+	if ( band == 0 ) {
+		/* This can happen for a 5GHz-only card with 5GHz
+		   scanning masked out by an active request. */
+		DBGC ( dev, "802.11 %p asked to prepare for scanning nothing\n",
+		       dev );
+		return -EINVAL_ACTIVE_SCAN;
+	}
+
+	dev->nr_channels = 0;
+
+	if ( active )
+		net80211_add_channels ( dev, 1, 11, NET80211_REG_TXPOWER );
+	else {
+		if ( band & NET80211_BAND_BIT_2GHZ )
+			net80211_add_channels ( dev, 1, 14,
+						NET80211_REG_TXPOWER );
+		if ( band & NET80211_BAND_BIT_5GHZ )
+			net80211_add_channels ( dev, 36, 8,
+						NET80211_REG_TXPOWER );
+	}
+
+	net80211_filter_hw_channels ( dev );
+
+	/* Use channel 1 for now */
+	dev->channel = 0;
+	dev->op->config ( dev, NET80211_CFG_CHANNEL );
+
+	/* Always do active probes at lowest (presumably first) speed */
+	dev->rate = 0;
+	dev->nr_rates = 1;
+	dev->rates[0] = dev->hw->rates[dev->channels[0].band][0];
+	dev->op->config ( dev, NET80211_CFG_RATE );
+
+	return 0;
+}
+
+/**
+ * Prepare 802.11 device channel and rate set for communication
+ *
+ * @v dev	802.11 device
+ * @v wlan	WLAN to prepare for communication with
+ * @ret rc	Return status code
+ */
+int net80211_prepare_assoc ( struct net80211_device *dev,
+			     struct net80211_wlan *wlan )
+{
+	struct ieee80211_frame *hdr = wlan->beacon->data;
+	struct ieee80211_beacon *beacon =
+		( struct ieee80211_beacon * ) hdr->data;
+	struct net80211_handshaker *handshaker;
+	int rc;
+
+	assert ( dev->netdev->state & NETDEV_OPEN );
+
+	net80211_set_state ( dev, NET80211_ASSOCIATED, 0, 0 );
+	memcpy ( dev->bssid, wlan->bssid, ETH_ALEN );
+	strcpy ( dev->essid, wlan->essid );
+
+	free ( dev->rsn_ie );
+	dev->rsn_ie = NULL;
+
+	dev->last_beacon_timestamp = beacon->timestamp;
+	dev->tx_beacon_interval = 1024 * beacon->beacon_interval;
+
+	/* Barring an IE that tells us the channel outright, assume
+	   the channel we heard this AP best on is the channel it's
+	   communicating on. */
+	net80211_change_channel ( dev, wlan->channel );
+
+	rc = net80211_process_capab ( dev, beacon->capability );
+	if ( rc )
+		return rc;
+
+	rc = net80211_process_ie ( dev, beacon->info_element,
+				   wlan->beacon->tail );
+	if ( rc )
+		return rc;
+
+	/* Associate at the lowest rate so we know it'll get through */
+	dev->rate = 0;
+	dev->op->config ( dev, NET80211_CFG_RATE );
+
+	/* Free old handshaker and crypto, if they exist */
+	if ( dev->handshaker && dev->handshaker->stop &&
+	     dev->handshaker->started )
+		dev->handshaker->stop ( dev );
+	free ( dev->handshaker );
+	dev->handshaker = NULL;
+	free ( dev->crypto );
+	free ( dev->gcrypto );
+	dev->crypto = dev->gcrypto = NULL;
+
+	/* Find new security handshaker to use */
+	for_each_table_entry ( handshaker, NET80211_HANDSHAKERS ) {
+		if ( handshaker->protocol == wlan->handshaking ) {
+			dev->handshaker = zalloc ( sizeof ( *handshaker ) +
+						   handshaker->priv_len );
+			if ( ! dev->handshaker )
+				return -ENOMEM;
+
+			memcpy ( dev->handshaker, handshaker,
+				 sizeof ( *handshaker ) );
+			dev->handshaker->priv = ( ( void * ) dev->handshaker +
+						  sizeof ( *handshaker ) );
+			break;
+		}
+	}
+
+	if ( ( wlan->handshaking != NET80211_SECPROT_NONE ) &&
+	     ! dev->handshaker ) {
+		DBGC ( dev, "802.11 %p no support for handshaking scheme %d\n",
+		       dev, wlan->handshaking );
+		return -( ENOTSUP | ( wlan->handshaking << 8 ) );
+	}
+
+	/* Initialize security handshaker */
+	if ( dev->handshaker ) {
+		rc = dev->handshaker->init ( dev );
+		if ( rc < 0 )
+			return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Send 802.11 initial authentication frame
+ *
+ * @v dev	802.11 device
+ * @v wlan	WLAN to authenticate with
+ * @v method	Authentication method
+ * @ret rc	Return status code
+ *
+ * @a method may be 0 for Open System authentication or 1 for Shared
+ * Key authentication. Open System provides no security in association
+ * whatsoever, relying on encryption for confidentiality, but Shared
+ * Key actively introduces security problems and is very rarely used.
+ */
+int net80211_send_auth ( struct net80211_device *dev,
+			 struct net80211_wlan *wlan, int method )
+{
+	struct io_buffer *iob = alloc_iob ( 64 );
+	struct ieee80211_auth *auth;
+
+	net80211_set_state ( dev, 0, NET80211_WAITING, 0 );
+	iob_reserve ( iob, IEEE80211_TYP_FRAME_HEADER_LEN );
+	auth = iob_put ( iob, sizeof ( *auth ) );
+	auth->algorithm = method;
+	auth->tx_seq = 1;
+	auth->status = 0;
+
+	return net80211_tx_mgmt ( dev, IEEE80211_STYPE_AUTH, wlan->bssid, iob );
+}
+
+/**
+ * Handle receipt of 802.11 authentication frame
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer
+ *
+ * If the authentication method being used is Shared Key, and the
+ * frame that was received included challenge text, the frame is
+ * encrypted using the cryptosystem currently in effect and sent back
+ * to the AP to complete the authentication.
+ */
+static void net80211_handle_auth ( struct net80211_device *dev,
+				   struct io_buffer *iob )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	struct ieee80211_auth *auth =
+	    ( struct ieee80211_auth * ) hdr->data;
+
+	if ( auth->tx_seq & 1 ) {
+		DBGC ( dev, "802.11 %p authentication received improperly "
+		       "directed frame (seq. %d)\n", dev, auth->tx_seq );
+		net80211_set_state ( dev, NET80211_WAITING, 0,
+				     IEEE80211_STATUS_FAILURE );
+		return;
+	}
+
+	if ( auth->status != IEEE80211_STATUS_SUCCESS ) {
+		DBGC ( dev, "802.11 %p authentication failed: status %d\n",
+		       dev, auth->status );
+		net80211_set_state ( dev, NET80211_WAITING, 0,
+				     auth->status );
+		return;
+	}
+
+	if ( auth->algorithm == IEEE80211_AUTH_SHARED_KEY && ! dev->crypto ) {
+		DBGC ( dev, "802.11 %p can't perform shared-key authentication "
+		       "without a cryptosystem\n", dev );
+		net80211_set_state ( dev, NET80211_WAITING, 0,
+				     IEEE80211_STATUS_FAILURE );
+		return;
+	}
+
+	if ( auth->algorithm == IEEE80211_AUTH_SHARED_KEY &&
+	     auth->tx_seq == 2 ) {
+		/* Since the iob we got is going to be freed as soon
+		   as we return, we can do some in-place
+		   modification. */
+		auth->tx_seq = 3;
+		auth->status = 0;
+
+		memcpy ( hdr->addr2, hdr->addr1, ETH_ALEN );
+		memcpy ( hdr->addr1, hdr->addr3, ETH_ALEN );
+
+		netdev_tx ( dev->netdev,
+			    dev->crypto->encrypt ( dev->crypto, iob ) );
+		return;
+	}
+
+	net80211_set_state ( dev, NET80211_WAITING, NET80211_AUTHENTICATED,
+			     IEEE80211_STATUS_SUCCESS );
+
+	return;
+}
+
+/**
+ * Send 802.11 association frame
+ *
+ * @v dev	802.11 device
+ * @v wlan	WLAN to associate with
+ * @ret rc	Return status code
+ */
+int net80211_send_assoc ( struct net80211_device *dev,
+			  struct net80211_wlan *wlan )
+{
+	struct io_buffer *iob = alloc_iob ( 128 );
+	struct ieee80211_assoc_req *assoc;
+	union ieee80211_ie *ie;
+
+	net80211_set_state ( dev, 0, NET80211_WAITING, 0 );
+
+	iob_reserve ( iob, IEEE80211_TYP_FRAME_HEADER_LEN );
+	assoc = iob->data;
+
+	assoc->capability = IEEE80211_CAPAB_MANAGED;
+	if ( ! ( dev->hw->flags & NET80211_HW_NO_SHORT_PREAMBLE ) )
+		assoc->capability |= IEEE80211_CAPAB_SHORT_PMBL;
+	if ( ! ( dev->hw->flags & NET80211_HW_NO_SHORT_SLOT ) )
+		assoc->capability |= IEEE80211_CAPAB_SHORT_SLOT;
+	if ( wlan->crypto )
+		assoc->capability |= IEEE80211_CAPAB_PRIVACY;
+
+	assoc->listen_interval = 1;
+
+	ie = net80211_marshal_request_info ( dev, assoc->info_element );
+
+	DBGP ( "802.11 %p about to send association request:\n", dev );
+	DBGP_HD ( iob->data, ( void * ) ie - iob->data );
+
+	iob_put ( iob, ( void * ) ie - iob->data );
+
+	return net80211_tx_mgmt ( dev, IEEE80211_STYPE_ASSOC_REQ,
+				  wlan->bssid, iob );
+}
+
+/**
+ * Handle receipt of 802.11 association reply frame
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer
+ */
+static void net80211_handle_assoc_reply ( struct net80211_device *dev,
+					  struct io_buffer *iob )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	struct ieee80211_assoc_resp *assoc =
+		( struct ieee80211_assoc_resp * ) hdr->data;
+
+	net80211_process_capab ( dev, assoc->capability );
+	net80211_process_ie ( dev, assoc->info_element, iob->tail );
+
+	if ( assoc->status != IEEE80211_STATUS_SUCCESS ) {
+		DBGC ( dev, "802.11 %p association failed: status %d\n",
+		       dev, assoc->status );
+		net80211_set_state ( dev, NET80211_WAITING, 0,
+				     assoc->status );
+		return;
+	}
+
+	/* ESSID was filled before the association request was sent */
+	memcpy ( dev->bssid, hdr->addr3, ETH_ALEN );
+	dev->aid = assoc->aid;
+
+	net80211_set_state ( dev, NET80211_WAITING, NET80211_ASSOCIATED,
+			     IEEE80211_STATUS_SUCCESS );
+}
+
+
+/**
+ * Send 802.11 disassociation frame
+ *
+ * @v dev	802.11 device
+ * @v reason	Reason for disassociation
+ * @v deauth	If TRUE, send deauthentication instead of disassociation
+ * @ret rc	Return status code
+ */
+static int net80211_send_disassoc ( struct net80211_device *dev, int reason,
+				    int deauth )
+{
+	struct io_buffer *iob = alloc_iob ( 64 );
+	struct ieee80211_disassoc *disassoc;
+
+	if ( ! ( dev->state & NET80211_ASSOCIATED ) )
+		return -EINVAL;
+
+	net80211_set_state ( dev, NET80211_ASSOCIATED, 0, 0 );
+	iob_reserve ( iob, IEEE80211_TYP_FRAME_HEADER_LEN );
+	disassoc = iob_put ( iob, sizeof ( *disassoc ) );
+	disassoc->reason = reason;
+
+	return net80211_tx_mgmt ( dev, deauth ? IEEE80211_STYPE_DEAUTH :
+				  IEEE80211_STYPE_DISASSOC, dev->bssid, iob );
+}
+
+
+/**
+ * Deauthenticate from current network and try again
+ *
+ * @v dev	802.11 device
+ * @v rc	Return status code indicating reason
+ *
+ * The deauthentication will be sent using an 802.11 "unspecified
+ * reason", as is common, but @a rc will be set as a link-up
+ * error to aid the user in debugging.
+ */
+void net80211_deauthenticate ( struct net80211_device *dev, int rc )
+{
+	net80211_send_disassoc ( dev, IEEE80211_REASON_UNSPECIFIED, 1 );
+	dev->assoc_rc = rc;
+	netdev_link_err ( dev->netdev, rc );
+
+	net80211_autoassociate ( dev );
+}
+
+
+/** Smoothing factor (1-7) for link quality calculation */
+#define LQ_SMOOTH	7
+
+/**
+ * Update link quality information based on received beacon
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer containing beacon
+ * @ret rc	Return status code
+ */
+static void net80211_update_link_quality ( struct net80211_device *dev,
+					   struct io_buffer *iob )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	struct ieee80211_beacon *beacon;
+	u32 dt, rxi;
+
+	if ( ! ( dev->state & NET80211_ASSOCIATED ) )
+		return;
+
+	beacon = ( struct ieee80211_beacon * ) hdr->data;
+	dt = ( u32 ) ( beacon->timestamp - dev->last_beacon_timestamp );
+	rxi = dev->rx_beacon_interval;
+
+	rxi = ( LQ_SMOOTH * rxi ) + ( ( 8 - LQ_SMOOTH ) * dt );
+	dev->rx_beacon_interval = rxi >> 3;
+
+	dev->last_beacon_timestamp = beacon->timestamp;
+}
+
+
+/**
+ * Handle receipt of 802.11 management frame
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer
+ * @v signal	Signal strength of received frame
+ */
+static void net80211_handle_mgmt ( struct net80211_device *dev,
+				   struct io_buffer *iob, int signal )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	struct ieee80211_disassoc *disassoc;
+	u16 stype = hdr->fc & IEEE80211_FC_SUBTYPE;
+	int keep = 0;
+	int is_deauth = ( stype == IEEE80211_STYPE_DEAUTH );
+
+	if ( ( hdr->fc & IEEE80211_FC_TYPE ) != IEEE80211_TYPE_MGMT ) {
+		free_iob ( iob );
+		return;		/* only handle management frames */
+	}
+
+	switch ( stype ) {
+		/* We reconnect on deauthentication and disassociation. */
+	case IEEE80211_STYPE_DEAUTH:
+	case IEEE80211_STYPE_DISASSOC:
+		disassoc = ( struct ieee80211_disassoc * ) hdr->data;
+		net80211_set_state ( dev, is_deauth ? NET80211_AUTHENTICATED :
+				     NET80211_ASSOCIATED, 0,
+				     NET80211_IS_REASON | disassoc->reason );
+		DBGC ( dev, "802.11 %p %s: reason %d\n",
+		       dev, is_deauth ? "deauthenticated" : "disassociated",
+		       disassoc->reason );
+
+		/* Try to reassociate, in case it's transient. */
+		net80211_autoassociate ( dev );
+
+		break;
+
+		/* We handle authentication and association. */
+	case IEEE80211_STYPE_AUTH:
+		if ( ! ( dev->state & NET80211_AUTHENTICATED ) )
+			net80211_handle_auth ( dev, iob );
+		break;
+
+	case IEEE80211_STYPE_ASSOC_RESP:
+	case IEEE80211_STYPE_REASSOC_RESP:
+		if ( ! ( dev->state & NET80211_ASSOCIATED ) )
+			net80211_handle_assoc_reply ( dev, iob );
+		break;
+
+		/* We pass probes and beacons onto network scanning
+		   code. Pass actions for future extensibility. */
+	case IEEE80211_STYPE_BEACON:
+		net80211_update_link_quality ( dev, iob );
+		/* fall through */
+	case IEEE80211_STYPE_PROBE_RESP:
+	case IEEE80211_STYPE_ACTION:
+		if ( dev->keep_mgmt ) {
+			struct net80211_rx_info *rxinf;
+			rxinf = zalloc ( sizeof ( *rxinf ) );
+			if ( ! rxinf ) {
+				DBGC ( dev, "802.11 %p out of memory\n", dev );
+				break;
+			}
+			rxinf->signal = signal;
+			list_add_tail ( &iob->list, &dev->mgmt_queue );
+			list_add_tail ( &rxinf->list, &dev->mgmt_info_queue );
+			keep = 1;
+		}
+		break;
+
+	case IEEE80211_STYPE_PROBE_REQ:
+		/* Some nodes send these broadcast. Ignore them. */
+		break;
+
+	case IEEE80211_STYPE_ASSOC_REQ:
+	case IEEE80211_STYPE_REASSOC_REQ:
+		/* We should never receive these, only send them. */
+		DBGC ( dev, "802.11 %p received strange management request "
+		       "(%04x)\n", dev, stype );
+		break;
+
+	default:
+		DBGC ( dev, "802.11 %p received unimplemented management "
+		       "packet (%04x)\n", dev, stype );
+		break;
+	}
+
+	if ( ! keep )
+		free_iob ( iob );
+}
+
+/* ---------- Packet handling functions ---------- */
+
+/**
+ * Free buffers used by 802.11 fragment cache entry
+ *
+ * @v dev	802.11 device
+ * @v fcid	Fragment cache entry index
+ *
+ * After this function, the referenced entry will be marked unused.
+ */
+static void net80211_free_frags ( struct net80211_device *dev, int fcid )
+{
+	int j;
+	struct net80211_frag_cache *frag = &dev->frags[fcid];
+
+	for ( j = 0; j < 16; j++ ) {
+		if ( frag->iob[j] ) {
+			free_iob ( frag->iob[j] );
+			frag->iob[j] = NULL;
+		}
+	}
+
+	frag->seqnr = 0;
+	frag->start_ticks = 0;
+	frag->in_use = 0;
+}
+
+/**
+ * Accumulate 802.11 fragments into one I/O buffer
+ *
+ * @v dev	802.11 device
+ * @v fcid	Fragment cache entry index
+ * @v nfrags	Number of fragments received
+ * @v size	Sum of sizes of all fragments, including headers
+ * @ret iob	I/O buffer containing reassembled packet
+ *
+ * This function does not free the fragment buffers.
+ */
+static struct io_buffer *net80211_accum_frags ( struct net80211_device *dev,
+						int fcid, int nfrags, int size )
+{
+	struct net80211_frag_cache *frag = &dev->frags[fcid];
+	int hdrsize = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int nsize = size - hdrsize * ( nfrags - 1 );
+	int i;
+
+	struct io_buffer *niob = alloc_iob ( nsize );
+	struct ieee80211_frame *hdr;
+
+	/* Add the header from the first one... */
+	memcpy ( iob_put ( niob, hdrsize ), frag->iob[0]->data, hdrsize );
+
+	/* ... and all the data from all of them. */
+	for ( i = 0; i < nfrags; i++ ) {
+		int len = iob_len ( frag->iob[i] ) - hdrsize;
+		memcpy ( iob_put ( niob, len ),
+			 frag->iob[i]->data + hdrsize, len );
+	}
+
+	/* Turn off the fragment bit. */
+	hdr = niob->data;
+	hdr->fc &= ~IEEE80211_FC_MORE_FRAG;
+
+	return niob;
+}
+
+/**
+ * Handle receipt of 802.11 fragment
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer containing fragment
+ * @v signal	Signal strength with which fragment was received
+ */
+static void net80211_rx_frag ( struct net80211_device *dev,
+			       struct io_buffer *iob, int signal )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	int fragnr = IEEE80211_FRAG ( hdr->seq );
+
+	if ( fragnr == 0 && ( hdr->fc & IEEE80211_FC_MORE_FRAG ) ) {
+		/* start a frag cache entry */
+		int i, newest = -1;
+		u32 curr_ticks = currticks(), newest_ticks = 0;
+		u32 timeout = ticks_per_sec() * NET80211_FRAG_TIMEOUT;
+
+		for ( i = 0; i < NET80211_NR_CONCURRENT_FRAGS; i++ ) {
+			if ( dev->frags[i].in_use == 0 )
+				break;
+
+			if ( dev->frags[i].start_ticks + timeout >=
+			     curr_ticks ) {
+				net80211_free_frags ( dev, i );
+				break;
+			}
+
+			if ( dev->frags[i].start_ticks > newest_ticks ) {
+				newest = i;
+				newest_ticks = dev->frags[i].start_ticks;
+			}
+		}
+
+		/* If we're being sent more concurrent fragmented
+		   packets than we can handle, drop the newest so the
+		   older ones have time to complete. */
+		if ( i == NET80211_NR_CONCURRENT_FRAGS ) {
+			i = newest;
+			net80211_free_frags ( dev, i );
+		}
+
+		dev->frags[i].in_use = 1;
+		dev->frags[i].seqnr = IEEE80211_SEQNR ( hdr->seq );
+		dev->frags[i].start_ticks = currticks();
+		dev->frags[i].iob[0] = iob;
+		return;
+	} else {
+		int i;
+		for ( i = 0; i < NET80211_NR_CONCURRENT_FRAGS; i++ ) {
+			if ( dev->frags[i].in_use && dev->frags[i].seqnr ==
+			     IEEE80211_SEQNR ( hdr->seq ) )
+				break;
+		}
+		if ( i == NET80211_NR_CONCURRENT_FRAGS ) {
+			/* Drop non-first not-in-cache fragments */
+			DBGC ( dev, "802.11 %p dropped fragment fc=%04x "
+			       "seq=%04x\n", dev, hdr->fc, hdr->seq );
+			free_iob ( iob );
+			return;
+		}
+
+		dev->frags[i].iob[fragnr] = iob;
+
+		if ( ! ( hdr->fc & IEEE80211_FC_MORE_FRAG ) ) {
+			int j, size = 0;
+			for ( j = 0; j < fragnr; j++ ) {
+				size += iob_len ( dev->frags[i].iob[j] );
+				if ( dev->frags[i].iob[j] == NULL )
+					break;
+			}
+			if ( j == fragnr ) {
+				/* We've got everything */
+				struct io_buffer *niob =
+				    net80211_accum_frags ( dev, i, fragnr,
+							   size );
+				net80211_free_frags ( dev, i );
+				net80211_rx ( dev, niob, signal, 0 );
+			} else {
+				DBGC ( dev, "802.11 %p dropping fragmented "
+				       "packet due to out-of-order arrival, "
+				       "fc=%04x seq=%04x\n", dev, hdr->fc,
+				       hdr->seq );
+				net80211_free_frags ( dev, i );
+			}
+		}
+	}
+}
+
+/**
+ * Handle receipt of 802.11 frame
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer
+ * @v signal	Received signal strength
+ * @v rate	Bitrate at which frame was received, in 100 kbps units
+ *
+ * If the rate or signal is unknown, 0 should be passed.
+ */
+void net80211_rx ( struct net80211_device *dev, struct io_buffer *iob,
+		   int signal, u16 rate )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	u16 type = hdr->fc & IEEE80211_FC_TYPE;
+	if ( ( hdr->fc & IEEE80211_FC_VERSION ) != IEEE80211_THIS_VERSION )
+		goto drop;	/* drop invalid-version packets */
+
+	if ( type == IEEE80211_TYPE_CTRL )
+		goto drop;	/* we don't handle control packets,
+				   the hardware does */
+
+	if ( dev->last_rx_seq == hdr->seq )
+		goto drop;	/* avoid duplicate packet */
+	dev->last_rx_seq = hdr->seq;
+
+	if ( dev->hw->flags & NET80211_HW_RX_HAS_FCS ) {
+		/* discard the FCS */
+		iob_unput ( iob, 4 );
+	}
+
+	/* Only decrypt packets from our BSSID, to avoid spurious errors */
+	if ( ( hdr->fc & IEEE80211_FC_PROTECTED ) &&
+	     ! memcmp ( hdr->addr2, dev->bssid, ETH_ALEN ) ) {
+		/* Decrypt packet; record and drop if it fails */
+		struct io_buffer *niob;
+		struct net80211_crypto *crypto = dev->crypto;
+
+		if ( ! dev->crypto ) {
+			DBGC ( dev, "802.11 %p cannot decrypt packet "
+			       "without a cryptosystem\n", dev );
+			goto drop_crypt;
+		}
+
+		if ( ( hdr->addr1[0] & 1 ) && dev->gcrypto ) {
+			/* Use group decryption if needed */
+			crypto = dev->gcrypto;
+		}
+
+		niob = crypto->decrypt ( crypto, iob );
+		if ( ! niob ) {
+			DBGC ( dev, "802.11 %p decryption error\n", dev );
+			goto drop_crypt;
+		}
+		free_iob ( iob );
+		iob = niob;
+	}
+
+	dev->last_signal = signal;
+
+	/* Fragments go into the frag cache or get dropped. */
+	if ( IEEE80211_FRAG ( hdr->seq ) != 0
+	     || ( hdr->fc & IEEE80211_FC_MORE_FRAG ) ) {
+		net80211_rx_frag ( dev, iob, signal );
+		return;
+	}
+
+	/* Management frames get handled, enqueued, or dropped. */
+	if ( type == IEEE80211_TYPE_MGMT ) {
+		net80211_handle_mgmt ( dev, iob, signal );
+		return;
+	}
+
+	/* Data frames get dropped or sent to the net_device. */
+	if ( ( hdr->fc & IEEE80211_FC_SUBTYPE ) != IEEE80211_STYPE_DATA )
+		goto drop;	/* drop QoS, CFP, or null data packets */
+
+	/* Update rate-control algorithm */
+	if ( dev->rctl )
+		rc80211_update_rx ( dev, hdr->fc & IEEE80211_FC_RETRY, rate );
+
+	/* Pass packet onward */
+	if ( dev->state & NET80211_ASSOCIATED ) {
+		netdev_rx ( dev->netdev, iob );
+		return;
+	}
+
+	/* No association? Drop it. */
+	goto drop;
+
+ drop_crypt:
+	netdev_rx_err ( dev->netdev, NULL, EINVAL_CRYPTO_REQUEST );
+ drop:
+	DBGC2 ( dev, "802.11 %p dropped packet fc=%04x seq=%04x\n", dev,
+		hdr->fc, hdr->seq );
+	free_iob ( iob );
+	return;
+}
+
+/** Indicate an error in receiving a packet
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer with received packet, or NULL
+ * @v rc	Error code
+ *
+ * This logs the error with the wrapping net_device, and frees iob if
+ * it is passed.
+ */
+void net80211_rx_err ( struct net80211_device *dev,
+		       struct io_buffer *iob, int rc )
+{
+	netdev_rx_err ( dev->netdev, iob, rc );
+}
+
+/** Indicate the completed transmission of a packet
+ *
+ * @v dev	802.11 device
+ * @v iob	I/O buffer of transmitted packet
+ * @v retries	Number of times this packet was retransmitted
+ * @v rc	Error code, or 0 for success
+ *
+ * This logs an error with the wrapping net_device if one occurred,
+ * and removes and frees the I/O buffer from its TX queue. The
+ * provided retry information is used to tune our transmission rate.
+ *
+ * If the packet did not need to be retransmitted because it was
+ * properly ACKed the first time, @a retries should be 0.
+ */
+void net80211_tx_complete ( struct net80211_device *dev,
+			    struct io_buffer *iob, int retries, int rc )
+{
+	/* Update rate-control algorithm */
+	if ( dev->rctl )
+		rc80211_update_tx ( dev, retries, rc );
+
+	/* Pass completion onward */
+	netdev_tx_complete_err ( dev->netdev, iob, rc );
+}
diff --git a/gpxe/src/net/80211/rc80211.c b/gpxe/src/net/80211/rc80211.c
new file mode 100644
index 00000000..5bd19143
--- /dev/null
+++ b/gpxe/src/net/80211/rc80211.c
@@ -0,0 +1,371 @@
+/*
+ * Simple 802.11 rate-control algorithm for gPXE.
+ *
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdlib.h>
+#include <gpxe/net80211.h>
+
+/**
+ * @file
+ *
+ * Simple 802.11 rate-control algorithm
+ */
+
+/** @page rc80211 Rate control philosophy
+ *
+ * We want to maximize our transmission speed, to the extent that we
+ * can do that without dropping undue numbers of packets. We also
+ * don't want to take up very much code space, so our algorithm has to
+ * be pretty simple
+ *
+ * When we receive a packet, we know what rate it was transmitted at,
+ * and whether it had to be retransmitted to get to us.
+ *
+ * When we send a packet, we hear back how many times it had to be
+ * retried to get through, and whether it got through at all.
+ *
+ * Indications of TX success are more reliable than RX success, but RX
+ * information helps us know where to start.
+ *
+ * To handle all of this, we keep for each rate and each direction (TX
+ * and RX separately) some state information for the most recent
+ * packets on that rate and the number of packets for which we have
+ * information. The state is a 32-bit unsigned integer in which two
+ * bits represent a packet: 11 if it went through well, 10 if it went
+ * through with one retry, 01 if it went through with more than one
+ * retry, or 00 if it didn't go through at all. We define the
+ * "goodness" for a particular (rate, direction) combination as the
+ * sum of all the 2-bit fields, times 33, divided by the number of
+ * 2-bit fields containing valid information (16 except when we're
+ * starting out). The number produced is between 0 and 99; we use -1
+ * for rates with less than 4 RX packets or 1 TX, as an indicator that
+ * we do not have enough information to rely on them.
+ *
+ * In deciding which rates are best, we find the weighted average of
+ * TX and RX goodness, where the weighting is by number of packets
+ * with data and TX packets are worth 4 times as much as RX packets.
+ * The weighted average is called "net goodness" and is also a number
+ * between 0 and 99.  If 3 consecutive packets fail transmission
+ * outright, we automatically ratchet down the rate; otherwise, we
+ * switch to the best rate whenever the current rate's goodness falls
+ * below some threshold, and try increasing our rate when the goodness
+ * is very high.
+ *
+ * This system is optimized for gPXE's style of usage. Because normal
+ * operation always involves receiving something, we'll make our way
+ * to the best rate pretty quickly. We tend to follow the lead of the
+ * sending AP in choosing rates, but we won't use rates for long that
+ * don't work well for us in transmission. We assume gPXE won't be
+ * running for long enough that rate patterns will change much, so we
+ * don't have to keep time counters or the like.  And if this doesn't
+ * work well in practice there are many ways it could be tweaked.
+ *
+ * To avoid staying at 1Mbps for a long time, we don't track any
+ * transmitted packets until we've set our rate based on received
+ * packets.
+ */
+
+/** Two-bit packet status indicator for a packet with no retries */
+#define RC_PKT_OK		0x3
+
+/** Two-bit packet status indicator for a packet with one retry */
+#define RC_PKT_RETRIED_ONCE	0x2
+
+/** Two-bit packet status indicator for a TX packet with multiple retries
+ *
+ * It is not possible to tell whether an RX packet had one or multiple
+ * retries; we rely instead on the fact that failed RX packets won't
+ * get to us at all, so if we receive a lot of RX packets on a certain
+ * rate it must be pretty good.
+ */
+#define RC_PKT_RETRIED_MULTI	0x1
+
+/** Two-bit packet status indicator for a TX packet that was never ACKed
+ *
+ * It is not possible to tell whether an RX packet was setn if it
+ * didn't get through to us, but if we don't see one we won't increase
+ * the goodness for its rate. This asymmetry is part of why TX packets
+ * are weighted much more heavily than RX.
+ */
+#define RC_PKT_FAILED		0x0
+
+/** Number of times to weight TX packets more heavily than RX packets */
+#define RC_TX_FACTOR		4
+
+/** Number of consecutive failed TX packets that cause an automatic rate drop */
+#define RC_TX_EMERG_FAIL	3
+
+/** Minimum net goodness below which we will search for a better rate */
+#define RC_GOODNESS_MIN		85
+
+/** Maximum net goodness above which we will try to increase our rate */
+#define RC_GOODNESS_MAX		95
+
+/** Minimum (num RX + @c RC_TX_FACTOR * num TX) to use a certain rate */
+#define RC_UNCERTAINTY_THRESH	4
+
+/** TX direction */
+#define TX	0
+
+/** RX direction */
+#define RX	1
+
+/** A rate control context */
+struct rc80211_ctx
+{
+	/** Goodness state for each rate, TX and RX */
+	u32 goodness[2][NET80211_MAX_RATES];
+
+	/** Number of packets recorded for each rate */
+	u8 count[2][NET80211_MAX_RATES];
+
+	/** Indication of whether we've set the device rate yet */
+	int started;
+
+	/** Counter of all packets sent and received */
+	int packets;
+};
+
+/**
+ * Initialize rate-control algorithm
+ *
+ * @v dev	802.11 device
+ * @ret ctx	Rate-control context, to be stored in @c dev->rctl
+ */
+struct rc80211_ctx * rc80211_init ( struct net80211_device *dev __unused )
+{
+	struct rc80211_ctx *ret = zalloc ( sizeof ( *ret ) );
+	return ret;
+}
+
+/**
+ * Calculate net goodness for a certain rate
+ *
+ * @v ctx	Rate-control context
+ * @v rate_idx	Index of rate to calculate net goodness for
+ */
+static int rc80211_calc_net_goodness ( struct rc80211_ctx *ctx,
+				       int rate_idx )
+{
+	int sum[2], num[2], dir, pkt;
+
+	for ( dir = 0; dir < 2; dir++ ) {
+		u32 good = ctx->goodness[dir][rate_idx];
+
+		num[dir] = ctx->count[dir][rate_idx];
+		sum[dir] = 0;
+
+		for ( pkt = 0; pkt < num[dir]; pkt++ )
+			sum[dir] += ( good >> ( 2 * pkt ) ) & 0x3;
+	}
+
+	if ( ( num[TX] * RC_TX_FACTOR + num[RX] ) < RC_UNCERTAINTY_THRESH )
+		return -1;
+
+	return ( 33 * ( sum[TX] * RC_TX_FACTOR + sum[RX] ) /
+		      ( num[TX] * RC_TX_FACTOR + num[RX] ) );
+}
+
+/**
+ * Determine the best rate to switch to and return it
+ *
+ * @v dev		802.11 device
+ * @ret rate_idx	Index of the best rate to switch to
+ */
+static int rc80211_pick_best ( struct net80211_device *dev )
+{
+	struct rc80211_ctx *ctx = dev->rctl;
+	int best_net_good = 0, best_rate = -1, i;
+
+	for ( i = 0; i < dev->nr_rates; i++ ) {
+		int net_good = rc80211_calc_net_goodness ( ctx, i );
+
+		if ( net_good > best_net_good ||
+		     ( best_net_good > RC_GOODNESS_MIN &&
+		       net_good > RC_GOODNESS_MIN ) ) {
+			best_net_good = net_good;
+			best_rate = i;
+		}
+	}
+
+	if ( best_rate >= 0 ) {
+		int old_good = rc80211_calc_net_goodness ( ctx, dev->rate );
+		if ( old_good != best_net_good )
+			DBGC ( ctx, "802.11 RC %p switching from goodness "
+			       "%d to %d\n", ctx, old_good, best_net_good );
+
+		ctx->started = 1;
+		return best_rate;
+	}
+
+	return dev->rate;
+}
+
+/**
+ * Set 802.11 device rate
+ *
+ * @v dev	802.11 device
+ * @v rate_idx	Index of rate to switch to
+ *
+ * This is a thin wrapper around net80211_set_rate_idx to insert a
+ * debugging message where appropriate.
+ */
+static inline void rc80211_set_rate ( struct net80211_device *dev,
+				      int rate_idx )
+{
+	DBGC ( dev->rctl, "802.11 RC %p changing rate %d->%d Mbps\n", dev->rctl,
+	       dev->rates[dev->rate] / 10, dev->rates[rate_idx] / 10 );
+
+	net80211_set_rate_idx ( dev, rate_idx );
+}
+
+/**
+ * Check rate-control state and change rate if necessary
+ *
+ * @v dev	802.11 device
+ */
+static void rc80211_maybe_set_new ( struct net80211_device *dev )
+{
+	struct rc80211_ctx *ctx = dev->rctl;
+	int net_good;
+
+	net_good = rc80211_calc_net_goodness ( ctx, dev->rate );
+
+	if ( ! ctx->started ) {
+		rc80211_set_rate ( dev, rc80211_pick_best ( dev ) );
+		return;
+	}
+
+	if ( net_good < 0 )	/* insufficient data */
+		return;
+
+	if ( net_good > RC_GOODNESS_MAX && dev->rate + 1 < dev->nr_rates ) {
+		int higher = rc80211_calc_net_goodness ( ctx, dev->rate + 1 );
+		if ( higher > net_good || higher < 0 )
+			rc80211_set_rate ( dev, dev->rate + 1 );
+		else
+			rc80211_set_rate ( dev, rc80211_pick_best ( dev ) );
+	}
+
+	if ( net_good < RC_GOODNESS_MIN ) {
+		rc80211_set_rate ( dev, rc80211_pick_best ( dev ) );
+	}
+}
+
+/**
+ * Update rate-control state
+ *
+ * @v dev		802.11 device
+ * @v direction		One of the direction constants TX or RX
+ * @v rate_idx		Index of rate at which packet was sent or received
+ * @v retries		Number of times packet was retried before success
+ * @v failed		If nonzero, the packet failed to get through
+ */
+static void rc80211_update ( struct net80211_device *dev, int direction,
+			     int rate_idx, int retries, int failed )
+{
+	struct rc80211_ctx *ctx = dev->rctl;
+	u32 goodness = ctx->goodness[direction][rate_idx];
+
+	if ( ctx->count[direction][rate_idx] < 16 )
+		ctx->count[direction][rate_idx]++;
+
+	goodness <<= 2;
+	if ( failed )
+		goodness |= RC_PKT_FAILED;
+	else if ( retries > 1 )
+		goodness |= RC_PKT_RETRIED_MULTI;
+	else if ( retries )
+		goodness |= RC_PKT_RETRIED_ONCE;
+	else
+		goodness |= RC_PKT_OK;
+
+	ctx->goodness[direction][rate_idx] = goodness;
+
+	ctx->packets++;
+
+	rc80211_maybe_set_new ( dev );
+}
+
+/**
+ * Update rate-control state for transmitted packet
+ *
+ * @v dev	802.11 device
+ * @v retries	Number of times packet was transmitted before success
+ * @v rc	Return status code for transmission
+ */
+void rc80211_update_tx ( struct net80211_device *dev, int retries, int rc )
+{
+	struct rc80211_ctx *ctx = dev->rctl;
+
+	if ( ! ctx->started )
+		return;
+
+	rc80211_update ( dev, TX, dev->rate, retries, rc );
+
+	/* Check if the last RC_TX_EMERG_FAIL packets have all failed */
+	if ( ! ( ctx->goodness[TX][dev->rate] &
+		 ( ( 1 << ( 2 * RC_TX_EMERG_FAIL ) ) - 1 ) ) ) {
+		if ( dev->rate == 0 )
+			DBGC ( dev->rctl, "802.11 RC %p saw %d consecutive "
+			       "failed TX, but cannot lower rate any further\n",
+			       dev->rctl, RC_TX_EMERG_FAIL );
+		else {
+			DBGC ( dev->rctl, "802.11 RC %p lowering rate (%d->%d "
+			       "Mbps) due to %d consecutive TX failures\n",
+			       dev->rctl, dev->rates[dev->rate] / 10,
+			       dev->rates[dev->rate - 1] / 10,
+			       RC_TX_EMERG_FAIL );
+
+			rc80211_set_rate ( dev, dev->rate - 1 );
+		}
+	}
+}
+
+/**
+ * Update rate-control state for received packet
+ *
+ * @v dev	802.11 device
+ * @v retry	Whether the received packet had been retransmitted
+ * @v rate	Rate at which packet was received, in 100 kbps units
+ */
+void rc80211_update_rx ( struct net80211_device *dev, int retry, u16 rate )
+{
+	int ridx;
+
+	for ( ridx = 0; ridx < dev->nr_rates && dev->rates[ridx] != rate;
+	      ridx++ )
+		;
+	if ( ridx >= dev->nr_rates )
+		return;		/* couldn't find the rate */
+
+	rc80211_update ( dev, RX, ridx, retry, 0 );
+}
+
+/**
+ * Free rate-control context
+ *
+ * @v ctx	Rate-control context
+ */
+void rc80211_free ( struct rc80211_ctx *ctx )
+{
+	free ( ctx );
+}
diff --git a/gpxe/src/net/80211/sec80211.c b/gpxe/src/net/80211/sec80211.c
new file mode 100644
index 00000000..c5aa1183
--- /dev/null
+++ b/gpxe/src/net/80211/sec80211.c
@@ -0,0 +1,503 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <gpxe/ieee80211.h>
+#include <gpxe/net80211.h>
+#include <gpxe/sec80211.h>
+
+/** @file
+ *
+ * General secured-network routines required whenever any secure
+ * network support at all is compiled in. This involves things like
+ * installing keys, determining the type of security used by a probed
+ * network, and some small helper functions that take advantage of
+ * static data in this file.
+ */
+
+/** Mapping from net80211 crypto/secprot types to RSN OUI descriptors */
+struct descriptor_map {
+	/** Value of net80211_crypto_alg or net80211_security_proto */
+	u32 net80211_type;
+
+	/** OUI+type in appropriate byte order, masked to exclude vendor */
+	u32 oui_type;
+};
+
+/** Magic number in @a oui_type showing end of list */
+#define END_MAGIC	0xFFFFFFFF
+
+/** Mapping between net80211 cryptosystems and 802.11i cipher IDs */
+static struct descriptor_map rsn_cipher_map[] = {
+	{ .net80211_type = NET80211_CRYPT_WEP,
+	  .oui_type = IEEE80211_RSN_CTYPE_WEP40 },
+
+	{ .net80211_type = NET80211_CRYPT_WEP,
+	  .oui_type = IEEE80211_RSN_CTYPE_WEP104 },
+
+	{ .net80211_type = NET80211_CRYPT_TKIP,
+	  .oui_type = IEEE80211_RSN_CTYPE_TKIP },
+
+	{ .net80211_type = NET80211_CRYPT_CCMP,
+	  .oui_type = IEEE80211_RSN_CTYPE_CCMP },
+
+	{ .net80211_type = NET80211_CRYPT_UNKNOWN,
+	  .oui_type = END_MAGIC },
+};
+
+/** Mapping between net80211 handshakers and 802.11i AKM IDs */
+static struct descriptor_map rsn_akm_map[] = {
+	{ .net80211_type = NET80211_SECPROT_EAP,
+	  .oui_type = IEEE80211_RSN_ATYPE_8021X },
+
+	{ .net80211_type = NET80211_SECPROT_PSK,
+	  .oui_type = IEEE80211_RSN_ATYPE_PSK },
+
+	{ .net80211_type = NET80211_SECPROT_UNKNOWN,
+	  .oui_type = END_MAGIC },
+};
+
+
+/**
+ * Install 802.11 cryptosystem
+ *
+ * @v which	Pointer to the cryptosystem structure to install in
+ * @v crypt	Cryptosystem ID number
+ * @v key	Encryption key to use
+ * @v len	Length of encryption key
+ * @v rsc	Initial receive sequence counter, if applicable
+ * @ret rc	Return status code
+ *
+ * The encryption key will not be accessed via the provided pointer
+ * after this function returns, so you may keep it on the stack.
+ *
+ * @a which must point to either @c dev->crypto (for the normal case
+ * of installing a unicast cryptosystem) or @c dev->gcrypto (to
+ * install a cryptosystem that will be used only for decrypting
+ * group-source frames).
+ */
+int sec80211_install ( struct net80211_crypto **which,
+		       enum net80211_crypto_alg crypt,
+		       const void *key, int len, const void *rsc )
+{
+	struct net80211_crypto *crypto = *which;
+	struct net80211_crypto *tbl_crypto;
+
+	/* Remove old crypto if it exists */
+	free ( *which );
+	*which = NULL;
+
+	if ( crypt == NET80211_CRYPT_NONE ) {
+		DBG ( "802.11-Sec not installing null cryptography\n" );
+		return 0;
+	}
+
+	/* Find cryptosystem to use */
+	for_each_table_entry ( tbl_crypto, NET80211_CRYPTOS ) {
+		if ( tbl_crypto->algorithm == crypt ) {
+			crypto = zalloc ( sizeof ( *crypto ) +
+					  tbl_crypto->priv_len );
+			if ( ! crypto ) {
+				DBG ( "802.11-Sec out of memory\n" );
+				return -ENOMEM;
+			}
+
+			memcpy ( crypto, tbl_crypto, sizeof ( *crypto ) );
+			crypto->priv = ( ( void * ) crypto +
+					 sizeof ( *crypto ) );
+			break;
+		}
+	}
+
+	if ( ! crypto ) {
+		DBG ( "802.11-Sec no support for cryptosystem %d\n", crypt );
+		return -( ENOTSUP | EUNIQ_10 | ( crypt << 8 ) );
+	}
+
+	*which = crypto;
+
+	DBG ( "802.11-Sec installing cryptosystem %d as %p with key of "
+	      "length %d\n", crypt, crypto, len );
+
+	return crypto->init ( crypto, key, len, rsc );
+}
+
+
+/**
+ * Determine net80211 crypto or handshaking type value to return for RSN info
+ *
+ * @v rsnp		Pointer to next descriptor count field in RSN IE
+ * @v rsn_end		Pointer to end of RSN IE
+ * @v map		Descriptor map to use
+ * @v tbl_start		Start of linker table to examine for gPXE support
+ * @v tbl_end		End of linker table to examine for gPXE support
+ * @ret rsnp		Updated to point to first byte after descriptors
+ * @ret map_ent		Descriptor map entry of translation to use
+ *
+ * The entries in the linker table must be either net80211_crypto or
+ * net80211_handshaker structures, and @a tbl_stride must be set to
+ * sizeof() the appropriate one.
+ *
+ * This function expects @a rsnp to point at a two-byte descriptor
+ * count followed by a list of four-byte cipher or AKM descriptors; it
+ * will return @c NULL if the input packet is malformed, and otherwise
+ * set @a rsnp to the first byte it has not looked at. It will return
+ * the first cipher in the list that is supported by the current build
+ * of gPXE, or the first of all if none are supported.
+ *
+ * We play rather fast and loose with type checking, because this
+ * function is only called from two well-defined places in the
+ * RSN-checking code. Don't try to use it for anything else.
+ */
+static struct descriptor_map * rsn_pick_desc ( u8 **rsnp, u8 *rsn_end,
+					       struct descriptor_map *map,
+					       void *tbl_start, void *tbl_end )
+{
+	int ndesc;
+	int ok = 0;
+	struct descriptor_map *map_ent, *map_ret = NULL;
+	u8 *rsn = *rsnp;
+	void *tblp;
+	size_t tbl_stride = ( map == rsn_cipher_map ?
+			      sizeof ( struct net80211_crypto ) :
+			      sizeof ( struct net80211_handshaker ) );
+
+	if ( map != rsn_cipher_map && map != rsn_akm_map )
+		return NULL;
+
+	/* Determine which types we support */
