[IA64-SGI] SGI Altix cross partition functionality [3/3]
authorDean Nelson <dcn@sgi.com>
Thu, 24 Mar 2005 03:50:00 +0000 (20:50 -0700)
committerTony Luck <tony.luck@intel.com>
Tue, 3 May 2005 19:37:19 +0000 (12:37 -0700)
This patch contains the cross partition pseudo-ethernet driver (XPNET)
functional support module.

Signed-off-by: Dean Nelson <dcn@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
arch/ia64/sn/kernel/Makefile
arch/ia64/sn/kernel/xpnet.c [new file with mode: 0644]

index 6959736eadea6370eef7ecb1f4263419f557ef80..4351c4ff984582470403bfe4333e8f693be5dde2 100644 (file)
@@ -15,3 +15,4 @@ obj-$(CONFIG_IA64_SGI_SN_XP)  += xp.o
 xp-y                           := xp_main.o xp_nofault.o
 obj-$(CONFIG_IA64_SGI_SN_XP)   += xpc.o
 xpc-y                          := xpc_main.o xpc_channel.o xpc_partition.o
+obj-$(CONFIG_IA64_SGI_SN_XP)   += xpnet.o
diff --git a/arch/ia64/sn/kernel/xpnet.c b/arch/ia64/sn/kernel/xpnet.c
new file mode 100644 (file)
index 0000000..78c13d6
--- /dev/null
@@ -0,0 +1,715 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 1999,2001-2005 Silicon Graphics, Inc. All rights reserved.
+ */
+
+
+/*
+ * Cross Partition Network Interface (XPNET) support
+ *
+ *     XPNET provides a virtual network layered on top of the Cross
+ *     Partition communication layer.
+ *
+ *     XPNET provides direct point-to-point and broadcast-like support
+ *     for an ethernet-like device.  The ethernet broadcast medium is
+ *     replaced with a point-to-point message structure which passes
+ *     pointers to a DMA-capable block that a remote partition should
+ *     retrieve and