+	for ( tblp = tbl_start; tblp < tbl_end; tblp += tbl_stride ) {
+		struct net80211_crypto *crypto = tblp;
+		struct net80211_handshaker *hs = tblp;
+
+		if ( map == rsn_cipher_map )
+			ok |= ( 1 << crypto->algorithm );
+		else
+			ok |= ( 1 << hs->protocol );
+	}
+
+	/* RSN sanity checks */
+	if ( rsn + 2 > rsn_end ) {
+		DBG ( "RSN detect: malformed descriptor count\n" );
+		return NULL;
+	}
+
+	ndesc = *( u16 * ) rsn;
+	rsn += 2;
+
+	if ( ! ndesc ) {
+		DBG ( "RSN detect: no descriptors\n" );
+		return NULL;
+	}
+
+	/* Determine which net80211 crypto types are listed */
+	while ( ndesc-- ) {
+		u32 desc;
+
+		if ( rsn + 4 > rsn_end ) {
+			DBG ( "RSN detect: malformed descriptor (%d left)\n",
+			      ndesc );
+			return NULL;
+		}
+
+		desc = *( u32 * ) rsn;
+		rsn += 4;
+
+		for ( map_ent = map; map_ent->oui_type != END_MAGIC; map_ent++ )
+			if ( map_ent->oui_type == ( desc & OUI_TYPE_MASK ) )
+				break;
+
+		/* Use first cipher as a fallback */
+		if ( ! map_ret )
+			map_ret = map_ent;
+
+		/* Once we find one we support, use it */
+		if ( ok & ( 1 << map_ent->net80211_type ) ) {
+			map_ret = map_ent;
+			break;
+		}
+	}
+
+	if ( ndesc > 0 )
+		rsn += 4 * ndesc;
+
+	*rsnp = rsn;
+	return map_ret;
+}
+
+
+/**
+ * Find the RSN or WPA information element in the provided beacon frame
+ *
+ * @v ie	Pointer to first information element to check
+ * @v ie_end	Pointer to end of information element space
+ * @ret is_rsn	TRUE if returned IE is RSN, FALSE if it's WPA
+ * @ret end	Pointer to byte immediately after last byte of data
+ * @ret data	Pointer to first byte of data (the `version' field)
+ *
+ * If both an RSN and a WPA information element are found, this
+ * function will return the first one seen, which by ordering rules
+ * should always prefer the newer RSN IE.
+ *
+ * If no RSN or WPA infomration element is found, returns @c NULL and
+ * leaves @a is_rsn and @a end in an undefined state.
+ *
+ * This function will not return a pointer to an information element
+ * that states it extends past the tail of the io_buffer, or whose @a
+ * version field is incorrect.
+ */
+u8 * sec80211_find_rsn ( union ieee80211_ie *ie, void *ie_end,
+			 int *is_rsn, u8 **end )
+{
+	u8 *rsn = NULL;
+
+	if ( ! ieee80211_ie_bound ( ie, ie_end ) )
+		return NULL;
+
+	while ( ie ) {
+		if ( ie->id == IEEE80211_IE_VENDOR &&
+		     ie->vendor.oui == IEEE80211_WPA_OUI_VEN ) {
+			DBG ( "RSN detect: old-style WPA IE found\n" );
+			rsn = &ie->vendor.data[0];
+			*end = rsn + ie->len - 4;
+			*is_rsn = 0;
+		} else if ( ie->id == IEEE80211_IE_RSN ) {
+			DBG ( "RSN detect: 802.11i RSN IE found\n" );
+			rsn = ( u8 * ) &ie->rsn.version;
+			*end = rsn + ie->len;
+			*is_rsn = 1;
+		}
+
+		if ( rsn && ( *end > ( u8 * ) ie_end || rsn >= *end ||
+			      *( u16 * ) rsn != IEEE80211_RSN_VERSION ) ) {
+			DBG ( "RSN detect: malformed RSN IE or unknown "
+			      "version, keep trying\n" );
+			rsn = NULL;
+		}
+
+		if ( rsn )
+			break;
+
+		ie = ieee80211_next_ie ( ie, ie_end );
+	}
+
+	if ( ! ie ) {
+		DBG ( "RSN detect: no RSN IE found\n" );
+		return NULL;
+	}
+
+	return rsn;
+}
+
+
+/**
+ * Detect crypto and AKM types from RSN information element
+ *
+ * @v is_rsn	If TRUE, IE is a new-style RSN information element
+ * @v start	Pointer to first byte of @a version field
+ * @v end	Pointer to first byte not in the RSN IE
+ * @ret secprot	Security handshaking protocol used by network
+ * @ret crypt	Cryptosystem used by network
+ * @ret rc	Return status code
+ *
+ * If the IE cannot be parsed, returns an error indication and leaves
+ * @a secprot and @a crypt unchanged.
+ */
+int sec80211_detect_ie ( int is_rsn, u8 *start, u8 *end,
+			 enum net80211_security_proto *secprot,
+			 enum net80211_crypto_alg *crypt )
+{
+	enum net80211_security_proto sp;
+	enum net80211_crypto_alg cr;
+	struct descriptor_map *map;
+	u8 *rsn = start;
+
+	/* Set some defaults */
+	cr = ( is_rsn ? NET80211_CRYPT_CCMP : NET80211_CRYPT_TKIP );
+	sp = NET80211_SECPROT_EAP;
+
+	rsn += 2;		/* version - already checked */
+	rsn += 4;		/* group cipher - we don't use it here */
+
+	if ( rsn >= end )
+		goto done;
+
+	/* Pick crypto algorithm */
+	map = rsn_pick_desc ( &rsn, end, rsn_cipher_map,
+			      table_start ( NET80211_CRYPTOS ),
+			      table_end ( NET80211_CRYPTOS ) );
+	if ( ! map )
+		goto invalid_rsn;
+
+	cr = map->net80211_type;
+
+	if ( rsn >= end )
+		goto done;
+
+	/* Pick handshaking algorithm */
+	map = rsn_pick_desc ( &rsn, end, rsn_akm_map,
+			      table_start ( NET80211_HANDSHAKERS ),
+			      table_end ( NET80211_HANDSHAKERS ) );
+	if ( ! map )
+		goto invalid_rsn;
+
+	sp = map->net80211_type;
+
+ done:
+	DBG ( "RSN detect: OK, crypto type %d, secprot type %d\n", cr, sp );
+	*secprot = sp;
+	*crypt = cr;
+	return 0;
+
+ invalid_rsn:
+	DBG ( "RSN detect: invalid RSN IE\n" );
+	return -EINVAL;
+}
+
+
+/**
+ * Detect the cryptosystem and handshaking protocol used by an 802.11 network
+ *
+ * @v iob	I/O buffer containing beacon frame
+ * @ret secprot	Security handshaking protocol used by network
+ * @ret crypt	Cryptosystem used by network
+ * @ret rc	Return status code
+ *
+ * This function uses weak linkage, as it must be called from generic
+ * contexts but should only be linked in if some encryption is
+ * supported; you must test its address against @c NULL before calling
+ * it. If it does not exist, any network with the PRIVACY bit set in
+ * beacon->capab should be considered unknown.
+ */
+int _sec80211_detect ( struct io_buffer *iob,
+		       enum net80211_security_proto *secprot,
+		       enum net80211_crypto_alg *crypt )
+{
+	struct ieee80211_frame *hdr = iob->data;
+	struct ieee80211_beacon *beacon =
+		( struct ieee80211_beacon * ) hdr->data;
+	u8 *rsn, *rsn_end;
+	int is_rsn, rc;
+
+	*crypt = NET80211_CRYPT_UNKNOWN;
+	*secprot = NET80211_SECPROT_UNKNOWN;
+
+	/* Find RSN or WPA IE */
+	if ( ! ( rsn = sec80211_find_rsn ( beacon->info_element, iob->tail,
+					   &is_rsn, &rsn_end ) ) ) {
+		/* No security IE at all; either WEP or no security. */
+		*secprot = NET80211_SECPROT_NONE;
+
+		if ( beacon->capability & IEEE80211_CAPAB_PRIVACY )
+			*crypt = NET80211_CRYPT_WEP;
+		else
+			*crypt = NET80211_CRYPT_NONE;
+
+		return 0;
+	}
+
+	/* Determine type of security */
+	if ( ( rc = sec80211_detect_ie ( is_rsn, rsn, rsn_end, secprot,
+					 crypt ) ) == 0 )
+		return 0;
+
+	/* If we get here, the RSN IE was invalid */
+
+	*crypt = NET80211_CRYPT_UNKNOWN;
+	*secprot = NET80211_SECPROT_UNKNOWN;
+	DBG ( "Failed to handle RSN IE:\n" );
+	DBG_HD ( rsn, rsn_end - rsn );
+	return rc;
+}
+
+
+/**
+ * Determine RSN descriptor for specified net80211 ID
+ *
+ * @v id	net80211 ID value
+ * @v rsnie	Whether to return a new-format (RSN IE) descriptor
+ * @v map	Map to use in translation
+ * @ret desc	RSN descriptor, or 0 on error
+ *
+ * If @a rsnie is false, returns an old-format (WPA vendor IE)
+ * descriptor.
+ */
+static u32 rsn_get_desc ( unsigned id, int rsnie, struct descriptor_map *map )
+{
+	u32 vendor = ( rsnie ? IEEE80211_RSN_OUI : IEEE80211_WPA_OUI );
+
+	for ( ; map->oui_type != END_MAGIC; map++ ) {
+		if ( map->net80211_type == id )
+			return map->oui_type | vendor;
+	}
+
+	return 0;
+}
+
+/**
+ * Determine RSN descriptor for specified net80211 cryptosystem number
+ *
+ * @v crypt	Cryptosystem number
+ * @v rsnie	Whether to return a new-format (RSN IE) descriptor
+ * @ret desc	RSN descriptor
+ *
+ * If @a rsnie is false, returns an old-format (WPA vendor IE)
+ * descriptor.
+ */
+u32 sec80211_rsn_get_crypto_desc ( enum net80211_crypto_alg crypt, int rsnie )
+{
+	return rsn_get_desc ( crypt, rsnie, rsn_cipher_map );
+}
+
+/**
+ * Determine RSN descriptor for specified net80211 handshaker number
+ *
+ * @v secprot	Handshaker number
+ * @v rsnie	Whether to return a new-format (RSN IE) descriptor
+ * @ret desc	RSN descriptor
+ *
+ * If @a rsnie is false, returns an old-format (WPA vendor IE)
+ * descriptor.
+ */
+u32 sec80211_rsn_get_akm_desc ( enum net80211_security_proto secprot,
+				int rsnie )
+{
+	return rsn_get_desc ( secprot, rsnie, rsn_akm_map );
+}
+
+/**
+ * Determine net80211 cryptosystem number from RSN descriptor
+ *
+ * @v desc	RSN descriptor
+ * @ret crypt	net80211 cryptosystem enumeration value
+ */
+enum net80211_crypto_alg sec80211_rsn_get_net80211_crypt ( u32 desc )
+{
+	struct descriptor_map *map = rsn_cipher_map;
+
+	for ( ; map->oui_type != END_MAGIC; map++ ) {
+		if ( map->oui_type == ( desc & OUI_TYPE_MASK ) )
+			break;
+	}
+
+	return map->net80211_type;
+}
diff --git a/gpxe/src/net/80211/wep.c b/gpxe/src/net/80211/wep.c
new file mode 100644
index 00000000..1c37e0c3
--- /dev/null
+++ b/gpxe/src/net/80211/wep.c
@@ -0,0 +1,303 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <gpxe/net80211.h>
+#include <gpxe/sec80211.h>
+#include <gpxe/crypto.h>
+#include <gpxe/arc4.h>
+#include <gpxe/crc32.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+
+/** @file
+ *
+ * The WEP wireless encryption method (insecure!)
+ *
+ * The data field in a WEP-encrypted packet contains a 3-byte
+ * initialisation vector, one-byte Key ID field (only the bottom two
+ * bits are ever used), encrypted data, and a 4-byte encrypted CRC of
+ * the plaintext data, called the ICV. To decrypt it, the IV is
+ * prepended to the shared key and the data stream (including ICV) is
+ * run through the ARC4 stream cipher; if the ICV matches a CRC32
+ * calculated on the plaintext, the packet is valid.
+ *
+ * For efficiency and code-size reasons, this file assumes it is
+ * running on a little-endian machine.
+ */
+
+/** Length of WEP initialisation vector */
+#define WEP_IV_LEN	3
+
+/** Length of WEP key ID byte */
+#define WEP_KID_LEN	1
+
+/** Length of WEP ICV checksum */
+#define WEP_ICV_LEN	4
+
+/** Maximum length of WEP key */
+#define WEP_MAX_KEY	16
+
+/** Amount of data placed before the encrypted bytes */
+#define WEP_HEADER_LEN	4
+
+/** Amount of data placed after the encrypted bytes */
+#define WEP_TRAILER_LEN	4
+
+/** Total WEP overhead bytes */
+#define WEP_OVERHEAD	8
+
+/** Context for WEP encryption and decryption */
+struct wep_ctx
+{
+	/** Encoded WEP key
+	 *
+	 * The actual key bytes are stored beginning at offset 3, to
+	 * leave room for easily inserting the IV before a particular
+	 * operation.
+	 */
+	u8 key[WEP_IV_LEN + WEP_MAX_KEY];
+
+	/** Length of WEP key (not including IV bytes) */
+	int keylen;
+
+	/** ARC4 context */
+	struct arc4_ctx arc4;
+};
+
+/**
+ * Initialize WEP algorithm
+ *
+ * @v crypto	802.11 cryptographic algorithm
+ * @v key	WEP key to use
+ * @v keylen	Length of WEP key
+ * @v rsc	Initial receive sequence counter (unused)
+ * @ret rc	Return status code
+ *
+ * Standard key lengths are 5 and 13 bytes; 16-byte keys are
+ * occasionally supported as an extension to the standard.
+ */
+static int wep_init ( struct net80211_crypto *crypto, const void *key,
+		      int keylen, const void *rsc __unused )
+{
+	struct wep_ctx *ctx = crypto->priv;
+
+	ctx->keylen = ( keylen > WEP_MAX_KEY ? WEP_MAX_KEY : keylen );
+	memcpy ( ctx->key + WEP_IV_LEN, key, ctx->keylen );
+
+	return 0;
+}
+
+/**
+ * Encrypt packet using WEP
+ *
+ * @v crypto	802.11 cryptographic algorithm
+ * @v iob	I/O buffer of plaintext packet
+ * @ret eiob	Newly allocated I/O buffer for encrypted packet, or NULL
+ *
+ * If memory allocation fails, @c NULL is returned.
+ */
+static struct io_buffer * wep_encrypt ( struct net80211_crypto *crypto,
+					struct io_buffer *iob )
+{
+	struct wep_ctx *ctx = crypto->priv;
+	struct io_buffer *eiob;
+	struct ieee80211_frame *hdr;
+	const int hdrlen = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int datalen = iob_len ( iob ) - hdrlen;
+	int newlen = hdrlen + datalen + WEP_OVERHEAD;
+	u32 iv, icv;
+
+	eiob = alloc_iob ( newlen );
+	if ( ! eiob )
+		return NULL;
+
+	memcpy ( iob_put ( eiob, hdrlen ), iob->data, hdrlen );
+	hdr = eiob->data;
+	hdr->fc |= IEEE80211_FC_PROTECTED;
+
+	/* Calculate IV, put it in the header (with key ID byte = 0), and
+	   set it up at the start of the encryption key. */
+	iv = random() & 0xffffff; /* IV in bottom 3 bytes, top byte = KID = 0 */
+	memcpy ( iob_put ( eiob, WEP_HEADER_LEN ), &iv, WEP_HEADER_LEN );
+	memcpy ( ctx->key, &iv, WEP_IV_LEN );
+
+	/* Encrypt the data using RC4 */
+	cipher_setkey ( &arc4_algorithm, &ctx->arc4, ctx->key,
+			ctx->keylen + WEP_IV_LEN );
+	cipher_encrypt ( &arc4_algorithm, &ctx->arc4, iob->data + hdrlen,
+			 iob_put ( eiob, datalen ), datalen );
+
+	/* Add ICV */
+	icv = ~crc32_le ( ~0, iob->data + hdrlen, datalen );
+	cipher_encrypt ( &arc4_algorithm, &ctx->arc4, &icv,
+			 iob_put ( eiob, WEP_ICV_LEN ), WEP_ICV_LEN );
+
+	return eiob;
+}
+
+/**
+ * Decrypt packet using WEP
+ *
+ * @v crypto	802.11 cryptographic algorithm
+ * @v eiob	I/O buffer of encrypted packet
+ * @ret iob	Newly allocated I/O buffer for plaintext packet, or NULL
+ *
+ * If a consistency check for the decryption fails (usually indicating
+ * an invalid key), @c NULL is returned.
+ */
+static struct io_buffer * wep_decrypt ( struct net80211_crypto *crypto,
+					struct io_buffer *eiob )
+{
+	struct wep_ctx *ctx = crypto->priv;
+	struct io_buffer *iob;
+	struct ieee80211_frame *hdr;
+	const int hdrlen = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int datalen = iob_len ( eiob ) - hdrlen - WEP_OVERHEAD;
+	int newlen = hdrlen + datalen;
+	u32 iv, icv, crc;
+
+	iob = alloc_iob ( newlen );
+	if ( ! iob )
+		return NULL;
+
+	memcpy ( iob_put ( iob, hdrlen ), eiob->data, hdrlen );
+	hdr = iob->data;
+	hdr->fc &= ~IEEE80211_FC_PROTECTED;
+
+	/* Strip off IV and use it to initialize cryptosystem */
+	memcpy ( &iv, eiob->data + hdrlen, 4 );
+	iv &= 0xffffff;		/* ignore key ID byte */
+	memcpy ( ctx->key, &iv, WEP_IV_LEN );
+
+	/* Decrypt the data using RC4 */
+	cipher_setkey ( &arc4_algorithm, &ctx->arc4, ctx->key,
+			ctx->keylen + WEP_IV_LEN );
+	cipher_decrypt ( &arc4_algorithm, &ctx->arc4, eiob->data + hdrlen +
+			 WEP_HEADER_LEN, iob_put ( iob, datalen ), datalen );
+
+	/* Strip off ICV and verify it */
+	cipher_decrypt ( &arc4_algorithm, &ctx->arc4, eiob->data + hdrlen +
+			 WEP_HEADER_LEN + datalen, &icv, WEP_ICV_LEN );
+	crc = ~crc32_le ( ~0, iob->data + hdrlen, datalen );
+	if ( crc != icv ) {
+		DBGC ( crypto, "WEP %p CRC mismatch: expect %08x, get %08x\n",
+		       crypto, icv, crc );
+		free_iob ( iob );
+		return NULL;
+	}
+	return iob;
+}
+
+/** WEP cryptosystem for 802.11 */
+struct net80211_crypto wep_crypto __net80211_crypto = {
+	.algorithm = NET80211_CRYPT_WEP,
+	.init = wep_init,
+	.encrypt = wep_encrypt,
+	.decrypt = wep_decrypt,
+	.priv_len = sizeof ( struct wep_ctx ),
+};
+
+/**
+ * Initialize trivial 802.11 security handshaker
+ *
+ * @v dev	802.11 device
+ * @v ctx	Security handshaker
+ *
+ * This simply fetches a WEP key from netX/key, and if it exists,
+ * installs WEP cryptography on the 802.11 device. No real handshaking
+ * is performed.
+ */
+static int trivial_init ( struct net80211_device *dev )
+{
+	u8 key[WEP_MAX_KEY];	/* support up to 128-bit keys */
+	int len;
+	int rc;
+
+	if ( dev->associating &&
+	     dev->associating->crypto == NET80211_CRYPT_NONE )
+		return 0;	/* no crypto? OK. */
+
+	len = fetch_setting ( netdev_settings ( dev->netdev ),
+			      &net80211_key_setting, key, WEP_MAX_KEY );
+
+	if ( len <= 0 ) {
+		DBGC ( dev, "802.11 %p cannot do WEP without a key\n", dev );
+		return -EACCES;
+	}
+
+	/* Full 128-bit keys are a nonstandard extension, but they're
+	   utterly trivial to support, so we do. */
+	if ( len != 5 && len != 13 && len != 16 ) {
+		DBGC ( dev, "802.11 %p invalid WEP key length %d\n",
+		       dev, len );
+		return -EINVAL;
+	}
+
+	DBGC ( dev, "802.11 %p installing %d-bit WEP\n", dev, len * 8 );
+
+	rc = sec80211_install ( &dev->crypto, NET80211_CRYPT_WEP, key, len,
+				NULL );
+	if ( rc < 0 )
+		return rc;
+
+	return 0;
+}
+
+/**
+ * Check for key change on trivial 802.11 security handshaker
+ *
+ * @v dev	802.11 device
+ * @v ctx	Security handshaker
+ */
+static int trivial_change_key ( struct net80211_device *dev )
+{
+	u8 key[WEP_MAX_KEY];
+	int len;
+	int change = 0;
+
+	/* If going from WEP to clear, or something else to WEP, reassociate. */
+	if ( ! dev->crypto || ( dev->crypto->init != wep_init ) )
+		change ^= 1;
+
+	len = fetch_setting ( netdev_settings ( dev->netdev ),
+			      &net80211_key_setting, key, WEP_MAX_KEY );
+	if ( len <= 0 )
+		change ^= 1;
+
+	/* Changing crypto type => return nonzero to reassociate. */
+	if ( change )
+		return -EINVAL;
+
+	/* Going from no crypto to still no crypto => nothing to do. */
+	if ( len <= 0 )
+		return 0;
+
+	/* Otherwise, reinitialise WEP with new key. */
+	return wep_init ( dev->crypto, key, len, NULL );
+}
+
+/** Trivial 802.11 security handshaker */
+struct net80211_handshaker trivial_handshaker __net80211_handshaker = {
+	.protocol = NET80211_SECPROT_NONE,
+	.init = trivial_init,
+	.change_key = trivial_change_key,
+	.priv_len = 0,
+};
diff --git a/gpxe/src/net/80211/wpa.c b/gpxe/src/net/80211/wpa.c
new file mode 100644
index 00000000..9bac8fe7
--- /dev/null
+++ b/gpxe/src/net/80211/wpa.c
@@ -0,0 +1,973 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <gpxe/net80211.h>
+#include <gpxe/sec80211.h>
+#include <gpxe/wpa.h>
+#include <gpxe/eapol.h>
+#include <gpxe/crypto.h>
+#include <gpxe/arc4.h>
+#include <gpxe/crc32.h>
+#include <gpxe/sha1.h>
+#include <gpxe/hmac.h>
+#include <gpxe/list.h>
+#include <gpxe/ethernet.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+
+/** @file
+ *
+ * Handler for the aspects of WPA handshaking that are independent of
+ * 802.1X/PSK or TKIP/CCMP; this mostly involves the 4-Way Handshake.
+ */
+
+/** List of WPA contexts in active use. */
+struct list_head wpa_contexts = LIST_HEAD_INIT ( wpa_contexts );
+
+
+/**
+ * Return an error code and deauthenticate
+ *
+ * @v ctx	WPA common context
+ * @v rc	Return status code
+ * @ret rc	The passed return status code
+ */
+static int wpa_fail ( struct wpa_common_ctx *ctx, int rc )
+{
+	net80211_deauthenticate ( ctx->dev, rc );
+	return rc;
+}
+
+
+/**
+ * Find a cryptosystem handler structure from a crypto ID
+ *
+ * @v crypt	Cryptosystem ID
+ * @ret crypto	Cryptosystem handler structure
+ *
+ * If support for @a crypt is not compiled in to gPXE, or if @a crypt
+ * is NET80211_CRYPT_UNKNOWN, returns @c NULL.
+ */
+static struct net80211_crypto *
+wpa_find_cryptosystem ( enum net80211_crypto_alg crypt )
+{
+	struct net80211_crypto *crypto;
+
+	for_each_table_entry ( crypto, NET80211_CRYPTOS ) {
+		if ( crypto->algorithm == crypt )
+			return crypto;
+	}
+
+	return NULL;
+}
+
+
+/**
+ * Find WPA key integrity and encryption handler from key version field
+ *
+ * @v ver	Version bits of EAPOL-Key info field
+ * @ret kie	Key integrity and encryption handler
+ */
+struct wpa_kie * wpa_find_kie ( int version )
+{
+	struct wpa_kie *kie;
+
+	for_each_table_entry ( kie, WPA_KIES ) {
+		if ( kie->version == version )
+			return kie;
+	}
+
+	return NULL;
+}
+
+
+/**
+ * Construct RSN or WPA information element
+ *
+ * @v dev	802.11 device
+ * @ret ie_ret	RSN or WPA information element
+ * @ret rc	Return status code
+ *
+ * This function allocates, fills, and returns a RSN or WPA
+ * information element suitable for including in an association
+ * request frame to the network identified by @c dev->associating.
+ * If it is impossible to construct an information element consistent
+ * with gPXE's capabilities that is compatible with that network, or
+ * if none should be sent because that network's beacon included no
+ * security information, returns an error indication and leaves
+ * @a ie_ret unchanged.
+ *
+ * The returned IE will be of the same type (RSN or WPA) as was
+ * included in the beacon for the network it is destined for.
+ */
+int wpa_make_rsn_ie ( struct net80211_device *dev, union ieee80211_ie **ie_ret )
+{
+	u8 *rsn, *rsn_end;
+	int is_rsn;
+	u32 group_cipher;
+	enum net80211_crypto_alg gcrypt;
+	int ie_len;
+	u8 *iep;
+	struct ieee80211_ie_rsn *ie;
+	struct ieee80211_frame *hdr;
+	struct ieee80211_beacon *beacon;
+
+	if ( ! dev->associating ) {
+		DBG ( "WPA: Can't make RSN IE for a non-associating device\n" );
+		return -EINVAL;
+	}
+
+	hdr = dev->associating->beacon->data;
+	beacon = ( struct ieee80211_beacon * ) hdr->data;
+	rsn = sec80211_find_rsn ( beacon->info_element,
+				  dev->associating->beacon->tail, &is_rsn,
+				  &rsn_end );
+	if ( ! rsn ) {
+		DBG ( "WPA: Can't make RSN IE when we didn't get one\n" );
+		return -EINVAL;
+	}
+
+	rsn += 2;		/* skip version */
+	group_cipher = *( u32 * ) rsn;
+	gcrypt = sec80211_rsn_get_net80211_crypt ( group_cipher );
+
+	if ( ! wpa_find_cryptosystem ( gcrypt ) ||
+	     ! wpa_find_cryptosystem ( dev->associating->crypto ) ) {
+		DBG ( "WPA: No support for (GC:%d, PC:%d)\n",
+		      gcrypt, dev->associating->crypto );
+		return -ENOTSUP;
+	}
+
+	/* Everything looks good - make our IE. */
+
+	/* WPA IEs need 4 more bytes for the OUI+type */
+	ie_len = ieee80211_rsn_size ( 1, 1, 0, is_rsn ) + ( 4 * ! is_rsn );
+	iep = malloc ( ie_len );
+	if ( ! iep )
+		return -ENOMEM;
+
+	*ie_ret = ( union ieee80211_ie * ) iep;
+
+	/* Store ID and length bytes. */
+	*iep++ = ( is_rsn ? IEEE80211_IE_RSN : IEEE80211_IE_VENDOR );
+	*iep++ = ie_len - 2;
+
+	/* Store OUI+type for WPA IEs. */
+	if ( ! is_rsn ) {
+		*( u32 * ) iep = IEEE80211_WPA_OUI_VEN;
+		iep += 4;
+	}
+
+	/* If this is a WPA IE, the id and len bytes in the
+	   ieee80211_ie_rsn structure will not be valid, but by doing
+	   the cast we can fill all the other fields much more
+	   readily. */
+
+	ie = ( struct ieee80211_ie_rsn * ) ( iep - 2 );
+	ie->version = IEEE80211_RSN_VERSION;
+	ie->group_cipher = group_cipher;
+	ie->pairwise_count = 1;
+	ie->pairwise_cipher[0] =
+		sec80211_rsn_get_crypto_desc ( dev->associating->crypto,
+					       is_rsn );
+	ie->akm_count = 1;
+	ie->akm_list[0] =
+		sec80211_rsn_get_akm_desc ( dev->associating->handshaking,
+					    is_rsn );
+	if ( is_rsn ) {
+		ie->rsn_capab = 0;
+		ie->pmkid_count = 0;
+	}
+
+	return 0;
+}
+
+
+/**
+ * Set up generic WPA support to handle 4-Way Handshake
+ *
+ * @v dev	802.11 device
+ * @v ctx	WPA common context
+ * @v pmk	Pairwise Master Key to use for session
+ * @v pmk_len	Length of PMK, almost always 32
+ * @ret rc	Return status code
+ */
+int wpa_start ( struct net80211_device *dev, struct wpa_common_ctx *ctx,
+		const void *pmk, size_t pmk_len )
+{
+	struct io_buffer *iob;
+	struct ieee80211_frame *hdr;
+	struct ieee80211_beacon *beacon;
+	u8 *ap_rsn_ie = NULL, *ap_rsn_ie_end;
+
+	if ( ! dev->rsn_ie || ! dev->associating )
+		return -EINVAL;
+
+	ctx->dev = dev;
+	memcpy ( ctx->pmk, pmk, ctx->pmk_len = pmk_len );
+	ctx->state = WPA_READY;
+	ctx->replay = ~0ULL;
+
+	iob = dev->associating->beacon;
+	hdr = iob->data;
+	beacon = ( struct ieee80211_beacon * ) hdr->data;
+	ap_rsn_ie = sec80211_find_rsn ( beacon->info_element, iob->tail,
+					&ctx->ap_rsn_is_rsn, &ap_rsn_ie_end );
+	if ( ap_rsn_ie ) {
+		ctx->ap_rsn_ie = malloc ( ap_rsn_ie_end - ap_rsn_ie );
+		if ( ! ctx->ap_rsn_ie )
+			return -ENOMEM;
+		memcpy ( ctx->ap_rsn_ie, ap_rsn_ie, ap_rsn_ie_end - ap_rsn_ie );
+		ctx->ap_rsn_ie_len = ap_rsn_ie_end - ap_rsn_ie;
+	} else {
+		return -ENOENT;
+	}
+
+	ctx->crypt = dev->associating->crypto;
+	ctx->gcrypt = NET80211_CRYPT_UNKNOWN;
+
+	list_add_tail ( &ctx->list, &wpa_contexts );
+	return 0;
+}
+
+
+/**
+ * Disable handling of received WPA handshake frames
+ *
+ * @v dev	802.11 device
+ */
+void wpa_stop ( struct net80211_device *dev )
+{
+	struct wpa_common_ctx *ctx, *tmp;
+
+	list_for_each_entry_safe ( ctx, tmp, &wpa_contexts, list ) {
+		if ( ctx->dev == dev ) {
+			free ( ctx->ap_rsn_ie );
+			ctx->ap_rsn_ie = NULL;
+			list_del ( &ctx->list );
+		}
+	}
+}
+
+
+/**
+ * Check PMKID consistency
+ *
+ * @v ctx	WPA common context
+ * @v pmkid	PMKID to check against (16 bytes long)
+ * @ret rc	Zero if they match, or a negative error code if not
+ */
+int wpa_check_pmkid ( struct wpa_common_ctx *ctx, const u8 *pmkid )
+{
+	u8 sha1_ctx[SHA1_CTX_SIZE];
+	u8 my_pmkid[SHA1_SIZE];
+	u8 pmk[ctx->pmk_len];
+	size_t pmk_len;
+	struct {
+		char name[8];
+		u8 aa[ETH_ALEN];
+		u8 spa[ETH_ALEN];
+	} __attribute__ (( packed )) pmkid_data;
+
+	memcpy ( pmk, ctx->pmk, ctx->pmk_len );
+	pmk_len = ctx->pmk_len;
+
+	memcpy ( pmkid_data.name, "PMK Name", 8 );
+	memcpy ( pmkid_data.aa, ctx->dev->bssid, ETH_ALEN );
+	memcpy ( pmkid_data.spa, ctx->dev->netdev->ll_addr, ETH_ALEN );
+
+	hmac_init ( &sha1_algorithm, sha1_ctx, pmk, &pmk_len );
+	hmac_update ( &sha1_algorithm, sha1_ctx, &pmkid_data,
+		      sizeof ( pmkid_data ) );
+	hmac_final ( &sha1_algorithm, sha1_ctx, pmk, &pmk_len, my_pmkid );
+
+	if ( memcmp ( my_pmkid, pmkid, WPA_PMKID_LEN ) != 0 )
+		return -EACCES;
+
+	return 0;
+}
+
+
+/**
+ * Derive pairwise transient key
+ *
+ * @v ctx	WPA common context
+ */
+static void wpa_derive_ptk ( struct wpa_common_ctx *ctx )
+{
+	struct {
+		u8 mac1[ETH_ALEN];
+		u8 mac2[ETH_ALEN];
+		u8 nonce1[WPA_NONCE_LEN];
+		u8 nonce2[WPA_NONCE_LEN];
+	} __attribute__ (( packed )) ptk_data;
+
+	/* The addresses and nonces are stored in numerical order (!) */
+
+	if ( memcmp ( ctx->dev->netdev->ll_addr, ctx->dev->bssid,
+		      ETH_ALEN ) < 0 ) {
+		memcpy ( ptk_data.mac1, ctx->dev->netdev->ll_addr, ETH_ALEN );
+		memcpy ( ptk_data.mac2, ctx->dev->bssid, ETH_ALEN );
+	} else {
+		memcpy ( ptk_data.mac1, ctx->dev->bssid, ETH_ALEN );
+		memcpy ( ptk_data.mac2, ctx->dev->netdev->ll_addr, ETH_ALEN );
+	}
+
+	if ( memcmp ( ctx->Anonce, ctx->Snonce, WPA_NONCE_LEN ) < 0 ) {
+		memcpy ( ptk_data.nonce1, ctx->Anonce, WPA_NONCE_LEN );
+		memcpy ( ptk_data.nonce2, ctx->Snonce, WPA_NONCE_LEN );
+	} else {
+		memcpy ( ptk_data.nonce1, ctx->Snonce, WPA_NONCE_LEN );
+		memcpy ( ptk_data.nonce2, ctx->Anonce, WPA_NONCE_LEN );
+	}
+
+	DBGC2 ( ctx, "WPA %p A1 %s, A2 %s\n", ctx, eth_ntoa ( ptk_data.mac1 ),
+	       eth_ntoa ( ptk_data.mac2 ) );
+	DBGC2 ( ctx, "WPA %p Nonce1, Nonce2:\n", ctx );
+	DBGC2_HD ( ctx, ptk_data.nonce1, WPA_NONCE_LEN );
+	DBGC2_HD ( ctx, ptk_data.nonce2, WPA_NONCE_LEN );
+
+	prf_sha1 ( ctx->pmk, ctx->pmk_len,
+		   "Pairwise key expansion",
+		   &ptk_data, sizeof ( ptk_data ),
+		   &ctx->ptk, sizeof ( ctx->ptk ) );
+
+	DBGC2 ( ctx, "WPA %p PTK:\n", ctx );
+	DBGC2_HD ( ctx, &ctx->ptk, sizeof ( ctx->ptk ) );
+}
+
+
+/**
+ * Install pairwise transient key
+ *
+ * @v ctx	WPA common context
+ * @v len	Key length (16 for CCMP, 32 for TKIP)
+ * @ret rc	Return status code
+ */
+static inline int wpa_install_ptk ( struct wpa_common_ctx *ctx, int len )
+{
+	DBGC ( ctx, "WPA %p: installing %d-byte pairwise transient key\n",
+	       ctx, len );
+	DBGC2_HD ( ctx, &ctx->ptk.tk, len );
+
+	return sec80211_install ( &ctx->dev->crypto, ctx->crypt,
+				  &ctx->ptk.tk, len, NULL );
+}
+
+/**
+ * Install group transient key
+ *
+ * @v ctx	WPA common context
+ * @v len	Key length (16 for CCMP, 32 for TKIP)
+ * @v rsc	Receive sequence counter field in EAPOL-Key packet
+ * @ret rc	Return status code
+ */
+static inline int wpa_install_gtk ( struct wpa_common_ctx *ctx, int len,
+				    const void *rsc )
+{
+	DBGC ( ctx, "WPA %p: installing %d-byte group transient key\n",
+	       ctx, len );
+	DBGC2_HD ( ctx, &ctx->gtk.tk, len );
+
+	return sec80211_install ( &ctx->dev->gcrypto, ctx->gcrypt,
+				  &ctx->gtk.tk, len, rsc );
+}
+
+/**
+ * Search for group transient key, and install it if found
+ *
+ * @v ctx	WPA common context
+ * @v ie	Pointer to first IE in key data field
+ * @v ie_end	Pointer to first byte not in key data field
+ * @v rsc	Receive sequence counter field in EAPOL-Key packet
+ * @ret rc	Return status code
+ */
+static int wpa_maybe_install_gtk ( struct wpa_common_ctx *ctx,
+				   union ieee80211_ie *ie, void *ie_end,
+				   const void *rsc )
+{
+	struct wpa_kde *kde;
+
+	if ( ! ieee80211_ie_bound ( ie, ie_end ) )
+		return -ENOENT;
+
+	while ( ie ) {
+		if ( ie->id == IEEE80211_IE_VENDOR &&
+		     ie->vendor.oui == WPA_KDE_GTK )
+			break;
+
+		ie = ieee80211_next_ie ( ie, ie_end );
+	}
+
+	if ( ! ie )
+		return -ENOENT;
+
+	if ( ie->len - 6u > sizeof ( ctx->gtk.tk ) ) {
+		DBGC ( ctx, "WPA %p: GTK KDE is too long (%d bytes, max %d)\n",
+		       ctx, ie->len - 4, sizeof ( ctx->gtk.tk ) );
+		return -EINVAL;
+	}
+
+	/* XXX We ignore key ID for now. */
+	kde = ( struct wpa_kde * ) ie;
+	memcpy ( &ctx->gtk.tk, &kde->gtk_encap.gtk, kde->len - 6 );
+
+	return wpa_install_gtk ( ctx, kde->len - 6, rsc );
+}
+
+
+/**
+ * Allocate I/O buffer for construction of outgoing EAPOL-Key frame
+ *
+ * @v kdlen	Maximum number of bytes in the Key Data field
+ * @ret iob	Newly allocated I/O buffer
+ *
+ * The returned buffer will have space reserved for the link-layer and
+ * EAPOL headers, and will have @c iob->tail pointing to the start of
+ * the Key Data field. Thus, it is necessary to use iob_put() in
+ * filling the Key Data.
+ */
+static struct io_buffer * wpa_alloc_frame ( int kdlen )
+{
+	struct io_buffer *ret = alloc_iob ( sizeof ( struct eapol_key_pkt ) +
+					    kdlen + EAPOL_HDR_LEN +
+					    MAX_LL_HEADER_LEN );
+	if ( ! ret )
+		return NULL;
+
+	iob_reserve ( ret, MAX_LL_HEADER_LEN + EAPOL_HDR_LEN );
+	memset ( iob_put ( ret, sizeof ( struct eapol_key_pkt ) ), 0,
+		 sizeof ( struct eapol_key_pkt ) );
+
+	return ret;
+}
+
+
+/**
+ * Send EAPOL-Key packet
+ *
+ * @v iob	I/O buffer, with sufficient headroom for headers
+ * @v dev	802.11 device
+ * @v kie	Key integrity and encryption handler
+ * @v is_rsn	If TRUE, handshake uses new RSN format
+ * @ret rc	Return status code
+ *
+ * If a KIE is specified, the MIC will be filled in before transmission.
+ */
+static int wpa_send_eapol ( struct io_buffer *iob, struct wpa_common_ctx *ctx,
+			    struct wpa_kie *kie )
+{
+	struct eapol_key_pkt *pkt = iob->data;
+	struct eapol_frame *eapol = iob_push ( iob, EAPOL_HDR_LEN );
+
+	pkt->info = htons ( pkt->info );
+	pkt->keysize = htons ( pkt->keysize );
+	pkt->datalen = htons ( pkt->datalen );
+	pkt->replay = cpu_to_be64 ( pkt->replay );
+	eapol->version = EAPOL_THIS_VERSION;
+	eapol->type = EAPOL_TYPE_KEY;
+	eapol->length = htons ( iob->tail - iob->data - sizeof ( *eapol ) );
+
+	memset ( pkt->mic, 0, sizeof ( pkt->mic ) );
+	if ( kie )
+		kie->mic ( &ctx->ptk.kck, eapol, EAPOL_HDR_LEN +
+			   sizeof ( *pkt ) + ntohs ( pkt->datalen ),
+			   pkt->mic );
+
+	return net_tx ( iob, ctx->dev->netdev, &eapol_protocol,
+			ctx->dev->bssid );
+}
+
+
+/**
+ * Send second frame in 4-Way Handshake
+ *
+ * @v ctx	WPA common context
+ * @v pkt	First frame, to which this is a reply
+ * @v is_rsn	If TRUE, handshake uses new RSN format
+ * @v kie	Key integrity and encryption handler
+ * @ret rc	Return status code
+ */
+static int wpa_send_2_of_4 ( struct wpa_common_ctx *ctx,
+			     struct eapol_key_pkt *pkt, int is_rsn,
+			     struct wpa_kie *kie )
+{
+	struct io_buffer *iob = wpa_alloc_frame ( ctx->dev->rsn_ie->len + 2 );
+	struct eapol_key_pkt *npkt;
+
+	if ( ! iob )
+		return -ENOMEM;
+
+	npkt = iob->data;
+	memcpy ( npkt, pkt, sizeof ( *pkt ) );
+	npkt->info &= ~EAPOL_KEY_INFO_KEY_ACK;
+	npkt->info |= EAPOL_KEY_INFO_KEY_MIC;
+	if ( is_rsn )
+		npkt->keysize = 0;
+	memcpy ( npkt->nonce, ctx->Snonce, sizeof ( npkt->nonce ) );
+	npkt->datalen = ctx->dev->rsn_ie->len + 2;
+	memcpy ( iob_put ( iob, npkt->datalen ), ctx->dev->rsn_ie,
+		 npkt->datalen );
+
+	DBGC ( ctx, "WPA %p: sending 2/4\n", ctx );
+
+	return wpa_send_eapol ( iob, ctx, kie );
+}
+
+
+/**
+ * Handle receipt of first frame in 4-Way Handshake
+ *
+ * @v ctx	WPA common context
+ * @v pkt	EAPOL-Key packet
+ * @v is_rsn	If TRUE, frame uses new RSN format
+ * @v kie	Key integrity and encryption handler
+ * @ret rc	Return status code
+ */
+static int wpa_handle_1_of_4 ( struct wpa_common_ctx *ctx,
+			       struct eapol_key_pkt *pkt, int is_rsn,
+			       struct wpa_kie *kie )
+{
+	int rc;
+
+	if ( ctx->state == WPA_WAITING )
+		return -EINVAL;
+
+	ctx->state = WPA_WORKING;
+	memcpy ( ctx->Anonce, pkt->nonce, sizeof ( ctx->Anonce ) );
+	if ( ! ctx->have_Snonce ) {
+		get_random_bytes ( ctx->Snonce, sizeof ( ctx->Snonce ) );
+		ctx->have_Snonce = 1;
+	}
+
+	if ( is_rsn && pkt->datalen ) {
+		union ieee80211_ie *ie = ( union ieee80211_ie * ) pkt->data;
+		void *ie_end = pkt->data + pkt->datalen;
+
+		if ( ! ieee80211_ie_bound ( ie, ie_end ) ) {
+			DBGC ( ctx, "WPA %p: malformed PMKID KDE\n", ctx );
+			return wpa_fail ( ctx, -EINVAL );
+		}
+
+		while ( ie ) {
+			if ( ie->id == IEEE80211_IE_VENDOR &&
+			     ie->vendor.oui == WPA_KDE_PMKID ) {
+				rc = wpa_check_pmkid ( ctx, ie->vendor.data );
+				if ( rc < 0 ) {
+					DBGC ( ctx, "WPA %p ALERT: PMKID "
+					       "mismatch in 1/4\n", ctx );
+					return wpa_fail ( ctx, rc );
+				}
+			}
+
+			ie = ieee80211_next_ie ( ie, ie_end );
+		}
+	}
+
+	DBGC ( ctx, "WPA %p: received 1/4, looks OK\n", ctx );
+
+	wpa_derive_ptk ( ctx );
+
+	return wpa_send_2_of_4 ( ctx, pkt, is_rsn, kie );
+}
+
+
+/**
+ * Send fourth frame in 4-Way Handshake, or second in Group Key Handshake
+ *
+ * @v ctx	WPA common context
+ * @v pkt	EAPOL-Key packet for frame to which we're replying
+ * @v is_rsn	If TRUE, frame uses new RSN format
+ * @v kie	Key integrity and encryption handler
+ * @ret rc	Return status code
+ */
+static int wpa_send_final ( struct wpa_common_ctx *ctx,
+			    struct eapol_key_pkt *pkt, int is_rsn,
+			    struct wpa_kie *kie )
+{
+	struct io_buffer *iob = wpa_alloc_frame ( 0 );
+	struct eapol_key_pkt *npkt;
+
+	if ( ! iob )
+		return -ENOMEM;
+
+	npkt = iob->data;
+	memcpy ( npkt, pkt, sizeof ( *pkt ) );
+	npkt->info &= ~( EAPOL_KEY_INFO_KEY_ACK | EAPOL_KEY_INFO_INSTALL |
+			 EAPOL_KEY_INFO_KEY_ENC );
+	if ( is_rsn )
+		npkt->keysize = 0;
+	memset ( npkt->nonce, 0, sizeof ( npkt->nonce ) );
+	memset ( npkt->iv, 0, sizeof ( npkt->iv ) );
+	npkt->datalen = 0;
+
+	if ( npkt->info & EAPOL_KEY_INFO_TYPE )
+		DBGC ( ctx, "WPA %p: sending 4/4\n", ctx );
+	else
+		DBGC ( ctx, "WPA %p: sending 2/2\n", ctx );
+
+	return wpa_send_eapol ( iob, ctx, kie );
+
+}
+
+
+/**
+ * Handle receipt of third frame in 4-Way Handshake
+ *
+ * @v ctx	WPA common context
+ * @v pkt	EAPOL-Key packet
+ * @v is_rsn	If TRUE, frame uses new RSN format
+ * @v kie	Key integrity and encryption handler
+ * @ret rc	Return status code
+ */
+static int wpa_handle_3_of_4 ( struct wpa_common_ctx *ctx,
+			       struct eapol_key_pkt *pkt, int is_rsn,
+			       struct wpa_kie *kie )
+{
+	int rc;
+	u8 *this_rsn, *this_rsn_end;
+	u8 *new_rsn, *new_rsn_end;
+	int this_is_rsn, new_is_rsn;
+
+	if ( ctx->state == WPA_WAITING )
+		return -EINVAL;
+
+	ctx->state = WPA_WORKING;
+
+	/* Check nonce */
+	if ( memcmp ( ctx->Anonce, pkt->nonce, WPA_NONCE_LEN ) != 0 ) {
+		DBGC ( ctx, "WPA %p ALERT: nonce mismatch in 3/4\n", ctx );
+		return wpa_fail ( ctx, -EACCES );
+	}
+
+	/* Check RSN IE */
+	this_rsn = sec80211_find_rsn ( ( union ieee80211_ie * ) pkt->data,
+				       pkt->data + pkt->datalen,
+				       &this_is_rsn, &this_rsn_end );
+	if ( this_rsn )
+		new_rsn = sec80211_find_rsn ( ( union ieee80211_ie * )
+					              this_rsn_end,
+					      pkt->data + pkt->datalen,
+					      &new_is_rsn, &new_rsn_end );
+	else
+		new_rsn = NULL;
+
+	if ( ! ctx->ap_rsn_ie || ! this_rsn ||
+	     ctx->ap_rsn_ie_len != ( this_rsn_end - this_rsn ) ||
+	     ctx->ap_rsn_is_rsn != this_is_rsn ||
+	     memcmp ( ctx->ap_rsn_ie, this_rsn, ctx->ap_rsn_ie_len ) != 0 ) {
+		DBGC ( ctx, "WPA %p ALERT: RSN mismatch in 3/4\n", ctx );
+		DBGC2 ( ctx, "WPA %p RSNs (in 3/4, in beacon):\n", ctx );
+		DBGC2_HD ( ctx, this_rsn, this_rsn_end - this_rsn );
+		DBGC2_HD ( ctx, ctx->ap_rsn_ie, ctx->ap_rsn_ie_len );
+		return wpa_fail ( ctx, -EACCES );
+	}
+
+	/* Don't switch if they just supplied both styles of IE
+	   simultaneously; we need two RSN IEs or two WPA IEs to
+	   switch ciphers. They'll be immediately consecutive because
+	   of ordering guarantees. */
+	if ( new_rsn && this_is_rsn == new_is_rsn ) {
+		struct net80211_wlan *assoc = ctx->dev->associating;
+		DBGC ( ctx, "WPA %p: accommodating bait-and-switch tactics\n",
+		       ctx );
+		DBGC2 ( ctx, "WPA %p RSNs (in 3/4+beacon, new in 3/4):\n",
+			ctx );
+		DBGC2_HD ( ctx, this_rsn, this_rsn_end - this_rsn );
+		DBGC2_HD ( ctx, new_rsn, new_rsn_end - new_rsn );
+
+		if ( ( rc = sec80211_detect_ie ( new_is_rsn, new_rsn,
+						 new_rsn_end,
+						 &assoc->handshaking,
+						 &assoc->crypto ) ) != 0 )
+			DBGC ( ctx, "WPA %p: bait-and-switch invalid, staying "
+			       "with original request\n", ctx );
+	} else {
+		new_rsn = this_rsn;
+		new_is_rsn = this_is_rsn;
+		new_rsn_end = this_rsn_end;
+	}
+
+	/* Grab group cryptosystem ID */
+	ctx->gcrypt = sec80211_rsn_get_net80211_crypt ( *( u32 * )
+							( new_rsn + 2 ) );
+
+	/* Check for a GTK, if info field is encrypted */
+	if ( pkt->info & EAPOL_KEY_INFO_KEY_ENC ) {
+		rc = wpa_maybe_install_gtk ( ctx,
+					     ( union ieee80211_ie * ) pkt->data,
+					     pkt->data + pkt->datalen,
+					     pkt->rsc );
+		if ( rc < 0 ) {
+			DBGC ( ctx, "WPA %p did not install GTK in 3/4: %s\n",
+			       ctx, strerror ( rc ) );
+			if ( rc != -ENOENT )
+				return wpa_fail ( ctx, rc );
+		}
+	}
+
+	DBGC ( ctx, "WPA %p: received 3/4, looks OK\n", ctx );
+
+	/* Send final message */
+	rc = wpa_send_final ( ctx, pkt, is_rsn, kie );
+	if ( rc < 0 )
+		return wpa_fail ( ctx, rc );
+
+	/* Install PTK */
+	rc = wpa_install_ptk ( ctx, pkt->keysize );
+	if ( rc < 0 ) {
+		DBGC ( ctx, "WPA %p failed to install PTK: %s\n", ctx,
+		       strerror ( rc ) );
+		return wpa_fail ( ctx, rc );
+	}
+
+	/* Mark us as needing a new Snonce if we rekey */
+	ctx->have_Snonce = 0;
+
+	/* Done! */
+	ctx->state = WPA_SUCCESS;
+	return 0;
+}
+
+
+/**
+ * Handle receipt of first frame in Group Key Handshake
+ *
+ * @v ctx	WPA common context
+ * @v pkt	EAPOL-Key packet
+ * @v is_rsn	If TRUE, frame uses new RSN format
+ * @v kie	Key integrity and encryption handler
+ * @ret rc	Return status code
+ */
+static int wpa_handle_1_of_2 ( struct wpa_common_ctx *ctx,
+			       struct eapol_key_pkt *pkt, int is_rsn,
+			       struct wpa_kie *kie )
+{
+	int rc;
+
+	/*
+	 * WPA and RSN do this completely differently.
+	 *
+	 * The idea of encoding the GTK (or PMKID, or various other
+	 * things) into a KDE that looks like an information element
+	 * is an RSN innovation; old WPA code never encapsulates
+	 * things like that. If it looks like an info element, it
+	 * really is (for the WPA IE check in frames 2/4 and 3/4). The
+	 * "key data encrypted" bit in the info field is also specific
+	 * to RSN.
+	 *
+	 * So from an old WPA host, 3/4 does not contain an
+	 * encapsulated GTK. The first frame of the GK handshake
+	 * contains it, encrypted, but without a KDE wrapper, and with
+	 * the key ID field (which gPXE doesn't use) shoved away in
+	 * the reserved bits in the info field, and the TxRx bit
+	 * stealing the Install bit's spot.
+	 */
+
+	if ( is_rsn && ( pkt->info & EAPOL_KEY_INFO_KEY_ENC ) ) {
+		rc = wpa_maybe_install_gtk ( ctx,
+					     ( union ieee80211_ie * ) pkt->data,
+					     pkt->data + pkt->datalen,
+					     pkt->rsc );
+		if ( rc < 0 ) {
+			DBGC ( ctx, "WPA %p: failed to install GTK in 1/2: "
+			       "%s\n", ctx, strerror ( rc ) );
+			return wpa_fail ( ctx, rc );
+		}
+	} else {
+		rc = kie->decrypt ( &ctx->ptk.kek, pkt->iv, pkt->data,
+				    &pkt->datalen );
+		if ( rc < 0 ) {
+			DBGC ( ctx, "WPA %p: failed to decrypt GTK: %s\n",
+			       ctx, strerror ( rc ) );
+			return rc; /* non-fatal */
+		}
+		if ( pkt->datalen > sizeof ( ctx->gtk.tk ) ) {
+			DBGC ( ctx, "WPA %p: too much GTK data (%d > %d)\n",
+			       ctx, pkt->datalen, sizeof ( ctx->gtk.tk ) );
+			return wpa_fail ( ctx, -EINVAL );
+		}
+
+		memcpy ( &ctx->gtk.tk, pkt->data, pkt->datalen );
+		wpa_install_gtk ( ctx, pkt->datalen, pkt->rsc );
+	}
+
+	DBGC ( ctx, "WPA %p: received 1/2, looks OK\n", ctx );
+
+	return wpa_send_final ( ctx, pkt, is_rsn, kie );
+}
+
+
+/**
+ * Handle receipt of EAPOL-Key frame for WPA
+ *
+ * @v iob	I/O buffer
+ * @v netdev	Network device
+ * @v ll_source	Source link-layer address
+ */
+static int eapol_key_rx ( struct io_buffer *iob, struct net_device *netdev,
+			  const void *ll_source )
+{
+	struct net80211_device *dev = net80211_get ( netdev );
+	struct eapol_key_pkt *pkt = iob->data;
+	int is_rsn, found_ctx;
+	struct wpa_common_ctx *ctx;
+	int rc = 0;
+	struct wpa_kie *kie;
+	u8 their_mic[16], our_mic[16];
+
+	if ( pkt->type != EAPOL_KEY_TYPE_WPA &&
+	     pkt->type != EAPOL_KEY_TYPE_RSN ) {
+		DBG ( "EAPOL-Key: packet not of 802.11 type\n" );
+		rc = -EINVAL;
+		goto drop;
+	}
+
+	is_rsn = ( pkt->type == EAPOL_KEY_TYPE_RSN );
+
+	if ( ! dev ) {
+		DBG ( "EAPOL-Key: packet not from 802.11\n" );
+		rc = -EINVAL;
+		goto drop;
+	}
+
+	if ( memcmp ( dev->bssid, ll_source, ETH_ALEN ) != 0 ) {
+		DBG ( "EAPOL-Key: packet not from associated AP\n" );
+		rc = -EINVAL;
+		goto drop;
+	}
+
+	if ( ! ( ntohs ( pkt->info ) & EAPOL_KEY_INFO_KEY_ACK ) ) {
+		DBG ( "EAPOL-Key: packet sent in wrong direction\n" );
+		rc = -EINVAL;
+		goto drop;
+	}
+
+	found_ctx = 0;
+	list_for_each_entry ( ctx, &wpa_contexts, list ) {
+		if ( ctx->dev == dev ) {
+			found_ctx = 1;
+			break;
+		}
+	}
+
+	if ( ! found_ctx ) {
+		DBG ( "EAPOL-Key: no WPA context to handle packet for %p\n",
+		      dev );
+		rc = -ENOENT;
+		goto drop;
+	}
+
+	if ( ( void * ) ( pkt + 1 ) + ntohs ( pkt->datalen ) > iob->tail ) {
+		DBGC ( ctx, "WPA %p: packet truncated (has %d extra bytes, "
+		       "states %d)\n", ctx, iob->tail - ( void * ) ( pkt + 1 ),
+		       ntohs ( pkt->datalen ) );
+		rc = -EINVAL;
+		goto drop;
+	}
+
+	/* Get a handle on key integrity/encryption handler */
+	kie = wpa_find_kie ( ntohs ( pkt->info ) & EAPOL_KEY_INFO_VERSION );
+	if ( ! kie ) {
+		DBGC ( ctx, "WPA %p: no support for packet version %d\n", ctx,
+		       ntohs ( pkt->info ) & EAPOL_KEY_INFO_VERSION );
+		rc = wpa_fail ( ctx, -ENOTSUP );
+		goto drop;
+	}
+
+	/* Check MIC */
+	if ( ntohs ( pkt->info ) & EAPOL_KEY_INFO_KEY_MIC ) {
+		memcpy ( their_mic, pkt->mic, sizeof ( pkt->mic ) );
+		memset ( pkt->mic, 0, sizeof ( pkt->mic ) );
+		kie->mic ( &ctx->ptk.kck, ( void * ) pkt - EAPOL_HDR_LEN,
+			   EAPOL_HDR_LEN + sizeof ( *pkt ) +
+			   ntohs ( pkt->datalen ), our_mic );
+		DBGC2 ( ctx, "WPA %p MIC comparison (theirs, ours):\n", ctx );
+		DBGC2_HD ( ctx, their_mic, 16 );
+		DBGC2_HD ( ctx, our_mic, 16 );
+		if ( memcmp ( their_mic, our_mic, sizeof ( pkt->mic ) ) != 0 ) {
+			DBGC ( ctx, "WPA %p: EAPOL MIC failure\n", ctx );
+			goto drop;
+		}
+	}
+
+	/* Fix byte order to local */
+	pkt->info = ntohs ( pkt->info );
+	pkt->keysize = ntohs ( pkt->keysize );
+	pkt->datalen = ntohs ( pkt->datalen );
+	pkt->replay = be64_to_cpu ( pkt->replay );
+
+	/* Check replay counter */
+	if ( ctx->replay != ~0ULL && ctx->replay >= pkt->replay ) {
+		DBGC ( ctx, "WPA %p ALERT: Replay detected! "
+		       "(%08x:%08x >= %08x:%08x)\n", ctx,
+		       ( u32 ) ( ctx->replay >> 32 ), ( u32 ) ctx->replay,
+		       ( u32 ) ( pkt->replay >> 32 ), ( u32 ) pkt->replay );
+		rc = 0;		/* ignore without error */
+		goto drop;
+	}
+	ctx->replay = pkt->replay;
+
+	/* Decrypt key data */
+	if ( pkt->info & EAPOL_KEY_INFO_KEY_ENC ) {
+		rc = kie->decrypt ( &ctx->ptk.kek, pkt->iv, pkt->data,
+				    &pkt->datalen );
+		if ( rc < 0 ) {
+			DBGC ( ctx, "WPA %p: failed to decrypt packet: %s\n",
+			       ctx, strerror ( rc ) );
+			goto drop;
+		}
+	}
+
+	/* Hand it off to appropriate handler */
+	switch ( pkt->info & ( EAPOL_KEY_INFO_TYPE |
+			       EAPOL_KEY_INFO_KEY_MIC ) ) {
+	case EAPOL_KEY_TYPE_PTK:
+		rc = wpa_handle_1_of_4 ( ctx, pkt, is_rsn, kie );
+		break;
+
+	case EAPOL_KEY_TYPE_PTK | EAPOL_KEY_INFO_KEY_MIC:
+		rc = wpa_handle_3_of_4 ( ctx, pkt, is_rsn, kie );
+		break;
+
+	case EAPOL_KEY_TYPE_GTK | EAPOL_KEY_INFO_KEY_MIC:
+		rc = wpa_handle_1_of_2 ( ctx, pkt, is_rsn, kie );
+		break;
+
+	default:
+		DBGC ( ctx, "WPA %p: Invalid combination of key flags %04x\n",
+		       ctx, pkt->info );
+		rc = -EINVAL;
+		break;
+	}
+
+ drop:
+	free_iob ( iob );
+	return rc;
+}
+
+struct eapol_handler eapol_key_handler __eapol_handler = {
+	.type = EAPOL_TYPE_KEY,
+	.rx = eapol_key_rx,
+};
+
+/* WPA always needs EAPOL in order to be useful */
+REQUIRE_OBJECT ( eapol );
diff --git a/gpxe/src/net/80211/wpa_ccmp.c b/gpxe/src/net/80211/wpa_ccmp.c
new file mode 100644
index 00000000..08b1e17b
--- /dev/null
+++ b/gpxe/src/net/80211/wpa_ccmp.c
@@ -0,0 +1,528 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <gpxe/net80211.h>
+#include <gpxe/crypto.h>
+#include <gpxe/hmac.h>
+#include <gpxe/sha1.h>
+#include <gpxe/aes.h>
+#include <gpxe/wpa.h>
+#include <byteswap.h>
+#include <errno.h>
+
+/** @file
+ *
+ * Backend for WPA using the CCMP encryption method
+ */
+
+/** Context for CCMP encryption and decryption */
+struct ccmp_ctx
+{
+	/** AES context - only ever used for encryption */
+	u8 aes_ctx[AES_CTX_SIZE];
+
+	/** Most recently sent packet number */
+	u64 tx_seq;
+
+	/** Most recently received packet number */
+	u64 rx_seq;
+};
+
+/** Header structure at the beginning of CCMP frame data */
+struct ccmp_head
+{
+	u8 pn_lo[2];		/**< Bytes 0 and 1 of packet number */
+	u8 _rsvd;		/**< Reserved byte */
+	u8 kid;			/**< Key ID and ExtIV byte */
+	u8 pn_hi[4];		/**< Bytes 2-5 (2 first) of packet number */
+} __attribute__ (( packed ));
+
+
+/** CCMP header overhead */
+#define CCMP_HEAD_LEN	8
+
+/** CCMP MIC trailer overhead */
+#define CCMP_MIC_LEN	8
+
+/** CCMP nonce length */
+#define CCMP_NONCE_LEN	13
+
+/** CCMP nonce structure */
+struct ccmp_nonce
+{
+	u8 prio;		/**< Packet priority, 0 for non-QoS */
+	u8 a2[ETH_ALEN];	/**< Address 2 from packet header (sender) */
+	u8 pn[6];		/**< Packet number */
+} __attribute__ (( packed ));
+
+/** CCMP additional authentication data length (for non-QoS, non-WDS frames) */
+#define CCMP_AAD_LEN	22
+
+/** CCMP additional authentication data structure */
+struct ccmp_aad
+{
+	u16 fc;			/**< Frame Control field */
+	u8 a1[6];		/**< Address 1 */
+	u8 a2[6];		/**< Address 2 */
+	u8 a3[6];		/**< Address 3 */
+	u16 seq;		/**< Sequence Control field */
+	/* Address 4 and QoS Control are included if present */
+} __attribute__ (( packed ));
+
+/** Mask for Frame Control field in AAD */
+#define CCMP_AAD_FC_MASK	0xC38F
+
+/** Mask for Sequence Control field in AAD */
+#define CCMP_AAD_SEQ_MASK	0x000F
+
+
+/**
+ * Convert 6-byte LSB packet number to 64-bit integer
+ *
+ * @v pn	Pointer to 6-byte packet number
+ * @ret v	64-bit integer value of @a pn
+ */
+static u64 pn_to_u64 ( const u8 *pn )
+{
+	int i;
+	u64 ret = 0;
+
+	for ( i = 5; i >= 0; i-- ) {
+		ret <<= 8;
+		ret |= pn[i];
+	}
+
+	return ret;
+}
+
+/**
+ * Convert 64-bit integer to 6-byte packet number
+ *
+ * @v v		64-bit integer
+ * @v msb	If TRUE, reverse the output PN to be in MSB order
+ * @ret pn	6-byte packet number
+ *
+ * The PN is stored in LSB order in the packet header and in MSB order
+ * in the nonce. WHYYYYY?
+ */
+static void u64_to_pn ( u64 v, u8 *pn, int msb )
+{
+	int i;
+	u8 *pnp = pn + ( msb ? 5 : 0 );
+	int delta = ( msb ? -1 : +1 );
+
+	for ( i = 0; i < 6; i++ ) {
+		*pnp = v & 0xFF;
+		pnp += delta;
+		v >>= 8;
+	}
+}
+
+/** Value for @a msb argument of u64_to_pn() for MSB output */
+#define PN_MSB	1
+
+/** Value for @a msb argument of u64_to_pn() for LSB output */
+#define PN_LSB	0
+
+
+
+/**
+ * Initialise CCMP state and install key
+ *
+ * @v crypto	CCMP cryptosystem structure
+ * @v key	Pointer to 16-byte temporal key to install
+ * @v keylen	Length of key (16 bytes)
+ * @v rsc	Initial receive sequence counter
+ */
+static int ccmp_init ( struct net80211_crypto *crypto, const void *key,
+		       int keylen, const void *rsc )
+{
+	struct ccmp_ctx *ctx = crypto->priv;
+
+	if ( keylen != 16 )
+		return -EINVAL;
+
+	if ( rsc )
+		ctx->rx_seq = pn_to_u64 ( rsc );
+
+	cipher_setkey ( &aes_algorithm, ctx->aes_ctx, key, keylen );
+
+	return 0;
+}
+
+
+/**
+ * Encrypt or decrypt data stream using AES in Counter mode
+ *
+ * @v ctx	CCMP cryptosystem context
+ * @v nonce	Nonce value, 13 bytes
+ * @v srcv	Data to encrypt or decrypt
+ * @v len	Number of bytes pointed to by @a src
+ * @v msrcv	MIC value to encrypt or decrypt (may be NULL)
+ * @ret destv	Encrypted or decrypted data
+ * @ret mdestv	Encrypted or decrypted MIC value
+ *
+ * This assumes CCMP parameters of L=2 and M=8. The algorithm is
+ * defined in RFC 3610.
+ */
+static void ccmp_ctr_xor ( struct ccmp_ctx *ctx, const void *nonce,
+			   const void *srcv, void *destv, int len,
+			   const void *msrcv, void *mdestv )
+{
+	u8 A[16], S[16];
+	u16 ctr;
+	int i;
+	const u8 *src = srcv, *msrc = msrcv;
+	u8 *dest = destv, *mdest = mdestv;
+
+	A[0] = 0x01;		/* flags, L' = L - 1 = 1, other bits rsvd */
+	memcpy ( A + 1, nonce, CCMP_NONCE_LEN );
+
+	if ( msrcv ) {
+		A[14] = A[15] = 0;
+
+		cipher_encrypt ( &aes_algorithm, ctx->aes_ctx, A, S, 16 );
+
+		for ( i = 0; i < 8; i++ ) {
+			*mdest++ = *msrc++ ^ S[i];
+		}
+	}
+
+	for ( ctr = 1 ;; ctr++ ) {
+		A[14] = ctr >> 8;
+		A[15] = ctr & 0xFF;
+
+		cipher_encrypt ( &aes_algorithm, ctx->aes_ctx, A, S, 16 );
+
+		for ( i = 0; i < len && i < 16; i++ )
+			*dest++ = *src++ ^ S[i];
+
+		if ( len <= 16 )
+			break;	/* we're done */
+
+		len -= 16;
+	}
+}
+
+
+/**
+ * Advance one block in CBC-MAC calculation
+ *
+ * @v aes_ctx	AES encryption context with key set
+ * @v B		Cleartext block to incorporate (16 bytes)
+ * @v X		Previous ciphertext block (16 bytes)
+ * @ret B	Clobbered
+ * @ret X	New ciphertext block (16 bytes)
+ *
+ * This function does X := E[key] ( X ^ B ).
+ */
+static void ccmp_feed_cbc_mac ( void *aes_ctx, u8 *B, u8 *X )
+{
+	int i;
+	for ( i = 0; i < 16; i++ )
+		B[i] ^= X[i];
+	cipher_encrypt ( &aes_algorithm, aes_ctx, B, X, 16 );
+}
+
+
+/**
+ * Calculate MIC on plaintext data using CBC-MAC
+ *
+ * @v ctx	CCMP cryptosystem context
+ * @v nonce	Nonce value, 13 bytes
+ * @v data	Data to calculate MIC over
+ * @v datalen	Length of @a data
+ * @v aad	Additional authentication data, for MIC but not encryption
+ * @ret mic	MIC value (unencrypted), 8 bytes
+ *
+ * @a aadlen is assumed to be 22 bytes long, as it always is for
+ * 802.11 use when transmitting non-QoS, not-between-APs frames (the
+ * only type we deal with).
+ */
+static void ccmp_cbc_mac ( struct ccmp_ctx *ctx, const void *nonce,
+			   const void *data, u16 datalen,
+			   const void *aad, void *mic )
+{
+	u8 X[16], B[16];
+
+	/* Zeroth block: flags, nonce, length */
+
+	/* Rsv AAD - M'-  - L'-
+	 *  0   1  0 1 1  0 0 1   for an 8-byte MAC and 2-byte message length
+	 */
+	B[0] = 0x59;
+	memcpy ( B + 1, nonce, CCMP_NONCE_LEN );
+	B[14] = datalen >> 8;
+	B[15] = datalen & 0xFF;
+
+	cipher_encrypt ( &aes_algorithm, ctx->aes_ctx, B, X, 16 );
+
+	/* First block: AAD length field and 14 bytes of AAD */
+	B[0] = 0;
+	B[1] = CCMP_AAD_LEN;
+	memcpy ( B + 2, aad, 14 );
+
+	ccmp_feed_cbc_mac ( ctx->aes_ctx, B, X );
+
+	/* Second block: Remaining 8 bytes of AAD, 8 bytes zero pad */
+	memcpy ( B, aad + 14, 8 );
+	memset ( B + 8, 0, 8 );
+
+	ccmp_feed_cbc_mac ( ctx->aes_ctx, B, X );
+
+	/* Message blocks */
+	while ( datalen ) {
+		if ( datalen >= 16 ) {
+			memcpy ( B, data, 16 );
+			datalen -= 16;
+		} else {
+			memcpy ( B, data, datalen );
+			memset ( B + datalen, 0, 16 - datalen );
+			datalen = 0;
+		}
+
+		ccmp_feed_cbc_mac ( ctx->aes_ctx, B, X );
+
+		data += 16;
+	}
+
+	/* Get MIC from final value of X */
+	memcpy ( mic, X, 8 );
+}
+
+
+/**
+ * Encapsulate and encrypt a packet using CCMP
+ *
+ * @v crypto	CCMP cryptosystem
+ * @v iob	I/O buffer containing cleartext packet
+ * @ret eiob	I/O buffer containing encrypted packet
+ */
+struct io_buffer * ccmp_encrypt ( struct net80211_crypto *crypto,
+				  struct io_buffer *iob )
+{
+	struct ccmp_ctx *ctx = crypto->priv;
+	struct ieee80211_frame *hdr = iob->data;
+	struct io_buffer *eiob;
+	const int hdrlen = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int datalen = iob_len ( iob ) - hdrlen;
+	struct ccmp_head head;
+	struct ccmp_nonce nonce;
+	struct ccmp_aad aad;
+	u8 mic[8], tx_pn[6];
+	void *edata, *emic;
+
+	ctx->tx_seq++;
+	u64_to_pn ( ctx->tx_seq, tx_pn, PN_LSB );
+
+	/* Allocate memory */
+	eiob = alloc_iob ( iob_len ( iob ) + CCMP_HEAD_LEN + CCMP_MIC_LEN );
+	if ( ! eiob )
+		return NULL;
+
+	/* Copy frame header */
+	memcpy ( iob_put ( eiob, hdrlen ), iob->data, hdrlen );
+	hdr = eiob->data;
+	hdr->fc |= IEEE80211_FC_PROTECTED;
+
+	/* Fill in packet number and extended IV */
+	memcpy ( head.pn_lo, tx_pn, 2 );
+	memcpy ( head.pn_hi, tx_pn + 2, 4 );
+	head.kid = 0x20;	/* have Extended IV, key ID 0 */
+	head._rsvd = 0;
+	memcpy ( iob_put ( eiob, sizeof ( head ) ), &head, sizeof ( head ) );
+
+	/* Form nonce */
+	nonce.prio = 0;
+	memcpy ( nonce.a2, hdr->addr2, ETH_ALEN );
+	u64_to_pn ( ctx->tx_seq, nonce.pn, PN_MSB );
+
+	/* Form additional authentication data */
+	aad.fc = hdr->fc & CCMP_AAD_FC_MASK;
+	memcpy ( aad.a1, hdr->addr1, 3 * ETH_ALEN ); /* all 3 at once */
+	aad.seq = hdr->seq & CCMP_AAD_SEQ_MASK;
+
+	/* Calculate MIC over the data */
+	ccmp_cbc_mac ( ctx, &nonce, iob->data + hdrlen, datalen, &aad, mic );
+
+	/* Copy and encrypt data and MIC */
+	edata = iob_put ( eiob, datalen );
+	emic = iob_put ( eiob, CCMP_MIC_LEN );
+	ccmp_ctr_xor ( ctx, &nonce,
+		       iob->data + hdrlen, edata, datalen,
+		       mic, emic );
+
+	/* Done! */
+	DBGC2 ( ctx, "WPA-CCMP %p: encrypted packet %p -> %p\n", ctx,
+		iob, eiob );
+
+	return eiob;
+}
+
+/**
+ * Decrypt a packet using CCMP
+ *
+ * @v crypto	CCMP cryptosystem
+ * @v eiob	I/O buffer containing encrypted packet
+ * @ret iob	I/O buffer containing cleartext packet
+ */
+static struct io_buffer * ccmp_decrypt ( struct net80211_crypto *crypto,
+					 struct io_buffer *eiob )
+{
+	struct ccmp_ctx *ctx = crypto->priv;
+	struct ieee80211_frame *hdr;
+	struct io_buffer *iob;
+	const int hdrlen = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int datalen = iob_len ( eiob ) - hdrlen - CCMP_HEAD_LEN - CCMP_MIC_LEN;
+	struct ccmp_head *head;
+	struct ccmp_nonce nonce;
+	struct ccmp_aad aad;
+	u8 rx_pn[6], their_mic[8], our_mic[8];
+
+	iob = alloc_iob ( hdrlen + datalen );
+	if ( ! iob )
+		return NULL;
+
+	/* Copy frame header */
+	memcpy ( iob_put ( iob, hdrlen ), eiob->data, hdrlen );
+	hdr = iob->data;
+	hdr->fc &= ~IEEE80211_FC_PROTECTED;
+
+	/* Check and update RX packet number */
+	head = eiob->data + hdrlen;
+	memcpy ( rx_pn, head->pn_lo, 2 );
+	memcpy ( rx_pn + 2, head->pn_hi, 4 );
+
+	if ( pn_to_u64 ( rx_pn ) <= ctx->rx_seq ) {
+		DBGC ( ctx, "WPA-CCMP %p: packet received out of order "
+		       "(%012llx <= %012llx)\n", ctx, pn_to_u64 ( rx_pn ),
+		       ctx->rx_seq );
+		free_iob ( iob );
+		return NULL;
+	}
+
+	ctx->rx_seq = pn_to_u64 ( rx_pn );
+	DBGC2 ( ctx, "WPA-CCMP %p: RX packet number %012llx\n", ctx, ctx->rx_seq );
+
+	/* Form nonce */
+	nonce.prio = 0;
+	memcpy ( nonce.a2, hdr->addr2, ETH_ALEN );
+	u64_to_pn ( ctx->rx_seq, nonce.pn, PN_MSB );
+
+	/* Form additional authentication data */
+	aad.fc = ( hdr->fc & CCMP_AAD_FC_MASK ) | IEEE80211_FC_PROTECTED;
+	memcpy ( aad.a1, hdr->addr1, 3 * ETH_ALEN ); /* all 3 at once */
+	aad.seq = hdr->seq & CCMP_AAD_SEQ_MASK;
+
+	/* Copy-decrypt data and MIC */
+	ccmp_ctr_xor ( ctx, &nonce, eiob->data + hdrlen + sizeof ( *head ),
+		       iob_put ( iob, datalen ), datalen,
+		       eiob->tail - CCMP_MIC_LEN, their_mic );
+
+	/* Check MIC */
+	ccmp_cbc_mac ( ctx, &nonce, iob->data + hdrlen, datalen, &aad,
+		       our_mic );
+
+	if ( memcmp ( their_mic, our_mic, CCMP_MIC_LEN ) != 0 ) {
+		DBGC2 ( ctx, "WPA-CCMP %p: MIC failure\n", ctx );
+		free_iob ( iob );
+		return NULL;
+	}
+
+	DBGC2 ( ctx, "WPA-CCMP %p: decrypted packet %p -> %p\n", ctx,
+		eiob, iob );
+
+	return iob;
+}
+
+
+/** CCMP cryptosystem */
+struct net80211_crypto ccmp_crypto __net80211_crypto = {
+	.algorithm = NET80211_CRYPT_CCMP,
+	.init = ccmp_init,
+	.encrypt = ccmp_encrypt,
+	.decrypt = ccmp_decrypt,
+	.priv_len = sizeof ( struct ccmp_ctx ),
+};
+
+
+
+
+/**
+ * Calculate HMAC-SHA1 MIC for EAPOL-Key frame
+ *
+ * @v kck	Key Confirmation Key, 16 bytes
+ * @v msg	Message to calculate MIC over
+ * @v len	Number of bytes to calculate MIC over
+ * @ret mic	Calculated MIC, 16 bytes long
+ */
+static void ccmp_kie_mic ( const void *kck, const void *msg, size_t len,
+			   void *mic )
+{
+	u8 sha1_ctx[SHA1_CTX_SIZE];
+	u8 kckb[16];
+	u8 hash[SHA1_SIZE];
+	size_t kck_len = 16;
+
+	memcpy ( kckb, kck, kck_len );
+
+	hmac_init ( &sha1_algorithm, sha1_ctx, kckb, &kck_len );
+	hmac_update ( &sha1_algorithm, sha1_ctx, msg, len );
+	hmac_final ( &sha1_algorithm, sha1_ctx, kckb, &kck_len, hash );
+
+	memcpy ( mic, hash, 16 );
+}
+
+/**
+ * Decrypt key data in EAPOL-Key frame
+ *
+ * @v kek	Key Encryption Key, 16 bytes
+ * @v iv	Initialisation vector, 16 bytes (unused)
+ * @v msg	Message to decrypt
+ * @v len	Length of message
+ * @ret msg	Decrypted message in place of original
+ * @ret len	Adjusted downward for 8 bytes of overhead
+ * @ret rc	Return status code
+ *
+ * The returned message may still contain padding of 0xDD followed by
+ * zero or more 0x00 octets. It is impossible to remove the padding
+ * without parsing the IEs in the packet (another design decision that
+ * tends to make one question the 802.11i committee's intelligence...)
+ */
+static int ccmp_kie_decrypt ( const void *kek, const void *iv __unused,
+			      void *msg, u16 *len )
+{
+	if ( *len % 8 != 0 )
+		return -EINVAL;
+
+	if ( aes_unwrap ( kek, msg, msg, *len / 8 - 1 ) != 0 )
+		return -EINVAL;
+
+	*len -= 8;
+
+	return 0;
+}
+
+/** CCMP-style key integrity and encryption handler */
+struct wpa_kie ccmp_kie __wpa_kie = {
+	.version = EAPOL_KEY_VERSION_WPA2,
+	.mic = ccmp_kie_mic,
+	.decrypt = ccmp_kie_decrypt,
+};
diff --git a/gpxe/src/net/80211/wpa_psk.c b/gpxe/src/net/80211/wpa_psk.c
new file mode 100644
index 00000000..e7521682
--- /dev/null
+++ b/gpxe/src/net/80211/wpa_psk.c
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <gpxe/net80211.h>
+#include <gpxe/sha1.h>
+#include <gpxe/wpa.h>
+#include <errno.h>
+
+/** @file
+ *
+ * Frontend for WPA using a pre-shared key.
+ */
+
+/**
+ * Initialise WPA-PSK state
+ *
+ * @v dev	802.11 device
+ * @ret rc	Return status code
+ */
+static int wpa_psk_init ( struct net80211_device *dev )
+{
+	return wpa_make_rsn_ie ( dev, &dev->rsn_ie );
+}
+
+/**
+ * Start WPA-PSK authentication
+ *
+ * @v dev	802.11 device
+ * @ret rc	Return status code
+ */
+static int wpa_psk_start ( struct net80211_device *dev )
+{
+	char passphrase[64+1];
+	u8 pmk[WPA_PMK_LEN];
+	int len;
+	struct wpa_common_ctx *ctx = dev->handshaker->priv;
+
+	len = fetch_string_setting ( netdev_settings ( dev->netdev ),
+				     &net80211_key_setting, passphrase,
+				     64 + 1 );
+
+	if ( len <= 0 ) {
+		DBGC ( ctx, "WPA-PSK %p: no passphrase provided!\n", ctx );
+		net80211_deauthenticate ( dev, -EACCES );
+		return -EACCES;
+	}
+
+	pbkdf2_sha1 ( passphrase, len, dev->essid, strlen ( dev->essid ),
+		      4096, pmk, WPA_PMK_LEN );
+
+	DBGC ( ctx, "WPA-PSK %p: derived PMK from passphrase `%s':\n", ctx,
+	       passphrase );
+	DBGC_HD ( ctx, pmk, WPA_PMK_LEN );
+
+	return wpa_start ( dev, ctx, pmk, WPA_PMK_LEN );
+}
+
+/**
+ * Step WPA-PSK authentication
+ *
+ * @v dev	802.11 device
+ * @ret rc	Return status code
+ */
+static int wpa_psk_step ( struct net80211_device *dev )
+{
+	struct wpa_common_ctx *ctx = dev->handshaker->priv;
+
+	switch ( ctx->state ) {
+	case WPA_SUCCESS:
+		return 1;
+	case WPA_FAILURE:
+		return -EACCES;
+	default:
+		return 0;
+	}
+}
+
+/**
+ * Do-nothing function; you can't change a WPA key post-authentication
+ *
+ * @v dev	802.11 device
+ * @ret rc	Return status code
+ */
+static int wpa_psk_no_change_key ( struct net80211_device *dev __unused )
+{
+	return 0;
+}
+
+/**
+ * Disable handling of received WPA authentication frames
+ *
+ * @v dev	802.11 device
+ */
+static void wpa_psk_stop ( struct net80211_device *dev )
+{
+	wpa_stop ( dev );
+}
+
+/** WPA-PSK security handshaker */
+struct net80211_handshaker wpa_psk_handshaker __net80211_handshaker = {
+	.protocol = NET80211_SECPROT_PSK,
+	.init = wpa_psk_init,
+	.start = wpa_psk_start,
+	.step = wpa_psk_step,
+	.change_key = wpa_psk_no_change_key,
+	.stop = wpa_psk_stop,
+	.priv_len = sizeof ( struct wpa_common_ctx ),
+};
diff --git a/gpxe/src/net/80211/wpa_tkip.c b/gpxe/src/net/80211/wpa_tkip.c
new file mode 100644
index 00000000..0cb697fa
--- /dev/null
+++ b/gpxe/src/net/80211/wpa_tkip.c
@@ -0,0 +1,586 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <gpxe/net80211.h>
+#include <gpxe/crypto.h>
+#include <gpxe/hmac.h>
+#include <gpxe/sha1.h>
+#include <gpxe/md5.h>
+#include <gpxe/crc32.h>
+#include <gpxe/arc4.h>
+#include <gpxe/wpa.h>
+#include <byteswap.h>
+#include <errno.h>
+
+/** @file
+ *
+ * Backend for WPA using the TKIP encryption standard.
+ */
+
+/** Context for one direction of TKIP, either encryption or decryption */
+struct tkip_dir_ctx
+{
+	/** High 32 bits of last sequence counter value used */
+	u32 tsc_hi;
+
+	/** Low 32 bits of last sequence counter value used */
+	u16 tsc_lo;
+
+	/** MAC address used to derive TTAK */
+	u8 mac[ETH_ALEN];
+
+	/** If TRUE, TTAK is valid */
+	u16 ttak_ok;
+
+	/** TKIP-mixed transmit address and key, depends on tsc_hi and MAC */
+	u16 ttak[5];
+};
+
+/** Context for TKIP encryption and decryption */
+struct tkip_ctx
+{
+	/** Temporal key to use */
+	struct tkip_tk tk;
+
+	/** State for encryption */
+	struct tkip_dir_ctx enc;
+
+	/** State for decryption */
+	struct tkip_dir_ctx dec;
+};
+
+/** Header structure at the beginning of TKIP frame data */
+struct tkip_head
+{
+	u8 tsc1;		/**< High byte of low 16 bits of TSC */
+	u8 seed1;		/**< Second byte of WEP seed */
+	u8 tsc0;		/**< Low byte of TSC */
+	u8 kid;			/**< Key ID and ExtIV byte */
+	u32 tsc_hi;		/**< High 32 bits of TSC, as an ExtIV */
+} __attribute__ (( packed ));
+
+
+/** TKIP header overhead (IV + KID + ExtIV) */
+#define TKIP_HEAD_LEN	8
+
+/** TKIP trailer overhead (MIC + ICV) [assumes unfragmented] */
+#define TKIP_FOOT_LEN	12
+
+/** TKIP MIC length */
+#define TKIP_MIC_LEN	8
+
+/** TKIP ICV length */
+#define TKIP_ICV_LEN	4
+
+
+/** TKIP S-box */
+static const u16 Sbox[256] = {
+	0xC6A5, 0xF884, 0xEE99, 0xF68D, 0xFF0D, 0xD6BD, 0xDEB1, 0x9154,
+	0x6050, 0x0203, 0xCEA9, 0x567D, 0xE719, 0xB562, 0x4DE6, 0xEC9A,
+	0x8F45, 0x1F9D, 0x8940, 0xFA87, 0xEF15, 0xB2EB, 0x8EC9, 0xFB0B,
+	0x41EC, 0xB367, 0x5FFD, 0x45EA, 0x23BF, 0x53F7, 0xE496, 0x9B5B,
+	0x75C2, 0xE11C, 0x3DAE, 0x4C6A, 0x6C5A, 0x7E41, 0xF502, 0x834F,
+	0x685C, 0x51F4, 0xD134, 0xF908, 0xE293, 0xAB73, 0x6253, 0x2A3F,
+	0x080C, 0x9552, 0x4665, 0x9D5E, 0x3028, 0x37A1, 0x0A0F, 0x2FB5,
+	0x0E09, 0x2436, 0x1B9B, 0xDF3D, 0xCD26, 0x4E69, 0x7FCD, 0xEA9F,
+	0x121B, 0x1D9E, 0x5874, 0x342E, 0x362D, 0xDCB2, 0xB4EE, 0x5BFB,
+	0xA4F6, 0x764D, 0xB761, 0x7DCE, 0x527B, 0xDD3E, 0x5E71, 0x1397,
+	0xA6F5, 0xB968, 0x0000, 0xC12C, 0x4060, 0xE31F, 0x79C8, 0xB6ED,
+	0xD4BE, 0x8D46, 0x67D9, 0x724B, 0x94DE, 0x98D4, 0xB0E8, 0x854A,
+	0xBB6B, 0xC52A, 0x4FE5, 0xED16, 0x86C5, 0x9AD7, 0x6655, 0x1194,
+	0x8ACF, 0xE910, 0x0406, 0xFE81, 0xA0F0, 0x7844, 0x25BA, 0x4BE3,
+	0xA2F3, 0x5DFE, 0x80C0, 0x058A, 0x3FAD, 0x21BC, 0x7048, 0xF104,
+	0x63DF, 0x77C1, 0xAF75, 0x4263, 0x2030, 0xE51A, 0xFD0E, 0xBF6D,
+	0x814C, 0x1814, 0x2635, 0xC32F, 0xBEE1, 0x35A2, 0x88CC, 0x2E39,
+	0x9357, 0x55F2, 0xFC82, 0x7A47, 0xC8AC, 0xBAE7, 0x322B, 0xE695,
+	0xC0A0, 0x1998, 0x9ED1, 0xA37F, 0x4466, 0x547E, 0x3BAB, 0x0B83,
+	0x8CCA, 0xC729, 0x6BD3, 0x283C, 0xA779, 0xBCE2, 0x161D, 0xAD76,
+	0xDB3B, 0x6456, 0x744E, 0x141E, 0x92DB, 0x0C0A, 0x486C, 0xB8E4,
+	0x9F5D, 0xBD6E, 0x43EF, 0xC4A6, 0x39A8, 0x31A4, 0xD337, 0xF28B,
+	0xD532, 0x8B43, 0x6E59, 0xDAB7, 0x018C, 0xB164, 0x9CD2, 0x49E0,
+	0xD8B4, 0xACFA, 0xF307, 0xCF25, 0xCAAF, 0xF48E, 0x47E9, 0x1018,
+	0x6FD5, 0xF088, 0x4A6F, 0x5C72, 0x3824, 0x57F1, 0x73C7, 0x9751,
+	0xCB23, 0xA17C, 0xE89C, 0x3E21, 0x96DD, 0x61DC, 0x0D86, 0x0F85,
+	0xE090, 0x7C42, 0x71C4, 0xCCAA, 0x90D8, 0x0605, 0xF701, 0x1C12,
+	0xC2A3, 0x6A5F, 0xAEF9, 0x69D0, 0x1791, 0x9958, 0x3A27, 0x27B9,
+	0xD938, 0xEB13, 0x2BB3, 0x2233, 0xD2BB, 0xA970, 0x0789, 0x33A7,
+	0x2DB6, 0x3C22, 0x1592, 0xC920, 0x8749, 0xAAFF, 0x5078, 0xA57A,
+	0x038F, 0x59F8, 0x0980, 0x1A17, 0x65DA, 0xD731, 0x84C6, 0xD0B8,
+	0x82C3, 0x29B0, 0x5A77, 0x1E11, 0x7BCB, 0xA8FC, 0x6DD6, 0x2C3A,
+};
+
+/**
+ * Perform S-box mapping on a 16-bit value
+ *
+ * @v v		Value to perform S-box mapping on
+ * @ret Sv	S-box mapped value
+ */
+static inline u16 S ( u16 v )
+{
+	return Sbox[v & 0xFF] ^ swap16 ( Sbox[v >> 8] );
+}
+
+/**
+ * Rotate 16-bit value right
+ *
+ * @v v		Value to rotate
+ * @v bits	Number of bits to rotate by
+ * @ret rotv	Rotated value
+ */
+static inline u16 ror16 ( u16 v, int bits )
+{
+	return ( v >> bits ) | ( v << ( 16 - bits ) );
+}
+
+/**
+ * Rotate 32-bit value right
+ *
+ * @v v		Value to rotate
+ * @v bits	Number of bits to rotate by
+ * @ret rotv	Rotated value
+ */
+static inline u32 ror32 ( u32 v, int bits )
+{
+	return ( v >> bits ) | ( v << ( 32 - bits ) );
+}
+
+/**
+ * Rotate 32-bit value left
+ *
+ * @v v		Value to rotate
+ * @v bits	Number of bits to rotate by
+ * @ret rotv	Rotated value
+ */
+static inline u32 rol32 ( u32 v, int bits )
+{
+	return ( v << bits ) | ( v >> ( 32 - bits ) );
+}
+
+
+/**
+ * Initialise TKIP state and install key
+ *
+ * @v crypto	TKIP cryptosystem structure
+ * @v key	Pointer to tkip_tk to install
+ * @v keylen	Length of key (32 bytes)
+ * @v rsc	Initial receive sequence counter
+ */
+static int tkip_init ( struct net80211_crypto *crypto, const void *key,
+		       int keylen, const void *rsc )
+{
+	struct tkip_ctx *ctx = crypto->priv;
+	const u8 *rscb = rsc;
+
+	if ( keylen != sizeof ( ctx->tk ) )
+		return -EINVAL;
+
+	if ( rscb ) {
+		ctx->dec.tsc_lo =   ( rscb[1] <<  8 ) |   rscb[0];
+		ctx->dec.tsc_hi = ( ( rscb[5] << 24 ) | ( rscb[4] << 16 ) |
+				    ( rscb[3] <<  8 ) |   rscb[2] );
+	}
+
+	memcpy ( &ctx->tk, key, sizeof ( ctx->tk ) );
+
+	return 0;
+}
+
+/**
+ * Perform TKIP key mixing, phase 1
+ *
+ * @v dctx	TKIP directional context
+ * @v tk	TKIP temporal key
+ * @v mac	MAC address of transmitter
+ *
+ * This recomputes the TTAK in @a dctx if necessary, and sets
+ * @c dctx->ttak_ok.
+ */
+static void tkip_mix_1 ( struct tkip_dir_ctx *dctx, struct tkip_tk *tk, u8 *mac )
+{
+	int i, j;
+
+	if ( dctx->ttak_ok && ! memcmp ( mac, dctx->mac, ETH_ALEN ) )
+		return;
+
+	memcpy ( dctx->mac, mac, ETH_ALEN );
+
+	dctx->ttak[0] = dctx->tsc_hi & 0xFFFF;
+	dctx->ttak[1] = dctx->tsc_hi >> 16;
+	dctx->ttak[2] = ( mac[1] << 8 ) | mac[0];
+	dctx->ttak[3] = ( mac[3] << 8 ) | mac[2];
+	dctx->ttak[4] = ( mac[5] << 8 ) | mac[4];
+
+	for ( i = 0; i < 8; i++ ) {
+		j = 2 * ( i & 1 );
+
+		dctx->ttak[0] += S ( dctx->ttak[4] ^ ( ( tk->key[1 + j] << 8 ) |
+						         tk->key[0 + j] ) );
+		dctx->ttak[1] += S ( dctx->ttak[0] ^ ( ( tk->key[5 + j] << 8 ) |
+						         tk->key[4 + j] ) );
+		dctx->ttak[2] += S ( dctx->ttak[1] ^ ( ( tk->key[9 + j] << 8 ) |
+						         tk->key[8 + j] ) );
+		dctx->ttak[3] += S ( dctx->ttak[2] ^ ( ( tk->key[13+ j] << 8 ) |
+						         tk->key[12+ j] ) );
+		dctx->ttak[4] += S ( dctx->ttak[3] ^ ( ( tk->key[1 + j] << 8 ) |
+						         tk->key[0 + j] ) ) + i;
+	}
+
+	dctx->ttak_ok = 1;
+}
+
+/**
+ * Perform TKIP key mixing, phase 2
+ *
+ * @v dctx	TKIP directional context
+ * @v tk	TKIP temporal key
+ * @ret key	ARC4 key, 16 bytes long
+ */
+static void tkip_mix_2 ( struct tkip_dir_ctx *dctx, struct tkip_tk *tk,
+			 void *key )
+{
+	u8 *kb = key;
+	u16 ppk[6];
+	int i;
+
+	memcpy ( ppk, dctx->ttak, sizeof ( dctx->ttak ) );
+	ppk[5] = dctx->ttak[4] + dctx->tsc_lo;
+
+	ppk[0] += S ( ppk[5] ^ ( ( tk->key[1] << 8 ) | tk->key[0] ) );
+	ppk[1] += S ( ppk[0] ^ ( ( tk->key[3] << 8 ) | tk->key[2] ) );
+	ppk[2] += S ( ppk[1] ^ ( ( tk->key[5] << 8 ) | tk->key[4] ) );
+	ppk[3] += S ( ppk[2] ^ ( ( tk->key[7] << 8 ) | tk->key[6] ) );
+	ppk[4] += S ( ppk[3] ^ ( ( tk->key[9] << 8 ) | tk->key[8] ) );
+	ppk[5] += S ( ppk[4] ^ ( ( tk->key[11] << 8 ) | tk->key[10] ) );
+
+	ppk[0] += ror16 ( ppk[5] ^ ( ( tk->key[13] << 8 ) | tk->key[12] ), 1 );
+	ppk[1] += ror16 ( ppk[0] ^ ( ( tk->key[15] << 8 ) | tk->key[14] ), 1 );
+	ppk[2] += ror16 ( ppk[1], 1 );
+	ppk[3] += ror16 ( ppk[2], 1 );
+	ppk[4] += ror16 ( ppk[3], 1 );
+	ppk[5] += ror16 ( ppk[4], 1 );
+
+	kb[0] = dctx->tsc_lo >> 8;
+	kb[1] = ( ( dctx->tsc_lo >> 8 ) | 0x20 ) & 0x7F;
+	kb[2] = dctx->tsc_lo & 0xFF;
+	kb[3] = ( ( ppk[5] ^ ( ( tk->key[1] << 8 ) | tk->key[0] ) ) >> 1 )
+		& 0xFF;
+
+	for ( i = 0; i < 6; i++ ) {
+		kb[4 + 2*i] = ppk[i] & 0xFF;
+		kb[5 + 2*i] = ppk[i] >> 8;
+	}
+}
+
+/**
+ * Update Michael message integrity code based on next 32-bit word of data
+ *
+ * @v V		Michael code state (two 32-bit words)
+ * @v word	Next 32-bit word of data
+ */
+static void tkip_feed_michael ( u32 *V, u32 word )
+{
+	V[0] ^= word;
+	V[1] ^= rol32 ( V[0], 17 );
+	V[0] += V[1];
+	V[1] ^= ( ( V[0] & 0xFF00FF00 ) >> 8 ) | ( ( V[0] & 0x00FF00FF ) << 8 );
+	V[0] += V[1];
+	V[1] ^= rol32 ( V[0], 3 );
+	V[0] += V[1];
+	V[1] ^= ror32 ( V[0], 2 );
+	V[0] += V[1];
+}
+
+/**
+ * Calculate Michael message integrity code
+ *
+ * @v key	MIC key to use (8 bytes)
+ * @v da	Destination link-layer address
+ * @v sa	Source link-layer address
+ * @v data	Start of data to calculate over
+ * @v len	Length of header + data
+ * @ret mic	Calculated Michael MIC (8 bytes)
+ */
+static void tkip_michael ( const void *key, const void *da, const void *sa,
+			   const void *data, size_t len, void *mic )
+{
+	u32 V[2];		/* V[0] = "l", V[1] = "r" in 802.11 */
+	union {
+		u8 byte[12];
+		u32 word[3];
+	} cap;
+	const u8 *ptr = data;
+	const u8 *end = ptr + len;
+	int i;
+
+	memcpy ( V, key, sizeof ( V ) );
+	V[0] = le32_to_cpu ( V[0] );
+	V[1] = le32_to_cpu ( V[1] );
+
+	/* Feed in header (we assume non-QoS, so Priority = 0) */
+	memcpy ( &cap.byte[0], da, ETH_ALEN );
+	memcpy ( &cap.byte[6], sa, ETH_ALEN );
+	tkip_feed_michael ( V, le32_to_cpu ( cap.word[0] ) );
+	tkip_feed_michael ( V, le32_to_cpu ( cap.word[1] ) );
+	tkip_feed_michael ( V, le32_to_cpu ( cap.word[2] ) );
+	tkip_feed_michael ( V, 0 );
+
+	/* Feed in data */
+	while ( ptr + 4 <= end ) {
+		tkip_feed_michael ( V, le32_to_cpu ( *( u32 * ) ptr ) );
+		ptr += 4;
+	}
+
+	/* Add unaligned part and padding */
+	for ( i = 0; ptr < end; i++ )
+		cap.byte[i] = *ptr++;
+	cap.byte[i++] = 0x5a;
+	for ( ; i < 8; i++ )
+		cap.byte[i] = 0;
+
+	/* Feed in padding */
+	tkip_feed_michael ( V, le32_to_cpu ( cap.word[0] ) );
+	tkip_feed_michael ( V, le32_to_cpu ( cap.word[1] ) );
+
+	/* Output MIC */
+	V[0] = cpu_to_le32 ( V[0] );
+	V[1] = cpu_to_le32 ( V[1] );
+	memcpy ( mic, V, sizeof ( V ) );
+}
+
+/**
+ * Encrypt a packet using TKIP
+ *
+ * @v crypto	TKIP cryptosystem
+ * @v iob	I/O buffer containing cleartext packet
+ * @ret eiob	I/O buffer containing encrypted packet
+ */
+static struct io_buffer * tkip_encrypt ( struct net80211_crypto *crypto,
+					 struct io_buffer *iob )
+{
+	struct tkip_ctx *ctx = crypto->priv;
+	struct ieee80211_frame *hdr = iob->data;
+	struct io_buffer *eiob;
+	struct arc4_ctx arc4;
+	u8 key[16];
+	struct tkip_head head;
+	u8 mic[8];
+	u32 icv;
+	const int hdrlen = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int datalen = iob_len ( iob ) - hdrlen;
+
+	ctx->enc.tsc_lo++;
+	if ( ctx->enc.tsc_lo == 0 ) {
+		ctx->enc.tsc_hi++;
+		ctx->enc.ttak_ok = 0;
+	}
+
+	tkip_mix_1 ( &ctx->enc, &ctx->tk, hdr->addr2 );
+	tkip_mix_2 ( &ctx->enc, &ctx->tk, key );
+
+	eiob = alloc_iob ( iob_len ( iob ) + TKIP_HEAD_LEN + TKIP_FOOT_LEN );
+	if ( ! eiob )
+		return NULL;
+
+	/* Copy frame header */
+	memcpy ( iob_put ( eiob, hdrlen ), iob->data, hdrlen );
+	hdr = eiob->data;
+	hdr->fc |= IEEE80211_FC_PROTECTED;
+
+	/* Fill in IV and key ID byte, and extended IV */
+	memcpy ( &head, key, 3 );
+	head.kid = 0x20;		/* have Extended IV, key ID 0 */
+	head.tsc_hi = cpu_to_le32 ( ctx->enc.tsc_hi );
+	memcpy ( iob_put ( eiob, sizeof ( head ) ), &head, sizeof ( head ) );
+
+	/* Copy and encrypt the data */
+	cipher_setkey ( &arc4_algorithm, &arc4, key, 16 );
+	cipher_encrypt ( &arc4_algorithm, &arc4, iob->data + hdrlen,
+			 iob_put ( eiob, datalen ), datalen );
+
+	/* Add MIC */
+	hdr = iob->data;
+	tkip_michael ( &ctx->tk.mic.tx, hdr->addr3, hdr->addr2,
+		       iob->data + hdrlen, datalen, mic );
+	cipher_encrypt ( &arc4_algorithm, &arc4, mic,
+			 iob_put ( eiob, sizeof ( mic ) ), sizeof ( mic ) );
+
+	/* Add ICV */
+	icv = crc32_le ( ~0, iob->data + hdrlen, datalen );
+	icv = crc32_le ( icv, mic, sizeof ( mic ) );
+	icv = cpu_to_le32 ( ~icv );
+	cipher_encrypt ( &arc4_algorithm, &arc4, &icv,
+			 iob_put ( eiob, TKIP_ICV_LEN ), TKIP_ICV_LEN );
+
+	DBGC2 ( ctx, "WPA-TKIP %p: encrypted packet %p -> %p\n", ctx,
+		iob, eiob );
+
+	return eiob;
+}
+
+/**
+ * Decrypt a packet using TKIP
+ *
+ * @v crypto	TKIP cryptosystem
+ * @v eiob	I/O buffer containing encrypted packet
+ * @ret iob	I/O buffer containing cleartext packet
+ */
+static struct io_buffer * tkip_decrypt ( struct net80211_crypto *crypto,
+					 struct io_buffer *eiob )
+{
+	struct tkip_ctx *ctx = crypto->priv;
+	struct ieee80211_frame *hdr;
+	struct io_buffer *iob;
+	const int hdrlen = IEEE80211_TYP_FRAME_HEADER_LEN;
+	int datalen = iob_len ( eiob ) - hdrlen - TKIP_HEAD_LEN - TKIP_FOOT_LEN;
+	struct tkip_head *head;
+	struct arc4_ctx arc4;
+	u16 rx_tsc_lo;
+	u8 key[16];
+	u8 mic[8];
+	u32 icv, crc;
+
+	iob = alloc_iob ( hdrlen + datalen + TKIP_FOOT_LEN );
+	if ( ! iob )
+		return NULL;
+
+	/* Copy frame header */
+	memcpy ( iob_put ( iob, hdrlen ), eiob->data, hdrlen );
+	hdr = iob->data;
+	hdr->fc &= ~IEEE80211_FC_PROTECTED;
+
+	/* Check and update TSC */
+	head = eiob->data + hdrlen;
+	rx_tsc_lo = ( head->tsc1 << 8 ) | head->tsc0;
+
+	if ( head->tsc_hi < ctx->dec.tsc_hi ||
+	     ( head->tsc_hi == ctx->dec.tsc_hi &&
+	       rx_tsc_lo <= ctx->dec.tsc_lo ) ) {
+		DBGC ( ctx, "WPA-TKIP %p: packet received out of order "
+		       "(%08x:%04x <= %08x:%04x)\n", ctx, head->tsc_hi,
+		       rx_tsc_lo, ctx->dec.tsc_hi, ctx->dec.tsc_lo );
+		free_iob ( iob );
+		return NULL;
+	}
+	ctx->dec.tsc_lo = rx_tsc_lo;
+	if ( ctx->dec.tsc_hi != head->tsc_hi ) {
+		ctx->dec.ttak_ok = 0;
+		ctx->dec.tsc_hi = head->tsc_hi;
+	}
+
+	/* Calculate key */
+	tkip_mix_1 ( &ctx->dec, &ctx->tk, hdr->addr2 );
+	tkip_mix_2 ( &ctx->dec, &ctx->tk, key );
+
+	/* Copy-decrypt data, MIC, ICV */
+	cipher_setkey ( &arc4_algorithm, &arc4, key, 16 );
+	cipher_decrypt ( &arc4_algorithm, &arc4,
+			 eiob->data + hdrlen + TKIP_HEAD_LEN,
+			 iob_put ( iob, datalen ), datalen + TKIP_FOOT_LEN );
+
+	/* Check ICV */
+	icv = le32_to_cpu ( *( u32 * ) ( iob->tail + TKIP_MIC_LEN ) );
+	crc = ~crc32_le ( ~0, iob->data + hdrlen, datalen + TKIP_MIC_LEN );
+	if ( crc != icv ) {
+		DBGC ( ctx, "WPA-TKIP %p CRC mismatch: expect %08x, get %08x\n",
+		       ctx, icv, crc );
+		free_iob ( iob );
+		return NULL;
+	}
+
+	/* Check MIC */
+	tkip_michael ( &ctx->tk.mic.rx, hdr->addr1, hdr->addr3,
+		       iob->data + hdrlen, datalen, mic );
+	if ( memcmp ( mic, iob->tail, TKIP_MIC_LEN ) != 0 ) {
+		DBGC ( ctx, "WPA-TKIP %p ALERT! MIC failure\n", ctx );
+		/* XXX we should do the countermeasures here */
+		free_iob ( iob );
+		return NULL;
+	}
+
+	DBGC2 ( ctx, "WPA-TKIP %p: decrypted packet %p -> %p\n", ctx,
+		eiob, iob );
+
+	return iob;
+}
+
+/** TKIP cryptosystem */
+struct net80211_crypto tkip_crypto __net80211_crypto = {
+	.algorithm = NET80211_CRYPT_TKIP,
+	.init = tkip_init,
+	.encrypt = tkip_encrypt,
+	.decrypt = tkip_decrypt,
+	.priv_len = sizeof ( struct tkip_ctx ),
+};
+
+
+
+
+/**
+ * Calculate HMAC-MD5 MIC for EAPOL-Key frame
+ *
+ * @v kck	Key Confirmation Key, 16 bytes
+ * @v msg	Message to calculate MIC over
+ * @v len	Number of bytes to calculate MIC over
+ * @ret mic	Calculated MIC, 16 bytes long
+ */
+static void tkip_kie_mic ( const void *kck, const void *msg, size_t len,
+			   void *mic )
+{
+	struct md5_ctx md5;
+	u8 kckb[16];
+	size_t kck_len = 16;
+
+	memcpy ( kckb, kck, kck_len );
+
+	hmac_init ( &md5_algorithm, &md5, kckb, &kck_len );
+	hmac_update ( &md5_algorithm, &md5, msg, len );
+	hmac_final ( &md5_algorithm, &md5, kckb, &kck_len, mic );
+}
+
+/**
+ * Decrypt key data in EAPOL-Key frame
+ *
+ * @v kek	Key Encryption Key, 16 bytes
+ * @v iv	Initialisation vector, 16 bytes
+ * @v msg	Message to decrypt
+ * @v len	Length of message
+ * @ret msg	Decrypted message in place of original
+ * @ret len	Unchanged
+ * @ret rc	Always 0 for success
+ */
+static int tkip_kie_decrypt ( const void *kek, const void *iv,
+			      void *msg, u16 *len )
+{
+	u8 key[32];
+	memcpy ( key, iv, 16 );
+	memcpy ( key + 16, kek, 16 );
+
+	arc4_skip ( key, 32, 256, msg, msg, *len );
+
+	return 0;
+}
+
+
+/** TKIP-style key integrity and encryption handler */
+struct wpa_kie tkip_kie __wpa_kie = {
+	.version = EAPOL_KEY_VERSION_WPA,
+	.mic = tkip_kie_mic,
+	.decrypt = tkip_kie_decrypt,
+};
diff --git a/gpxe/src/net/aoe.c b/gpxe/src/net/aoe.c
index 08887fe0..839a875b 100644
--- a/gpxe/src/net/aoe.c
+++ b/gpxe/src/net/aoe.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stddef.h>
 #include <string.h>
 #include <stdio.h>
@@ -66,6 +68,7 @@ static void aoe_done ( struct aoe_session *aoe, int rc ) {
 	/* Record overall command status */
 	if ( aoe->command ) {
 		aoe->command->cb.cmd_stat = aoe->status;
+		aoe->command->rc = rc;
 		aoe->command = NULL;
 	}
 
@@ -354,7 +357,6 @@ static int aoe_command ( struct ata_device *ata,
 			 struct ata_command *command ) {
 	struct aoe_session *aoe =
 		container_of ( ata->backend, struct aoe_session, refcnt );
-	int rc;
 
 	aoe->command = command;
 	aoe->status = 0;
@@ -363,15 +365,9 @@ static int aoe_command ( struct ata_device *ata,
 
 	aoe_send_command ( aoe );
 
-	aoe->rc = -EINPROGRESS;
-	while ( aoe->rc == -EINPROGRESS )
-		step();
-	rc = aoe->rc;
-
-	return rc;
+	return 0;
 }
 
-
 /**
  * Issue AoE config query for AoE target discovery
  *
@@ -444,8 +440,7 @@ int aoe_attach ( struct ata_device *ata, struct net_device *netdev,
 		return -ENOMEM;
 	aoe->refcnt.free = aoe_free;
 	aoe->netdev = netdev_get ( netdev );
-	memcpy ( aoe->target, ethernet_protocol.ll_broadcast,
-		 sizeof ( aoe->target ) );
+	memcpy ( aoe->target, netdev->ll_broadcast, sizeof ( aoe->target ) );
 	aoe->tag = AOE_TAG_MAGIC;
 	aoe->timer.expired = aoe_timer_expired;
 
diff --git a/gpxe/src/net/arp.c b/gpxe/src/net/arp.c
index ba9ebf48..124a856e 100644
--- a/gpxe/src/net/arp.c
+++ b/gpxe/src/net/arp.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <string.h>
 #include <byteswap.h>
@@ -36,12 +38,6 @@
  *
  */
 
-/** Registered ARP protocols */
-static struct arp_net_protocol arp_net_protocols[0]
-	__table_start ( struct arp_net_protocol, arp_net_protocols );
-static struct arp_net_protocol arp_net_protocols_end[0]
-	__table_end ( struct arp_net_protocol, arp_net_protocols );
-
 /** An ARP cache entry */
 struct arp_entry {
 	/** Network-layer protocol */
@@ -160,7 +156,7 @@ int arp_resolve ( struct net_device *netdev, struct net_protocol *net_protocol,
 
 	/* Transmit ARP request */
 	if ( ( rc = net_tx ( iobuf, netdev, &arp_protocol, 
-			     ll_protocol->ll_broadcast ) ) != 0 )
+			     netdev->ll_broadcast ) ) != 0 )
 		return rc;
 
 	return -ENOENT;
@@ -176,8 +172,7 @@ int arp_resolve ( struct net_device *netdev, struct net_protocol *net_protocol,
 static struct arp_net_protocol * arp_find_protocol ( uint16_t net_proto ) {
 	struct arp_net_protocol *arp_net_protocol;
 
-	for ( arp_net_protocol = arp_net_protocols ;
-	      arp_net_protocol < arp_net_protocols_end ; arp_net_protocol++ ) {
+	for_each_table_entry ( arp_net_protocol, ARP_NET_PROTOCOLS ) {
 		if ( arp_net_protocol->net_protocol->net_proto == net_proto ) {
 			return arp_net_protocol;
 		}
diff --git a/gpxe/src/net/cachedhcp.c b/gpxe/src/net/cachedhcp.c
new file mode 100644
index 00000000..37f344b6
--- /dev/null
+++ b/gpxe/src/net/cachedhcp.c
@@ -0,0 +1,76 @@
+/*
+ * Copyright (C) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <gpxe/dhcp.h>
+#include <gpxe/dhcppkt.h>
+#include <gpxe/netdevice.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/uaccess.h>
+
+/** @file
+ *
+ * Cached DHCP packet handling
+ *
+ */
+
+/**
+ * Store cached DHCPACK packet
+ *
+ * @v data		User pointer to cached DHCP packet data
+ * @v len		Length of cached DHCP packet data
+ * @ret rc		Return status code
+ *
+ * This function should be called by the architecture-specific
+ * get_cached_dhcpack() handler.
+ */
+void store_cached_dhcpack ( userptr_t data, size_t len ) {
+	struct dhcp_packet *dhcppkt;
+	struct dhcphdr *dhcphdr;
+	struct settings *parent;
+	int rc;
+
+	/* Create DHCP packet */
+	dhcppkt = zalloc ( sizeof ( *dhcppkt ) + len );
+	if ( ! dhcppkt )
+		return;
+
+	/* Fill in data for DHCP packet */
+	dhcphdr = ( ( ( void * ) dhcppkt ) + sizeof ( * dhcppkt ) );
+	copy_from_user ( dhcphdr, data, 0, len );
+	dhcppkt_init ( dhcppkt, dhcphdr, len );
+	DBG_HD ( dhcppkt->options.data, dhcppkt->options.len );
+
+	/* Register settings on the last opened network device.
+	 * This will have the effect of registering cached settings
+	 * with a network device when "dhcp netX" is performed for that
+	 * device, which is usually what we want.
+	 */
+	parent = netdev_settings ( last_opened_netdev() );
+	if ( ( rc = register_settings ( &dhcppkt->settings, parent ) ) != 0 )
+		DBG ( "DHCP could not register cached settings: %s\n",
+		      strerror ( rc ) );
+
+	dhcppkt_put ( dhcppkt );
+
+	DBG ( "DHCP registered cached settings\n" );
+}
diff --git a/gpxe/src/net/dhcpopts.c b/gpxe/src/net/dhcpopts.c
index 1898011a..6482c627 100644
--- a/gpxe/src/net/dhcpopts.c
+++ b/gpxe/src/net/dhcpopts.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -101,7 +103,7 @@ static unsigned int dhcp_option_len ( struct dhcp_option *option ) {
  * DHCP option block.  Encapsulated options may be searched for by
  * using DHCP_ENCAP_OPT() to construct the tag value.
  *
- * If the option is encapsulated, and @c encapsulator is non-NULL, it
+ * If the option is encapsulated, and @c encap_offset is non-NULL, it
  * will be filled in with the offset of the encapsulating option.
  *
  * This routine is designed to be paranoid.  It does not assume that
@@ -134,8 +136,15 @@ static int find_dhcp_option_with_encap ( struct dhcp_options *options,
 		if ( remaining < 0 )
 			break;
 		/* Check for explicit end marker */
-		if ( option->tag == DHCP_END )
-			break;
+		if ( option->tag == DHCP_END ) {
+			if ( tag == DHCP_END )
+				/* Special case where the caller is interested
+				 * in whether we have this marker or not.
+				 */
+				return offset;
+			else
+				break;
+		}
 		/* Check for matching tag */
 		if ( option->tag == tag ) {
 			DBGC ( options, "DHCPOPT %p found %s (length %d)\n",
@@ -254,7 +263,7 @@ static int set_dhcp_option ( struct dhcp_options *options, unsigned int tag,
 	static const uint8_t empty_encapsulator[] = { DHCP_END };
 	int offset;
 	int encap_offset = -1;
-	int creation_offset = 0;
+	int creation_offset;
 	struct dhcp_option *option;
 	unsigned int encap_tag = DHCP_ENCAPSULATOR ( tag );
 	size_t old_len = 0;
@@ -265,6 +274,10 @@ static int set_dhcp_option ( struct dhcp_options *options, unsigned int tag,
 	if ( tag == DHCP_PAD )
 		return -ENOTTY;
 
+	creation_offset = find_dhcp_option_with_encap ( options, DHCP_END,
+							NULL );
+	if ( creation_offset < 0 )
+		creation_offset = options->len;
 	/* Find old instance of this option, if any */
 	offset = find_dhcp_option_with_encap ( options, tag, &encap_offset );
 	if ( offset >= 0 ) {
diff --git a/gpxe/src/net/dhcppkt.c b/gpxe/src/net/dhcppkt.c
index 1f2d373c..20a0e666 100644
--- a/gpxe/src/net/dhcppkt.c
+++ b/gpxe/src/net/dhcppkt.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -155,6 +157,8 @@ int dhcppkt_store ( struct dhcp_packet *dhcppkt, unsigned int tag,
 		memset ( field_data, 0, field->len );
 		memcpy ( dhcp_packet_field ( dhcppkt->dhcphdr, field ),
 			 data, len );
+		/* Erase any equivalent option from the options block */
+		dhcpopt_store ( &dhcppkt->options, tag, NULL, 0 );
 		return 0;
 	}
 
@@ -181,14 +185,16 @@ int dhcppkt_fetch ( struct dhcp_packet *dhcppkt, unsigned int tag,
 		    void *data, size_t len ) {
 	struct dhcp_packet_field *field;
 	void *field_data;
-	size_t field_len;
+	size_t field_len = 0;
 	
-	/* If this is a special field, return it */
+	/* Identify special field, if any */
 	if ( ( field = find_dhcp_packet_field ( tag ) ) != NULL ) {
 		field_data = dhcp_packet_field ( dhcppkt->dhcphdr, field );
 		field_len = field->used_len ( field_data, field->len );
-		if ( ! field_len )
-			return -ENOENT;
+	}
+
+	/* Return special field, if it exists and is populated */
+	if ( field_len ) {
 		if ( len > field_len )
 			len = field_len;
 		memcpy ( data, field_data, len );
diff --git a/gpxe/src/net/eapol.c b/gpxe/src/net/eapol.c
new file mode 100644
index 00000000..507c8ce2
--- /dev/null
+++ b/gpxe/src/net/eapol.c
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2009 Joshua Oreman <oremanj@rwcr.net>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+/** @file
+ *
+ * 802.1X Extensible Authentication Protocol over LANs demultiplexer
+ *
+ */
+
+#include <gpxe/netdevice.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/if_ether.h>
+#include <gpxe/eapol.h>
+#include <errno.h>
+#include <byteswap.h>
+
+/**
+ * Receive EAPOL network-layer packet
+ *
+ * @v iob	I/O buffer
+ * @v netdev	Network device
+ * @v ll_source	Link-layer source address
+ *
+ * This function takes ownership of the I/O buffer passed to it.
+ */
+static int eapol_rx ( struct io_buffer *iob, struct net_device *netdev,
+		      const void *ll_source )
+{
+	struct eapol_frame *eapol = iob->data;
+	struct eapol_handler *handler;
+
+	if ( iob_len ( iob ) < EAPOL_HDR_LEN ) {
+		free_iob ( iob );
+		return -EINVAL;
+	}
+
+	for_each_table_entry ( handler, EAPOL_HANDLERS ) {
+		if ( handler->type == eapol->type ) {
+			iob_pull ( iob, EAPOL_HDR_LEN );
+			return handler->rx ( iob, netdev, ll_source );
+		}
+	}
+
+	free_iob ( iob );
+	return -( ENOTSUP | ( ( eapol->type & 0x1f ) << 8 ) );
+}
+
+/**
+ * Transcribe EAPOL network-layer address
+ *
+ * @v net_addr	Network-layer address
+ * @ret str	String representation of network-layer address
+ *
+ * EAPOL doesn't have network-layer addresses, so we just return the
+ * string @c "<EAPOL>".
+ */
+static const char * eapol_ntoa ( const void *net_addr __unused )
+{
+	return "<EAPOL>";
+}
+
+/** EAPOL network protocol */
+struct net_protocol eapol_protocol __net_protocol = {
+	.name = "EAPOL",
+	.rx = eapol_rx,
+	.ntoa = eapol_ntoa,
+	.net_proto = htons ( ETH_P_EAPOL ),
+};
diff --git a/gpxe/src/net/ethernet.c b/gpxe/src/net/ethernet.c
index b16135a9..79ed1dc6 100644
--- a/gpxe/src/net/ethernet.c
+++ b/gpxe/src/net/ethernet.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdio.h>
 #include <string.h>
@@ -41,13 +43,15 @@ static uint8_t eth_broadcast[ETH_ALEN] = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff};
 /**
  * Add Ethernet link-layer header
  *
+ * @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
  */
-static int eth_push ( struct io_buffer *iobuf, const void *ll_dest,
+static int eth_push ( struct net_device *netdev __unused,
+		      struct io_buffer *iobuf, const void *ll_dest,
 		      const void *ll_source, uint16_t net_proto ) {
 	struct ethhdr *ethhdr = iob_push ( iobuf, sizeof ( *ethhdr ) );
 
@@ -62,13 +66,15 @@ static int eth_push ( struct io_buffer *iobuf, const void *ll_dest,
 /**
  * Remove Ethernet link-layer header
  *
+ * @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 rc		Return status code
  */
-static int eth_pull ( struct io_buffer *iobuf, const void **ll_dest,
+static int eth_pull ( struct net_device *netdev __unused, 
+		      struct io_buffer *iobuf, const void **ll_dest,
 		      const void **ll_source, uint16_t *net_proto ) {
 	struct ethhdr *ethhdr = iobuf->data;
 
@@ -91,6 +97,16 @@ static int eth_pull ( struct io_buffer *iobuf, const void **ll_dest,
 }
 
 /**
+ * Initialise Ethernet address
+ *
+ * @v hw_addr		Hardware address
+ * @v ll_addr		Link-layer address
+ */
+void eth_init_addr ( const void *hw_addr, void *ll_addr ) {
+	memcpy ( ll_addr, hw_addr, ETH_ALEN );
+}
+
+/**
  * Transcribe Ethernet address
  *
  * @v ll_addr		Link-layer address
@@ -114,8 +130,7 @@ const char * eth_ntoa ( const void *ll_addr ) {
  * @v ll_addr		Link-layer address to fill in
  * @ret rc		Return status code
  */
-static int eth_mc_hash ( unsigned int af, const void *net_addr,
-			 void *ll_addr ) {
+int eth_mc_hash ( unsigned int af, const void *net_addr, void *ll_addr ) {
 	const uint8_t *net_addr_bytes = net_addr;
 	uint8_t *ll_addr_bytes = ll_addr;
 
@@ -133,15 +148,46 @@ static int eth_mc_hash ( unsigned int af, const void *net_addr,
 	}
 }
 
+/**
+ * Generate Ethernet-compatible compressed link-layer address
+ *
+ * @v ll_addr		Link-layer address
+ * @v eth_addr		Ethernet-compatible address to fill in
+ */
+int eth_eth_addr ( const void *ll_addr, void *eth_addr ) {
+	memcpy ( eth_addr, ll_addr, ETH_ALEN );
+	return 0;
+}
+
 /** Ethernet protocol */
 struct ll_protocol ethernet_protocol __ll_protocol = {
 	.name		= "Ethernet",
 	.ll_proto	= htons ( ARPHRD_ETHER ),
+	.hw_addr_len	= ETH_ALEN,
 	.ll_addr_len	= ETH_ALEN,
 	.ll_header_len	= ETH_HLEN,
-	.ll_broadcast	= eth_broadcast,
 	.push		= eth_push,
 	.pull		= eth_pull,
+	.init_addr	= eth_init_addr,
 	.ntoa		= eth_ntoa,
 	.mc_hash	= eth_mc_hash,
+	.eth_addr	= eth_eth_addr,
 };
+
+/**
+ * Allocate Ethernet device
+ *
+ * @v priv_size		Size of driver private data
+ * @ret netdev		Network device, or NULL
+ */
+struct net_device * alloc_etherdev ( size_t priv_size ) {
+	struct net_device *netdev;
+
+	netdev = alloc_netdev ( priv_size );
+	if ( netdev ) {
+		netdev->ll_protocol = &ethernet_protocol;
+		netdev->ll_broadcast = eth_broadcast;
+		netdev->max_pkt_len = ETH_FRAME_LEN;
+	}
+	return netdev;
+}
diff --git a/gpxe/src/net/fakedhcp.c b/gpxe/src/net/fakedhcp.c
index 0518789c..ad3f046f 100644
--- a/gpxe/src/net/fakedhcp.c
+++ b/gpxe/src/net/fakedhcp.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
diff --git a/gpxe/src/net/icmp.c b/gpxe/src/net/icmp.c
index 3e45c1f6..749c3454 100644
--- a/gpxe/src/net/icmp.c
+++ b/gpxe/src/net/icmp.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <string.h>
 #include <errno.h>
 #include <gpxe/iobuf.h>
diff --git a/gpxe/src/net/infiniband.c b/gpxe/src/net/infiniband.c
index d79bdc2c..d7813249 100644
--- a/gpxe/src/net/infiniband.c
+++ b/gpxe/src/net/infiniband.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -25,12 +27,15 @@
 #include <errno.h>
 #include <assert.h>
 #include <gpxe/list.h>
+#include <gpxe/errortab.h>
 #include <gpxe/if_arp.h>
 #include <gpxe/netdevice.h>
 #include <gpxe/iobuf.h>
 #include <gpxe/ipoib.h>
 #include <gpxe/process.h>
 #include <gpxe/infiniband.h>
+#include <gpxe/ib_mi.h>
+#include <gpxe/ib_sma.h>
 
 /** @file
  *
@@ -41,6 +46,26 @@
 /** List of Infiniband devices */
 struct list_head ib_devices = LIST_HEAD_INIT ( ib_devices );
 
+/** List of open Infiniband devices, in reverse order of opening */
+static struct list_head open_ib_devices = LIST_HEAD_INIT ( open_ib_devices );
+
+/* Disambiguate the various possible EINPROGRESSes */
+#define EINPROGRESS_INIT ( EINPROGRESS | EUNIQ_01 )
+#define EINPROGRESS_ARMED ( EINPROGRESS | EUNIQ_02 )
+
+/** Human-readable message for the link statuses */
+struct errortab infiniband_errors[] __errortab = {
+	{ EINPROGRESS_INIT, "Initialising" },
+	{ EINPROGRESS_ARMED, "Armed" },
+};
+
+/***************************************************************************
+ *
+ * Completion queues
+ *
+ ***************************************************************************
+ */
+
 /**
  * Create completion queue
  *
@@ -61,6 +86,8 @@ ib_create_cq ( struct ib_device *ibdev, unsigned int num_cqes,
 	cq = zalloc ( sizeof ( *cq ) );
 	if ( ! cq )
 		goto err_alloc_cq;
+	cq->ibdev = ibdev;
+	list_add ( &cq->list, &ibdev->cqs );
 	cq->num_cqes = num_cqes;
 	INIT_LIST_HEAD ( &cq->work_queues );
 	cq->op = op;
@@ -79,6 +106,7 @@ ib_create_cq ( struct ib_device *ibdev, unsigned int num_cqes,
 
 	ibdev->op->destroy_cq ( ibdev, cq );
  err_dev_create_cq:
+	list_del ( &cq->list );
 	free ( cq );
  err_alloc_cq:
 	return NULL;
@@ -96,26 +124,57 @@ void ib_destroy_cq ( struct ib_device *ibdev,
 	       ibdev, cq->cqn );
 	assert ( list_empty ( &cq->work_queues ) );
 	ibdev->op->destroy_cq ( ibdev, cq );
+	list_del ( &cq->list );
 	free ( cq );
 }
 
 /**
+ * Poll completion queue
+ *
+ * @v ibdev		Infiniband device
+ * @v cq		Completion queue
+ */
+void ib_poll_cq ( struct ib_device *ibdev,
+		  struct ib_completion_queue *cq ) {
+	struct ib_work_queue *wq;
+
+	/* Poll completion queue */
+	ibdev->op->poll_cq ( ibdev, cq );
+
+	/* Refill receive work queues */
+	list_for_each_entry ( wq, &cq->work_queues, list ) {
+		if ( ! wq->is_send )
+			ib_refill_recv ( ibdev, wq->qp );
+	}
+}
+
+/***************************************************************************
+ *
+ * Work queues
+ *
+ ***************************************************************************
+ */
+
+/**
  * Create queue pair
  *
  * @v ibdev		Infiniband device
+ * @v type		Queue pair type
  * @v num_send_wqes	Number of send work queue entries
  * @v send_cq		Send completion queue
  * @v num_recv_wqes	Number of receive work queue entries
  * @v recv_cq		Receive completion queue
- * @v qkey		Queue key
  * @ret qp		Queue pair
+ *
+ * The queue pair will be left in the INIT state; you must call
+ * ib_modify_qp() before it is ready to use for sending and receiving.
  */
 struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
+				      enum ib_queue_pair_type type,
 				      unsigned int num_send_wqes,
 				      struct ib_completion_queue *send_cq,
 				      unsigned int num_recv_wqes,
-				      struct ib_completion_queue *recv_cq,
-				      unsigned long qkey ) {
+				      struct ib_completion_queue *recv_cq ) {
 	struct ib_queue_pair *qp;
 	size_t total_size;
 	int rc;
@@ -131,16 +190,18 @@ struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
 		goto err_alloc_qp;
 	qp->ibdev = ibdev;
 	list_add ( &qp->list, &ibdev->qps );
-	qp->qkey = qkey;
+	qp->type = type;
 	qp->send.qp = qp;
 	qp->send.is_send = 1;
 	qp->send.cq = send_cq;
 	list_add ( &qp->send.list, &send_cq->work_queues );
+	qp->send.psn = ( random() & 0xffffffUL );
 	qp->send.num_wqes = num_send_wqes;
 	qp->send.iobufs = ( ( ( void * ) qp ) + sizeof ( *qp ) );
 	qp->recv.qp = qp;
 	qp->recv.cq = recv_cq;
 	list_add ( &qp->recv.list, &recv_cq->work_queues );
+	qp->recv.psn = ( random() & 0xffffffUL );
 	qp->recv.num_wqes = num_recv_wqes;
 	qp->recv.iobufs = ( ( ( void * ) qp ) + sizeof ( *qp ) +
 			    ( num_send_wqes * sizeof ( qp->send.iobufs[0] ) ));
@@ -152,7 +213,6 @@ struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
 		       "%s\n", ibdev, strerror ( rc ) );
 		goto err_dev_create_qp;
 	}
-
 	DBGC ( ibdev, "IBDEV %p created queue pair %p (%p) with QPN %#lx\n",
 	       ibdev, qp, ib_qp_get_drvdata ( qp ), qp->qpn );
 	DBGC ( ibdev, "IBDEV %p QPN %#lx has %d send entries at [%p,%p)\n",
@@ -161,6 +221,24 @@ struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
 	DBGC ( ibdev, "IBDEV %p QPN %#lx has %d receive entries at [%p,%p)\n",
 	       ibdev, qp->qpn, num_recv_wqes, qp->recv.iobufs,
 	       ( ( ( void * ) qp ) + total_size ) );
+
+	/* Calculate externally-visible QPN */
+	switch ( type ) {
+	case IB_QPT_SMI:
+		qp->ext_qpn = IB_QPN_SMI;
+		break;
+	case IB_QPT_GSI:
+		qp->ext_qpn = IB_QPN_GSI;
+		break;
+	default:
+		qp->ext_qpn = qp->qpn;
+		break;
+	}
+	if ( qp->ext_qpn != qp->qpn ) {
+		DBGC ( ibdev, "IBDEV %p QPN %#lx has external QPN %#lx\n",
+		       ibdev, qp->qpn, qp->ext_qpn );
+	}
+
 	return qp;
 
 	ibdev->op->destroy_qp ( ibdev, qp );
@@ -178,20 +256,15 @@ struct ib_queue_pair * ib_create_qp ( struct ib_device *ibdev,
  *
  * @v ibdev		Infiniband device
  * @v qp		Queue pair
- * @v mod_list		Modification list
- * @v qkey		New queue key, if applicable
+ * @v av		New address vector, if applicable
  * @ret rc		Return status code
  */
-int ib_modify_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp,
-		   unsigned long mod_list, unsigned long qkey ) {
+int ib_modify_qp ( struct ib_device *ibdev, struct ib_queue_pair *qp ) {
 	int rc;
 
 	DBGC ( ibdev, "IBDEV %p modifying QPN %#lx\n", ibdev, qp->qpn );
 
-	if ( mod_list & IB_MODIFY_QKEY )
-		qp->qkey = qkey;
-
-	if ( ( rc = ibdev->op->modify_qp ( ibdev, qp, mod_list ) ) != 0 ) {
+	if ( ( rc = ibdev->op->modify_qp ( ibdev, qp ) ) != 0 ) {
 		DBGC ( ibdev, "IBDEV %p could not modify QPN %#lx: %s\n",
 		       ibdev, qp->qpn, strerror ( rc ) );
 		return rc;
@@ -251,7 +324,7 @@ struct ib_queue_pair * ib_find_qp_qpn ( struct ib_device *ibdev,
 	struct ib_queue_pair *qp;
 
 	list_for_each_entry ( qp, &ibdev->qps, list ) {
-		if ( qp->qpn == qpn )
+		if ( ( qpn == qp->qpn ) || ( qpn == qp->ext_qpn ) )
 			return qp;
 	}
 	return NULL;
@@ -311,6 +384,7 @@ struct ib_work_queue * ib_find_wq ( struct ib_completion_queue *cq,
 int ib_post_send ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 		   struct ib_address_vector *av,
 		   struct io_buffer *iobuf ) {
+	struct ib_address_vector av_copy;
 	int rc;
 
 	/* Check queue fill level */
@@ -320,6 +394,20 @@ int ib_post_send ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 		return -ENOBUFS;
 	}
 
+	/* Use default address vector if none specified */
+	if ( ! av )
+		av = &qp->av;
+
+	/* Make modifiable copy of address vector */
+	memcpy ( &av_copy, av, sizeof ( av_copy ) );
+	av = &av_copy;
+
+	/* Fill in optional parameters in address vector */
+	if ( ! av->qkey )
+		av->qkey = qp->qkey;
+	if ( ! av->rate )
+		av->rate = IB_RATE_2_5;
+
 	/* Post to hardware */
 	if ( ( rc = ibdev->op->post_send ( ibdev, qp, av, iobuf ) ) != 0 ) {
 		DBGC ( ibdev, "IBDEV %p QPN %#lx could not post send WQE: "
@@ -343,6 +431,13 @@ int ib_post_recv ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 		   struct io_buffer *iobuf ) {
 	int rc;
 
+	/* Check packet length */
+	if ( iob_tailroom ( iobuf ) < IB_MAX_PAYLOAD_SIZE ) {
+		DBGC ( ibdev, "IBDEV %p QPN %#lx wrong RX buffer size (%zd)\n",
+		       ibdev, qp->qpn, iob_tailroom ( iobuf ) );
+		return -EINVAL;
+	}
+
 	/* Check queue fill level */
 	if ( qp->recv.fill >= qp->recv.num_wqes ) {
 		DBGC ( ibdev, "IBDEV %p QPN %#lx receive queue full\n",
@@ -371,7 +466,12 @@ int ib_post_recv ( struct ib_device *ibdev, struct ib_queue_pair *qp,
  */
 void ib_complete_send ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 			struct io_buffer *iobuf, int rc ) {
-	qp->send.cq->op->complete_send ( ibdev, qp, iobuf, rc );
+
+	if ( qp->send.cq->op->complete_send ) {
+		qp->send.cq->op->complete_send ( ibdev, qp, iobuf, rc );
+	} else {
+		free_iob ( iobuf );
+	}
 	qp->send.fill--;
 }
 
@@ -387,11 +487,54 @@ void ib_complete_send ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 void ib_complete_recv ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 			struct ib_address_vector *av,
 			struct io_buffer *iobuf, int rc ) {
-	qp->recv.cq->op->complete_recv ( ibdev, qp, av, iobuf, rc );
+
+	if ( qp->recv.cq->op->complete_recv ) {
+		qp->recv.cq->op->complete_recv ( ibdev, qp, av, iobuf, rc );
+	} else {
+		free_iob ( iobuf );
+	}
 	qp->recv.fill--;
 }
 
 /**
+ * Refill receive work queue
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ */
+void ib_refill_recv ( struct ib_device *ibdev, struct ib_queue_pair *qp ) {
+	struct io_buffer *iobuf;
+	int rc;
+
+	/* Keep filling while unfilled entries remain */
+	while ( qp->recv.fill < qp->recv.num_wqes ) {
+
+		/* Allocate I/O buffer */
+		iobuf = alloc_iob ( IB_MAX_PAYLOAD_SIZE );
+		if ( ! iobuf ) {
+			/* Non-fatal; we will refill on next attempt */
+			return;
+		}
+
+		/* Post I/O buffer */
+		if ( ( rc = ib_post_recv ( ibdev, qp, iobuf ) ) != 0 ) {
+			DBGC ( ibdev, "IBDEV %p could not refill: %s\n",
+			       ibdev, strerror ( rc ) );
+			free_iob ( iobuf );
+			/* Give up */
+			return;
+		}
+	}
+}
+
+/***************************************************************************
+ *
+ * Link control
+ *
+ ***************************************************************************
+ */
+
+/**
  * Open port
  *
  * @v ibdev		Infiniband device
@@ -400,16 +543,59 @@ void ib_complete_recv ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 int ib_open ( struct ib_device *ibdev ) {
 	int rc;
 
-	/* Open device if this is the first requested opening */
-	if ( ibdev->open_count == 0 ) {
-		if ( ( rc = ibdev->op->open ( ibdev ) ) != 0 )
-			return rc;
+	/* Increment device open request counter */
+	if ( ibdev->open_count++ > 0 ) {
+		/* Device was already open; do nothing */
+		return 0;
 	}
 
-	/* Increment device open request counter */
-	ibdev->open_count++;
+	/* Create subnet management interface */
+	ibdev->smi = ib_create_mi ( ibdev, IB_QPT_SMI );
+	if ( ! ibdev->smi ) {
+		DBGC ( ibdev, "IBDEV %p could not create SMI\n", ibdev );
+		rc = -ENOMEM;
+		goto err_create_smi;
+	}
 
+	/* Create subnet management agent */
+	if ( ( rc = ib_create_sma ( ibdev, ibdev->smi ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not create SMA: %s\n",
+		       ibdev, strerror ( rc ) );
+		goto err_create_sma;
+	}
+
+	/* Create general services interface */
+	ibdev->gsi = ib_create_mi ( ibdev, IB_QPT_GSI );
+	if ( ! ibdev->gsi ) {
+		DBGC ( ibdev, "IBDEV %p could not create GSI\n", ibdev );
+		rc = -ENOMEM;
+		goto err_create_gsi;
+	}
+
+	/* Open device */
+	if ( ( rc = ibdev->op->open ( ibdev ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not open: %s\n",
+		       ibdev, strerror ( rc ) );
+		goto err_open;
+	}
+
+	/* Add to head of open devices list */
+	list_add ( &ibdev->open_list, &open_ib_devices );
+
+	assert ( ibdev->open_count == 1 );
 	return 0;
+
+	ibdev->op->close ( ibdev );
+ err_open:
+	ib_destroy_mi ( ibdev, ibdev->gsi );
+ err_create_gsi:
+	ib_destroy_sma ( ibdev, ibdev->smi );
+ err_create_sma:
+	ib_destroy_mi ( ibdev, ibdev->smi );
+ err_create_smi:
+	assert ( ibdev->open_count == 1 );
+	ibdev->open_count = 0;
+	return rc;
 }
 
 /**
@@ -423,10 +609,38 @@ void ib_close ( struct ib_device *ibdev ) {
 	ibdev->open_count--;
 
 	/* Close device if this was the last remaining requested opening */
-	if ( ibdev->open_count == 0 )
+	if ( ibdev->open_count == 0 ) {
+		list_del ( &ibdev->open_list );
+		ib_destroy_mi ( ibdev, ibdev->gsi );
+		ib_destroy_sma ( ibdev, ibdev->smi );
+		ib_destroy_mi ( ibdev, ibdev->smi );
 		ibdev->op->close ( ibdev );
+	}
+}
+
+/**
+ * Get link state
+ *
+ * @v ibdev		Infiniband device
+ * @ret rc		Link status code
+ */
+int ib_link_rc ( struct ib_device *ibdev ) {
+	switch ( ibdev->port_state ) {
+	case IB_PORT_STATE_DOWN:	return -ENOTCONN;
+	case IB_PORT_STATE_INIT:	return -EINPROGRESS_INIT;
+	case IB_PORT_STATE_ARMED:	return -EINPROGRESS_ARMED;
+	case IB_PORT_STATE_ACTIVE:	return 0;
+	default:			return -EINVAL;
+	}
 }
 
+/***************************************************************************
+ *
+ * Multicast
+ *
+ ***************************************************************************
+ */
+
 /**
  * Attach to multicast group
  *
@@ -434,6 +648,10 @@ void ib_close ( struct ib_device *ibdev ) {
  * @v qp		Queue pair
  * @v gid		Multicast GID
  * @ret rc		Return status code
+ *
+ * Note that this function handles only the local device's attachment
+ * to the multicast GID; it does not issue the relevant MADs to join
+ * the multicast group on the subnet.
  */
 int ib_mcast_attach ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 		      struct ib_gid *gid ) {
@@ -486,6 +704,89 @@ void ib_mcast_detach ( struct ib_device *ibdev, struct ib_queue_pair *qp,
 	}
 }
 
+/***************************************************************************
+ *
+ * Miscellaneous
+ *
+ ***************************************************************************
+ */
+
+/**
+ * Get Infiniband HCA information
+ *
+ * @v ibdev		Infiniband device
+ * @ret hca_guid	HCA GUID
+ * @ret num_ports	Number of ports
+ */
+int ib_get_hca_info ( struct ib_device *ibdev,
+		      struct ib_gid_half *hca_guid ) {
+	struct ib_device *tmp;
+	int num_ports = 0;
+
+	/* Search for IB devices with the same physical device to
+	 * identify port count and a suitable Node GUID.
+	 */
+	for_each_ibdev ( tmp ) {
+		if ( tmp->dev != ibdev->dev )
+			continue;
+		if ( num_ports == 0 ) {
+			memcpy ( hca_guid, &tmp->gid.u.half[1],
+				 sizeof ( *hca_guid ) );
+		}
+		num_ports++;
+	}
+	return num_ports;
+}
+
+/**
+ * Set port information
+ *
+ * @v ibdev		Infiniband device
+ * @v mad		Set port information MAD
+ */
+int ib_set_port_info ( struct ib_device *ibdev, union ib_mad *mad ) {
+	int rc;
+
+	/* Adapters with embedded SMAs do not need to support this method */
+	if ( ! ibdev->op->set_port_info ) {
+		DBGC ( ibdev, "IBDEV %p does not support setting port "
+		       "information\n", ibdev );
+		return -ENOTSUP;
+	}
+
+	if ( ( rc = ibdev->op->set_port_info ( ibdev, mad ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not set port information: %s\n",
+		       ibdev, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+};
+
+/**
+ * Set partition key table
+ *
+ * @v ibdev		Infiniband device
+ * @v mad		Set partition key table MAD
+ */
+int ib_set_pkey_table ( struct ib_device *ibdev, union ib_mad *mad ) {
+	int rc;
+
+	/* Adapters with embedded SMAs do not need to support this method */
+	if ( ! ibdev->op->set_pkey_table ) {
+		DBGC ( ibdev, "IBDEV %p does not support setting partition "
+		       "key table\n", ibdev );
+		return -ENOTSUP;
+	}
+
+	if ( ( rc = ibdev->op->set_pkey_table ( ibdev, mad ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not set partition key table: "
+		       "%s\n", ibdev, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+};
 
 /***************************************************************************
  *
@@ -506,6 +807,22 @@ void ib_link_state_changed ( struct ib_device *ibdev ) {
 }
 
 /**
+ * Poll event queue
+ *
+ * @v ibdev		Infiniband device
+ */
+void ib_poll_eq ( struct ib_device *ibdev ) {
+	struct ib_completion_queue *cq;
+
+	/* Poll device's event queue */
+	ibdev->op->poll_eq ( ibdev );
+
+	/* Poll all completion queues */
+	list_for_each_entry ( cq, &ibdev->cqs, list )
+		ib_poll_cq ( ibdev, cq );
+}
+
+/**
  * Single-step the Infiniband event queue
  *
  * @v process		Infiniband event queue process
@@ -513,13 +830,13 @@ void ib_link_state_changed ( struct ib_device *ibdev ) {
 static void ib_step ( struct process *process __unused ) {
 	struct ib_device *ibdev;
 
-	list_for_each_entry ( ibdev, &ib_devices, list ) {
-		ibdev->op->poll_eq ( ibdev );
-	}
+	for_each_ibdev ( ibdev )
+		ib_poll_eq ( ibdev );
 }
 
 /** Infiniband event queue process */
 struct process ib_process __permanent_process = {
+	.list = LIST_HEAD_INIT ( ib_process.list ),
 	.step = ib_step,
 };
 
@@ -546,9 +863,11 @@ struct ib_device * alloc_ibdev ( size_t priv_size ) {
 	if ( ibdev ) {
 		drv_priv = ( ( ( void * ) ibdev ) + sizeof ( *ibdev ) );
 		ib_set_drvdata ( ibdev, drv_priv );
+		INIT_LIST_HEAD ( &ibdev->cqs );
 		INIT_LIST_HEAD ( &ibdev->qps );
+		ibdev->port_state = IB_PORT_STATE_DOWN;
 		ibdev->lid = IB_LID_NONE;
-		ibdev->pkey = IB_PKEY_NONE;
+		ibdev->pkey = IB_PKEY_DEFAULT;
 	}
 	return ibdev;
 }
@@ -598,3 +917,35 @@ void unregister_ibdev ( struct ib_device *ibdev ) {
 	ibdev_put ( ibdev );
 	DBGC ( ibdev, "IBDEV %p unregistered\n", ibdev );
 }
+
+/**
+ * Find Infiniband device by GID
+ *
+ * @v gid		GID
+ * @ret ibdev		Infiniband device, or NULL
+ */
+struct ib_device * find_ibdev ( struct ib_gid *gid ) {
+	struct ib_device *ibdev;
+
+	for_each_ibdev ( ibdev ) {
+		if ( memcmp ( gid, &ibdev->gid, sizeof ( *gid ) ) == 0 )
+			return ibdev;
+	}
+	return NULL;
+}
+
+/**
+ * Get most recently opened Infiniband device
+ *
+ * @ret ibdev		Most recently opened Infiniband device, or NULL
+ */
+struct ib_device * last_opened_ibdev ( void ) {
+	struct ib_device *ibdev;
+
+	list_for_each_entry ( ibdev, &open_ib_devices, open_list ) {
+		assert ( ibdev->open_count != 0 );
+		return ibdev;
+	}
+
+	return NULL;
+}
diff --git a/gpxe/src/net/infiniband/ib_cm.c b/gpxe/src/net/infiniband/ib_cm.c
new file mode 100644
index 00000000..ebe65b33
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_cm.c
@@ -0,0 +1,413 @@
+/*
+ * Copyright (C) 2009 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <byteswap.h>
+#include <errno.h>
+#include <assert.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_mi.h>
+#include <gpxe/ib_pathrec.h>
+#include <gpxe/ib_cm.h>
+
+/**
+ * @file
+ *
+ * Infiniband communication management
+ *
+ */
+
+/** List of connections */
+static LIST_HEAD ( ib_cm_conns );
+
+/**
+ * Send "ready to use" response
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v conn		Connection
+ * @v av		Address vector
+ * @ret rc		Return status code
+ */
+static int ib_cm_send_rtu ( struct ib_device *ibdev,
+			    struct ib_mad_interface *mi,
+			    struct ib_connection *conn,
+			    struct ib_address_vector *av ) {
+	union ib_mad mad;
+	struct ib_cm_ready_to_use *ready =
+		&mad.cm.cm_data.ready_to_use;
+	int rc;
+
+	/* Construct "ready to use" response */
+	memset ( &mad, 0, sizeof ( mad ) );
+	mad.hdr.mgmt_class = IB_MGMT_CLASS_CM;
+	mad.hdr.class_version = IB_CM_CLASS_VERSION;
+	mad.hdr.method = IB_MGMT_METHOD_SEND;
+	mad.hdr.attr_id = htons ( IB_CM_ATTR_READY_TO_USE );
+	ready->local_id = htonl ( conn->local_id );
+	ready->remote_id = htonl ( conn->remote_id );
+	if ( ( rc = ib_mi_send ( ibdev, mi, &mad, av ) ) != 0 ){
+		DBGC ( conn, "CM %p could not send RTU: %s\n",
+		       conn, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Handle duplicate connection replies
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ * @ret rc		Return status code
+ *
+ * If a "ready to use" MAD is lost, the peer may resend the connection
+ * reply.  We have to respond to these with duplicate "ready to use"
+ * MADs, otherwise the peer may time out and drop the connection.
+ */
+static void ib_cm_connect_rep ( struct ib_device *ibdev,
+				struct ib_mad_interface *mi,
+				union ib_mad *mad,
+				struct ib_address_vector *av ) {
+	struct ib_cm_connect_reply *connect_rep =
+		&mad->cm.cm_data.connect_reply;
+	struct ib_connection *conn;
+	int rc;
+
+	/* Identify connection */
+	list_for_each_entry ( conn, &ib_cm_conns, list ) {
+		if ( ntohl ( connect_rep->remote_id ) != conn->local_id )
+			continue;
+		/* Try to send "ready to use" reply */
+		if ( ( rc = ib_cm_send_rtu ( ibdev, mi, conn, av ) ) != 0 ) {
+			/* Ignore errors */
+			return;
+		}
+		return;
+	}
+
+	DBG ( "CM unidentified connection %08x\n",
+	      ntohl ( connect_rep->remote_id ) );
+}
+
+/** Communication management agents */
+struct ib_mad_agent ib_cm_agent[] __ib_mad_agent = {
+	{
+		.mgmt_class = IB_MGMT_CLASS_CM,
+		.class_version = IB_CM_CLASS_VERSION,
+		.attr_id = htons ( IB_CM_ATTR_CONNECT_REPLY ),
+		.handle = ib_cm_connect_rep,
+	},
+};
+
+/**
+ * Convert connection rejection reason to return status code
+ *
+ * @v reason		Rejection reason (in network byte order)
+ * @ret rc		Return status code
+ */
+static int ib_cm_rejection_reason_to_rc ( uint16_t reason ) {
+	switch ( reason ) {
+	case htons ( IB_CM_REJECT_BAD_SERVICE_ID ) :
+		return -ENODEV;
+	case htons ( IB_CM_REJECT_STALE_CONN ) :
+		return -EALREADY;
+	case htons ( IB_CM_REJECT_CONSUMER ) :
+		return -ENOTTY;
+	default:
+		return -EPERM;
+	}
+}
+
+/**
+ * Handle connection request transaction completion
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v madx		Management transaction
+ * @v rc		Status code
+ * @v mad		Received MAD (or NULL on error)
+ * @v av		Source address vector (or NULL on error)
+ */
+static void ib_cm_req_complete ( struct ib_device *ibdev,
+				 struct ib_mad_interface *mi,
+				 struct ib_mad_transaction *madx,
+				 int rc, union ib_mad *mad,
+				 struct ib_address_vector *av ) {
+	struct ib_connection *conn = ib_madx_get_ownerdata ( madx );
+	struct ib_queue_pair *qp = conn->qp;
+	struct ib_cm_common *common = &mad->cm.cm_data.common;
+	struct ib_cm_connect_reply *connect_rep =
+		&mad->cm.cm_data.connect_reply;
+	struct ib_cm_connect_reject *connect_rej =
+		&mad->cm.cm_data.connect_reject;
+	void *private_data = NULL;
+	size_t private_data_len = 0;
+
+	/* Report failures */
+	if ( ( rc == 0 ) && ( mad->hdr.status != htons ( IB_MGMT_STATUS_OK ) ))
+		rc = -EIO;
+	if ( rc != 0 ) {
+		DBGC ( conn, "CM %p connection request failed: %s\n",
+		       conn, strerror ( rc ) );
+		goto out;
+	}
+
+	/* Record remote communication ID */
+	conn->remote_id = ntohl ( common->local_id );
+
+	/* Handle response */
+	switch ( mad->hdr.attr_id ) {
+
+	case htons ( IB_CM_ATTR_CONNECT_REPLY ) :
+		/* Extract fields */
+		qp->av.qpn = ( ntohl ( connect_rep->local_qpn ) >> 8 );
+		qp->send.psn = ( ntohl ( connect_rep->starting_psn ) >> 8 );
+		private_data = &connect_rep->private_data;
+		private_data_len = sizeof ( connect_rep->private_data );
+		DBGC ( conn, "CM %p connected to QPN %lx PSN %x\n",
+		       conn, qp->av.qpn, qp->send.psn );
+
+		/* Modify queue pair */
+		if ( ( rc = ib_modify_qp ( ibdev, qp ) ) != 0 ) {
+			DBGC ( conn, "CM %p could not modify queue pair: %s\n",
+			       conn, strerror ( rc ) );
+			goto out;
+		}
+
+		/* Send "ready to use" reply */
+		if ( ( rc = ib_cm_send_rtu ( ibdev, mi, conn, av ) ) != 0 ) {
+			/* Treat as non-fatal */
+			rc = 0;
+		}
+		break;
+
+	case htons ( IB_CM_ATTR_CONNECT_REJECT ) :
+		/* Extract fields */
+		DBGC ( conn, "CM %p connection rejected (reason %d)\n",
+		       conn, ntohs ( connect_rej->reason ) );
+		/* Private data is valid only for a Consumer Reject */
+		if ( connect_rej->reason == htons ( IB_CM_REJECT_CONSUMER ) ) {
+			private_data = &connect_rej->private_data;
+			private_data_len = sizeof (connect_rej->private_data);
+		}
+		rc = ib_cm_rejection_reason_to_rc ( connect_rej->reason );
+		break;
+
+	default:
+		DBGC ( conn, "CM %p unexpected response (attribute %04x)\n",
+		       conn, ntohs ( mad->hdr.attr_id ) );
+		rc = -ENOTSUP;
+		break;
+	}
+
+ out:
+	/* Destroy the completed transaction */
+	ib_destroy_madx ( ibdev, ibdev->gsi, madx );
+	conn->madx = NULL;
+
+	/* Hand off to the upper completion handler */
+	conn->op->changed ( ibdev, qp, conn, rc, private_data,
+			    private_data_len );
+}
+
+/** Connection request operations */
+static struct ib_mad_transaction_operations ib_cm_req_op = {
+	.complete = ib_cm_req_complete,
+};
+
+/**
+ * Handle connection path transaction completion
+ *
+ * @v ibdev		Infiniband device
+ * @v path		Path
+ * @v rc		Status code
+ * @v av		Address vector, or NULL on error
+ */
+static void ib_cm_path_complete ( struct ib_device *ibdev,
+				  struct ib_path *path, int rc,
+				  struct ib_address_vector *av ) {
+	struct ib_connection *conn = ib_path_get_ownerdata ( path );
+	struct ib_queue_pair *qp = conn->qp;
+	union ib_mad mad;
+	struct ib_cm_connect_request *connect_req =
+		&mad.cm.cm_data.connect_request;
+	size_t private_data_len;
+
+	/* Report failures */
+	if ( rc != 0 ) {
+		DBGC ( conn, "CM %p path lookup failed: %s\n",
+		       conn, strerror ( rc ) );
+		conn->op->changed ( ibdev, qp, conn, rc, NULL, 0 );
+		goto out;
+	}
+
+	/* Update queue pair peer path */
+	memcpy ( &qp->av, av, sizeof ( qp->av ) );
+
+	/* Construct connection request */
+	memset ( &mad, 0, sizeof ( mad ) );
+	mad.hdr.mgmt_class = IB_MGMT_CLASS_CM;
+	mad.hdr.class_version = IB_CM_CLASS_VERSION;
+	mad.hdr.method = IB_MGMT_METHOD_SEND;
+	mad.hdr.attr_id = htons ( IB_CM_ATTR_CONNECT_REQUEST );
+	connect_req->local_id = htonl ( conn->local_id );
+	memcpy ( &connect_req->service_id, &conn->service_id,
+		 sizeof ( connect_req->service_id ) );
+	ib_get_hca_info ( ibdev, &connect_req->local_ca );
+	connect_req->local_qpn__responder_resources =
+		htonl ( ( qp->qpn << 8 ) | 1 );
+	connect_req->local_eecn__initiator_depth = htonl ( ( 0 << 8 ) | 1 );
+	connect_req->remote_eecn__remote_timeout__service_type__ee_flow_ctrl =
+		htonl ( ( 0x14 << 3 ) | ( IB_CM_TRANSPORT_RC << 1 ) |
+			( 0 << 0 ) );
+	connect_req->starting_psn__local_timeout__retry_count =
+		htonl ( ( qp->recv.psn << 8 ) | ( 0x14 << 3 ) |
+			( 0x07 << 0 ) );
+	connect_req->pkey = htons ( ibdev->pkey );
+	connect_req->payload_mtu__rdc_exists__rnr_retry =
+		( ( IB_MTU_2048 << 4 ) | ( 1 << 3 ) | ( 0x07 << 0 ) );
+	connect_req->max_cm_retries__srq =
+		( ( 0x0f << 4 ) | ( 0 << 3 ) );
+	connect_req->primary.local_lid = htons ( ibdev->lid );
+	connect_req->primary.remote_lid = htons ( conn->qp->av.lid );
+	memcpy ( &connect_req->primary.local_gid, &ibdev->gid,
+		 sizeof ( connect_req->primary.local_gid ) );
+	memcpy ( &connect_req->primary.remote_gid, &conn->qp->av.gid,
+		 sizeof ( connect_req->primary.remote_gid ) );
+	connect_req->primary.flow_label__rate =
+		htonl ( ( 0 << 12 ) | ( conn->qp->av.rate << 0 ) );
+	connect_req->primary.hop_limit = 0;
+	connect_req->primary.sl__subnet_local =
+		( ( conn->qp->av.sl << 4 ) | ( 1 << 3 ) );
+	connect_req->primary.local_ack_timeout = ( 0x13 << 3 );
+	private_data_len = conn->private_data_len;
+	if ( private_data_len > sizeof ( connect_req->private_data ) )
+		private_data_len = sizeof ( connect_req->private_data );
+	memcpy ( &connect_req->private_data, &conn->private_data,
+		 private_data_len );
+
+	/* Create connection request */
+	av->qpn = IB_QPN_GSI;
+	av->qkey = IB_QKEY_GSI;
+	conn->madx = ib_create_madx ( ibdev, ibdev->gsi, &mad, av,
+				      &ib_cm_req_op );
+	if ( ! conn->madx ) {
+		DBGC ( conn, "CM %p could not create connection request\n",
+		       conn );
+		conn->op->changed ( ibdev, qp, conn, rc, NULL, 0 );
+		goto out;
+	}
+	ib_madx_set_ownerdata ( conn->madx, conn );
+
+ out:
+	/* Destroy the completed transaction */
+	ib_destroy_path ( ibdev, path );
+	conn->path = NULL;
+}
+
+/** Connection path operations */
+static struct ib_path_operations ib_cm_path_op = {
+	.complete = ib_cm_path_complete,
+};
+
+/**
+ * Create connection to remote QP
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v dgid		Target GID
+ * @v service_id	Target service ID
+ * @v private_data	Connection request private data
+ * @v private_data_len	Length of connection request private data
+ * @v op		Connection operations
+ * @ret conn		Connection
+ */
+struct ib_connection *
+ib_create_conn ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+		 struct ib_gid *dgid, struct ib_gid_half *service_id,
+		 void *private_data, size_t private_data_len,
+		 struct ib_connection_operations *op ) {
+	struct ib_connection *conn;
+
+	/* Allocate and initialise request */
+	conn = zalloc ( sizeof ( *conn ) + private_data_len );
+	if ( ! conn )
+		goto err_alloc_conn;
+	conn->ibdev = ibdev;
+	conn->qp = qp;
+	memset ( &qp->av, 0, sizeof ( qp->av ) );
+	qp->av.gid_present = 1;
+	memcpy ( &qp->av.gid, dgid, sizeof ( qp->av.gid ) );
+	conn->local_id = random();
+	memcpy ( &conn->service_id, service_id, sizeof ( conn->service_id ) );
+	conn->op = op;
+	conn->private_data_len = private_data_len;
+	memcpy ( &conn->private_data, private_data, private_data_len );
+
+	/* Create path */
+	conn->path = ib_create_path ( ibdev, &qp->av, &ib_cm_path_op );
+	if ( ! conn->path )
+		goto err_create_path;
+	ib_path_set_ownerdata ( conn->path, conn );
+
+	/* Add to list of connections */
+	list_add ( &conn->list, &ib_cm_conns );
+
+	DBGC ( conn, "CM %p created for IBDEV %p QPN %lx\n",
+	       conn, ibdev, qp->qpn );
+	DBGC ( conn, "CM %p connecting to %08x:%08x:%08x:%08x %08x:%08x\n",
+	       conn, ntohl ( dgid->u.dwords[0] ), ntohl ( dgid->u.dwords[1] ),
+	       ntohl ( dgid->u.dwords[2] ), ntohl ( dgid->u.dwords[3] ),
+	       ntohl ( service_id->u.dwords[0] ),
+	       ntohl ( service_id->u.dwords[1] ) );
+
+	return conn;
+
+	ib_destroy_path ( ibdev, conn->path );
+ err_create_path:
+	free ( conn );
+ err_alloc_conn:
+	return NULL;
+}
+
+/**
+ * Destroy connection to remote QP
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v conn		Connection
+ */
+void ib_destroy_conn ( struct ib_device *ibdev,
+		       struct ib_queue_pair *qp __unused,
+		       struct ib_connection *conn ) {
+
+	list_del ( &conn->list );
+	if ( conn->madx )
+		ib_destroy_madx ( ibdev, ibdev->gsi, conn->madx );
+	if ( conn->path )
+		ib_destroy_path ( ibdev, conn->path );
+	free ( conn );
+}
diff --git a/gpxe/src/net/infiniband/ib_cmrc.c b/gpxe/src/net/infiniband/ib_cmrc.c
new file mode 100644
index 00000000..2d648115
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_cmrc.c
@@ -0,0 +1,436 @@
+/*
+ * Copyright (C) 2009 Fen Systems Ltd <mbrown@fensystems.co.uk>.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *   Redistributions of source code must retain the above copyright
+ *   notice, this list of conditions and the following disclaimer.
+ *
+ *   Redistributions in binary form must reproduce the above copyright
+ *   notice, this list of conditions and the following disclaimer in
+ *   the documentation and/or other materials provided with the
+ *   distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
+ * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
+ * OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+FILE_LICENCE ( BSD2 );
+
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/xfer.h>
+#include <gpxe/process.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_cm.h>
+#include <gpxe/ib_cmrc.h>
+
+/**
+ * @file
+ *
+ * Infiniband Communication-managed Reliable Connections
+ *
+ */
+
+/** CMRC number of send WQEs
+ *
+ * This is a policy decision.
+ */
+#define IB_CMRC_NUM_SEND_WQES 4
+
+/** CMRC number of receive WQEs
+ *
+ * This is a policy decision.
+ */
+#define IB_CMRC_NUM_RECV_WQES 2
+
+/** CMRC number of completion queue entries
+ *
+ * This is a policy decision
+ */
+#define IB_CMRC_NUM_CQES 8
+
+/** An Infiniband Communication-Managed Reliable Connection */
+struct ib_cmrc_connection {
+	/** Reference count */
+	struct refcnt refcnt;
+	/** Data transfer interface */
+	struct xfer_interface xfer;
+	/** Infiniband device */
+	struct ib_device *ibdev;
+	/** Completion queue */
+	struct ib_completion_queue *cq;
+	/** Queue pair */
+	struct ib_queue_pair *qp;
+	/** Connection */
+	struct ib_connection *conn;
+	/** Destination GID */
+	struct ib_gid dgid;
+	/** Service ID */
+	struct ib_gid_half service_id;
+	/** QP is connected */
+	int connected;
+	/** Shutdown process */
+	struct process shutdown;
+};
+
+/**
+ * Shut down CMRC connection gracefully
+ *
+ * @v process		Process
+ *
+ * The Infiniband data structures are not reference-counted or
+ * guarded.  It is therefore unsafe to shut them down while we may be
+ * in the middle of a callback from the Infiniband stack (e.g. in a
+ * receive completion handler).
+ *
+ * This shutdown process will run some time after the call to
+ * ib_cmrc_close(), after control has returned out of the Infiniband
+ * core, and will shut down the Infiniband interfaces cleanly.
+ *
+ * The shutdown process holds an implicit reference on the CMRC
+ * connection, ensuring that the structure is not freed before the
+ * shutdown process has run.
+ */
+static void ib_cmrc_shutdown ( struct process *process ) {
+	struct ib_cmrc_connection *cmrc =
+		container_of ( process, struct ib_cmrc_connection, shutdown );
+
+	DBGC ( cmrc, "CMRC %p shutting down\n", cmrc );
+
+	/* Shut down Infiniband interface */
+	ib_destroy_conn ( cmrc->ibdev, cmrc->qp, cmrc->conn );
+	ib_destroy_qp ( cmrc->ibdev, cmrc->qp );
+	ib_destroy_cq ( cmrc->ibdev, cmrc->cq );
+	ib_close ( cmrc->ibdev );
+
+	/* Remove process from run queue */
+	process_del ( &cmrc->shutdown );
+
+	/* Drop the remaining reference */
+	ref_put ( &cmrc->refcnt );
+}
+
+/**
+ * Close CMRC connection
+ *
+ * @v cmrc		Communication-Managed Reliable Connection
+ * @v rc		Reason for close
+ */
+static void ib_cmrc_close ( struct ib_cmrc_connection *cmrc, int rc ) {
+
+	/* Close data transfer interface */
+	xfer_nullify ( &cmrc->xfer );
+	xfer_close ( &cmrc->xfer, rc );
+
+	/* Schedule shutdown process */
+	process_add ( &cmrc->shutdown );
+}
+
+/**
+ * Handle change of CMRC connection status
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v conn		Connection
+ * @v rc_cm		Connection status code
+ * @v private_data	Private data, if available
+ * @v private_data_len	Length of private data
+ */
+static void ib_cmrc_changed ( struct ib_device *ibdev __unused,
+			      struct ib_queue_pair *qp,
+			      struct ib_connection *conn __unused, int rc_cm,
+			      void *private_data, size_t private_data_len ) {
+	struct ib_cmrc_connection *cmrc = ib_qp_get_ownerdata ( qp );
+	int rc_xfer;
+
+	/* Record connection status */
+	if ( rc_cm == 0 ) {
+		DBGC ( cmrc, "CMRC %p connected\n", cmrc );
+		cmrc->connected = 1;
+	} else {
+		DBGC ( cmrc, "CMRC %p disconnected: %s\n",
+		       cmrc, strerror ( rc_cm ) );
+		cmrc->connected = 0;
+	}
+
+	/* Pass up any private data */
+	DBGC2 ( cmrc, "CMRC %p received private data:\n", cmrc );
+	DBGC2_HDA ( cmrc, 0, private_data, private_data_len );
+	if ( private_data &&
+	     ( rc_xfer = xfer_deliver_raw ( &cmrc->xfer, private_data,
+					    private_data_len ) ) != 0 ) {
+		DBGC ( cmrc, "CMRC %p could not deliver private data: %s\n",
+		       cmrc, strerror ( rc_xfer ) );
+		ib_cmrc_close ( cmrc, rc_xfer );
+		return;
+	}
+
+	/* If we are disconnected, close the upper connection */
+	if ( rc_cm != 0 ) {
+		ib_cmrc_close ( cmrc, rc_cm );
+		return;
+	}
+}
+
+/** CMRC connection operations */
+static struct ib_connection_operations ib_cmrc_conn_op = {
+	.changed = ib_cmrc_changed,
+};
+
+/**
+ * Handle CMRC send completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v iobuf		I/O buffer
+ * @v rc		Completion status code
+ */
+static void ib_cmrc_complete_send ( struct ib_device *ibdev __unused,
+				    struct ib_queue_pair *qp,
+				    struct io_buffer *iobuf, int rc ) {
+	struct ib_cmrc_connection *cmrc = ib_qp_get_ownerdata ( qp );
+
+	/* Free the completed I/O buffer */
+	free_iob ( iobuf );
+
+	/* Close the connection on any send errors */
+	if ( rc != 0 ) {
+		DBGC ( cmrc, "CMRC %p send error: %s\n",
+		       cmrc, strerror ( rc ) );
+		ib_cmrc_close ( cmrc, rc );
+		return;
+	}
+}
+
+/**
+ * Handle CMRC receive completion
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v av		Address vector, or NULL
+ * @v iobuf		I/O buffer
+ * @v rc		Completion status code
+ */
+static void ib_cmrc_complete_recv ( struct ib_device *ibdev __unused,
+				    struct ib_queue_pair *qp,
+				    struct ib_address_vector *av __unused,
+				    struct io_buffer *iobuf, int rc ) {
+	struct ib_cmrc_connection *cmrc = ib_qp_get_ownerdata ( qp );
+
+	/* Close the connection on any receive errors */
+	if ( rc != 0 ) {
+		DBGC ( cmrc, "CMRC %p receive error: %s\n",
+		       cmrc, strerror ( rc ) );
+		free_iob ( iobuf );
+		ib_cmrc_close ( cmrc, rc );
+		return;
+	}
+
+	DBGC2 ( cmrc, "CMRC %p received:\n", cmrc );
+	DBGC2_HDA ( cmrc, 0, iobuf->data, iob_len ( iobuf ) );
+
+	/* Pass up data */
+	if ( ( rc = xfer_deliver_iob ( &cmrc->xfer, iobuf ) ) != 0 ) {
+		DBGC ( cmrc, "CMRC %p could not deliver data: %s\n",
+		       cmrc, strerror ( rc ) );
+		ib_cmrc_close ( cmrc, rc );
+		return;
+	}
+}
+
+/** Infiniband CMRC completion operations */
+static struct ib_completion_queue_operations ib_cmrc_completion_ops = {
+	.complete_send = ib_cmrc_complete_send,
+	.complete_recv = ib_cmrc_complete_recv,
+};
+
+/**
+ * Send data via CMRC
+ *
+ * @v xfer		Data transfer interface
+ * @v iobuf		Datagram I/O buffer
+ * @v meta		Data transfer metadata
+ * @ret rc		Return status code
+ */
+static int ib_cmrc_xfer_deliver_iob ( struct xfer_interface *xfer,
+				      struct io_buffer *iobuf,
+				      struct xfer_metadata *meta __unused ) {
+	struct ib_cmrc_connection *cmrc =
+		container_of ( xfer, struct ib_cmrc_connection, xfer );
+	int rc;
+
+	/* If no connection has yet been attempted, send this datagram
+	 * as the CM REQ private data.  Otherwise, send it via the QP.
+	 */
+	if ( ! cmrc->connected ) {
+
+		/* Abort if we have already sent a CM connection request */
+		if ( cmrc->conn ) {
+			DBGC ( cmrc, "CMRC %p attempt to send before "
+			       "connection is complete\n", cmrc );
+			rc = -EIO;
+			goto out;
+		}
+
+		/* Send via CM connection request */
+		cmrc->conn = ib_create_conn ( cmrc->ibdev, cmrc->qp,
+					      &cmrc->dgid, &cmrc->service_id,
+					      iobuf->data, iob_len ( iobuf ),
+					      &ib_cmrc_conn_op );
+		if ( ! cmrc->conn ) {
+			DBGC ( cmrc, "CMRC %p could not connect\n", cmrc );
+			rc = -ENOMEM;
+			goto out;
+		}
+
+	} else {
+
+		/* Send via QP */
+		if ( ( rc = ib_post_send ( cmrc->ibdev, cmrc->qp, NULL,
+					   iob_disown ( iobuf ) ) ) != 0 ) {
+			DBGC ( cmrc, "CMRC %p could not send: %s\n",
+			       cmrc, strerror ( rc ) );
+			goto out;
+		}
+
+	}
+	return 0;
+
+ out:
+	/* Free the I/O buffer if necessary */
+	free_iob ( iobuf );
+
+	/* Close the connection on any errors */
+	if ( rc != 0 )
+		ib_cmrc_close ( cmrc, rc );
+
+	return rc;
+}
+
+/**
+ * Check CMRC flow control window
+ *
+ * @v xfer		Data transfer interface
+ * @ret len		Length of window
+ */
+static size_t ib_cmrc_xfer_window ( struct xfer_interface *xfer ) {
+	struct ib_cmrc_connection *cmrc =
+		container_of ( xfer, struct ib_cmrc_connection, xfer );
+
+	/* We indicate a window only when we are successfully
+	 * connected.
+	 */
+	return ( cmrc->connected ? IB_MAX_PAYLOAD_SIZE : 0 );
+}
+
+/**
+ * Close CMRC data-transfer interface
+ *
+ * @v xfer		Data transfer interface
+ * @v rc		Reason for close
+ */
+static void ib_cmrc_xfer_close ( struct xfer_interface *xfer, int rc ) {
+	struct ib_cmrc_connection *cmrc =
+		container_of ( xfer, struct ib_cmrc_connection, xfer );
+
+	DBGC ( cmrc, "CMRC %p closed: %s\n", cmrc, strerror ( rc ) );
+	ib_cmrc_close ( cmrc, rc );
+}
+
+/** CMRC data transfer interface operations */
+static struct xfer_interface_operations ib_cmrc_xfer_operations = {
+	.close		= ib_cmrc_xfer_close,
+	.vredirect	= ignore_xfer_vredirect,
+	.window		= ib_cmrc_xfer_window,
+	.alloc_iob	= default_xfer_alloc_iob,
+	.deliver_iob	= ib_cmrc_xfer_deliver_iob,
+	.deliver_raw	= xfer_deliver_as_iob,
+};
+
+/**
+ * Open CMRC connection
+ *
+ * @v xfer		Data transfer interface
+ * @v ibdev		Infiniband device
+ * @v dgid		Destination GID
+ * @v service_id	Service ID
+ * @ret rc		Returns status code
+ */
+int ib_cmrc_open ( struct xfer_interface *xfer, struct ib_device *ibdev,
+		   struct ib_gid *dgid, struct ib_gid_half *service_id ) {
+	struct ib_cmrc_connection *cmrc;
+	int rc;
+
+	/* Allocate and initialise structure */
+	cmrc = zalloc ( sizeof ( *cmrc ) );
+	if ( ! cmrc ) {
+		rc = -ENOMEM;
+		goto err_alloc;
+	}
+	xfer_init ( &cmrc->xfer, &ib_cmrc_xfer_operations, &cmrc->refcnt );
+	cmrc->ibdev = ibdev;
+	memcpy ( &cmrc->dgid, dgid, sizeof ( cmrc->dgid ) );
+	memcpy ( &cmrc->service_id, service_id, sizeof ( cmrc->service_id ) );
+	process_init_stopped ( &cmrc->shutdown, ib_cmrc_shutdown,
+			       &cmrc->refcnt );
+
+	/* Open Infiniband device */
+	if ( ( rc = ib_open ( ibdev ) ) != 0 ) {
+		DBGC ( cmrc, "CMRC %p could not open device: %s\n",
+		       cmrc, strerror ( rc ) );
+		goto err_open;
+	}
+
+	/* Create completion queue */
+	cmrc->cq = ib_create_cq ( ibdev, IB_CMRC_NUM_CQES,
+				  &ib_cmrc_completion_ops );
+	if ( ! cmrc->cq ) {
+		DBGC ( cmrc, "CMRC %p could not create completion queue\n",
+		       cmrc );
+		rc = -ENOMEM;
+		goto err_create_cq;
+	}
+
+	/* Create queue pair */
+	cmrc->qp = ib_create_qp ( ibdev, IB_QPT_RC, IB_CMRC_NUM_SEND_WQES,
+				  cmrc->cq, IB_CMRC_NUM_RECV_WQES, cmrc->cq );
+	if ( ! cmrc->qp ) {
+		DBGC ( cmrc, "CMRC %p could not create queue pair\n", cmrc );
+		rc = -ENOMEM;
+		goto err_create_qp;
+	}
+	ib_qp_set_ownerdata ( cmrc->qp, cmrc );
+	DBGC ( cmrc, "CMRC %p using QPN %lx\n", cmrc, cmrc->qp->qpn );
+
+	/* Attach to parent interface, transfer reference (implicitly)
+	 * to our shutdown process, and return.
+	 */
+	xfer_plug_plug ( &cmrc->xfer, xfer );
+	return 0;
+
+	ib_destroy_qp ( ibdev, cmrc->qp );
+ err_create_qp:
+	ib_destroy_cq ( ibdev, cmrc->cq );
+ err_create_cq:
+	ib_close ( ibdev );
+ err_open:
+	ref_put ( &cmrc->refcnt );
+ err_alloc:
+	return rc;
+}
diff --git a/gpxe/src/net/infiniband/ib_mcast.c b/gpxe/src/net/infiniband/ib_mcast.c
new file mode 100644
index 00000000..5cb395de
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_mcast.c
@@ -0,0 +1,218 @@
+/*
+ * Copyright (C) 2007 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <string.h>
+#include <byteswap.h>
+#include <errno.h>
+#include <gpxe/list.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_mi.h>
+#include <gpxe/ib_mcast.h>
+
+/** @file
+ *
+ * Infiniband multicast groups
+ *
+ */
+
+/**
+ * Generate multicast membership MAD
+ *
+ * @v ibdev		Infiniband device
+ * @v gid		Multicast GID
+ * @v join		Join (rather than leave) group
+ * @v mad		MAD to fill in
+ */
+static void ib_mcast_mad ( struct ib_device *ibdev, struct ib_gid *gid,
+			   int join, union ib_mad *mad ) {
+	struct ib_mad_sa *sa = &mad->sa;
+
+	/* Construct multicast membership record request */
+	memset ( sa, 0, sizeof ( *sa ) );
+	sa->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	sa->mad_hdr.class_version = IB_SA_CLASS_VERSION;
+	sa->mad_hdr.method =
+		( join ? IB_MGMT_METHOD_SET : IB_MGMT_METHOD_DELETE );
+	sa->mad_hdr.attr_id = htons ( IB_SA_ATTR_MC_MEMBER_REC );
+	sa->sa_hdr.comp_mask[1] =
+		htonl ( IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
+			IB_SA_MCMEMBER_REC_JOIN_STATE );
+	sa->sa_data.mc_member_record.scope__join_state = 1;
+	memcpy ( &sa->sa_data.mc_member_record.mgid, gid,
+		 sizeof ( sa->sa_data.mc_member_record.mgid ) );
+	memcpy ( &sa->sa_data.mc_member_record.port_gid, &ibdev->gid,
+		 sizeof ( sa->sa_data.mc_member_record.port_gid ) );
+}
+
+/**
+ * Handle multicast membership record join response
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v madx		Management transaction
+ * @v rc		Status code
+ * @v mad		Received MAD (or NULL on error)
+ * @v av		Source address vector (or NULL on error)
+ */
+static void ib_mcast_complete ( struct ib_device *ibdev,
+				struct ib_mad_interface *mi __unused,
+				struct ib_mad_transaction *madx,
+				int rc, union ib_mad *mad,
+				struct ib_address_vector *av __unused ) {
+	struct ib_mc_membership *membership = ib_madx_get_ownerdata ( madx );
+	struct ib_queue_pair *qp = membership->qp;
+	struct ib_gid *gid = &membership->gid;
+	struct ib_mc_member_record *mc_member_record =
+		&mad->sa.sa_data.mc_member_record;
+	int joined;
+	unsigned long qkey;
+
+	/* Report failures */
+	if ( ( rc == 0 ) && ( mad->hdr.status != htons ( IB_MGMT_STATUS_OK ) ))
+		rc = -ENOTCONN;
+	if ( rc != 0 ) {
+		DBGC ( ibdev, "IBDEV %p QPN %lx join failed: %s\n",
+		       ibdev, qp->qpn, strerror ( rc ) );
+		goto out;
+	}
+
+	/* Extract values from MAD */
+	joined = ( mad->hdr.method == IB_MGMT_METHOD_GET_RESP );
+	qkey = ntohl ( mc_member_record->qkey );
+	DBGC ( ibdev, "IBDEV %p QPN %lx %s %08x:%08x:%08x:%08x qkey %lx\n",
+	       ibdev, qp->qpn, ( joined ? "joined" : "left" ),
+	       ntohl ( gid->u.dwords[0] ), ntohl ( gid->u.dwords[1] ),
+	       ntohl ( gid->u.dwords[2] ), ntohl ( gid->u.dwords[3] ),
+	       qkey );
+
+	/* Set queue key */
+	qp->qkey = qkey;
+	if ( ( rc = ib_modify_qp ( ibdev, qp ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p QPN %lx could not modify qkey: %s\n",
+		       ibdev, qp->qpn, strerror ( rc ) );
+		goto out;
+	}
+
+ out:
+	/* Destroy the completed transaction */
+	ib_destroy_madx ( ibdev, mi, madx );
+	membership->madx = NULL;
+
+	/* Hand off to upper completion handler */
+	membership->complete ( ibdev, qp, membership, rc, mad );
+}
+
+/** Multicast membership management transaction completion operations */
+static struct ib_mad_transaction_operations ib_mcast_op = {
+	.complete = ib_mcast_complete,
+};
+
+/**
+ * Join multicast group
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v membership	Multicast group membership
+ * @v gid		Multicast GID to join
+ * @v joined		Join completion handler
+ * @ret rc		Return status code
+ */
+int ib_mcast_join ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+		    struct ib_mc_membership *membership, struct ib_gid *gid,
+		    void ( * complete ) ( struct ib_device *ibdev,
+					  struct ib_queue_pair *qp,
+					  struct ib_mc_membership *membership,
+					  int rc, union ib_mad *mad ) ) {
+	union ib_mad mad;
+	int rc;
+
+	DBGC ( ibdev, "IBDEV %p QPN %lx joining %08x:%08x:%08x:%08x\n",
+	       ibdev, qp->qpn, ntohl ( gid->u.dwords[0] ),
+	       ntohl ( gid->u.dwords[1] ), ntohl ( gid->u.dwords[2] ),
+	       ntohl ( gid->u.dwords[3] ) );
+
+	/* Initialise structure */
+	membership->qp = qp;
+	memcpy ( &membership->gid, gid, sizeof ( membership->gid ) );
+	membership->complete = complete;
+
+	/* Attach queue pair to multicast GID */
+	if ( ( rc = ib_mcast_attach ( ibdev, qp, gid ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p QPN %lx could not attach: %s\n",
+		       ibdev, qp->qpn, strerror ( rc ) );
+		goto err_mcast_attach;
+	}
+
+	/* Initiate multicast membership join */
+	ib_mcast_mad ( ibdev, gid, 1, &mad );
+	membership->madx = ib_create_madx ( ibdev, ibdev->gsi, &mad, NULL,
+					    &ib_mcast_op );
+	if ( ! membership->madx ) {
+		DBGC ( ibdev, "IBDEV %p QPN %lx could not create join "
+		       "transaction\n", ibdev, qp->qpn );
+		rc = -ENOMEM;
+		goto err_create_madx;
+	}
+	ib_madx_set_ownerdata ( membership->madx, membership );
+
+	return 0;
+
+	ib_destroy_madx ( ibdev, ibdev->gsi, membership->madx );
+ err_create_madx:
+	ib_mcast_detach ( ibdev, qp, gid );
+ err_mcast_attach:
+	return rc;
+}
+
+/**
+ * Leave multicast group
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v membership	Multicast group membership
+ */
+void ib_mcast_leave ( struct ib_device *ibdev, struct ib_queue_pair *qp,
+		      struct ib_mc_membership *membership ) {
+	struct ib_gid *gid = &membership->gid;
+	union ib_mad mad;
+	int rc;
+
+	DBGC ( ibdev, "IBDEV %p QPN %lx leaving %08x:%08x:%08x:%08x\n",
+	       ibdev, qp->qpn, ntohl ( gid->u.dwords[0] ),
+	       ntohl ( gid->u.dwords[1] ), ntohl ( gid->u.dwords[2] ),
+	       ntohl ( gid->u.dwords[3] ) );
+
+	/* Detach from multicast GID */
+	ib_mcast_detach ( ibdev, qp, &membership->gid );
+
+	/* Cancel multicast membership join, if applicable */
+	if ( membership->madx ) {
+		ib_destroy_madx ( ibdev, ibdev->gsi, membership->madx );
+		membership->madx = NULL;
+	}
+
+	/* Send a single group leave MAD */
+	ib_mcast_mad ( ibdev, &membership->gid, 0, &mad );
+	if ( ( rc = ib_mi_send ( ibdev, ibdev->gsi, &mad, NULL ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p QPN %lx could not send leave request: "
+		       "%s\n", ibdev, qp->qpn, strerror ( rc ) );
+	}
+}
diff --git a/gpxe/src/net/infiniband/ib_mi.c b/gpxe/src/net/infiniband/ib_mi.c
new file mode 100644
index 00000000..7511fd87
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_mi.c
@@ -0,0 +1,406 @@
+/*
+ * Copyright (C) 2009 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <byteswap.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/ib_mi.h>
+
+/**
+ * @file
+ *
+ * Infiniband management interfaces
+ *
+ */
+
+/** Management interface number of send WQEs
+ *
+ * This is a policy decision.
+ */
+#define IB_MI_NUM_SEND_WQES 4
+
+/** Management interface number of receive WQEs
+ *
+ * This is a policy decision.
+ */
+#define IB_MI_NUM_RECV_WQES 2
+
+/** Management interface number of completion queue entries
+ *
+ * This is a policy decision
+ */
+#define IB_MI_NUM_CQES 8
+
+/** TID magic signature */
+#define IB_MI_TID_MAGIC ( ( 'g' << 24 ) | ( 'P' << 16 ) | ( 'X' << 8 ) | 'E' )
+
+/** TID to use for next MAD */
+static unsigned int next_tid;
+
+/**
+ * Handle received MAD
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ * @ret rc		Return status code
+ */
+static int ib_mi_handle ( struct ib_device *ibdev,
+			  struct ib_mad_interface *mi,
+			  union ib_mad *mad,
+			  struct ib_address_vector *av ) {
+	struct ib_mad_hdr *hdr = &mad->hdr;
+	struct ib_mad_transaction *madx;
+	struct ib_mad_agent *agent;
+
+	/* Look for a matching transaction by TID */
+	list_for_each_entry ( madx, &mi->madx, list ) {
+		if ( memcmp ( &hdr->tid, &madx->mad.hdr.tid,
+			      sizeof ( hdr->tid ) ) != 0 )
+			continue;
+		/* Found a matching transaction */
+		madx->op->complete ( ibdev, mi, madx, 0, mad, av );
+		return 0;
+	}
+
+	/* If there is no matching transaction, look for a listening agent */
+	for_each_table_entry ( agent, IB_MAD_AGENTS ) {
+		if ( ( ( agent->mgmt_class & IB_MGMT_CLASS_MASK ) !=
+		       ( hdr->mgmt_class & IB_MGMT_CLASS_MASK ) ) ||
+		     ( agent->class_version != hdr->class_version ) ||
+		     ( agent->attr_id != hdr->attr_id ) )
+			continue;
+		/* Found a matching agent */
+		agent->handle ( ibdev, mi, mad, av );
+		return 0;
+	}
+
+	/* Otherwise, ignore it */
+	DBGC ( mi, "MI %p RX TID %08x%08x ignored\n",
+	       mi, ntohl ( hdr->tid[0] ), ntohl ( hdr->tid[1] ) );
+	return -ENOTSUP;
+}
+
+/**
+ * Complete receive via management interface
+ *
+ *
+ * @v ibdev		Infiniband device
+ * @v qp		Queue pair
+ * @v av		Address vector
+ * @v iobuf		I/O buffer
+ * @v rc		Completion status code
+ */
+static void ib_mi_complete_recv ( struct ib_device *ibdev,
+				  struct ib_queue_pair *qp,
+				  struct ib_address_vector *av,
+				  struct io_buffer *iobuf, int rc ) {
+	struct ib_mad_interface *mi = ib_qp_get_ownerdata ( qp );
+	union ib_mad *mad;
+	struct ib_mad_hdr *hdr;
+
+	/* Ignore errors */
+	if ( rc != 0 ) {
+		DBGC ( mi, "MI %p RX error: %s\n", mi, strerror ( rc ) );
+		goto out;
+	}
+
+	/* Sanity checks */
+	if ( iob_len ( iobuf ) != sizeof ( *mad ) ) {
+		DBGC ( mi, "MI %p RX bad size (%zd bytes)\n",
+		       mi, iob_len ( iobuf ) );
+		DBGC_HDA ( mi, 0, iobuf->data, iob_len ( iobuf ) );
+		goto out;
+	}
+	mad = iobuf->data;
+	hdr = &mad->hdr;
+	if ( hdr->base_version != IB_MGMT_BASE_VERSION ) {
+		DBGC ( mi, "MI %p RX unsupported base version %x\n",
+		       mi, hdr->base_version );
+		DBGC_HDA ( mi, 0, mad, sizeof ( *mad ) );
+		goto out;
+	}
+	DBGC ( mi, "MI %p RX TID %08x%08x (%02x,%02x,%02x,%04x) status "
+	       "%04x\n", mi, ntohl ( hdr->tid[0] ), ntohl ( hdr->tid[1] ),
+	       hdr->mgmt_class, hdr->class_version, hdr->method,
+	       ntohs ( hdr->attr_id ), ntohs ( hdr->status ) );
+	DBGC2_HDA ( mi, 0, mad, sizeof ( *mad ) );
+
+	/* Handle MAD */
+	if ( ( rc = ib_mi_handle ( ibdev, mi, mad, av ) ) != 0 )
+		goto out;
+
+ out:
+	free_iob ( iobuf );
+}
+
+/** Management interface completion operations */
+static struct ib_completion_queue_operations ib_mi_completion_ops = {
+	.complete_recv = ib_mi_complete_recv,
+};
+
+/**
+ * Transmit MAD
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		MAD
+ * @v av		Destination address vector
+ * @ret rc		Return status code
+ */
+int ib_mi_send ( struct ib_device *ibdev, struct ib_mad_interface *mi,
+		 union ib_mad *mad, struct ib_address_vector *av ) {
+	struct ib_mad_hdr *hdr = &mad->hdr;
+	struct io_buffer *iobuf;
+	int rc;
+
+	/* Set common fields */
+	hdr->base_version = IB_MGMT_BASE_VERSION;
+	if ( ( hdr->tid[0] == 0 ) && ( hdr->tid[1] == 0 ) ) {
+		hdr->tid[0] = htonl ( IB_MI_TID_MAGIC );
+		hdr->tid[1] = htonl ( ++next_tid );
+	}
+	DBGC ( mi, "MI %p TX TID %08x%08x (%02x,%02x,%02x,%04x) status "
+	       "%04x\n", mi, ntohl ( hdr->tid[0] ), ntohl ( hdr->tid[1] ),
+	       hdr->mgmt_class, hdr->class_version, hdr->method,
+	       ntohs ( hdr->attr_id ), ntohs ( hdr->status ) );
+	DBGC2_HDA ( mi, 0, mad, sizeof ( *mad ) );
+
+	/* Construct directed route portion of response, if necessary */
+	if ( hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ) {
+		struct ib_mad_smp *smp = &mad->smp;
+		unsigned int hop_pointer;
+		unsigned int hop_count;
+
+		smp->mad_hdr.status |= htons ( IB_SMP_STATUS_D_INBOUND );
+		hop_pointer = smp->mad_hdr.class_specific.smp.hop_pointer;
+		hop_count = smp->mad_hdr.class_specific.smp.hop_count;
+		assert ( hop_count == hop_pointer );
+		if ( hop_pointer < ( sizeof ( smp->return_path.hops ) /
+				     sizeof ( smp->return_path.hops[0] ) ) ) {
+			smp->return_path.hops[hop_pointer] = ibdev->port;
+		} else {
+			DBGC ( mi, "MI %p TX TID %08x%08x invalid hop pointer "
+			       "%d\n", mi, ntohl ( hdr->tid[0] ),
+			       ntohl ( hdr->tid[1] ), hop_pointer );
+			return -EINVAL;
+		}
+	}
+
+	/* Construct I/O buffer */
+	iobuf = alloc_iob ( sizeof ( *mad ) );
+	if ( ! iobuf ) {
+		DBGC ( mi, "MI %p could not allocate buffer for TID "
+		       "%08x%08x\n",
+		       mi, ntohl ( hdr->tid[0] ), ntohl ( hdr->tid[1] ) );
+		return -ENOMEM;
+	}
+	memcpy ( iob_put ( iobuf, sizeof ( *mad ) ), mad, sizeof ( *mad ) );
+
+	/* Send I/O buffer */
+	if ( ( rc = ib_post_send ( ibdev, mi->qp, av, iobuf ) ) != 0 ) {
+		DBGC ( mi, "MI %p TX TID %08x%08x failed: %s\n",
+		       mi,  ntohl ( hdr->tid[0] ), ntohl ( hdr->tid[1] ),
+		       strerror ( rc ) );
+		free_iob ( iobuf );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Handle management transaction timer expiry
+ *
+ * @v timer		Retry timer
+ * @v expired		Failure indicator
+ */
+static void ib_mi_timer_expired ( struct retry_timer *timer, int expired ) {
+	struct ib_mad_transaction *madx =
+		container_of ( timer, struct ib_mad_transaction, timer );
+	struct ib_mad_interface *mi = madx->mi;
+	struct ib_device *ibdev = mi->ibdev;
+	struct ib_mad_hdr *hdr = &madx->mad.hdr;
+
+	/* Abandon transaction if we have tried too many times */
+	if ( expired ) {
+		DBGC ( mi, "MI %p abandoning TID %08x%08x\n",
+		       mi, ntohl ( hdr->tid[0] ), ntohl ( hdr->tid[1] ) );
+		madx->op->complete ( ibdev, mi, madx, -ETIMEDOUT, NULL, NULL );
+		return;
+	}
+
+	/* Restart retransmission timer */
+	start_timer ( timer );
+
+	/* Resend MAD */
+	ib_mi_send ( ibdev, mi, &madx->mad, &madx->av );
+}
+
+/**
+ * Create management transaction
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		MAD to send
+ * @v av		Destination address, or NULL to use SM's GSI
+ * @v op		Management transaction operations
+ * @ret madx		Management transaction, or NULL
+ */
+struct ib_mad_transaction *
+ib_create_madx ( struct ib_device *ibdev, struct ib_mad_interface *mi,
+		 union ib_mad *mad, struct ib_address_vector *av,
+		 struct ib_mad_transaction_operations *op ) {
+	struct ib_mad_transaction *madx;
+
+	/* Allocate and initialise structure */
+	madx = zalloc ( sizeof ( *madx ) );
+	if ( ! madx )
+		return NULL;
+	madx->mi = mi;
+	madx->timer.expired = ib_mi_timer_expired;
+	madx->op = op;
+
+	/* Determine address vector */
+	if ( av ) {
+		memcpy ( &madx->av, av, sizeof ( madx->av ) );
+	} else {
+		madx->av.lid = ibdev->sm_lid;
+		madx->av.sl = ibdev->sm_sl;
+		madx->av.qpn = IB_QPN_GSI;
+		madx->av.qkey = IB_QKEY_GSI;
+	}
+
+	/* Copy MAD */
+	memcpy ( &madx->mad, mad, sizeof ( madx->mad ) );
+
+	/* Add to list and start timer to send initial MAD */
+	list_add ( &madx->list, &mi->madx );
+	start_timer_nodelay ( &madx->timer );
+
+	return madx;
+}
+
+/**
+ * Destroy management transaction
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v madx		Management transaction
+ */
+void ib_destroy_madx ( struct ib_device *ibdev __unused,
+		       struct ib_mad_interface *mi __unused,
+		       struct ib_mad_transaction *madx ) {
+
+	/* Stop timer and remove from list */
+	stop_timer ( &madx->timer );
+	list_del ( &madx->list );
+
+	/* Free transaction */
+	free ( madx );
+}
+
+/**
+ * Create management interface
+ *
+ * @v ibdev		Infiniband device
+ * @v type		Queue pair type
+ * @ret mi		Management agent, or NULL
+ */
+struct ib_mad_interface * ib_create_mi ( struct ib_device *ibdev,
+					 enum ib_queue_pair_type type ) {
+	struct ib_mad_interface *mi;
+	int rc;
+
+	/* Allocate and initialise fields */
+	mi = zalloc ( sizeof ( *mi ) );
+	if ( ! mi )
+		goto err_alloc;
+	mi->ibdev = ibdev;
+	INIT_LIST_HEAD ( &mi->madx );
+
+	/* Create completion queue */
+	mi->cq = ib_create_cq ( ibdev, IB_MI_NUM_CQES, &ib_mi_completion_ops );
+	if ( ! mi->cq ) {
+		DBGC ( mi, "MI %p could not allocate completion queue\n", mi );
+		goto err_create_cq;
+	}
+
+	/* Create queue pair */
+	mi->qp = ib_create_qp ( ibdev, type, IB_MI_NUM_SEND_WQES, mi->cq,
+				IB_MI_NUM_RECV_WQES, mi->cq );
+	if ( ! mi->qp ) {
+		DBGC ( mi, "MI %p could not allocate queue pair\n", mi );
+		goto err_create_qp;
+	}
+	ib_qp_set_ownerdata ( mi->qp, mi );
+	DBGC ( mi, "MI %p (%s) running on QPN %#lx\n",
+	       mi, ( ( type == IB_QPT_SMI ) ? "SMI" : "GSI" ), mi->qp->qpn );
+
+	/* Set queue key */
+	mi->qp->qkey = ( ( type == IB_QPT_SMI ) ? IB_QKEY_SMI : IB_QKEY_GSI );
+	if ( ( rc = ib_modify_qp ( ibdev, mi->qp ) ) != 0 ) {
+		DBGC ( mi, "MI %p could not set queue key: %s\n",
+		       mi, strerror ( rc ) );
+		goto err_modify_qp;
+	}
+
+	/* Fill receive ring */
+	ib_refill_recv ( ibdev, mi->qp );
+	return mi;
+
+ err_modify_qp:
+	ib_destroy_qp ( ibdev, mi->qp );
+ err_create_qp:
+	ib_destroy_cq ( ibdev, mi->cq );
+ err_create_cq:
+	free ( mi );
+ err_alloc:
+	return NULL;
+}
+
+/**
+ * Destroy management interface
+ *
+ * @v mi		Management interface
+ */
+void ib_destroy_mi ( struct ib_device *ibdev, struct ib_mad_interface *mi ) {
+	struct ib_mad_transaction *madx;
+	struct ib_mad_transaction *tmp;
+
+	/* Flush any outstanding requests */
+	list_for_each_entry_safe ( madx, tmp, &mi->madx, list ) {
+		DBGC ( mi, "MI %p destroyed while TID %08x%08x in progress\n",
+		       mi, ntohl ( madx->mad.hdr.tid[0] ),
+		       ntohl ( madx->mad.hdr.tid[1] ) );
+		madx->op->complete ( ibdev, mi, madx, -ECANCELED, NULL, NULL );
+	}
+
+	ib_destroy_qp ( ibdev, mi->qp );
+	ib_destroy_cq ( ibdev, mi->cq );
+	free ( mi );
+}
diff --git a/gpxe/src/net/infiniband/ib_packet.c b/gpxe/src/net/infiniband/ib_packet.c
new file mode 100644
index 00000000..08820ef3
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_packet.c
@@ -0,0 +1,244 @@
+/*
+ * Copyright (C) 2008 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <byteswap.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_packet.h>
+
+/**
+ * @file
+ *
+ * Infiniband Packet Formats
+ *
+ */
+
+/**
+ * Add IB headers
+ *
+ * @v ibdev		Infiniband device
+ * @v iobuf		I/O buffer to contain headers
+ * @v qp		Queue pair
+ * @v payload_len	Payload length
+ * @v av		Address vector
+ */
+int ib_push ( struct ib_device *ibdev, struct io_buffer *iobuf,
+	      struct ib_queue_pair *qp, size_t payload_len,
+	      const struct ib_address_vector *av ) {
+	struct ib_local_route_header *lrh;
+	struct ib_global_route_header *grh;
+	struct ib_base_transport_header *bth;
+	struct ib_datagram_extended_transport_header *deth;
+	size_t orig_iob_len = iob_len ( iobuf );
+	size_t pad_len;
+	size_t lrh_len;
+	size_t grh_len;
+	unsigned int vl;
+	unsigned int lnh;
+
+	DBGC2 ( ibdev, "IBDEV %p TX %04x:%08lx => %04x:%08lx (key %08lx)\n",
+		ibdev, ibdev->lid, qp->ext_qpn, av->lid, av->qpn, av->qkey );
+
+	/* Calculate packet length */
+	pad_len = ( (-payload_len) & 0x3 );
+	payload_len += pad_len;
+	payload_len += 4; /* ICRC */
+
+	/* Reserve space for headers */
+	orig_iob_len = iob_len ( iobuf );
+	deth = iob_push ( iobuf, sizeof ( *deth ) );
+	bth = iob_push ( iobuf, sizeof ( *bth ) );
+	grh_len = ( payload_len + iob_len ( iobuf ) - orig_iob_len );
+	grh = ( av->gid_present ?
+		iob_push ( iobuf, sizeof ( *grh ) ) : NULL );
+	lrh = iob_push ( iobuf, sizeof ( *lrh ) );
+	lrh_len = ( payload_len + iob_len ( iobuf ) - orig_iob_len );
+
+	/* Construct LRH */
+	vl = ( ( qp->ext_qpn == IB_QPN_SMI ) ? IB_VL_SMP : IB_VL_DEFAULT );
+	lrh->vl__lver = ( vl << 4 );
+	lnh = ( grh ? IB_LNH_GRH : IB_LNH_BTH );
+	lrh->sl__lnh = ( ( av->sl << 4 ) | lnh );
+	lrh->dlid = htons ( av->lid );
+	lrh->length = htons ( lrh_len >> 2 );
+	lrh->slid = htons ( ibdev->lid );
+
+	/* Construct GRH, if required */
+	if ( grh ) {
+		grh->ipver__tclass__flowlabel =
+			htonl ( IB_GRH_IPVER_IPv6 << 28 );
+		grh->paylen = htons ( grh_len );
+		grh->nxthdr = IB_GRH_NXTHDR_IBA;
+		grh->hoplmt = 0;
+		memcpy ( &grh->sgid, &ibdev->gid, sizeof ( grh->sgid ) );
+		memcpy ( &grh->dgid, &av->gid, sizeof ( grh->dgid ) );
+	}
+
+	/* Construct BTH */
+	bth->opcode = BTH_OPCODE_UD_SEND;
+	bth->se__m__padcnt__tver = ( pad_len << 4 );
+	bth->pkey = htons ( ibdev->pkey );
+	bth->dest_qp = htonl ( av->qpn );
+	bth->ack__psn = htonl ( ( qp->send.psn++ ) & 0xffffffUL );
+
+	/* Construct DETH */
+	deth->qkey = htonl ( av->qkey );
+	deth->src_qp = htonl ( qp->ext_qpn );
+
+	DBGCP_HDA ( ibdev, 0, iobuf->data,
+		    ( iob_len ( iobuf ) - orig_iob_len ) );
+
+	return 0;
+}
+
+/**
+ * Remove IB headers
+ *
+ * @v ibdev		Infiniband device
+ * @v iobuf		I/O buffer containing headers
+ * @v qp		Queue pair to fill in, or NULL
+ * @v payload_len	Payload length to fill in, or NULL
+ * @v av		Address vector to fill in
+ */
+int ib_pull ( struct ib_device *ibdev, struct io_buffer *iobuf,
+	      struct ib_queue_pair **qp, size_t *payload_len,
+	      struct ib_address_vector *av ) {
+	struct ib_local_route_header *lrh;
+	struct ib_global_route_header *grh;
+	struct ib_base_transport_header *bth;
+	struct ib_datagram_extended_transport_header *deth;
+	size_t orig_iob_len = iob_len ( iobuf );
+	unsigned int lnh;
+	size_t pad_len;
+	unsigned long qpn;
+	unsigned int lid;
+
+	/* Clear return values */
+	if ( qp )
+		*qp = NULL;
+	if ( payload_len )
+		*payload_len = 0;
+	memset ( av, 0, sizeof ( *av ) );
+
+	/* Extract LRH */
+	if ( iob_len ( iobuf ) < sizeof ( *lrh ) ) {
+		DBGC ( ibdev, "IBDEV %p RX too short (%zd bytes) for LRH\n",
+		       ibdev, iob_len ( iobuf ) );
+		return -EINVAL;
+	}
+	lrh = iobuf->data;
+	iob_pull ( iobuf, sizeof ( *lrh ) );
+	av->lid = ntohs ( lrh->slid );
+	av->sl = ( lrh->sl__lnh >> 4 );
+	lnh = ( lrh->sl__lnh & 0x3 );
+	lid = ntohs ( lrh->dlid );
+
+	/* Reject unsupported packets */
+	if ( ! ( ( lnh == IB_LNH_BTH ) || ( lnh == IB_LNH_GRH ) ) ) {
+		DBGC ( ibdev, "IBDEV %p RX unsupported LNH %x\n",
+		       ibdev, lnh );
+		return -ENOTSUP;
+	}
+
+	/* Extract GRH, if present */
+	if ( lnh == IB_LNH_GRH ) {
+		if ( iob_len ( iobuf ) < sizeof ( *grh ) ) {
+			DBGC ( ibdev, "IBDEV %p RX too short (%zd bytes) "
+			       "for GRH\n", ibdev, iob_len ( iobuf ) );
+			return -EINVAL;
+		}
+		grh = iobuf->data;
+		iob_pull ( iobuf, sizeof ( *grh ) );
+		av->gid_present = 1;
+		memcpy ( &av->gid, &grh->sgid, sizeof ( av->gid ) );
+	} else {
+		grh = NULL;
+	}
+
+	/* Extract BTH */
+	if ( iob_len ( iobuf ) < sizeof ( *bth ) ) {
+		DBGC ( ibdev, "IBDEV %p RX too short (%zd bytes) for BTH\n",
+		       ibdev, iob_len ( iobuf ) );
+		return -EINVAL;
+	}
+	bth = iobuf->data;
+	iob_pull ( iobuf, sizeof ( *bth ) );
+	if ( bth->opcode != BTH_OPCODE_UD_SEND ) {
+		DBGC ( ibdev, "IBDEV %p unsupported BTH opcode %x\n",
+		       ibdev, bth->opcode );
+		return -ENOTSUP;
+	}
+	qpn = ntohl ( bth->dest_qp );
+
+	/* Extract DETH */
+	if ( iob_len ( iobuf ) < sizeof ( *deth ) ) {
+		DBGC ( ibdev, "IBDEV %p RX too short (%zd bytes) for DETH\n",
+		       ibdev, iob_len ( iobuf ) );
+		return -EINVAL;
+	}
+	deth = iobuf->data;
+	iob_pull ( iobuf, sizeof ( *deth ) );
+	av->qpn = ntohl ( deth->src_qp );
+	av->qkey = ntohl ( deth->qkey );
+
+	/* Calculate payload length, if applicable */
+	if ( payload_len ) {
+		pad_len = ( ( bth->se__m__padcnt__tver >> 4 ) & 0x3 );
+		*payload_len = ( ( ntohs ( lrh->length ) << 2 )
+				 - ( orig_iob_len - iob_len ( iobuf ) )
+				 - pad_len - 4 /* ICRC */ );
+	}
+
+	/* Determine destination QP, if applicable */
+	if ( qp ) {
+		if ( IB_LID_MULTICAST ( lid ) && grh ) {
+			if ( ! ( *qp = ib_find_qp_mgid ( ibdev, &grh->dgid ))){
+				DBGC ( ibdev, "IBDEV %p RX for unknown MGID "
+				       "%08x:%08x:%08x:%08x\n", ibdev,
+				       ntohl ( grh->dgid.u.dwords[0] ),
+				       ntohl ( grh->dgid.u.dwords[1] ),
+				       ntohl ( grh->dgid.u.dwords[2] ),
+				       ntohl ( grh->dgid.u.dwords[3] ) );
+				return -ENODEV;
+			}
+		} else {
+			if ( ! ( *qp = ib_find_qp_qpn ( ibdev, qpn ) ) ) {
+				DBGC ( ibdev, "IBDEV %p RX for nonexistent "
+				       "QPN %lx\n", ibdev, qpn );
+				return -ENODEV;
+			}
+		}
+		assert ( *qp );
+	}
+
+	DBGC2 ( ibdev, "IBDEV %p RX %04x:%08lx <= %04x:%08lx (key %08x)\n",
+		ibdev, lid, ( IB_LID_MULTICAST( lid ) ?
+			      ( qp ? (*qp)->ext_qpn : -1UL ) : qpn ),
+		av->lid, av->qpn, ntohl ( deth->qkey ) );
+	DBGCP_HDA ( ibdev, 0,
+		    ( iobuf->data - ( orig_iob_len - iob_len ( iobuf ) ) ),
+		    ( orig_iob_len - iob_len ( iobuf ) ) );
+
+	return 0;
+}
diff --git a/gpxe/src/net/infiniband/ib_pathrec.c b/gpxe/src/net/infiniband/ib_pathrec.c
new file mode 100644
index 00000000..136e628d
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_pathrec.c
@@ -0,0 +1,296 @@
+/*
+ * Copyright (C) 2009 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <byteswap.h>
+#include <errno.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_mi.h>
+#include <gpxe/ib_pathrec.h>
+
+/** @file
+ *
+ * Infiniband path lookups
+ *
+ */
+
+/**
+ * Handle path transaction completion
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v madx		Management transaction
+ * @v rc		Status code
+ * @v mad		Received MAD (or NULL on error)
+ * @v av		Source address vector (or NULL on error)
+ */
+static void ib_path_complete ( struct ib_device *ibdev,
+			       struct ib_mad_interface *mi,
+			       struct ib_mad_transaction *madx,
+			       int rc, union ib_mad *mad,
+			       struct ib_address_vector *av __unused ) {
+	struct ib_path *path = ib_madx_get_ownerdata ( madx );
+	struct ib_gid *dgid = &path->av.gid;
+	struct ib_path_record *pathrec = &mad->sa.sa_data.path_record;
+
+	/* Report failures */
+	if ( ( rc == 0 ) && ( mad->hdr.status != htons ( IB_MGMT_STATUS_OK ) ))
+		rc = -ENETUNREACH;
+	if ( rc != 0 ) {
+		DBGC ( ibdev, "IBDEV %p path lookup for %08x:%08x:%08x:%08x "
+		       "failed: %s\n", ibdev, htonl ( dgid->u.dwords[0] ),
+		       htonl ( dgid->u.dwords[1] ),
+		       htonl ( dgid->u.dwords[2] ),
+		       htonl ( dgid->u.dwords[3] ), strerror ( rc ) );
+		goto out;
+	}
+
+	/* Extract values from MAD */
+	path->av.lid = ntohs ( pathrec->dlid );
+	path->av.sl = ( pathrec->reserved__sl & 0x0f );
+	path->av.rate = ( pathrec->rate_selector__rate & 0x3f );
+	DBGC ( ibdev, "IBDEV %p path to %08x:%08x:%08x:%08x is %04x sl %d "
+	       "rate %d\n", ibdev, htonl ( dgid->u.dwords[0] ),
+	       htonl ( dgid->u.dwords[1] ), htonl ( dgid->u.dwords[2] ),
+	       htonl ( dgid->u.dwords[3] ), path->av.lid, path->av.sl,
+	       path->av.rate );
+
+ out:
+	/* Destroy the completed transaction */
+	ib_destroy_madx ( ibdev, mi, madx );
+	path->madx = NULL;
+
+	/* Hand off to upper completion handler */
+	path->op->complete ( ibdev, path, rc, &path->av );
+}
+
+/** Path transaction completion operations */
+static struct ib_mad_transaction_operations ib_path_op = {
+	.complete = ib_path_complete,
+};
+
+/**
+ * Create path
+ *
+ * @v ibdev		Infiniband device
+ * @v av		Address vector to complete
+ * @v op		Path operations
+ * @ret path		Path
+ */
+struct ib_path *
+ib_create_path ( struct ib_device *ibdev, struct ib_address_vector *av,
+		 struct ib_path_operations *op ) {
+	struct ib_path *path;
+	union ib_mad mad;
+	struct ib_mad_sa *sa = &mad.sa;
+
+	/* Allocate and initialise structure */
+	path = zalloc ( sizeof ( *path ) );
+	if ( ! path )
+		goto err_alloc_path;
+	path->ibdev = ibdev;
+	memcpy ( &path->av, av, sizeof ( path->av ) );
+	path->op = op;
+
+	/* Construct path request */
+	memset ( sa, 0, sizeof ( *sa ) );
+	sa->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
+	sa->mad_hdr.class_version = IB_SA_CLASS_VERSION;
+	sa->mad_hdr.method = IB_MGMT_METHOD_GET;
+	sa->mad_hdr.attr_id = htons ( IB_SA_ATTR_PATH_REC );
+	sa->sa_hdr.comp_mask[1] =
+		htonl ( IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID );
+	memcpy ( &sa->sa_data.path_record.dgid, &path->av.gid,
+		 sizeof ( sa->sa_data.path_record.dgid ) );
+	memcpy ( &sa->sa_data.path_record.sgid, &ibdev->gid,
+		 sizeof ( sa->sa_data.path_record.sgid ) );
+
+	/* Create management transaction */
+	path->madx = ib_create_madx ( ibdev, ibdev->gsi, &mad, NULL,
+				      &ib_path_op );
+	if ( ! path->madx )
+		goto err_create_madx;
+	ib_madx_set_ownerdata ( path->madx, path );
+
+	return path;
+
+	ib_destroy_madx ( ibdev, ibdev->gsi, path->madx );
+ err_create_madx:
+	free ( path );
+ err_alloc_path:
+	return NULL;
+}
+
+/**
+ * Destroy path
+ *
+ * @v ibdev		Infiniband device
+ * @v path		Path
+ */
+void ib_destroy_path ( struct ib_device *ibdev, struct ib_path *path ) {
+
+	if ( path->madx )
+		ib_destroy_madx ( ibdev, ibdev->gsi, path->madx );
+	free ( path );
+}
+
+/** Number of path cache entries
+ *
+ * Must be a power of two.
+ */
+#define IB_NUM_CACHED_PATHS 4
+
+/** A cached path */
+struct ib_cached_path {
+	/** Path */
+	struct ib_path *path;
+};
+
+/** Path cache */
+static struct ib_cached_path ib_path_cache[IB_NUM_CACHED_PATHS];
+
+/** Oldest path cache entry index */
+static unsigned int ib_path_cache_idx;
+
+/**
+ * Find path cache entry
+ *
+ * @v ibdev		Infiniband device
+ * @v dgid		Destination GID
+ * @ret path		Path cache entry, or NULL
+ */
+static struct ib_cached_path *
+ib_find_path_cache_entry ( struct ib_device *ibdev, struct ib_gid *dgid ) {
+	struct ib_cached_path *cached;
+	unsigned int i;
+
+	for ( i = 0 ; i < IB_NUM_CACHED_PATHS ; i++ ) {
+		cached = &ib_path_cache[i];
+		if ( ! cached->path )
+			continue;
+		if ( cached->path->ibdev != ibdev )
+			continue;
+		if ( memcmp ( &cached->path->av.gid, dgid,
+			      sizeof ( cached->path->av.gid ) ) != 0 )
+			continue;
+		return cached;
+	}
+
+	return NULL;
+}
+
+/**
+ * Handle cached path transaction completion
+ *
+ * @v ibdev		Infiniband device
+ * @v path		Path
+ * @v rc		Status code
+ * @v av		Address vector, or NULL on error
+ */
+static void ib_cached_path_complete ( struct ib_device *ibdev,
+				      struct ib_path *path, int rc,
+				      struct ib_address_vector *av __unused ) {
+	struct ib_cached_path *cached = ib_path_get_ownerdata ( path );
+
+	/* If the transaction failed, erase the cache entry */
+	if ( rc != 0 ) {
+		/* Destroy the old cache entry */
+		ib_destroy_path ( ibdev, path );
+		memset ( cached, 0, sizeof ( *cached ) );
+		return;
+	}
+
+	/* Do not destroy the completed transaction; we still need to
+	 * refer to the resolved path.
+	 */
+}
+
+/** Cached path transaction completion operations */
+static struct ib_path_operations ib_cached_path_op = {
+	.complete = ib_cached_path_complete,
+};
+
+/**
+ * Resolve path
+ *
+ * @v ibdev		Infiniband device
+ * @v av		Address vector to complete
+ * @ret rc		Return status code
+ *
+ * This provides a non-transactional way to resolve a path, via a
+ * cache similar to ARP.
+ */
+int ib_resolve_path ( struct ib_device *ibdev, struct ib_address_vector *av ) {
+	struct ib_gid *gid = &av->gid;
+	struct ib_cached_path *cached;
+	unsigned int cache_idx;
+
+	/* Sanity check */
+	if ( ! av->gid_present ) {
+		DBGC ( ibdev, "IBDEV %p attempt to look up path "
+		       "without GID\n", ibdev );
+		return -EINVAL;
+	}
+
+	/* Look in cache for a matching entry */
+	cached = ib_find_path_cache_entry ( ibdev, gid );
+	if ( cached && cached->path->av.lid ) {
+		/* Populated entry found */
+		av->lid = cached->path->av.lid;
+		av->rate = cached->path->av.rate;
+		av->sl = cached->path->av.sl;
+		DBGC2 ( ibdev, "IBDEV %p cache hit for %08x:%08x:%08x:%08x\n",
+			ibdev, htonl ( gid->u.dwords[0] ),
+			htonl ( gid->u.dwords[1] ), htonl ( gid->u.dwords[2] ),
+			htonl ( gid->u.dwords[3] ) );
+		return 0;
+	}
+	DBGC ( ibdev, "IBDEV %p cache miss for %08x:%08x:%08x:%08x%s\n",
+	       ibdev, htonl ( gid->u.dwords[0] ), htonl ( gid->u.dwords[1] ),
+	       htonl ( gid->u.dwords[2] ), htonl ( gid->u.dwords[3] ),
+	       ( cached ? " (in progress)" : "" ) );
+
+	/* If lookup is already in progress, do nothing */
+	if ( cached )
+		return -ENOENT;
+
+	/* Locate a new cache entry to use */
+	cache_idx = ( (ib_path_cache_idx++) % IB_NUM_CACHED_PATHS );
+	cached = &ib_path_cache[cache_idx];
+
+	/* Destroy the old cache entry */
+	if ( cached->path )
+		ib_destroy_path ( ibdev, cached->path );
+	memset ( cached, 0, sizeof ( *cached ) );
+
+	/* Create new path */
+	cached->path = ib_create_path ( ibdev, av, &ib_cached_path_op );
+	if ( ! cached->path ) {
+		DBGC ( ibdev, "IBDEV %p could not create path\n",
+		       ibdev );
+		return -ENOMEM;
+	}
+	ib_path_set_ownerdata ( cached->path, cached );
+
+	/* Not found yet */
+	return -ENOENT;
+}
diff --git a/gpxe/src/net/infiniband/ib_sma.c b/gpxe/src/net/infiniband/ib_sma.c
new file mode 100644
index 00000000..ff4cbbf6
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_sma.c
@@ -0,0 +1,369 @@
+/*
+ * Copyright (C) 2009 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <byteswap.h>
+#include <gpxe/settings.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/iobuf.h>
+#include <gpxe/ib_mi.h>
+#include <gpxe/ib_sma.h>
+
+/**
+ * @file
+ *
+ * Infiniband Subnet Management Agent
+ *
+ */
+
+/**
+ * Node information
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ */
+static void ib_sma_node_info ( struct ib_device *ibdev,
+			       struct ib_mad_interface *mi,
+			       union ib_mad *mad,
+			       struct ib_address_vector *av ) {
+	struct ib_node_info *node_info = &mad->smp.smp_data.node_info;
+	int rc;
+
+	/* Fill in information */
+	memset ( node_info, 0, sizeof ( *node_info ) );
+	node_info->base_version = IB_MGMT_BASE_VERSION;
+	node_info->class_version = IB_SMP_CLASS_VERSION;
+	node_info->node_type = IB_NODE_TYPE_HCA;
+	node_info->num_ports = ib_get_hca_info ( ibdev, &node_info->sys_guid );
+	memcpy ( &node_info->node_guid, &node_info->sys_guid,
+		 sizeof ( node_info->node_guid ) );
+	memcpy ( &node_info->port_guid, &ibdev->gid.u.half[1],
+		 sizeof ( node_info->port_guid ) );
+	node_info->partition_cap = htons ( 1 );
+	node_info->local_port_num = ibdev->port;
+
+	/* Send GetResponse */
+	mad->hdr.method = IB_MGMT_METHOD_GET_RESP;
+	if ( ( rc = ib_mi_send ( ibdev, mi, mad, av ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not send NodeInfo GetResponse: %s\n",
+		       mi, strerror ( rc ) );
+		return;
+	}
+}
+
+/**
+ * Node description
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ */
+static void ib_sma_node_desc ( struct ib_device *ibdev,
+			       struct ib_mad_interface *mi,
+			       union ib_mad *mad,
+			       struct ib_address_vector *av ) {
+	struct ib_node_desc *node_desc = &mad->smp.smp_data.node_desc;
+	struct ib_gid_half guid;
+	char hostname[ sizeof ( node_desc->node_string ) ];
+	int hostname_len;
+	int rc;
+
+	/* Fill in information */
+	memset ( node_desc, 0, sizeof ( *node_desc ) );
+	ib_get_hca_info ( ibdev, &guid );
+	hostname_len = fetch_string_setting ( NULL, &hostname_setting,
+					      hostname, sizeof ( hostname ) );
+	snprintf ( node_desc->node_string, sizeof ( node_desc->node_string ),
+		   "gPXE %s%s%02x:%02x:%02x:%02x:%02x:%02x:%02x:%02x (%s)",
+		   hostname, ( ( hostname_len >= 0 ) ? " " : "" ),
+		   guid.u.bytes[0], guid.u.bytes[1], guid.u.bytes[2],
+		   guid.u.bytes[3], guid.u.bytes[4], guid.u.bytes[5],
+		   guid.u.bytes[6], guid.u.bytes[7], ibdev->dev->name );
+
+	/* Send GetResponse */
+	mad->hdr.method = IB_MGMT_METHOD_GET_RESP;
+	if ( ( rc = ib_mi_send ( ibdev, mi, mad, av ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not send NodeDesc GetResponse: %s\n",
+		       mi, strerror ( rc ) );
+		return;
+	}
+}
+
+/**
+ * GUID information
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ */
+static void ib_sma_guid_info ( struct ib_device *ibdev,
+			       struct ib_mad_interface *mi,
+			       union ib_mad *mad,
+			       struct ib_address_vector *av ) {
+	struct ib_guid_info *guid_info = &mad->smp.smp_data.guid_info;
+	int rc;
+
+	/* Fill in information */
+	memset ( guid_info, 0, sizeof ( *guid_info ) );
+	memcpy ( guid_info->guid[0], &ibdev->gid.u.half[1],
+		 sizeof ( guid_info->guid[0] ) );
+
+	/* Send GetResponse */
+	mad->hdr.method = IB_MGMT_METHOD_GET_RESP;
+	if ( ( rc = ib_mi_send ( ibdev, mi, mad, av ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not send GuidInfo GetResponse: %s\n",
+		       mi, strerror ( rc ) );
+		return;
+	}
+}
+
+/**
+ * Set port information
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @ret rc		Return status code
+ */
+static int ib_sma_set_port_info ( struct ib_device *ibdev,
+				  struct ib_mad_interface *mi,
+				  union ib_mad *mad ) {
+	const struct ib_port_info *port_info = &mad->smp.smp_data.port_info;
+	unsigned int link_width_enabled;
+	unsigned int link_speed_enabled;
+	int rc;
+
+	/* Set parameters */
+	memcpy ( &ibdev->gid.u.half[0], port_info->gid_prefix,
+		 sizeof ( ibdev->gid.u.half[0] ) );
+	ibdev->lid = ntohs ( port_info->lid );
+	ibdev->sm_lid = ntohs ( port_info->mastersm_lid );
+	if ( ( link_width_enabled = port_info->link_width_enabled ) )
+		ibdev->link_width_enabled = link_width_enabled;
+	if ( ( link_speed_enabled =
+	       ( port_info->link_speed_active__link_speed_enabled & 0xf ) ) )
+		ibdev->link_speed_enabled = link_speed_enabled;
+	ibdev->sm_sl = ( port_info->neighbour_mtu__mastersm_sl & 0xf );
+	DBGC ( mi, "SMA %p set LID %04x SMLID %04x link width %02x speed "
+	       "%02x\n", mi, ibdev->lid, ibdev->sm_lid,
+	       ibdev->link_width_enabled, ibdev->link_speed_enabled );
+
+	/* Update parameters on device */
+	if ( ( rc = ib_set_port_info ( ibdev, mad ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not set port information: %s\n",
+		       mi, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Port information
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ */
+static void ib_sma_port_info ( struct ib_device *ibdev,
+			       struct ib_mad_interface *mi,
+			       union ib_mad *mad,
+			       struct ib_address_vector *av ) {
+	struct ib_port_info *port_info = &mad->smp.smp_data.port_info;
+	int rc;
+
+	/* Set parameters if applicable */
+	if ( mad->hdr.method == IB_MGMT_METHOD_SET ) {
+		if ( ( rc = ib_sma_set_port_info ( ibdev, mi, mad ) ) != 0 ) {
+			mad->hdr.status =
+			      htons ( IB_MGMT_STATUS_UNSUPPORTED_METHOD_ATTR );
+			/* Fall through to generate GetResponse */
+		}
+	}
+
+	/* Fill in information */
+	memset ( port_info, 0, sizeof ( *port_info ) );
+	memcpy ( port_info->gid_prefix, &ibdev->gid.u.half[0],
+		 sizeof ( port_info->gid_prefix ) );
+	port_info->lid = ntohs ( ibdev->lid );
+	port_info->mastersm_lid = ntohs ( ibdev->sm_lid );
+	port_info->local_port_num = ibdev->port;
+	port_info->link_width_enabled = ibdev->link_width_enabled;
+	port_info->link_width_supported = ibdev->link_width_supported;
+	port_info->link_width_active = ibdev->link_width_active;
+	port_info->link_speed_supported__port_state =
+		( ( ibdev->link_speed_supported << 4 ) | ibdev->port_state );
+	port_info->port_phys_state__link_down_def_state =
+		( ( IB_PORT_PHYS_STATE_POLLING << 4 ) |
+		  IB_PORT_PHYS_STATE_POLLING );
+	port_info->link_speed_active__link_speed_enabled =
+		( ( ibdev->link_speed_active << 4 ) |
+		  ibdev->link_speed_enabled );
+	port_info->neighbour_mtu__mastersm_sl =
+		( ( IB_MTU_2048 << 4 ) | ibdev->sm_sl );
+	port_info->vl_cap__init_type = ( IB_VL_0 << 4 );
+	port_info->init_type_reply__mtu_cap = IB_MTU_2048;
+	port_info->operational_vls__enforcement = ( IB_VL_0 << 4 );
+	port_info->guid_cap = 1;
+
+	/* Send GetResponse */
+	mad->hdr.method = IB_MGMT_METHOD_GET_RESP;
+	if ( ( rc = ib_mi_send ( ibdev, mi, mad, av ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not send PortInfo GetResponse: %s\n",
+		       mi, strerror ( rc ) );
+		return;
+	}
+}
+
+/**
+ * Set partition key table
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @ret rc		Return status code
+ */
+static int ib_sma_set_pkey_table ( struct ib_device *ibdev,
+				   struct ib_mad_interface *mi,
+				   union ib_mad *mad ) {
+	struct ib_pkey_table *pkey_table = &mad->smp.smp_data.pkey_table;
+	int rc;
+
+	/* Set parameters */
+	ibdev->pkey = ntohs ( pkey_table->pkey[0] );
+	DBGC ( mi, "SMA %p set pkey %04x\n", mi, ibdev->pkey );
+
+	/* Update parameters on device */
+	if ( ( rc = ib_set_pkey_table ( ibdev, mad ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not set pkey table: %s\n",
+		       mi, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
+ * Partition key table
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @v mad		Received MAD
+ * @v av		Source address vector
+ */
+static void ib_sma_pkey_table ( struct ib_device *ibdev,
+				struct ib_mad_interface *mi,
+				union ib_mad *mad,
+				struct ib_address_vector *av ) {
+	struct ib_pkey_table *pkey_table = &mad->smp.smp_data.pkey_table;
+	int rc;
+
+	/* Set parameters, if applicable */
+	if ( mad->hdr.method == IB_MGMT_METHOD_SET ) {
+		if ( ( rc = ib_sma_set_pkey_table ( ibdev, mi, mad ) ) != 0 ) {
+			mad->hdr.status =
+			      htons ( IB_MGMT_STATUS_UNSUPPORTED_METHOD_ATTR );
+			/* Fall through to generate GetResponse */
+		}
+	}
+
+	/* Fill in information */
+	mad->hdr.method = IB_MGMT_METHOD_GET_RESP;
+	memset ( pkey_table, 0, sizeof ( *pkey_table ) );
+	pkey_table->pkey[0] = htons ( ibdev->pkey );
+
+	/* Send GetResponse */
+	mad->hdr.method = IB_MGMT_METHOD_GET_RESP;
+	if ( ( rc = ib_mi_send ( ibdev, mi, mad, av ) ) != 0 ) {
+		DBGC ( mi, "SMA %p could not send PKeyTable GetResponse: %s\n",
+		       mi, strerror ( rc ) );
+		return;
+	}
+}
+
+/** Subnet management agent */
+struct ib_mad_agent ib_sma_agent[] __ib_mad_agent = {
+	{
+		.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED,
+		.class_version = IB_SMP_CLASS_VERSION,
+		.attr_id = htons ( IB_SMP_ATTR_NODE_INFO ),
+		.handle = ib_sma_node_info,
+	},
+	{
+		.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED,
+		.class_version = IB_SMP_CLASS_VERSION,
+		.attr_id = htons ( IB_SMP_ATTR_NODE_DESC ),
+		.handle = ib_sma_node_desc,
+	},
+	{
+		.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED,
+		.class_version = IB_SMP_CLASS_VERSION,
+		.attr_id = htons ( IB_SMP_ATTR_GUID_INFO ),
+		.handle = ib_sma_guid_info,
+	},
+	{
+		.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED,
+		.class_version = IB_SMP_CLASS_VERSION,
+		.attr_id = htons ( IB_SMP_ATTR_PORT_INFO ),
+		.handle = ib_sma_port_info,
+	},
+	{
+		.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED,
+		.class_version = IB_SMP_CLASS_VERSION,
+		.attr_id = htons ( IB_SMP_ATTR_PKEY_TABLE ),
+		.handle = ib_sma_pkey_table,
+	},
+};
+
+/**
+ * Create subnet management agent and interface
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ * @ret rc		Return status code
+ */
+int ib_create_sma ( struct ib_device *ibdev, struct ib_mad_interface *mi ) {
+
+	/* Nothing to do */
+	DBGC ( ibdev, "IBDEV %p SMA using SMI %p\n", ibdev, mi );
+
+	return 0;
+}
+
+/**
+ * Destroy subnet management agent and interface
+ *
+ * @v ibdev		Infiniband device
+ * @v mi		Management interface
+ */
+void ib_destroy_sma ( struct ib_device *ibdev __unused,
+		      struct ib_mad_interface *mi __unused ) {
+	/* Nothing to do */
+}
diff --git a/gpxe/src/net/infiniband/ib_smc.c b/gpxe/src/net/infiniband/ib_smc.c
new file mode 100644
index 00000000..d308dd9d
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_smc.c
@@ -0,0 +1,179 @@
+/*
+ * Copyright (C) 2008 Michael Brown <mbrown@fensystems.co.uk>.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ */
+
+FILE_LICENCE ( GPL2_OR_LATER );
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <unistd.h>
+#include <byteswap.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_smc.h>
+
+/**
+ * @file
+ *
+ * Infiniband Subnet Management Client
+ *
+ */
+
+/**
+ * Get port information
+ *
+ * @v ibdev		Infiniband device
+ * @v local_mad		Method for issuing local MADs
+ * @v mad		Management datagram to fill in
+ * @ret rc		Return status code
+ */
+static int ib_smc_get_port_info ( struct ib_device *ibdev,
+				  ib_local_mad_t local_mad,
+				  union ib_mad *mad ) {
+	int rc;
+
+	/* Construct MAD */
+	memset ( mad, 0, sizeof ( *mad ) );
+	mad->hdr.base_version = IB_MGMT_BASE_VERSION;
+	mad->hdr.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
+	mad->hdr.class_version = 1;
+	mad->hdr.method = IB_MGMT_METHOD_GET;
+	mad->hdr.attr_id = htons ( IB_SMP_ATTR_PORT_INFO );
+	mad->hdr.attr_mod = htonl ( ibdev->port );
+
+	if ( ( rc = local_mad ( ibdev, mad ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not get port info: %s\n",
+		       ibdev, strerror ( rc ) );
+		return rc;
+	}
+	return 0;
+}
+
+/**
+ * Get GUID information
+ *
+ * @v ibdev		Infiniband device
+ * @v local_mad		Method for issuing local MADs
+ * @v mad		Management datagram to fill in
+ * @ret rc		Return status code
+ */
+static int ib_smc_get_guid_info ( struct ib_device *ibdev,
+				  ib_local_mad_t local_mad,
+				  union ib_mad *mad ) {
+	int rc;
+
+	/* Construct MAD */
+	memset ( mad, 0, sizeof ( *mad ) );
+	mad->hdr.base_version = IB_MGMT_BASE_VERSION;
+	mad->hdr.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
+	mad->hdr.class_version = 1;
+	mad->hdr.method = IB_MGMT_METHOD_GET;
+	mad->hdr.attr_id = htons ( IB_SMP_ATTR_GUID_INFO );
+
+	if ( ( rc = local_mad ( ibdev, mad ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not get GUID info: %s\n",
+		       ibdev, strerror ( rc ) );
+		return rc;
+	}
+	return 0;
+}
+
+/**
+ * Get partition key table
+ *
+ * @v ibdev		Infiniband device
+ * @v local_mad		Method for issuing local MADs
+ * @v mad		Management datagram to fill in
+ * @ret rc		Return status code
+ */
+static int ib_smc_get_pkey_table ( struct ib_device *ibdev,
+				   ib_local_mad_t local_mad,
+				   union ib_mad *mad ) {
+	int rc;
+
+	/* Construct MAD */
+	memset ( mad, 0, sizeof ( *mad ) );
+	mad->hdr.base_version = IB_MGMT_BASE_VERSION;
+	mad->hdr.mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED;
+	mad->hdr.class_version = 1;
+	mad->hdr.method = IB_MGMT_METHOD_GET;
+	mad->hdr.attr_id = htons ( IB_SMP_ATTR_PKEY_TABLE );
+
+	if ( ( rc = local_mad ( ibdev, mad ) ) != 0 ) {
+		DBGC ( ibdev, "IBDEV %p could not get pkey table: %s\n",
+		       ibdev, strerror ( rc ) );
+		return rc;
+	}
+	return 0;
+}
+
+/**
+ * Get MAD parameters
+ *
+ * @v ibdev		Infiniband device
+ * @v local_mad		Method for issuing local MADs
+ * @ret rc		Return status code
+ */
+int ib_smc_update ( struct ib_device *ibdev, ib_local_mad_t local_mad ) {
+	union ib_mad mad;
+	struct ib_port_info *port_info = &mad.smp.smp_data.port_info;
+	struct ib_guid_info *guid_info = &mad.smp.smp_data.guid_info;
+	struct ib_pkey_table *pkey_table = &mad.smp.smp_data.pkey_table;
+	int rc;
+
+	/* Port info gives us the link state, the first half of the
+	 * port GID and the SM LID.
+	 */
+	if ( ( rc = ib_smc_get_port_info ( ibdev, local_mad, &mad ) ) != 0 )
+		return rc;
+	memcpy ( &ibdev->gid.u.half[0], port_info->gid_prefix,
+		 sizeof ( ibdev->gid.u.half[0] ) );
+	ibdev->lid = ntohs ( port_info->lid );
+	ibdev->sm_lid = ntohs ( port_info->mastersm_lid );
+	ibdev->link_width_enabled = port_info->link_width_enabled;
+	ibdev->link_width_supported = port_info->link_width_supported;
+	ibdev->link_width_active = port_info->link_width_active;
+	ibdev->link_speed_supported =
+		( port_info->link_speed_supported__port_state >> 4 );
+	ibdev->port_state =
+		( port_info->link_speed_supported__port_state & 0xf );
+	ibdev->link_speed_active =
+		( port_info->link_speed_active__link_speed_enabled >> 4 );
+	ibdev->link_speed_enabled =
+		( port_info->link_speed_active__link_speed_enabled & 0xf );
+	ibdev->sm_sl = ( port_info->neighbour_mtu__mastersm_sl & 0xf );
+
+	/* GUID info gives us the second half of the port GID */
+	if ( ( rc = ib_smc_get_guid_info ( ibdev, local_mad, &mad ) ) != 0 )
+		return rc;
+	memcpy ( &ibdev->gid.u.half[1], guid_info->guid[0],
+		 sizeof ( ibdev->gid.u.half[1] ) );
+
+	/* Get partition key */
+	if ( ( rc = ib_smc_get_pkey_table ( ibdev, local_mad, &mad ) ) != 0 )
+		return rc;
+	ibdev->pkey = ntohs ( pkey_table->pkey[0] );
+
+	DBGC ( ibdev, "IBDEV %p port GID is %08x:%08x:%08x:%08x\n", ibdev,
+	       htonl ( ibdev->gid.u.dwords[0] ),
+	       htonl ( ibdev->gid.u.dwords[1] ),
+	       htonl ( ibdev->gid.u.dwords[2] ),
+	       htonl ( ibdev->gid.u.dwords[3] ) );
+
+	return 0;
+}
diff --git a/gpxe/src/net/infiniband/ib_srp.c b/gpxe/src/net/infiniband/ib_srp.c
new file mode 100644
index 00000000..c156d3ae
--- /dev/null
+++ b/gpxe/src/net/infiniband/ib_srp.c
@@ -0,0 +1,406 @@
+/*
+ * Copyright (C) 2009 Fen Systems Ltd <mbrown@fensystems.co.uk>.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ *   Redistributions of source code must retain the above copyright
+ *   notice, this list of conditions and the following disclaimer.
+ *
+ *   Redistributions in binary form must reproduce the above copyright
+ *   notice, this list of conditions and the following disclaimer in
+ *   the documentation and/or other materials provided with the
+ *   distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
+ * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+ * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+ * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
+ * OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+FILE_LICENCE ( BSD2 );
+
+#include <stdlib.h>
+#include <errno.h>
+#include <gpxe/srp.h>
+#include <gpxe/infiniband.h>
+#include <gpxe/ib_cmrc.h>
+#include <gpxe/ib_srp.h>
+
+/**
+ * @file
+ *
+ * SCSI RDMA Protocol over Infiniband
+ *
+ */
+
+/* Disambiguate the various possible EINVALs */
+#define EINVAL_BYTE_STRING_LEN ( EINVAL | EUNIQ_01 )
+#define EINVAL_BYTE_STRING ( EINVAL | EUNIQ_02 )
+#define EINVAL_INTEGER ( EINVAL | EUNIQ_03 )
+#define EINVAL_RP_TOO_SHORT ( EINVAL | EUNIQ_04 )
+
+/** IB SRP parse flags */
+enum ib_srp_parse_flags {
+	IB_SRP_PARSE_REQUIRED = 0x0000,
+	IB_SRP_PARSE_OPTIONAL = 0x8000,
+	IB_SRP_PARSE_FLAG_MASK = 0xf000,
+};
+
+/** IB SRP root path parameters */
+struct ib_srp_root_path {
+	/** SCSI LUN */
+	struct scsi_lun *lun;
+	/** SRP port IDs */
+	struct srp_port_ids *port_ids;
+	/** IB SRP parameters */
+	struct ib_srp_parameters *ib;
+};
+
+/**
+ * Parse IB SRP root path byte-string value
+ *
+ * @v rp_comp		Root path component string
+ * @v default_value	Default value to use if component string is empty
+ * @ret value		Value
+ */
+static int ib_srp_parse_byte_string ( const char *rp_comp, uint8_t *bytes,
+				      unsigned int size_flags ) {
+	size_t size = ( size_flags & ~IB_SRP_PARSE_FLAG_MASK );
+	size_t rp_comp_len = strlen ( rp_comp );
+	char buf[3];
+	char *buf_end;
+
+	/* Allow optional components to be empty */
+	if ( ( rp_comp_len == 0 ) &&
+	     ( size_flags & IB_SRP_PARSE_OPTIONAL ) )
+		return 0;
+
+	/* Check string length */
+	if ( rp_comp_len != ( 2 * size ) )
+		return -EINVAL_BYTE_STRING_LEN;
+
+	/* Parse byte string */
+	for ( ; size ; size--, rp_comp += 2, bytes++ ) {
+		memcpy ( buf, rp_comp, 2 );
+		buf[2] = '\0';
+		*bytes = strtoul ( buf, &buf_end, 16 );
+		if ( buf_end != &buf[2] )
+			return -EINVAL_BYTE_STRING;
+	}
+	return 0;
+}
+
+/**
+ * Parse IB SRP root path integer value
+ *
+ * @v rp_comp		Root path component string
+ * @v default_value	Default value to use if component string is empty
+ * @ret value		Value
+ */
+static int ib_srp_parse_integer ( const char *rp_comp, int default_value ) {
+	int value;
+	char *end;
+
+	value = strtoul ( rp_comp, &end, 16 );
+	if ( *end )
+		return -EINVAL_INTEGER;
+
+	if ( end == rp_comp )
+		return default_value;
+
+	return value;
+}
+
+/**
+ * Parse IB SRP root path literal component
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_literal ( const char *rp_comp __unused,
+				  struct ib_srp_root_path *rp __unused ) {
+	/* Ignore */
+	return 0;
+}
+
+/**
+ * Parse IB SRP root path source GID
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_sgid ( const char *rp_comp,
+			       struct ib_srp_root_path *rp ) {
+	struct ib_device *ibdev;
+
+	/* Default to the GID of the last opened Infiniband device */
+	if ( ( ibdev = last_opened_ibdev() ) != NULL )
+		memcpy ( &rp->ib->sgid, &ibdev->gid, sizeof ( rp->ib->sgid ) );
+
+	return ib_srp_parse_byte_string ( rp_comp, rp->ib->sgid.u.bytes,
+					  ( sizeof ( rp->ib->sgid ) |
+					    IB_SRP_PARSE_OPTIONAL ) );
+}
+
+/**
+ * Parse IB SRP root path initiator identifier extension
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_initiator_id_ext ( const char *rp_comp,
+					   struct ib_srp_root_path *rp ) {
+	struct ib_srp_initiator_port_id *port_id =
+		ib_srp_initiator_port_id ( rp->port_ids );
+
+	return ib_srp_parse_byte_string ( rp_comp, port_id->id_ext.u.bytes,
+					  ( sizeof ( port_id->id_ext ) |
+					    IB_SRP_PARSE_OPTIONAL ) );
+}
+
+/**
+ * Parse IB SRP root path initiator HCA GUID
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_initiator_hca_guid ( const char *rp_comp,
+					     struct ib_srp_root_path *rp ) {
+	struct ib_srp_initiator_port_id *port_id =
+		ib_srp_initiator_port_id ( rp->port_ids );
+
+	/* Default to the GUID portion of the source GID */
+	memcpy ( &port_id->hca_guid, &rp->ib->sgid.u.half[1],
+		 sizeof ( port_id->hca_guid ) );
+
+	return ib_srp_parse_byte_string ( rp_comp, port_id->hca_guid.u.bytes,
+					  ( sizeof ( port_id->hca_guid ) |
+					    IB_SRP_PARSE_OPTIONAL ) );
+}
+
+/**
+ * Parse IB SRP root path destination GID
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_dgid ( const char *rp_comp,
+			       struct ib_srp_root_path *rp ) {
+	return ib_srp_parse_byte_string ( rp_comp, rp->ib->dgid.u.bytes,
+					  ( sizeof ( rp->ib->dgid ) |
+					    IB_SRP_PARSE_REQUIRED ) );
+}
+
+/**
+ * Parse IB SRP root path partition key
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_pkey ( const char *rp_comp,
+			       struct ib_srp_root_path *rp ) {
+	int pkey;
+
+	if ( ( pkey = ib_srp_parse_integer ( rp_comp, IB_PKEY_DEFAULT ) ) < 0 )
+		return pkey;
+	rp->ib->pkey = pkey;
+	return 0;
+}
+
+/**
+ * Parse IB SRP root path service ID
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_service_id ( const char *rp_comp,
+				     struct ib_srp_root_path *rp ) {
+	return ib_srp_parse_byte_string ( rp_comp, rp->ib->service_id.u.bytes,
+					  ( sizeof ( rp->ib->service_id ) |
+					    IB_SRP_PARSE_REQUIRED ) );
+}
+
+/**
+ * Parse IB SRP root path LUN
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_lun ( const char *rp_comp,
+			      struct ib_srp_root_path *rp ) {
+	return scsi_parse_lun ( rp_comp, rp->lun );
+}
+
+/**
+ * Parse IB SRP root path target identifier extension
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_target_id_ext ( const char *rp_comp,
+					struct ib_srp_root_path *rp ) {
+	struct ib_srp_target_port_id *port_id =
+		ib_srp_target_port_id ( rp->port_ids );
+
+	return ib_srp_parse_byte_string ( rp_comp, port_id->id_ext.u.bytes,
+					  ( sizeof ( port_id->id_ext ) |
+					    IB_SRP_PARSE_REQUIRED ) );
+}
+
+/**
+ * Parse IB SRP root path target I/O controller GUID
+ *
+ * @v rp_comp		Root path component string
+ * @v rp		IB SRP root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_target_ioc_guid ( const char *rp_comp,
+					  struct ib_srp_root_path *rp ) {
+	struct ib_srp_target_port_id *port_id =
+		ib_srp_target_port_id ( rp->port_ids );
+
+	return ib_srp_parse_byte_string ( rp_comp, port_id->ioc_guid.u.bytes,
+					  ( sizeof ( port_id->ioc_guid ) |
+					    IB_SRP_PARSE_REQUIRED ) );
+}
+
+/** IB SRP root path component parser */
+struct ib_srp_root_path_parser {
+	/**
+	 * Parse IB SRP root path component
+	 *
+	 * @v rp_comp		Root path component string
+	 * @v rp		IB SRP root path
+	 * @ret rc		Return status code
+	 */
+	int ( * parse ) ( const char *rp_comp, struct ib_srp_root_path *rp );
+};
+
+/** IB SRP root path components */
+static struct ib_srp_root_path_parser ib_srp_rp_parser[] = {
+	{ ib_srp_parse_literal },
+	{ ib_srp_parse_sgid },
+	{ ib_srp_parse_initiator_id_ext },
+	{ ib_srp_parse_initiator_hca_guid },
+	{ ib_srp_parse_dgid },
+	{ ib_srp_parse_pkey },
+	{ ib_srp_parse_service_id },
+	{ ib_srp_parse_lun },
+	{ ib_srp_parse_target_id_ext },
+	{ ib_srp_parse_target_ioc_guid },
+};
+
+/** Number of IB SRP root path components */
+#define IB_SRP_NUM_RP_COMPONENTS \
+	( sizeof ( ib_srp_rp_parser ) / sizeof ( ib_srp_rp_parser[0] ) )
+
+/**
+ * Parse IB SRP root path
+ *
+ * @v srp		SRP device
+ * @v rp_string		Root path
+ * @ret rc		Return status code
+ */
+static int ib_srp_parse_root_path ( struct srp_device *srp,
+				    const char *rp_string ) {
+	struct ib_srp_parameters *ib_params = ib_srp_params ( srp );
+	struct ib_srp_root_path rp = {
+		.lun = &srp->lun,
+		.port_ids = &srp->port_ids,
+		.ib = ib_params,
+	};
+	char rp_string_copy[ strlen ( rp_string ) + 1 ];
+	char *rp_comp[IB_SRP_NUM_RP_COMPONENTS];
+	char *rp_string_tmp = rp_string_copy;
+	unsigned int i = 0;
+	int rc;
+
+	/* Split root path into component parts */
+	strcpy ( rp_string_copy, rp_string );
+	while ( 1 ) {
+		rp_comp[i++] = rp_string_tmp;
+		if ( i == IB_SRP_NUM_RP_COMPONENTS )
+			break;
+		for ( ; *rp_string_tmp != ':' ; rp_string_tmp++ ) {
+			if ( ! *rp_string_tmp ) {
+				DBGC ( srp, "SRP %p root path \"%s\" too "
+				       "short\n", srp, rp_string );
+				return -EINVAL_RP_TOO_SHORT;
+			}
+		}
+		*(rp_string_tmp++) = '\0';
+	}
+
+	/* Parse root path components */
+	for ( i = 0 ; i < IB_SRP_NUM_RP_COMPONENTS ; i++ ) {
+		if ( ( rc = ib_srp_rp_parser[i].parse ( rp_comp[i],
+							&rp ) ) != 0 ) {
+			DBGC ( srp, "SRP %p could not parse \"%s\" in root "
+			       "path \"%s\": %s\n", srp, rp_comp[i],
+			       rp_string, strerror ( rc ) );
+			return rc;
+		}
+	}
+
+	return 0;
+}
+
+/**
+ * Connect IB SRP session
+ *
+ * @v srp		SRP device
+ * @ret rc		Return status code
+ */
+static int ib_srp_connect ( struct srp_device *srp ) {
+	struct ib_srp_parameters *ib_params = ib_srp_params ( srp );
+	struct ib_device *ibdev;
+	int rc;
+
+	/* Identify Infiniband device */
+	ibdev = find_ibdev ( &ib_params->sgid );
+	if ( ! ibdev ) {
+		DBGC ( srp, "SRP %p could not identify Infiniband device\n",
+		       srp );
+		return -ENODEV;
+	}
+
+	/* Configure remaining SRP parameters */
+	srp->memory_handle = ibdev->rdma_key;
+
+	/* Open CMRC socket */
+	if ( ( rc = ib_cmrc_open ( &srp->socket, ibdev, &ib_params->dgid,
+				   &ib_params->service_id ) ) != 0 ) {
+		DBGC ( srp, "SRP %p could not open CMRC socket: %s\n",
+		       srp, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
+/** IB SRP transport type */
+struct srp_transport_type ib_srp_transport = {
+	.priv_len = sizeof ( struct ib_srp_parameters ),
+	.parse_root_path = ib_srp_parse_root_path,
+	.connect = ib_srp_connect,
+};
diff --git a/gpxe/src/net/iobpad.c b/gpxe/src/net/iobpad.c
index 9961edca..cbae221a 100644
--- a/gpxe/src/net/iobpad.c
+++ b/gpxe/src/net/iobpad.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /**
  * @file
  *
diff --git a/gpxe/src/net/ipv4.c b/gpxe/src/net/ipv4.c
index 8668d44b..4c1393f2 100644
--- a/gpxe/src/net/ipv4.c
+++ b/gpxe/src/net/ipv4.c
@@ -21,6 +21,8 @@
  *
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /* Unique IP datagram identification number */
 static uint16_t next_ident = 0;
 
@@ -38,7 +40,7 @@ static LIST_HEAD ( frag_buffers );
  * @v netdev		Network device
  * @v address		IPv4 address
  * @v netmask		Subnet mask
- * @v gateway		Gateway address (or @c INADDR_NONE for no gateway)
+ * @v gateway		Gateway address (if any)
  * @ret miniroute	Routing table entry, or NULL
  */
 static struct ipv4_miniroute * __malloc
@@ -48,7 +50,7 @@ add_ipv4_miniroute ( struct net_device *netdev, struct in_addr address,
 
 	DBG ( "IPv4 add %s", inet_ntoa ( address ) );
 	DBG ( "/%s ", inet_ntoa ( netmask ) );
-	if ( gateway.s_addr != INADDR_NONE )
+	if ( gateway.s_addr )
 		DBG ( "gw %s ", inet_ntoa ( gateway ) );
 	DBG ( "via %s\n", netdev->name );
 
@@ -68,7 +70,7 @@ add_ipv4_miniroute ( struct net_device *netdev, struct in_addr address,
 	/* Add to end of list if we have a gateway, otherwise
 	 * to start of list.
 	 */
-	if ( gateway.s_addr != INADDR_NONE ) {
+	if ( gateway.s_addr ) {
 		list_add_tail ( &miniroute->list, &ipv4_miniroutes );
 	} else {
 		list_add ( &miniroute->list, &ipv4_miniroutes );
@@ -86,7 +88,7 @@ static void del_ipv4_miniroute ( struct ipv4_miniroute *miniroute ) {
 
 	DBG ( "IPv4 del %s", inet_ntoa ( miniroute->address ) );
 	DBG ( "/%s ", inet_ntoa ( miniroute->netmask ) );
-	if ( miniroute->gateway.s_addr != INADDR_NONE )
+	if ( miniroute->gateway.s_addr )
 		DBG ( "gw %s ", inet_ntoa ( miniroute->gateway ) );
 	DBG ( "via %s\n", miniroute->netdev->name );
 
@@ -116,9 +118,11 @@ static struct ipv4_miniroute * ipv4_route ( struct in_addr *dest ) {
 
 	/* Find first usable route in routing table */
 	list_for_each_entry ( miniroute, &ipv4_miniroutes, list ) {
+		if ( ! ( miniroute->netdev->state & NETDEV_OPEN ) )
+			continue;
 		local = ( ( ( dest->s_addr ^ miniroute->address.s_addr )
 			    & miniroute->netmask.s_addr ) == 0 );
-		has_gw = ( miniroute->gateway.s_addr != INADDR_NONE );
+		has_gw = ( miniroute->gateway.s_addr );
 		if ( local || has_gw ) {
 			if ( ! local )
 				*dest = miniroute->gateway;
@@ -269,7 +273,7 @@ static int ipv4_ll_addr ( struct in_addr dest, struct in_addr src,
 
 	if ( dest.s_addr == INADDR_BROADCAST ) {
 		/* Broadcast address */
-		memcpy ( ll_dest, ll_protocol->ll_broadcast,
+		memcpy ( ll_dest, netdev->ll_broadcast,
 			 ll_protocol->ll_addr_len );
 		return 0;
 	} else if ( IN_MULTICAST ( ntohl ( dest.s_addr ) ) ) {
@@ -584,7 +588,7 @@ static int ipv4_create_routes ( void ) {
 	struct settings *settings;
 	struct in_addr address = { 0 };
 	struct in_addr netmask = { 0 };
-	struct in_addr gateway = { INADDR_NONE };
+	struct in_addr gateway = { 0 };
 
 	/* Delete all existing routes */
 	list_for_each_entry_safe ( miniroute, tmp, &ipv4_miniroutes, list )
@@ -598,20 +602,19 @@ static int ipv4_create_routes ( void ) {
 		fetch_ipv4_setting ( settings, &ip_setting, &address );
 		if ( ! address.s_addr )
 			continue;
-		/* Calculate default netmask */
-		if ( IN_CLASSA ( ntohl ( address.s_addr ) ) ) {
-			netmask.s_addr = htonl ( IN_CLASSA_NET );
-		} else if ( IN_CLASSB ( ntohl ( address.s_addr ) ) ) {
-			netmask.s_addr = htonl ( IN_CLASSB_NET );
-		} else if ( IN_CLASSC ( ntohl ( address.s_addr ) ) ) {
-			netmask.s_addr = htonl ( IN_CLASSC_NET );
-		} else {
-			netmask.s_addr = 0;
-		}
-		/* Override with subnet mask, if present */
+		/* Get subnet mask */
 		fetch_ipv4_setting ( settings, &netmask_setting, &netmask );
+		/* Calculate default netmask, if necessary */
+		if ( ! netmask.s_addr ) {
+			if ( IN_CLASSA ( ntohl ( address.s_addr ) ) ) {
+				netmask.s_addr = htonl ( IN_CLASSA_NET );
+			} else if ( IN_CLASSB ( ntohl ( address.s_addr ) ) ) {
+				netmask.s_addr = htonl ( IN_CLASSB_NET );
+			} else if ( IN_CLASSC ( ntohl ( address.s_addr ) ) ) {
+				netmask.s_addr = htonl ( IN_CLASSC_NET );
+			}
+		}
 		/* Get default gateway, if present */
-		gateway.s_addr = INADDR_NONE;
 		fetch_ipv4_setting ( settings, &gateway_setting, &gateway );
 		/* Configure route */
 		miniroute = add_ipv4_miniroute ( netdev, address,
diff --git a/gpxe/src/net/mii.c b/gpxe/src/net/mii.c
new file mode 100644
index 00000000..0de64428
--- /dev/null
+++ b/gpxe/src/net/mii.c
@@ -0,0 +1,147 @@
+/*
+
+	mii.c: MII interface library
+
+	Ported to gPXE by Daniel Verkamp <daniel@drv.nu>
+	from Linux drivers/net/mii.c
+
+	Maintained by Jeff Garzik <jgarzik@pobox.com>
+	Copyright 2001,2002 Jeff Garzik
+
+	Various code came from myson803.c and other files by
+	Donald Becker.  Copyright:
+
+		Written 1998-2002 by Donald Becker.
+
+		This software may be used and distributed according
+		to the terms of the GNU General Public License (GPL),
+		incorporated herein by reference.  Drivers based on
+		or derived from this code fall under the GPL and must
+		retain the authorship, copyright and license notice.
+		This file is not a complete program and may only be
+		used when the entire operating system is licensed
+		under the GPL.
+
+		The author may be reached as becker@scyld.com, or C/O
+		Scyld Computing Corporation
+		410 Severn Ave., Suite 210
+		Annapolis MD 21403
+
+*/
+
+#include <mii.h>
+
+/**
+ * mii_link_ok - is link status up/ok
+ * @mii: the MII interface
+ *
+ * Returns 1 if the MII reports link status up/ok, 0 otherwise.
+ */
+int
+mii_link_ok ( struct mii_if_info *mii )
+{
+	/* first, a dummy read, needed to latch some MII phys */
+	mii->mdio_read ( mii->dev, mii->phy_id, MII_BMSR );
+	if ( mii->mdio_read ( mii->dev, mii->phy_id, MII_BMSR ) & BMSR_LSTATUS )
+		return 1;
+	return 0;
+}
+
+/**
+ * mii_check_link - check MII link status
+ * @mii: MII interface
+ *
+ * If the link status changed (previous != current), call
+ * netif_carrier_on() if current link status is Up or call
+ * netif_carrier_off() if current link status is Down.
+ */
+void
+mii_check_link ( struct mii_if_info *mii )
+{
+	int cur_link = mii_link_ok ( mii );
+	int prev_link = netdev_link_ok ( mii->dev );
+
+	if ( cur_link && !prev_link )
+		netdev_link_up ( mii->dev );
+	else if (prev_link && !cur_link)
+		netdev_link_down ( mii->dev );
+}
+
+
+/**
+ * mii_check_media - check the MII interface for a duplex change
+ * @mii: the MII interface
+ * @ok_to_print: OK to print link up/down messages
+ * @init_media: OK to save duplex mode in @mii
+ *
+ * Returns 1 if the duplex mode changed, 0 if not.
+ * If the media type is forced, always returns 0.
+ */
+unsigned int
+mii_check_media ( struct mii_if_info *mii,
+                  unsigned int ok_to_print,
+                  unsigned int init_media )
+{
+	unsigned int old_carrier, new_carrier;
+	int advertise, lpa, media, duplex;
+	int lpa2 = 0;
+
+	/* if forced media, go no further */
+	if (mii->force_media)
+		return 0; /* duplex did not change */
+
+	/* check current and old link status */
+	old_carrier = netdev_link_ok ( mii->dev ) ? 1 : 0;
+	new_carrier = (unsigned int) mii_link_ok ( mii );
+
+	/* if carrier state did not change, this is a "bounce",
+	 * just exit as everything is already set correctly
+	 */
+	if ( ( ! init_media ) && ( old_carrier == new_carrier ) )
+		return 0; /* duplex did not change */
+
+	/* no carrier, nothing much to do */
+	if ( ! new_carrier ) {
+		netdev_link_down ( mii->dev );
+		if ( ok_to_print )
+			DBG ( "%s: link down\n", mii->dev->name);
+		return 0; /* duplex did not change */
+	}
+
+	/*
+	 * we have carrier, see who's on the other end
+	 */
+	netdev_link_up ( mii->dev );
+
+	/* get MII advertise and LPA values */
+	if ( ( ! init_media ) && ( mii->advertising ) ) {
+		advertise = mii->advertising;
+	} else {
+		advertise = mii->mdio_read ( mii->dev, mii->phy_id, MII_ADVERTISE );
+		mii->advertising = advertise;
+	}
+	lpa = mii->mdio_read ( mii->dev, mii->phy_id, MII_LPA );
+	if ( mii->supports_gmii )
+		lpa2 = mii->mdio_read ( mii->dev, mii->phy_id, MII_STAT1000 );
+
+	/* figure out media and duplex from advertise and LPA values */
+	media = mii_nway_result ( lpa & advertise );
+	duplex = ( media & ADVERTISE_FULL ) ? 1 : 0;
+	if ( lpa2 & LPA_1000FULL )
+		duplex = 1;
+
+	if ( ok_to_print )
+		DBG ( "%s: link up, %sMbps, %s-duplex, lpa 0x%04X\n",
+		       mii->dev->name,
+		       lpa2 & ( LPA_1000FULL | LPA_1000HALF ) ? "1000" :
+		       media & ( ADVERTISE_100FULL | ADVERTISE_100HALF ) ? "100" : "10",
+		       duplex ? "full" : "half",
+		       lpa);
+
+	if ( ( init_media ) || ( mii->full_duplex != duplex ) ) {
+		mii->full_duplex = duplex;
+		return 1; /* duplex changed */
+	}
+
+	return 0; /* duplex did not change */
+}
diff --git a/gpxe/src/net/netdev_settings.c b/gpxe/src/net/netdev_settings.c
index 44aca7d8..d814193b 100644
--- a/gpxe/src/net/netdev_settings.c
+++ b/gpxe/src/net/netdev_settings.c
@@ -16,10 +16,14 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <string.h>
 #include <errno.h>
+#include <byteswap.h>
 #include <gpxe/dhcp.h>
 #include <gpxe/settings.h>
+#include <gpxe/device.h>
 #include <gpxe/netdevice.h>
 
 /** @file
@@ -34,6 +38,11 @@ struct setting mac_setting __setting = {
 	.description = "MAC address",
 	.type = &setting_type_hex,
 };
+struct setting busid_setting __setting = {
+	.name = "busid",
+	.description = "Bus ID",
+	.type = &setting_type_hex,
+};
 
 /**
  * Store value of network device setting
@@ -54,9 +63,9 @@ static int netdev_store ( struct settings *settings, struct setting *setting,
 			return -EINVAL;
 		memcpy ( netdev->ll_addr, data, len );
 		return 0;
-	} else {
-		return simple_settings_store ( settings, setting, data, len );
 	}
+
+	return generic_settings_store ( settings, setting, data, len );
 }
 
 /**
@@ -72,19 +81,40 @@ static int netdev_fetch ( struct settings *settings, struct setting *setting,
 			  void *data, size_t len ) {
 	struct net_device *netdev = container_of ( settings, struct net_device,
 						   settings.settings );
+	struct device_description *desc = &netdev->dev->desc;
+	struct dhcp_netdev_desc dhcp_desc;
 
 	if ( setting_cmp ( setting, &mac_setting ) == 0 ) {
 		if ( len > netdev->ll_protocol->ll_addr_len )
 			len = netdev->ll_protocol->ll_addr_len;
 		memcpy ( data, netdev->ll_addr, len );
 		return netdev->ll_protocol->ll_addr_len;
-	} else {
-		return simple_settings_fetch ( settings, setting, data, len );
 	}
+	if ( setting_cmp ( setting, &busid_setting ) == 0 ) {
+		dhcp_desc.type = desc->bus_type;
+		dhcp_desc.vendor = htons ( desc->vendor );
+		dhcp_desc.device = htons ( desc->device );
+		if ( len > sizeof ( dhcp_desc ) )
+			len = sizeof ( dhcp_desc );
+		memcpy ( data, &dhcp_desc, len );
+		return sizeof ( dhcp_desc );
+	}
+
+	return generic_settings_fetch ( settings, setting, data, len );
+}
+
+/**
+ * Clear network device settings
+ *
+ * @v settings		Settings block
+ */
+static void netdev_clear ( struct settings *settings ) {
+	generic_settings_clear ( settings );
 }
 
 /** Network device configuration settings operations */
 struct settings_operations netdev_settings_operations = {
 	.store = netdev_store,
 	.fetch = netdev_fetch,
+	.clear = netdev_clear,
 };
diff --git a/gpxe/src/net/netdevice.c b/gpxe/src/net/netdevice.c
index 9e142d27..ee0d0b72 100644
--- a/gpxe/src/net/netdevice.c
+++ b/gpxe/src/net/netdevice.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -28,6 +30,7 @@
 #include <gpxe/process.h>
 #include <gpxe/init.h>
 #include <gpxe/device.h>
+#include <gpxe/errortab.h>
 #include <gpxe/netdevice.h>
 
 /** @file
@@ -36,17 +39,39 @@
  *
  */
 
-/** Registered network-layer protocols */
-static struct net_protocol net_protocols[0]
-	__table_start ( struct net_protocol, net_protocols );
-static struct net_protocol net_protocols_end[0]
-	__table_end ( struct net_protocol, net_protocols );
-
 /** List of network devices */
 struct list_head net_devices = LIST_HEAD_INIT ( net_devices );
 
 /** List of open network devices, in reverse order of opening */
-struct list_head open_net_devices = LIST_HEAD_INIT ( open_net_devices );
+static struct list_head open_net_devices = LIST_HEAD_INIT ( open_net_devices );
+
+/** Default link status code */
+#define EUNKNOWN_LINK_STATUS EINPROGRESS
+
+/** Human-readable message for the default link status */
+struct errortab netdev_errors[] __errortab = {
+	{ EUNKNOWN_LINK_STATUS, "Unknown" },
+};
+
+/**
+ * Mark network device as having link down
+ *
+ * @v netdev		Network device
+ */
+void netdev_link_down ( struct net_device *netdev ) {
+
+	switch ( netdev->link_rc ) {
+	case 0:
+	case -EUNKNOWN_LINK_STATUS:
+		netdev->link_rc = -ENOTCONN;
+		break;
+	default:
+		/* Avoid clobbering a more detailed link status code,
+		 * if one is already set.
+		 */
+		break;
+	}
+}
 
 /**
  * Record network device statistic
@@ -286,6 +311,7 @@ static void free_netdev ( struct refcnt *refcnt ) {
 	
 	netdev_tx_flush ( netdev );
 	netdev_rx_flush ( netdev );
+	clear_settings ( netdev_settings ( netdev ) );
 	free ( netdev );
 }
 
@@ -305,11 +331,10 @@ struct net_device * alloc_netdev ( size_t priv_size ) {
 	netdev = zalloc ( total_len );
 	if ( netdev ) {
 		netdev->refcnt.free = free_netdev;
+		netdev->link_rc = -EUNKNOWN_LINK_STATUS;
 		INIT_LIST_HEAD ( &netdev->tx_queue );
 		INIT_LIST_HEAD ( &netdev->rx_queue );
-		settings_init ( netdev_settings ( netdev ),
-				&netdev_settings_operations, &netdev->refcnt,
-				netdev->name, 0 );
+		netdev_settings_init ( netdev );
 		netdev->priv = ( ( ( void * ) netdev ) + sizeof ( *netdev ) );
 	}
 	return netdev;
@@ -332,6 +357,9 @@ int register_netdev ( struct net_device *netdev ) {
 	snprintf ( netdev->name, sizeof ( netdev->name ), "net%d",
 		   ifindex++ );
 
+	/* Set initial link-layer address */
+	netdev->ll_protocol->init_addr ( netdev->hw_addr, netdev->ll_addr );
+
 	/* Register per-netdev configuration settings */
 	if ( ( rc = register_settings ( netdev_settings ( netdev ),
 					NULL ) ) != 0 ) {
@@ -345,7 +373,7 @@ int register_netdev ( struct net_device *netdev ) {
 	list_add_tail ( &netdev->list, &net_devices );
 	DBGC ( netdev, "NETDEV %p registered as %s (phys %s hwaddr %s)\n",
 	       netdev, netdev->name, netdev->dev->name,
-	       netdev_hwaddr ( netdev ) );
+	       netdev_addr ( netdev ) );
 
 	return 0;
 }
@@ -514,7 +542,7 @@ int net_tx ( struct io_buffer *iobuf, struct net_device *netdev,
 	netdev_poll ( netdev );
 
 	/* Add link-layer header */
-	if ( ( rc = ll_protocol->push ( iobuf, ll_dest, netdev->ll_addr,
+	if ( ( rc = ll_protocol->push ( netdev, iobuf, ll_dest, netdev->ll_addr,
 					net_protocol->net_proto ) ) != 0 ) {
 		free_iob ( iobuf );
 		return rc;
@@ -538,12 +566,13 @@ int net_rx ( struct io_buffer *iobuf, struct net_device *netdev,
 	struct net_protocol *net_protocol;
 
 	/* Hand off to network-layer protocol, if any */
-	for ( net_protocol = net_protocols ; net_protocol < net_protocols_end ;
-	      net_protocol++ ) {
-		if ( net_protocol->net_proto == net_proto ) {
+	for_each_table_entry ( net_protocol, NET_PROTOCOLS ) {
+		if ( net_protocol->net_proto == net_proto )
 			return net_protocol->rx ( iobuf, netdev, ll_source );
-		}
 	}
+
+	DBGC ( netdev, "NETDEV %p unknown network protocol %04x\n",
+	       netdev, ntohs ( net_proto ) );
 	free_iob ( iobuf );
 	return 0;
 }
@@ -585,8 +614,8 @@ static void net_step ( struct process *process __unused ) {
 
 			/* Remove link-layer header */
 			ll_protocol = netdev->ll_protocol;
-			if ( ( rc = ll_protocol->pull ( iobuf, &ll_dest,
-							&ll_source,
+			if ( ( rc = ll_protocol->pull ( netdev, iobuf,
+							&ll_dest, &ll_source,
 							&net_proto ) ) != 0 ) {
 				free_iob ( iobuf );
 				continue;
@@ -599,5 +628,6 @@ static void net_step ( struct process *process __unused ) {
 
 /** Networking stack process */
 struct process net_process __permanent_process = {
+	.list = LIST_HEAD_INIT ( net_process.list ),
 	.step = net_step,
 };
diff --git a/gpxe/src/net/nullnet.c b/gpxe/src/net/nullnet.c
index 7e199ce3..381f02a6 100644
--- a/gpxe/src/net/nullnet.c
+++ b/gpxe/src/net/nullnet.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <errno.h>
 #include <gpxe/iobuf.h>
diff --git a/gpxe/src/net/rarp.c b/gpxe/src/net/rarp.c
index bb5e6ad7..1d0dd961 100644
--- a/gpxe/src/net/rarp.c
+++ b/gpxe/src/net/rarp.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <byteswap.h>
 #include <gpxe/netdevice.h>
diff --git a/gpxe/src/net/retry.c b/gpxe/src/net/retry.c
index cd793a7f..40f656f2 100644
--- a/gpxe/src/net/retry.c
+++ b/gpxe/src/net/retry.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stddef.h>
 #include <gpxe/timer.h>
 #include <gpxe/list.h>
@@ -185,5 +187,6 @@ static void retry_step ( struct process *process __unused ) {
 
 /** Retry timer process */
 struct process retry_process __permanent_process = {
+	.list = LIST_HEAD_INIT ( retry_process.list ),
 	.step = retry_step,
 };
diff --git a/gpxe/src/net/tcp.c b/gpxe/src/net/tcp.c
index 6bcd193c..a0619622 100644
--- a/gpxe/src/net/tcp.c
+++ b/gpxe/src/net/tcp.c
@@ -21,6 +21,8 @@
  *
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /** A TCP connection */
 struct tcp_connection {
 	/** Reference counter */
@@ -142,15 +144,15 @@ tcp_dump_state ( struct tcp_connection *tcp ) {
 static inline __attribute__ (( always_inline )) void
 tcp_dump_flags ( struct tcp_connection *tcp, unsigned int flags ) {
 	if ( flags & TCP_RST )
-		DBGC ( tcp, " RST" );
+		DBGC2 ( tcp, " RST" );
 	if ( flags & TCP_SYN )
-		DBGC ( tcp, " SYN" );
+		DBGC2 ( tcp, " SYN" );
 	if ( flags & TCP_PSH )
-		DBGC ( tcp, " PSH" );
+		DBGC2 ( tcp, " PSH" );
 	if ( flags & TCP_FIN )
-		DBGC ( tcp, " FIN" );
+		DBGC2 ( tcp, " FIN" );
 	if ( flags & TCP_ACK )
-		DBGC ( tcp, " ACK" );
+		DBGC2 ( tcp, " ACK" );
 }
 
 /***************************************************************************
@@ -397,9 +399,10 @@ static int tcp_xmit ( struct tcp_connection *tcp, int force_send ) {
 	void *payload;
 	unsigned int flags;
 	size_t len = 0;
-	size_t seq_len;
-	size_t app_win;
-	size_t max_rcv_win;
+	uint32_t seq_len;
+	uint32_t app_win;
+	uint32_t max_rcv_win;
+	int rc;
 
 	/* If retransmission timer is already running, do nothing */
 	if ( timer_running ( &tcp->timer ) )
@@ -436,7 +439,9 @@ static int tcp_xmit ( struct tcp_connection *tcp, int force_send ) {
 	/* Allocate I/O buffer */
 	iobuf = alloc_iob ( len + MAX_HDR_LEN );
 	if ( ! iobuf ) {
-		DBGC ( tcp, "TCP %p could not allocate data buffer\n", tcp );
+		DBGC ( tcp, "TCP %p could not allocate iobuf for %08x..%08x "
+		       "%08x\n", tcp, tcp->snd_seq, ( tcp->snd_seq + seq_len ),
+		       tcp->rcv_ack );
 		return -ENOMEM;
 	}
 	iob_reserve ( iobuf, MAX_HDR_LEN );
@@ -471,6 +476,8 @@ static int tcp_xmit ( struct tcp_connection *tcp, int force_send ) {
 		tsopt->tsopt.tsval = ntohl ( currticks() );
 		tsopt->tsopt.tsecr = ntohl ( tcp->ts_recent );
 	}
+	if ( ! ( flags & TCP_SYN ) )
+		flags |= TCP_PSH;
 	tcphdr = iob_push ( iobuf, sizeof ( *tcphdr ) );
 	memset ( tcphdr, 0, sizeof ( *tcphdr ) );
 	tcphdr->src = tcp->local_port;
@@ -478,21 +485,28 @@ static int tcp_xmit ( struct tcp_connection *tcp, int force_send ) {
 	tcphdr->seq = htonl ( tcp->snd_seq );
 	tcphdr->ack = htonl ( tcp->rcv_ack );
 	tcphdr->hlen = ( ( payload - iobuf->data ) << 2 );
-	tcphdr->flags = ( flags | TCP_PSH );
+	tcphdr->flags = flags;
 	tcphdr->win = htons ( tcp->rcv_win );
 	tcphdr->csum = tcpip_chksum ( iobuf->data, iob_len ( iobuf ) );
 
 	/* Dump header */
-	DBGC ( tcp, "TCP %p TX %d->%d %08x..%08zx           %08x %4zd",
-	       tcp, ntohs ( tcphdr->src ), ntohs ( tcphdr->dest ),
-	       ntohl ( tcphdr->seq ), ( ntohl ( tcphdr->seq ) + seq_len ),
-	       ntohl ( tcphdr->ack ), len );
+	DBGC2 ( tcp, "TCP %p TX %d->%d %08x..%08x           %08x %4zd",
+		tcp, ntohs ( tcphdr->src ), ntohs ( tcphdr->dest ),
+		ntohl ( tcphdr->seq ), ( ntohl ( tcphdr->seq ) + seq_len ),
+		ntohl ( tcphdr->ack ), len );
 	tcp_dump_flags ( tcp, tcphdr->flags );
-	DBGC ( tcp, "\n" );
+	DBGC2 ( tcp, "\n" );
 
 	/* Transmit packet */
-	return tcpip_tx ( iobuf, &tcp_protocol, NULL, &tcp->peer, NULL,
-			  &tcphdr->csum );
+	if ( ( rc = tcpip_tx ( iobuf, &tcp_protocol, NULL, &tcp->peer, NULL,
+			       &tcphdr->csum ) ) != 0 ) {
+		DBGC ( tcp, "TCP %p could not transmit %08x..%08x %08x: %s\n",
+		       tcp, tcp->snd_seq, ( tcp->snd_seq + tcp->snd_sent ),
+		       tcp->rcv_ack, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
 }
 
 /**
@@ -506,8 +520,9 @@ static void tcp_expired ( struct retry_timer *timer, int over ) {
 		container_of ( timer, struct tcp_connection, timer );
 	int graceful_close = TCP_CLOSED_GRACEFULLY ( tcp->tcp_state );
 
-	DBGC ( tcp, "TCP %p timer %s in %s\n", tcp,
-	       ( over ? "expired" : "fired" ), tcp_state ( tcp->tcp_state ) );
+	DBGC ( tcp, "TCP %p timer %s in %s for %08x..%08x %08x\n", tcp,
+	       ( over ? "expired" : "fired" ), tcp_state ( tcp->tcp_state ),
+	       tcp->snd_seq, ( tcp->snd_seq + tcp->snd_sent ), tcp->rcv_ack );
 
 	assert ( ( tcp->tcp_state == TCP_SYN_SENT ) ||
 		 ( tcp->tcp_state == TCP_SYN_RCVD ) ||
@@ -542,11 +557,14 @@ static int tcp_xmit_reset ( struct tcp_connection *tcp,
 			    struct tcp_header *in_tcphdr ) {
 	struct io_buffer *iobuf;
 	struct tcp_header *tcphdr;
+	int rc;
 
 	/* Allocate space for dataless TX buffer */
 	iobuf = alloc_iob ( MAX_HDR_LEN );
 	if ( ! iobuf ) {
-		DBGC ( tcp, "TCP %p could not allocate data buffer\n", tcp );
+		DBGC ( tcp, "TCP %p could not allocate iobuf for RST "
+		       "%08x..%08x %08x\n", tcp, ntohl ( in_tcphdr->ack ),
+		       ntohl ( in_tcphdr->ack ), ntohl ( in_tcphdr->seq ) );
 		return -ENOMEM;
 	}
 	iob_reserve ( iobuf, MAX_HDR_LEN );
@@ -564,16 +582,24 @@ static int tcp_xmit_reset ( struct tcp_connection *tcp,
 	tcphdr->csum = tcpip_chksum ( iobuf->data, iob_len ( iobuf ) );
 
 	/* Dump header */
-	DBGC ( tcp, "TCP %p TX %d->%d %08x..%08x           %08x %4d",
-	       tcp, ntohs ( tcphdr->src ), ntohs ( tcphdr->dest ),
-	       ntohl ( tcphdr->seq ), ( ntohl ( tcphdr->seq ) ),
-	       ntohl ( tcphdr->ack ), 0 );
+	DBGC2 ( tcp, "TCP %p TX %d->%d %08x..%08x           %08x %4d",
+		tcp, ntohs ( tcphdr->src ), ntohs ( tcphdr->dest ),
+		ntohl ( tcphdr->seq ), ( ntohl ( tcphdr->seq ) ),
+		ntohl ( tcphdr->ack ), 0 );
 	tcp_dump_flags ( tcp, tcphdr->flags );
-	DBGC ( tcp, "\n" );
+	DBGC2 ( tcp, "\n" );
 
 	/* Transmit packet */
-	return tcpip_tx ( iobuf, &tcp_protocol, NULL, st_dest,
-			  NULL, &tcphdr->csum );
+	if ( ( rc = tcpip_tx ( iobuf, &tcp_protocol, NULL, st_dest,
+			       NULL, &tcphdr->csum ) ) != 0 ) {
+		DBGC ( tcp, "TCP %p could not transmit RST %08x..%08x %08x: "
+		       "%s\n", tcp, ntohl ( in_tcphdr->ack ),
+		       ntohl ( in_tcphdr->ack ), ntohl ( in_tcphdr->seq ),
+		       strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
 }
 
 /***************************************************************************
@@ -645,7 +671,7 @@ static void tcp_rx_opts ( struct tcp_connection *tcp, const void *data,
  * @v tcp		TCP connection
  * @v seq_len		Sequence space length to consume
  */
-static void tcp_rx_seq ( struct tcp_connection *tcp, size_t seq_len ) {
+static void tcp_rx_seq ( struct tcp_connection *tcp, uint32_t seq_len ) {
 	tcp->rcv_ack += seq_len;
 	if ( tcp->rcv_win > seq_len ) {
 		tcp->rcv_win -= seq_len;
@@ -696,20 +722,42 @@ static int tcp_rx_syn ( struct tcp_connection *tcp, uint32_t seq,
  */
 static int tcp_rx_ack ( struct tcp_connection *tcp, uint32_t ack,
 			uint32_t win ) {
-	size_t ack_len = ( ack - tcp->snd_seq );
+	uint32_t ack_len = ( ack - tcp->snd_seq );
 	size_t len;
 	unsigned int acked_flags;
 
-	/* Ignore duplicate or out-of-range ACK */
+	/* Check for out-of-range or old duplicate ACKs */
 	if ( ack_len > tcp->snd_sent ) {
-		DBGC ( tcp, "TCP %p received ACK for [%08x,%08zx), "
-		       "sent only [%08x,%08x)\n", tcp, tcp->snd_seq,
+		DBGC ( tcp, "TCP %p received ACK for %08x..%08x, "
+		       "sent only %08x..%08x\n", tcp, tcp->snd_seq,
 		       ( tcp->snd_seq + ack_len ), tcp->snd_seq,
 		       ( tcp->snd_seq + tcp->snd_sent ) );
-		return -EINVAL;
+
+		if ( TCP_HAS_BEEN_ESTABLISHED ( tcp->tcp_state ) ) {
+			/* Just ignore what might be old duplicate ACKs */
+			return 0;
+		} else {
+			/* Send RST if an out-of-range ACK is received
+			 * on a not-yet-established connection, as per
+			 * RFC 793.
+			 */
+			return -EINVAL;
+		}
 	}
 
-	/* Acknowledge any flags being sent */
+	/* Ignore ACKs that don't actually acknowledge any new data.
+	 * (In particular, do not stop the retransmission timer; this
+	 * avoids creating a sorceror's apprentice syndrome when a
+	 * duplicate ACK is received and we still have data in our
+	 * transmit queue.)
+	 */
+	if ( ack_len == 0 )
+		return 0;
+
+	/* Stop the retransmission timer */
+	stop_timer ( &tcp->timer );
+
+	/* Determine acknowledged flags and data length */
 	len = ack_len;
 	acked_flags = ( TCP_FLAGS_SENDING ( tcp->tcp_state ) &
 			( TCP_SYN | TCP_FIN ) );
@@ -721,9 +769,6 @@ static int tcp_rx_ack ( struct tcp_connection *tcp, uint32_t ack,
 	tcp->snd_sent = 0;
 	tcp->snd_win = win;
 
-	/* Stop the retransmission timer */
-	stop_timer ( &tcp->timer );
-
 	/* Remove any acknowledged data from transmit queue */
 	tcp_process_queue ( tcp, len, NULL, 1 );
 		
@@ -750,8 +795,8 @@ static int tcp_rx_ack ( struct tcp_connection *tcp, uint32_t ack,
  */
 static int tcp_rx_data ( struct tcp_connection *tcp, uint32_t seq,
 			 struct io_buffer *iobuf ) {
-	size_t already_rcvd;
-	size_t len;
+	uint32_t already_rcvd;
+	uint32_t len;
 	int rc;
 
 	/* Ignore duplicate or out-of-order data */
@@ -765,8 +810,11 @@ static int tcp_rx_data ( struct tcp_connection *tcp, uint32_t seq,
 	len -= already_rcvd;
 
 	/* Deliver data to application */
-	if ( ( rc = xfer_deliver_iob ( &tcp->xfer, iobuf ) ) != 0 )
+	if ( ( rc = xfer_deliver_iob ( &tcp->xfer, iobuf ) ) != 0 ) {
+		DBGC ( tcp, "TCP %p could not deliver %08x..%08x: %s\n",
+		       tcp, seq, ( seq + len ), strerror ( rc ) );
 		return rc;
+	}
 
 	/* Acknowledge new data */
 	tcp_rx_seq ( tcp, len );
@@ -824,6 +872,7 @@ static int tcp_rx_rst ( struct tcp_connection *tcp, uint32_t seq ) {
 	tcp_dump_state ( tcp );
 	tcp_close ( tcp, -ECONNRESET );
 
+	DBGC ( tcp, "TCP %p connection reset by peer\n", tcp );
 	return -ECONNRESET;
 }
 
@@ -894,13 +943,13 @@ static int tcp_rx ( struct io_buffer *iobuf,
 	len = iob_len ( iobuf );
 
 	/* Dump header */
-	DBGC ( tcp, "TCP %p RX %d<-%d           %08x %08x..%08zx %4zd",
-	       tcp, ntohs ( tcphdr->dest ), ntohs ( tcphdr->src ),
-	       ntohl ( tcphdr->ack ), ntohl ( tcphdr->seq ),
-	       ( ntohl ( tcphdr->seq ) + len +
-		 ( ( tcphdr->flags & ( TCP_SYN | TCP_FIN ) ) ? 1 : 0 ) ), len);
+	DBGC2 ( tcp, "TCP %p RX %d<-%d           %08x %08x..%08zx %4zd",
+		tcp, ntohs ( tcphdr->dest ), ntohs ( tcphdr->src ),
+		ntohl ( tcphdr->ack ), ntohl ( tcphdr->seq ),
+		( ntohl ( tcphdr->seq ) + len +
+		  ( ( tcphdr->flags & ( TCP_SYN | TCP_FIN ) ) ? 1 : 0 )), len);
 	tcp_dump_flags ( tcp, tcphdr->flags );
-	DBGC ( tcp, "\n" );
+	DBGC2 ( tcp, "\n" );
 
 	/* If no connection was found, send RST */
 	if ( ! tcp ) {
diff --git a/gpxe/src/net/tcp/ftp.c b/gpxe/src/net/tcp/ftp.c
index 445e32bb..920e537a 100644
--- a/gpxe/src/net/tcp/ftp.c
+++ b/gpxe/src/net/tcp/ftp.c
@@ -131,11 +131,33 @@ static const char * ftp_uri_path ( struct ftp_request *ftp ) {
 	return ftp->uri->path;
 }
 
+/**
+ * Retrieve FTP user
+ *
+ * @v ftp		FTP request
+ * @ret user		FTP user
+ */
+static const char * ftp_user ( struct ftp_request *ftp ) {
+	static char *ftp_default_user = "anonymous";
+	return ftp->uri->user ? ftp->uri->user : ftp_default_user;
+}
+
+/**
+ * Retrieve FTP password
+ *
+ * @v ftp		FTP request
+ * @ret password	FTP password
+ */
+static const char * ftp_password ( struct ftp_request *ftp ) {
+	static char *ftp_default_password = "etherboot@etherboot.org";
+	return ftp->uri->password ? ftp->uri->password : ftp_default_password;
+}
+
 /** FTP control channel strings */
 static struct ftp_control_string ftp_strings[] = {
 	[FTP_CONNECT]	= { NULL, NULL },
-	[FTP_USER]	= { "USER anonymous", NULL },
-	[FTP_PASS]	= { "PASS etherboot@etherboot.org", NULL },
+	[FTP_USER]	= { "USER ", ftp_user },
+	[FTP_PASS]	= { "PASS ", ftp_password },
 	[FTP_TYPE]	= { "TYPE I", NULL },
 	[FTP_PASV]	= { "PASV", NULL },
 	[FTP_RETR]	= { "RETR ", ftp_uri_path },
@@ -335,7 +357,7 @@ static int ftp_control_deliver_raw ( struct xfer_interface *control,
 /** FTP control channel operations */
 static struct xfer_interface_operations ftp_control_operations = {
 	.close		= ftp_control_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= xfer_deliver_as_raw,
@@ -402,7 +424,7 @@ static int ftp_data_deliver_iob ( struct xfer_interface *data,
 /** FTP data channel operations */
 static struct xfer_interface_operations ftp_data_operations = {
 	.close		= ftp_data_closed,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= ftp_data_deliver_iob,
diff --git a/gpxe/src/net/tcp/http.c b/gpxe/src/net/tcp/http.c
index 93ccfd3b..a365b2a4 100644
--- a/gpxe/src/net/tcp/http.c
+++ b/gpxe/src/net/tcp/http.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /**
  * @file
  *
@@ -138,6 +140,8 @@ static void http_done ( struct http_request *http, int rc ) {
 static int http_response_to_rc ( unsigned int response ) {
 	switch ( response ) {
 	case 200:
+	case 301:
+	case 302:
 		return 0;
 	case 404:
 		return -ENOENT;
@@ -181,6 +185,28 @@ static int http_rx_response ( struct http_request *http, char *response ) {
 }
 
 /**
+ * Handle HTTP Location header
+ *
+ * @v http		HTTP request
+ * @v value		HTTP header value
+ * @ret rc		Return status code
+ */
+static int http_rx_location ( struct http_request *http, const char *value ) {
+	int rc;
+
+	/* Redirect to new location */
+	DBGC ( http, "HTTP %p redirecting to %s\n", http, value );
+	if ( ( rc = xfer_redirect ( &http->xfer, LOCATION_URI_STRING,
+				    value ) ) != 0 ) {
+		DBGC ( http, "HTTP %p could not redirect: %s\n",
+		       http, strerror ( rc ) );
+		return rc;
+	}
+
+	return 0;
+}
+
+/**
  * Handle HTTP Content-Length header
  *
  * @v http		HTTP request
@@ -223,6 +249,10 @@ struct http_header_handler {
 /** List of HTTP header handlers */
 static struct http_header_handler http_header_handlers[] = {
 	{
+		.header = "Location",
+		.rx = http_rx_location,
+	},
+	{
 		.header = "Content-Length",
 		.rx = http_rx_content_length,
 	},
@@ -387,9 +417,7 @@ static int http_socket_deliver_iob ( struct xfer_interface *socket,
 static void http_step ( struct process *process ) {
 	struct http_request *http =
 		container_of ( process, struct http_request, process );
-	const char *path = http->uri->path;
 	const char *host = http->uri->host;
-	const char *query = http->uri->query;
 	const char *user = http->uri->user;
 	const char *password =
 		( http->uri->password ? http->uri->password : "" );
@@ -399,27 +427,24 @@ static void http_step ( struct process *process ) {
 	char user_pw[ user_pw_len + 1 /* NUL */ ];
 	char user_pw_base64[ user_pw_base64_len + 1 /* NUL */ ];
 	int rc;
+	int request_len = unparse_uri ( NULL, 0, http->uri,
+					URI_PATH_BIT | URI_QUERY_BIT );
 
 	if ( xfer_window ( &http->socket ) ) {
+		char request[request_len + 1];
+
+		/* Construct path?query request */
+		unparse_uri ( request, sizeof ( request ), http->uri,
+			      URI_PATH_BIT | URI_QUERY_BIT );
 
 		/* We want to execute only once */
 		process_del ( &http->process );
 
 		/* Construct authorisation, if applicable */
 		if ( user ) {
-			char *buf = user_pw;
-			ssize_t remaining = sizeof ( user_pw );
-			size_t len;
-
-			/* URI-decode the username and password */
-			len = uri_decode ( user, buf, remaining );
-			buf += len;
-			remaining -= len;
-			*(remaining--, buf++) = ':';
-			len = uri_decode ( password, buf, remaining );
-			buf += len;
-			remaining -= len;
-			assert ( remaining >= 0 );
+			/* Make "user:password" string from decoded fields */
+			snprintf ( user_pw, sizeof ( user_pw ), "%s:%s",
+				   user, password );
 
 			/* Base64-encode the "user:password" string */
 			base64_encode ( user_pw, user_pw_base64 );
@@ -427,14 +452,13 @@ static void http_step ( struct process *process ) {
 
 		/* Send GET request */
 		if ( ( rc = xfer_printf ( &http->socket,
-					  "GET %s%s%s HTTP/1.0\r\n"
+					  "GET %s%s HTTP/1.0\r\n"
 					  "User-Agent: gPXE/" VERSION "\r\n"
 					  "%s%s%s"
 					  "Host: %s\r\n"
 					  "\r\n",
-					  ( path ? path : "/" ),
-					  ( query ? "?" : "" ),
-					  ( query ? query : "" ),
+					  http->uri->path ? "" : "/",
+					  request,
 					  ( user ?
 					    "Authorization: Basic " : "" ),
 					  ( user ? user_pw_base64 : "" ),
@@ -464,7 +488,7 @@ static void http_socket_close ( struct xfer_interface *socket, int rc ) {
 /** HTTP socket operations */
 static struct xfer_interface_operations http_socket_operations = {
 	.close		= http_socket_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= http_socket_deliver_iob,
diff --git a/gpxe/src/net/tcp/https.c b/gpxe/src/net/tcp/https.c
index 15ab32ef..7a2961f2 100644
--- a/gpxe/src/net/tcp/https.c
+++ b/gpxe/src/net/tcp/https.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /**
  * @file
  *
diff --git a/gpxe/src/net/tcp/iscsi.c b/gpxe/src/net/tcp/iscsi.c
index 45519e66..771384b9 100644
--- a/gpxe/src/net/tcp/iscsi.c
+++ b/gpxe/src/net/tcp/iscsi.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stddef.h>
 #include <string.h>
 #include <stdlib.h>
@@ -180,9 +182,10 @@ static void iscsi_close_connection ( struct iscsi_session *iscsi, int rc ) {
 static void iscsi_scsi_done ( struct iscsi_session *iscsi, int rc ) {
 
 	assert ( iscsi->tx_state == ISCSI_TX_IDLE );
+	assert ( iscsi->command != NULL );
 
+	iscsi->command->rc = rc;
 	iscsi->command = NULL;
-	iscsi->rc = rc;
 }
 
 /****************************************************************************
@@ -1514,7 +1517,7 @@ static int iscsi_vredirect ( struct xfer_interface *socket, int type,
 		va_end ( tmp );
 	}
 
-	return xfer_vopen ( socket, type, args );
+	return xfer_vreopen ( socket, type, args );
 }
 			     
 
@@ -1548,37 +1551,24 @@ static int iscsi_command ( struct scsi_device *scsi,
 		container_of ( scsi->backend, struct iscsi_session, refcnt );
 	int rc;
 
+	/* Abort immediately if we have a recorded permanent failure */
+	if ( iscsi->instant_rc )
+		return iscsi->instant_rc;
+
 	/* Record SCSI command */
 	iscsi->command = command;
 
-	/* Abort immediately if we have a recorded permanent failure */
-	if ( iscsi->instant_rc ) {
-		rc = iscsi->instant_rc;
-		goto done;
-	}
-
 	/* Issue command or open connection as appropriate */
 	if ( iscsi->status ) {
 		iscsi_start_command ( iscsi );
 	} else {
-		if ( ( rc = iscsi_open_connection ( iscsi ) ) != 0 )
-			goto done;
+		if ( ( rc = iscsi_open_connection ( iscsi ) ) != 0 ) {
+			iscsi->command = NULL;
+			return rc;
+		}
 	}
 
-	/* Wait for command to complete */
-	iscsi->rc = -EINPROGRESS;
-	while ( iscsi->rc == -EINPROGRESS )
-		step();
-	rc = iscsi->rc;
-
- done:
-	iscsi->command = NULL;
-	return rc;
-}
-
-static int iscsi_detached_command ( struct scsi_device *scsi __unused,
-				    struct scsi_command *command __unused ) {
-	return -ENODEV;
+	return 0;
 }
 
 /**
@@ -1593,7 +1583,7 @@ void iscsi_detach ( struct scsi_device *scsi ) {
 	xfer_nullify ( &iscsi->socket );
 	iscsi_close_connection ( iscsi, 0 );
 	process_del ( &iscsi->process );
-	scsi->command = iscsi_detached_command;
+	scsi->command = scsi_detached_command;
 	ref_put ( scsi->backend );
 	scsi->backend = NULL;
 }
@@ -1616,42 +1606,6 @@ enum iscsi_root_path_component {
 };
 
 /**
- * Parse iSCSI LUN
- *
- * @v iscsi		iSCSI session
- * @v lun_string	LUN string representation (as per RFC4173)
- * @ret rc		Return status code
- */
-static int iscsi_parse_lun ( struct iscsi_session *iscsi,
-			     const char *lun_string ) {
-	union {
-		uint64_t u64;
-		uint16_t u16[4];
-	} lun;
-	char *p;
-	int i;
-
-	memset ( &lun, 0, sizeof ( lun ) );
-	if ( lun_string ) {
-		p = ( char * ) lun_string;
-		
-		for ( i = 0 ; i < 4 ; i++ ) {
-			lun.u16[i] = htons ( strtoul ( p, &p, 16 ) );
-			if ( *p == '\0' )
-				break;
-			if ( *p != '-' )
-				return -EINVAL;
-			p++;
-		}
-		if ( *p )
-			return -EINVAL;
-	}
-
-	iscsi->lun = lun.u64;
-	return 0;
-}
-
-/**
  * Parse iSCSI root path
  *
  * @v iscsi		iSCSI session
@@ -1689,7 +1643,7 @@ static int iscsi_parse_root_path ( struct iscsi_session *iscsi,
 	iscsi->target_port = strtoul ( rp_comp[RP_PORT], NULL, 10 );
 	if ( ! iscsi->target_port )
 		iscsi->target_port = ISCSI_PORT;
-	if ( ( rc = iscsi_parse_lun ( iscsi, rp_comp[RP_LUN] ) ) != 0 ) {
+	if ( ( rc = scsi_parse_lun ( rp_comp[RP_LUN], &iscsi->lun ) ) != 0 ) {
 		DBGC ( iscsi, "iSCSI %p invalid LUN \"%s\"\n",
 		       iscsi, rp_comp[RP_LUN] );
 		return rc;
@@ -1809,7 +1763,6 @@ int iscsi_attach ( struct scsi_device *scsi, const char *root_path ) {
 	/* Attach parent interface, mortalise self, and return */
 	scsi->backend = ref_get ( &iscsi->refcnt );
 	scsi->command = iscsi_command;
-	scsi->lun = iscsi->lun;
 	ref_put ( &iscsi->refcnt );
 	return 0;
 	
diff --git a/gpxe/src/net/tcpip.c b/gpxe/src/net/tcpip.c
index d4542b05..932fd482 100644
--- a/gpxe/src/net/tcpip.c
+++ b/gpxe/src/net/tcpip.c
@@ -14,17 +14,7 @@
  * TCP/IP transport-network layer interface
  */
 
-/** Registered network-layer protocols that support TCP/IP */
-static struct tcpip_net_protocol tcpip_net_protocols[0]
-	__table_start ( struct tcpip_net_protocol, tcpip_net_protocols );
-static struct tcpip_net_protocol tcpip_net_protocols_end[0]
-	__table_end ( struct tcpip_net_protocol, tcpip_net_protocols );
-
-/** Registered transport-layer protocols that support TCP/IP */
-static struct tcpip_protocol tcpip_protocols[0]
-	__table_start ( struct tcpip_protocol, tcpip_protocols );
-static struct tcpip_protocol tcpip_protocols_end[0]
-	__table_end ( struct tcpip_protocol, tcpip_protocols );
+FILE_LICENCE ( GPL2_OR_LATER );
 
 /** Process a received TCP/IP packet
  *
@@ -48,7 +38,7 @@ int tcpip_rx ( struct io_buffer *iobuf, uint8_t tcpip_proto,
 	struct tcpip_protocol *tcpip;
 
 	/* Hand off packet to the appropriate transport-layer protocol */
-	for ( tcpip = tcpip_protocols; tcpip < tcpip_protocols_end; tcpip++ ) {
+	for_each_table_entry ( tcpip, TCPIP_PROTOCOLS ) {
 		if ( tcpip->tcpip_proto == tcpip_proto ) {
 			DBG ( "TCP/IP received %s packet\n", tcpip->name );
 			return tcpip->rx ( iobuf, st_src, st_dest, pshdr_csum );
@@ -76,8 +66,7 @@ int tcpip_tx ( struct io_buffer *iobuf, struct tcpip_protocol *tcpip_protocol,
 	struct tcpip_net_protocol *tcpip_net;
 
 	/* Hand off packet to the appropriate network-layer protocol */
-	for ( tcpip_net = tcpip_net_protocols ;
-	      tcpip_net < tcpip_net_protocols_end ; tcpip_net++ ) {
+	for_each_table_entry ( tcpip_net, TCPIP_NET_PROTOCOLS ) {
 		if ( tcpip_net->sa_family == st_dest->st_family ) {
 			DBG ( "TCP/IP sending %s packet\n", tcpip_net->name );
 			return tcpip_net->tx ( iobuf, tcpip_protocol, st_src,
diff --git a/gpxe/src/net/tls.c b/gpxe/src/net/tls.c
index 73f9ad06..a5b126ed 100644
--- a/gpxe/src/net/tls.c
+++ b/gpxe/src/net/tls.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /**
  * @file
  *
@@ -1625,7 +1627,7 @@ static int tls_cipherstream_deliver_raw ( struct xfer_interface *xfer,
 /** TLS ciphertext stream operations */
 static struct xfer_interface_operations tls_cipherstream_operations = {
 	.close		= tls_cipherstream_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= filter_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= xfer_deliver_as_raw,
diff --git a/gpxe/src/net/udp.c b/gpxe/src/net/udp.c
index c3a1eba3..771655e0 100644
--- a/gpxe/src/net/udp.c
+++ b/gpxe/src/net/udp.c
@@ -16,6 +16,8 @@
  * UDP protocol
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 /**
  * A UDP connection
  *
@@ -238,7 +240,7 @@ static int udp_tx ( struct udp_connection *udp, struct io_buffer *iobuf,
  * @ret udp		UDP connection, or NULL
  */
 static struct udp_connection * udp_demux ( struct sockaddr_tcpip *local ) {
-	static const struct sockaddr_tcpip empty_sockaddr;
+	static const struct sockaddr_tcpip empty_sockaddr = { .pad = { 0, } };
 	struct udp_connection *udp;
 
 	list_for_each_entry ( udp, &udp_conns, list ) {
diff --git a/gpxe/src/net/udp/dhcp.c b/gpxe/src/net/udp/dhcp.c
index ab843ce1..ce2c8207 100644
--- a/gpxe/src/net/udp/dhcp.c
+++ b/gpxe/src/net/udp/dhcp.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <string.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -69,6 +71,7 @@ static const uint8_t dhcp_op[] = {
 
 /** Raw option data for options common to all DHCP requests */
 static uint8_t dhcp_request_options_data[] = {
+	DHCP_MESSAGE_TYPE, DHCP_BYTE ( 0 ),
 	DHCP_MAX_MESSAGE_SIZE,
 	DHCP_WORD ( ETH_MAX_MTU - 20 /* IP header */ - 8 /* UDP header */ ),
 	DHCP_CLIENT_ARCHITECTURE, DHCP_WORD ( 0 ),
@@ -88,10 +91,6 @@ static uint8_t dhcp_request_options_data[] = {
 	DHCP_END
 };
 
-/** DHCP feature codes */
-static uint8_t dhcp_features[0] __table_start ( uint8_t, dhcp_features );
-static uint8_t dhcp_features_end[0] __table_end ( uint8_t, dhcp_features );
-
 /** Version number feature */
 FEATURE_VERSION ( VERSION_MAJOR, VERSION_MINOR, VERSION_PATCH );
 
@@ -111,6 +110,14 @@ struct setting user_class_setting __setting = {
 	.type = &setting_type_string,
 };
 
+/** Use cached network settings */
+struct setting use_cached_setting __setting = {
+	.name = "use-cached",
+	.description = "Use cached network settings",
+	.tag = DHCP_EB_USE_CACHED,
+	.type = &setting_type_uint8,
+};
+
 /**
  * Name a DHCP packet type
  *
@@ -199,6 +206,36 @@ static struct dhcp_session_state dhcp_state_request;
 static struct dhcp_session_state dhcp_state_proxy;
 static struct dhcp_session_state dhcp_state_pxebs;
 
+/** DHCP offer is valid for IP lease */
+#define DHCP_OFFER_IP	1
+
+/** DHCP offer is valid for PXE options */
+#define DHCP_OFFER_PXE	2
+
+/** A DHCP offer */
+struct dhcp_offer {
+	/** IP address of server granting offer */
+	struct in_addr server;
+
+	/** IP address being offered, or 0.0.0.0 for a pure proxy */
+	struct in_addr ip;
+
+	/** DHCP packet containing PXE options; NULL if missing or proxied */
+	struct dhcp_packet *pxe;
+
+	/** Valid uses for this offer, a combination of DHCP_OFFER bits */
+	uint8_t valid;
+
+	/** Priority of this offer */
+	int8_t priority;
+
+	/** Whether to ignore PXE DHCP extensions */
+	uint8_t no_pxedhcp;
+};
+
+/** Maximum number of DHCP offers to queue */
+#define DHCP_MAX_OFFERS   6
+
 /** A DHCP session */
 struct dhcp_session {
 	/** Reference counter */
@@ -215,20 +252,6 @@ struct dhcp_session {
 	/** State of the session */
 	struct dhcp_session_state *state;
 
-	/** Offered IP address */
-	struct in_addr offer;
-	/** DHCP server */
-	struct in_addr server;
-	/** DHCP offer priority */
-	int priority;
-
-	/** ProxyDHCP protocol extensions should be ignored */
-	int no_pxedhcp;
-	/** ProxyDHCP server */
-	struct in_addr proxy_server;
-	/** ProxyDHCP server priority */
-	int proxy_priority;
-
 	/** PXE Boot Server type */
 	uint16_t pxe_type;
 	/** List of PXE Boot Servers to attempt */
@@ -240,6 +263,11 @@ struct dhcp_session {
 	struct retry_timer timer;
 	/** Start time of the current state (in ticks) */
 	unsigned long start;
+
+	/** DHCP offer just requested */
+	struct dhcp_offer *current_offer;
+	/** List of DHCP offers received */
+	struct dhcp_offer offers[DHCP_MAX_OFFERS];
 };
 
 /**
@@ -250,6 +278,12 @@ struct dhcp_session {
 static void dhcp_free ( struct refcnt *refcnt ) {
 	struct dhcp_session *dhcp =
 		container_of ( refcnt, struct dhcp_session, refcnt );
+	int i;
+
+	for ( i = 0 ; i < DHCP_MAX_OFFERS ; i++ ) {
+		if ( dhcp->offers[i].pxe )
+			dhcppkt_put ( dhcp->offers[i].pxe );
+	}
 
 	netdev_put ( dhcp->netdev );
 	free ( dhcp );
@@ -294,6 +328,35 @@ static void dhcp_set_state ( struct dhcp_session *dhcp,
 	start_timer_nodelay ( &dhcp->timer );
 }
 
+/**
+ * Determine next DHCP offer to try
+ *
+ * @v dhcp		DHCP session
+ * @v type		DHCP offer type
+ * @ret offer		Next DHCP offer to try
+ *
+ * Offers are ranked by priority, then by completeness (combined
+ * IP+PXE are tried before @a type alone), then by order of receipt.
+ */
+static struct dhcp_offer * dhcp_next_offer ( struct dhcp_session *dhcp,
+					     uint8_t type ) {
+
+	struct dhcp_offer *offer;
+	struct dhcp_offer *best = NULL;
+
+	for ( offer = dhcp->offers ; offer < dhcp->offers + DHCP_MAX_OFFERS ;
+	      offer++ ) {
+		if ( ( offer->valid & type ) &&
+		     ( ( best == NULL ) ||
+		       ( offer->priority > best->priority ) ||
+		       ( ( offer->priority == best->priority ) &&
+			 ( offer->valid & ~best->valid ) ) ) )
+			best = offer;
+	}
+
+	return best;
+}
+
 /****************************************************************************
  *
  * DHCP state machine
@@ -321,7 +384,7 @@ static int dhcp_discovery_tx ( struct dhcp_session *dhcp,
 }
 
 /**
- * Handle received packet during DHCP discovery
+ * Handle received DHCPOFFER during any state
  *
  * @v dhcp		DHCP session
  * @v dhcppkt		DHCP packet
@@ -329,17 +392,17 @@ static int dhcp_discovery_tx ( struct dhcp_session *dhcp,
  * @v msgtype		DHCP message type
  * @v server_id		DHCP server ID
  */
-static void dhcp_discovery_rx ( struct dhcp_session *dhcp,
-				struct dhcp_packet *dhcppkt,
-				struct sockaddr_in *peer, uint8_t msgtype,
-				struct in_addr server_id ) {
-	struct in_addr ip;
+static void dhcp_rx_offer ( struct dhcp_session *dhcp,
+			    struct dhcp_packet *dhcppkt,
+			    struct sockaddr_in *peer, uint8_t msgtype,
+			    struct in_addr server_id ) {
 	char vci[9]; /* "PXEClient" */
 	int vci_len;
 	int has_pxeclient;
-	int8_t priority = 0;
-	uint8_t no_pxedhcp = 0;
-	unsigned long elapsed;
+	int pxeopts_len;
+	int has_pxeopts;
+	struct dhcp_offer *offer;
+	int i;
 
 	DBGC ( dhcp, "DHCP %p %s from %s:%d", dhcp,
 	       dhcp_msgtype_name ( msgtype ), inet_ntoa ( peer->sin_addr ),
@@ -348,47 +411,87 @@ static void dhcp_discovery_rx ( struct dhcp_session *dhcp,
 		DBGC ( dhcp, " (%s)", inet_ntoa ( server_id ) );
 
 	/* Identify offered IP address */
-	ip = dhcppkt->dhcphdr->yiaddr;
-	if ( ip.s_addr )
-		DBGC ( dhcp, " for %s", inet_ntoa ( ip ) );
+	if ( dhcppkt->dhcphdr->yiaddr.s_addr )
+		DBGC ( dhcp, " for %s", inet_ntoa ( dhcppkt->dhcphdr->yiaddr ));
+
+	/* Enqueue an offer to be filled in */
+	for ( i = 0 ; i < DHCP_MAX_OFFERS ; i++ ) {
+		if ( dhcp->offers[i].server.s_addr == server_id.s_addr ) {
+			DBGC ( dhcp, " dup\n" );
+			return;
+		}
+
+		if ( ! dhcp->offers[i].valid )
+			break;
+	}
+	if ( i == DHCP_MAX_OFFERS ) {
+		DBGC ( dhcp, " dropped\n" );
+		return;
+	}
+
+	offer = &dhcp->offers[i];
+	offer->server = server_id;
+	offer->ip = dhcppkt->dhcphdr->yiaddr;
 
 	/* Identify "PXEClient" vendor class */
 	vci_len = dhcppkt_fetch ( dhcppkt, DHCP_VENDOR_CLASS_ID,
 				  vci, sizeof ( vci ) );
 	has_pxeclient = ( ( vci_len >= ( int ) sizeof ( vci ) ) &&
 			  ( strncmp ( "PXEClient", vci, sizeof (vci) ) == 0 ));
+
+	/* Identify presence of PXE-specific options */
+	pxeopts_len = dhcppkt_fetch ( dhcppkt, DHCP_PXE_BOOT_MENU, NULL, 0 );
+	has_pxeopts = ( pxeopts_len >= 0 );
 	if ( has_pxeclient )
-		DBGC ( dhcp, " pxe" );
+		DBGC ( dhcp, "%s", ( has_pxeopts ? " pxe" : " proxy" ) );
+
+	if ( has_pxeclient && has_pxeopts ) {
+		/* Save reference to packet for future use */
+		if ( offer->pxe )
+			dhcppkt_put ( offer->pxe );
+		offer->pxe = dhcppkt_get ( dhcppkt );
+	}
 
 	/* Identify priority */
-	dhcppkt_fetch ( dhcppkt, DHCP_EB_PRIORITY, &priority,
-			sizeof ( priority ) );
-	if ( priority )
-		DBGC ( dhcp, " pri %d", priority );
+	dhcppkt_fetch ( dhcppkt, DHCP_EB_PRIORITY, &offer->priority,
+			sizeof ( offer->priority ) );
+	if ( offer->priority )
+		DBGC ( dhcp, " pri %d", offer->priority );
 
 	/* Identify ignore-PXE flag */
-	dhcppkt_fetch ( dhcppkt, DHCP_EB_NO_PXEDHCP, &no_pxedhcp,
-			sizeof ( no_pxedhcp ) );
-	if ( no_pxedhcp )
+	dhcppkt_fetch ( dhcppkt, DHCP_EB_NO_PXEDHCP, &offer->no_pxedhcp,
+			sizeof ( offer->no_pxedhcp ) );
+	if ( offer->no_pxedhcp )
 		DBGC ( dhcp, " nopxe" );
 	DBGC ( dhcp, "\n" );
 
-	/* Select as DHCP offer, if applicable */
-	if ( ip.s_addr && ( peer->sin_port == htons ( BOOTPS_PORT ) ) &&
-	     ( ( msgtype == DHCPOFFER ) || ( ! msgtype /* BOOTP */ ) ) &&
-	     ( priority >= dhcp->priority ) ) {
-		dhcp->offer = ip;
-		dhcp->server = server_id;
-		dhcp->priority = priority;
-		dhcp->no_pxedhcp = no_pxedhcp;
-	}
+	/* Determine roles this offer can fill */
+	if ( offer->ip.s_addr &&
+	     ( peer->sin_port == htons ( BOOTPS_PORT ) ) &&
+	     ( ( msgtype == DHCPOFFER ) || ( ! msgtype /* BOOTP */ ) ) )
+		offer->valid |= DHCP_OFFER_IP;
 
-	/* Select as ProxyDHCP offer, if applicable */
-	if ( has_pxeclient && ( msgtype == DHCPOFFER ) &&
-	     ( priority >= dhcp->proxy_priority ) ) {
-		dhcp->proxy_server = server_id;
-		dhcp->proxy_priority = priority;
-	}
+	if ( has_pxeclient && ( msgtype == DHCPOFFER ) )
+		offer->valid |= DHCP_OFFER_PXE;
+}
+
+/**
+ * Handle received packet during DHCP discovery
+ *
+ * @v dhcp		DHCP session
+ * @v dhcppkt		DHCP packet
+ * @v peer		DHCP server address
+ * @v msgtype		DHCP message type
+ * @v server_id		DHCP server ID
+ */
+static void dhcp_discovery_rx ( struct dhcp_session *dhcp,
+				struct dhcp_packet *dhcppkt,
+				struct sockaddr_in *peer, uint8_t msgtype,
+				struct in_addr server_id ) {
+	unsigned long elapsed;
+	struct dhcp_offer *ip_offer;
+
+	dhcp_rx_offer ( dhcp, dhcppkt, peer, msgtype, server_id );
 
 	/* We can exit the discovery state when we have a valid
 	 * DHCPOFFER, and either:
@@ -399,12 +502,14 @@ static void dhcp_discovery_rx ( struct dhcp_session *dhcp,
 	 */
 
 	/* If we don't yet have a DHCPOFFER, do nothing */
-	if ( ! dhcp->offer.s_addr )
+	ip_offer = dhcp_next_offer ( dhcp, DHCP_OFFER_IP );
+	if ( ! ip_offer )
 		return;
 
 	/* If we can't yet transition to DHCPREQUEST, do nothing */
 	elapsed = ( currticks() - dhcp->start );
-	if ( ! ( dhcp->no_pxedhcp || dhcp->proxy_server.s_addr ||
+	if ( ! ( ip_offer->no_pxedhcp ||
+		 dhcp_next_offer ( dhcp, DHCP_OFFER_PXE ) ||
 		 ( elapsed > PROXYDHCP_MAX_TIMEOUT ) ) )
 		return;
 
@@ -421,7 +526,8 @@ static void dhcp_discovery_expired ( struct dhcp_session *dhcp ) {
 	unsigned long elapsed = ( currticks() - dhcp->start );
 
 	/* Give up waiting for ProxyDHCP before we reach the failure point */
-	if ( dhcp->offer.s_addr && ( elapsed > PROXYDHCP_MAX_TIMEOUT ) ) {
+	if ( dhcp_next_offer ( dhcp, DHCP_OFFER_IP ) &&
+	     ( elapsed > PROXYDHCP_MAX_TIMEOUT ) ) {
 		dhcp_set_state ( dhcp, &dhcp_state_request );
 		return;
 	}
@@ -451,21 +557,23 @@ static int dhcp_request_tx ( struct dhcp_session *dhcp,
 			     struct dhcp_packet *dhcppkt,
 			     struct sockaddr_in *peer ) {
 	int rc;
+	struct dhcp_offer *offer;
+
+	offer = dhcp->current_offer = dhcp_next_offer ( dhcp, DHCP_OFFER_IP );
 
 	DBGC ( dhcp, "DHCP %p DHCPREQUEST to %s:%d",
-	       dhcp, inet_ntoa ( dhcp->server ), BOOTPS_PORT );
-	DBGC ( dhcp, " for %s\n", inet_ntoa ( dhcp->offer ) );
+	       dhcp, inet_ntoa ( offer->server ), BOOTPS_PORT );
+	DBGC ( dhcp, " for %s\n", inet_ntoa ( offer->ip ) );
 
 	/* Set server ID */
 	if ( ( rc = dhcppkt_store ( dhcppkt, DHCP_SERVER_IDENTIFIER,
-				    &dhcp->server,
-				    sizeof ( dhcp->server ) ) ) != 0 )
+				    &offer->server,
+				    sizeof ( offer->server ) ) ) != 0 )
 		return rc;
 
 	/* Set requested IP address */
 	if ( ( rc = dhcppkt_store ( dhcppkt, DHCP_REQUESTED_ADDRESS,
-				    &dhcp->offer,
-				    sizeof ( dhcp->offer ) ) ) != 0 )
+				    &offer->ip, sizeof ( offer->ip ) ) ) != 0 )
 		return rc;
 
 	/* Set server address */
@@ -491,6 +599,18 @@ static void dhcp_request_rx ( struct dhcp_session *dhcp,
 	struct in_addr ip;
 	struct settings *parent;
 	int rc;
+	struct dhcp_offer *pxe_offer;
+
+	if ( msgtype == DHCPOFFER ) {
+		dhcp_rx_offer ( dhcp, dhcppkt, peer, msgtype, server_id );
+		if ( dhcp_next_offer ( dhcp, DHCP_OFFER_IP ) !=
+		     dhcp->current_offer ) {
+			/* Restart due to higher-priority offer received */
+			DBGC ( dhcp, "DHCP %p re-requesting\n", dhcp );
+			dhcp_set_state ( dhcp, &dhcp_state_request );
+		}
+		return;
+	}
 
 	DBGC ( dhcp, "DHCP %p %s from %s:%d", dhcp,
 	       dhcp_msgtype_name ( msgtype ), inet_ntoa ( peer->sin_addr ),
@@ -509,7 +629,7 @@ static void dhcp_request_rx ( struct dhcp_session *dhcp,
 		return;
 	if ( msgtype /* BOOTP */ && ( msgtype != DHCPACK ) )
 		return;
-	if ( server_id.s_addr != dhcp->server.s_addr )
+	if ( server_id.s_addr != dhcp->current_offer->server.s_addr )
 		return;
 
 	/* Record assigned address */
@@ -524,14 +644,31 @@ static void dhcp_request_rx ( struct dhcp_session *dhcp,
 		return;
 	}
 
-	/* Start ProxyDHCPREQUEST if applicable */
-	if ( dhcp->proxy_server.s_addr && ( ! dhcp->no_pxedhcp ) ) {
+	/* Locate best source of PXE settings */
+	pxe_offer = dhcp_next_offer ( dhcp, DHCP_OFFER_PXE );
+
+	if ( ( ! pxe_offer ) || /* No PXE available */
+	     /* IP offer instructs us to ignore PXE */
+	     dhcp->current_offer->no_pxedhcp ||
+	     /* PXE settings already registered with IP offer */
+	     ( ( dhcp->current_offer == pxe_offer ) && ( pxe_offer->pxe ) ) ) {
+
+		/* Terminate DHCP */
+		dhcp_finished ( dhcp, 0 );
+
+	} else if ( pxe_offer->pxe ) {
+		/* Register PXE settings and terminate DHCP */
+		pxe_offer->pxe->settings.name = PROXYDHCP_SETTINGS_NAME;
+		if ( ( rc = register_settings ( &pxe_offer->pxe->settings,
+						NULL ) ) != 0 ) {
+			DBGC ( dhcp, "DHCP %p could not register settings: "
+			       "%s\n", dhcp, strerror ( rc ) );
+		}
+		dhcp_finished ( dhcp, rc );
+	} else {
+		/* Start ProxyDHCP */
 		dhcp_set_state ( dhcp, &dhcp_state_proxy );
-		return;
 	}
-
-	/* Terminate DHCP */
-	dhcp_finished ( dhcp, 0 );
 }
 
 /**
@@ -566,18 +703,21 @@ static int dhcp_proxy_tx ( struct dhcp_session *dhcp,
 			   struct dhcp_packet *dhcppkt,
 			   struct sockaddr_in *peer ) {
 	int rc;
+	struct dhcp_offer *offer;
+
+	offer = dhcp->current_offer = dhcp_next_offer ( dhcp, DHCP_OFFER_PXE );
 
-	DBGC ( dhcp, "DHCP %p ProxyDHCP REQUEST to %s:%d\n",
-	       dhcp, inet_ntoa ( dhcp->proxy_server ), PXE_PORT );
+	DBGC ( dhcp, "DHCP %p ProxyDHCP REQUEST to %s:%d\n", dhcp,
+	       inet_ntoa ( offer->server ), PXE_PORT );
 
 	/* Set server ID */
 	if ( ( rc = dhcppkt_store ( dhcppkt, DHCP_SERVER_IDENTIFIER,
-				    &dhcp->proxy_server,
-				    sizeof ( dhcp->proxy_server ) ) ) != 0 )
+				    &offer->server,
+				    sizeof ( offer->server ) ) )  != 0 )
 		return rc;
 
 	/* Set server address */
-	peer->sin_addr = dhcp->proxy_server;
+	peer->sin_addr = offer->server;
 	peer->sin_port = htons ( PXE_PORT );
 
 	return 0;
@@ -598,6 +738,13 @@ static void dhcp_proxy_rx ( struct dhcp_session *dhcp,
 			    struct in_addr server_id ) {
 	int rc;
 
+	/* Enqueue last-minute DHCPOFFERs for use in case of failure */
+	if ( peer->sin_port == htons ( BOOTPS_PORT ) &&
+	     msgtype == DHCPOFFER ) {
+		dhcp_rx_offer ( dhcp, dhcppkt, peer, msgtype, server_id );
+		return;
+	}
+
 	DBGC ( dhcp, "DHCP %p %s from %s:%d", dhcp,
 	       dhcp_msgtype_name ( msgtype ), inet_ntoa ( peer->sin_addr ),
 	       ntohs ( peer->sin_port ) );
@@ -608,10 +755,10 @@ static void dhcp_proxy_rx ( struct dhcp_session *dhcp,
 	/* Filter out unacceptable responses */
 	if ( peer->sin_port != htons ( PXE_PORT ) )
 		return;
-	if ( msgtype != DHCPACK )
+	if ( msgtype != DHCPACK && msgtype != DHCPOFFER )
 		return;
 	if ( server_id.s_addr /* Linux PXE server omits server ID */ &&
-	     ( server_id.s_addr != dhcp->proxy_server.s_addr ) )
+	     ( server_id.s_addr != dhcp->current_offer->server.s_addr ) )
 		return;
 
 	/* Register settings */
@@ -637,6 +784,28 @@ static void dhcp_proxy_expired ( struct dhcp_session *dhcp ) {
 
 	/* Give up waiting for ProxyDHCP before we reach the failure point */
 	if ( elapsed > PROXYDHCP_MAX_TIMEOUT ) {
+
+		/* Mark failed offer as unsuitable for ProxyDHCP */
+		dhcp->current_offer->valid &= ~DHCP_OFFER_PXE;
+
+		/* Prefer not to use only half of a PXE+IP offer if we
+		 * have other offers available
+		 */
+		dhcp->current_offer->priority = -1;
+
+		/* If we have any other PXE offers we can try, go back
+		 * to DHCPREQUEST (since they might not be proxied
+		 * offers, or might be coupled to a new IP address).
+		 * We should probably DHCPRELEASE our old IP, but the
+		 * standard does not require it.
+		 */
+		if ( dhcp_next_offer ( dhcp, DHCP_OFFER_PXE ) ) {
+			dhcp->local.sin_addr.s_addr = 0;
+			dhcp_set_state ( dhcp, &dhcp_state_request );
+			return;
+		}
+
+		/* No possibilities left; finish without PXE options */
 		dhcp_finished ( dhcp, 0 );
 		return;
 	}
@@ -668,9 +837,14 @@ static int dhcp_pxebs_tx ( struct dhcp_session *dhcp,
 	struct dhcp_pxe_boot_menu_item menu_item = { 0, 0 };
 	int rc;
 
+	/* Set server address */
+	peer->sin_addr = *(dhcp->pxe_attempt);
+	peer->sin_port = ( ( peer->sin_addr.s_addr == INADDR_BROADCAST ) ?
+			   htons ( BOOTPS_PORT ) : htons ( PXE_PORT ) );
+
 	DBGC ( dhcp, "DHCP %p PXEBS REQUEST to %s:%d for type %d\n",
-	       dhcp, inet_ntoa ( *(dhcp->pxe_attempt) ), PXE_PORT,
-	       ntohs ( dhcp->pxe_type ) );
+	       dhcp, inet_ntoa ( peer->sin_addr ), ntohs ( peer->sin_port ),
+	       le16_to_cpu ( dhcp->pxe_type ) );
 
 	/* Set boot menu item */
 	menu_item.type = dhcp->pxe_type;
@@ -678,10 +852,6 @@ static int dhcp_pxebs_tx ( struct dhcp_session *dhcp,
 				    &menu_item, sizeof ( menu_item ) ) ) != 0 )
 		return rc;
 
-	/* Set server address */
-	peer->sin_addr = *(dhcp->pxe_attempt);
-	peer->sin_port = htons ( PXE_PORT );
-
 	return 0;
 }
 
@@ -741,7 +911,8 @@ static void dhcp_pxebs_rx ( struct dhcp_session *dhcp,
 	DBGC ( dhcp, "\n" );
 
 	/* Filter out unacceptable responses */
-	if ( peer->sin_port != htons ( PXE_PORT ) )
+	if ( ( peer->sin_port != htons ( BOOTPS_PORT ) ) &&
+	     ( peer->sin_port != htons ( PXE_PORT ) ) )
 		return;
 	if ( msgtype != DHCPACK )
 		return;
@@ -807,6 +978,46 @@ static struct dhcp_session_state dhcp_state_pxebs = {
  */
 
 /**
+ * Construct DHCP client hardware address field and broadcast flag
+ *
+ * @v netdev		Network device
+ * @v hlen		DHCP hardware address length to fill in
+ * @v flags		DHCP flags to fill in
+ * @ret chaddr		DHCP client hardware address
+ */
+void * dhcp_chaddr ( struct net_device *netdev, uint8_t *hlen,
+		     uint16_t *flags ) {
+	struct ll_protocol *ll_protocol = netdev->ll_protocol;
+	typeof ( ( ( struct dhcphdr * ) NULL )->chaddr ) chaddr;
+
+	/* If the link-layer address cannot fit into the chaddr field
+	 * (as is the case for IPoIB) then try using the hardware
+	 * address instead.  If we do this, set the broadcast flag,
+	 * since chaddr then does not represent a valid link-layer
+	 * address for the return path.
+	 *
+	 * If even the hardware address is too large, use an empty
+	 * chaddr field and set the broadcast flag.
+	 *
+	 * This goes against RFC4390, but RFC4390 mandates that we use
+	 * a DHCP client identifier that conforms with RFC4361, which
+	 * we cannot do without either persistent (NIC-independent)
+	 * storage, or by eliminating the hardware address completely
+	 * from the DHCP packet, which seems unfriendly to users.
+	 */
+	if ( ( *hlen = ll_protocol->ll_addr_len ) <= sizeof ( chaddr ) ) {
+		return netdev->ll_addr;
+	}
+	*flags = htons ( BOOTP_FL_BROADCAST );
+	if ( ( *hlen = ll_protocol->hw_addr_len ) <= sizeof ( chaddr ) ) {
+		return netdev->hw_addr;
+	} else {
+		*hlen = 0;
+		return NULL;
+	}
+}
+
+/**
  * Create a DHCP packet
  *
  * @v dhcppkt		DHCP packet structure to fill in
@@ -826,7 +1037,7 @@ int dhcp_create_packet ( struct dhcp_packet *dhcppkt,
 			 const void *options, size_t options_len,
 			 void *data, size_t max_len ) {
 	struct dhcphdr *dhcphdr = data;
-	unsigned int hlen;
+	void *chaddr;
 	int rc;
 
 	/* Sanity check */
@@ -839,16 +1050,8 @@ int dhcp_create_packet ( struct dhcp_packet *dhcppkt,
 	dhcphdr->magic = htonl ( DHCP_MAGIC_COOKIE );
 	dhcphdr->htype = ntohs ( netdev->ll_protocol->ll_proto );
 	dhcphdr->op = dhcp_op[msgtype];
-	/* If hardware length exceeds the chaddr field length, don't
-	 * use the chaddr field.  This is as per RFC4390.
-	 */
-	hlen = netdev->ll_protocol->ll_addr_len;
-	if ( hlen > sizeof ( dhcphdr->chaddr ) ) {
-		hlen = 0;
-		dhcphdr->flags = htons ( BOOTP_FL_BROADCAST );
-	}
-	dhcphdr->hlen = hlen;
-	memcpy ( dhcphdr->chaddr, netdev->ll_addr, hlen );
+	chaddr = dhcp_chaddr ( netdev, &dhcphdr->hlen, &dhcphdr->flags );
+	memcpy ( dhcphdr->chaddr, chaddr, dhcphdr->hlen );
 	memcpy ( dhcphdr->options, options, options_len );
 
 	/* Initialise DHCP packet structure */
@@ -880,10 +1083,10 @@ int dhcp_create_packet ( struct dhcp_packet *dhcppkt,
 int dhcp_create_request ( struct dhcp_packet *dhcppkt,
 			  struct net_device *netdev, unsigned int msgtype,
 			  struct in_addr ciaddr, void *data, size_t max_len ) {
-	struct device_description *desc = &netdev->dev->desc;
 	struct dhcp_netdev_desc dhcp_desc;
 	struct dhcp_client_id client_id;
 	struct dhcp_client_uuid client_uuid;
+	uint8_t *dhcp_features;
 	size_t dhcp_features_len;
 	size_t ll_addr_len;
 	ssize_t len;
@@ -903,7 +1106,8 @@ int dhcp_create_request ( struct dhcp_packet *dhcppkt,
 	dhcppkt->dhcphdr->ciaddr = ciaddr;
 
 	/* Add options to identify the feature list */
-	dhcp_features_len = ( dhcp_features_end - dhcp_features );
+	dhcp_features = table_start ( DHCP_FEATURES );
+	dhcp_features_len = table_num_entries ( DHCP_FEATURES );
 	if ( ( rc = dhcppkt_store ( dhcppkt, DHCP_EB_ENCAP, dhcp_features,
 				    dhcp_features_len ) ) != 0 ) {
 		DBG ( "DHCP could not set features list option: %s\n",
@@ -912,9 +1116,8 @@ int dhcp_create_request ( struct dhcp_packet *dhcppkt,
 	}
 
 	/* Add options to identify the network device */
-	dhcp_desc.type = desc->bus_type;
-	dhcp_desc.vendor = htons ( desc->vendor );
-	dhcp_desc.device = htons ( desc->device );
+	fetch_setting ( &netdev->settings.settings, &busid_setting, &dhcp_desc,
+		sizeof ( dhcp_desc ) );
 	if ( ( rc = dhcppkt_store ( dhcppkt, DHCP_EB_BUS_ID, &dhcp_desc,
 				    sizeof ( dhcp_desc ) ) ) != 0 ) {
 		DBG ( "DHCP could not set bus ID option: %s\n",
@@ -1109,7 +1312,7 @@ static int dhcp_deliver_iob ( struct xfer_interface *xfer,
 /** DHCP data transfer interface operations */
 static struct xfer_interface_operations dhcp_xfer_operations = {
 	.close		= ignore_xfer_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= dhcp_deliver_iob,
@@ -1184,16 +1387,28 @@ static struct sockaddr dhcp_peer = {
  *
  * @v job		Job control interface
  * @v netdev		Network device
- * @ret rc		Return status code
+ * @ret rc		Return status code, or positive if cached
  *
  * Starts DHCP on the specified network device.  If successful, the
  * DHCPACK (and ProxyDHCPACK, if applicable) will be registered as
  * option sources.
+ *
+ * On a return of 0, a background job has been started to perform the
+ * DHCP request. Any nonzero return means the job has not been
+ * started; a positive return value indicates the success condition of
+ * having fetched the appropriate data from cached information.
  */
 int start_dhcp ( struct job_interface *job, struct net_device *netdev ) {
 	struct dhcp_session *dhcp;
 	int rc;
 
+	/* Check for cached DHCP information */
+	get_cached_dhcpack();
+	if ( fetch_uintz_setting ( NULL, &use_cached_setting ) ) {
+		DBG ( "DHCP using cached network settings\n" );
+		return 1;
+	}
+
 	/* Allocate and initialise structure */
 	dhcp = zalloc ( sizeof ( *dhcp ) );
 	if ( ! dhcp )
@@ -1309,7 +1524,7 @@ int start_pxebs ( struct job_interface *job, struct net_device *netdev,
 	fetch_ipv4_setting ( netdev_settings ( netdev ), &ip_setting,
 			     &dhcp->local.sin_addr );
 	dhcp->local.sin_port = htons ( BOOTPC_PORT );
-	dhcp->pxe_type = htons ( pxe_type );
+	dhcp->pxe_type = cpu_to_le16 ( pxe_type );
 	dhcp->timer.expired = dhcp_timer_expired;
 
 	/* Construct PXE boot server IP address lists */
diff --git a/gpxe/src/net/udp/dns.c b/gpxe/src/net/udp/dns.c
index a498aefc..f94094aa 100644
--- a/gpxe/src/net/udp/dns.c
+++ b/gpxe/src/net/udp/dns.c
@@ -19,6 +19,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <string.h>
@@ -459,7 +461,7 @@ static void dns_xfer_close ( struct xfer_interface *socket, int rc ) {
 /** DNS socket operations */
 static struct xfer_interface_operations dns_socket_operations = {
 	.close		= dns_xfer_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= xfer_deliver_as_raw,
diff --git a/gpxe/src/net/udp/slam.c b/gpxe/src/net/udp/slam.c
index 6add99bc..396f69b0 100644
--- a/gpxe/src/net/udp/slam.c
+++ b/gpxe/src/net/udp/slam.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <string.h>
@@ -614,7 +616,7 @@ static void slam_socket_close ( struct xfer_interface *socket, int rc ) {
 /** SLAM unicast socket data transfer operations */
 static struct xfer_interface_operations slam_socket_operations = {
 	.close		= slam_socket_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= slam_socket_deliver,
@@ -640,7 +642,7 @@ static void slam_mc_socket_close ( struct xfer_interface *mc_socket, int rc ){
 /** SLAM multicast socket data transfer operations */
 static struct xfer_interface_operations slam_mc_socket_operations = {
 	.close		= slam_mc_socket_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= slam_mc_socket_deliver,
diff --git a/gpxe/src/net/udp/tftp.c b/gpxe/src/net/udp/tftp.c
index 19525f79..3de2fb9b 100644
--- a/gpxe/src/net/udp/tftp.c
+++ b/gpxe/src/net/udp/tftp.c
@@ -16,6 +16,8 @@
  * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  */
 
+FILE_LICENCE ( GPL2_OR_LATER );
+
 #include <stdint.h>
 #include <stdlib.h>
 #include <stdio.h>
@@ -131,6 +133,8 @@ enum {
 	TFTP_FL_RRQ_MULTICAST = 0x0004,
 	/** Perform MTFTP recovery on timeout */
 	TFTP_FL_MTFTP_RECOVERY = 0x0008,
+	/** Only get filesize and then abort the transfer */
+	TFTP_FL_SIZEONLY = 0x0010,
 };
 
 /** Maximum number of MTFTP open requests before falling back to TFTP */
@@ -409,6 +413,42 @@ static int tftp_send_ack ( struct tftp_request *tftp ) {
 }
 
 /**
+ * Transmit ERROR (Abort)
+ *
+ * @v tftp		TFTP connection
+ * @v errcode		TFTP error code
+ * @v errmsg		Error message string
+ * @ret rc		Return status code
+ */
+static int tftp_send_error ( struct tftp_request *tftp, int errcode,
+			     const char *errmsg ) {
+	struct tftp_error *err;
+	struct io_buffer *iobuf;
+	struct xfer_metadata meta = {
+		.dest = ( struct sockaddr * ) &tftp->peer,
+	};
+	size_t msglen;
+
+	DBGC2 ( tftp, "TFTP %p sending ERROR %d: %s\n", tftp, errcode,
+		errmsg );
+
+	/* Allocate buffer */
+	msglen = sizeof ( *err ) + strlen ( errmsg ) + 1 /* NUL */;
+	iobuf = xfer_alloc_iob ( &tftp->socket, msglen );
+	if ( ! iobuf )
+		return -ENOMEM;
+
+	/* Build ERROR */
+	err = iob_put ( iobuf, msglen );
+	err->opcode = htons ( TFTP_ERROR );
+	err->errcode = htons ( errcode );
+	strcpy ( err->errmsg, errmsg );
+
+	/* ERR always goes to the peer recorded from the RRQ response */
+	return xfer_deliver_iob_meta ( &tftp->socket, iobuf, &meta );
+}
+
+/**
  * Transmit next relevant packet
  *
  * @v tftp		TFTP connection
@@ -416,9 +456,16 @@ static int tftp_send_ack ( struct tftp_request *tftp ) {
  */
 static int tftp_send_packet ( struct tftp_request *tftp ) {
 
-	/* Update retransmission timer */
+	/* Update retransmission timer.  While name resolution takes place the
+	 * window is zero.  Avoid unnecessary delay after name resolution
+	 * completes by retrying immediately.
+	 */
 	stop_timer ( &tftp->timer );
-	start_timer ( &tftp->timer );
+	if ( xfer_window ( &tftp->socket ) ) {
+		start_timer ( &tftp->timer );
+	} else {
+		start_timer_nodelay ( &tftp->timer );
+	}
 
 	/* Send RRQ or ACK as appropriate */
 	if ( ! tftp->peer.st_family ) {
@@ -670,6 +717,7 @@ static int tftp_rx_oack ( struct tftp_request *tftp, void *buf, size_t len ) {
 	char *end = buf + len;
 	char *name;
 	char *value;
+	char *next;
 	int rc = 0;
 
 	/* Sanity check */
@@ -679,26 +727,41 @@ static int tftp_rx_oack ( struct tftp_request *tftp, void *buf, size_t len ) {
 		rc = -EINVAL;
 		goto done;
 	}
-	if ( end[-1] != '\0' ) {
-		DBGC ( tftp, "TFTP %p received OACK missing final NUL\n",
-		       tftp );
-		rc = -EINVAL;
-		goto done;
-	}
 
 	/* Process each option in turn */
-	name = oack->data;
-	while ( name < end ) {
-		value = ( name + strlen ( name ) + 1 );
+	for ( name = oack->data ; name < end ; name = next ) {
+
+		/* Parse option name and value
+		 *
+		 * We treat parsing errors as non-fatal, because there
+		 * exists at least one TFTP server (IBM Tivoli PXE
+		 * Server 5.1.0.3) that has been observed to send
+		 * malformed OACKs containing trailing garbage bytes.
+		 */
+		value = ( name + strnlen ( name, ( end - name ) ) + 1 );
+		if ( value > end ) {
+			DBGC ( tftp, "TFTP %p received OACK with malformed "
+			       "option name:\n", tftp );
+			DBGC_HD ( tftp, oack, len );
+			break;
+		}
 		if ( value == end ) {
 			DBGC ( tftp, "TFTP %p received OACK missing value "
 			       "for option \"%s\"\n", tftp, name );
-			rc = -EINVAL;
-			goto done;
+			DBGC_HD ( tftp, oack, len );
+			break;
 		}
+		next = ( value + strnlen ( value, ( end - value ) ) + 1 );
+		if ( next > end ) {
+			DBGC ( tftp, "TFTP %p received OACK with malformed "
+			       "value for option \"%s\":\n", tftp, name );
+			DBGC_HD ( tftp, oack, len );
+			break;
+		}
+
+		/* Process option */
 		if ( ( rc = tftp_process_option ( tftp, name, value ) ) != 0 )
 			goto done;
-		name = ( value + strlen ( value ) + 1 );
 	}
 
 	/* Process tsize information, if available */
@@ -707,6 +770,14 @@ static int tftp_rx_oack ( struct tftp_request *tftp, void *buf, size_t len ) {
 			goto done;
 	}
 
+	/* Abort request if only trying to determine file size */
+	if ( tftp->flags & TFTP_FL_SIZEONLY ) {
+		rc = 0;
+		tftp_send_error ( tftp, 0, "TFTP Aborted" );
+		tftp_done ( tftp, rc );
+		return rc;
+	}
+
 	/* Request next data block */
 	tftp_send_packet ( tftp );
 
@@ -729,11 +800,18 @@ static int tftp_rx_data ( struct tftp_request *tftp,
 			  struct io_buffer *iobuf ) {
 	struct tftp_data *data = iobuf->data;
 	struct xfer_metadata meta;
-	int block;
+	unsigned int block;
 	off_t offset;
 	size_t data_len;
 	int rc;
 
+	if ( tftp->flags & TFTP_FL_SIZEONLY ) {
+		/* If we get here then server doesn't support SIZE option */
+		rc = -ENOTSUP;
+		tftp_send_error ( tftp, 0, "TFTP Aborted" );
+		goto done;
+	}
+
 	/* Sanity check */
 	if ( iob_len ( iobuf ) < sizeof ( *data ) ) {
 		DBGC ( tftp, "TFTP %p received underlength DATA packet "
@@ -741,14 +819,17 @@ static int tftp_rx_data ( struct tftp_request *tftp,
 		rc = -EINVAL;
 		goto done;
 	}
-	if ( data->block == 0 ) {
+
+	/* Calculate block number */
+	block = ( ( bitmap_first_gap ( &tftp->bitmap ) + 1 ) & ~0xffff );
+	if ( data->block == 0 && block == 0 ) {
 		DBGC ( tftp, "TFTP %p received data block 0\n", tftp );
 		rc = -EINVAL;
 		goto done;
 	}
+	block += ( ntohs ( data->block ) - 1 );
 
 	/* Extract data */
-	block = ( ntohs ( data->block ) - 1 );
 	offset = ( block * tftp->blksize );
 	iob_pull ( iobuf, sizeof ( *data ) );
 	data_len = iob_len ( iobuf );
@@ -934,7 +1015,7 @@ static int tftp_socket_deliver_iob ( struct xfer_interface *socket,
 /** TFTP socket operations */
 static struct xfer_interface_operations tftp_socket_operations = {
 	.close		= ignore_xfer_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= tftp_socket_deliver_iob,
@@ -961,7 +1042,7 @@ static int tftp_mc_socket_deliver_iob ( struct xfer_interface *mc_socket,
 /** TFTP multicast socket operations */
 static struct xfer_interface_operations tftp_mc_socket_operations = {
 	.close		= ignore_xfer_close,
-	.vredirect	= xfer_vopen,
+	.vredirect	= xfer_vreopen,
 	.window		= unlimited_xfer_window,
 	.alloc_iob	= default_xfer_alloc_iob,
 	.deliver_iob	= tftp_mc_socket_deliver_iob,
@@ -1093,6 +1174,26 @@ struct uri_opener tftp_uri_opener __uri_opener = {
 };
 
 /**
+ * Initiate TFTP-size request
+ *
+ * @v xfer		Data transfer interface
+ * @v uri		Uniform Resource Identifier
+ * @ret rc		Return status code
+ */
+static int tftpsize_open ( struct xfer_interface *xfer, struct uri *uri ) {
+	return tftp_core_open ( xfer, uri, TFTP_PORT, NULL,
+				( TFTP_FL_RRQ_SIZES |
+				  TFTP_FL_SIZEONLY ) );
+
+}
+
+/** TFTP URI opener */
+struct uri_opener tftpsize_uri_opener __uri_opener = {
+	.scheme	= "tftpsize",
+	.open	= tftpsize_open,
+};
+
+/**
  * Initiate TFTM download
  *
  * @v xfer		Data transfer interface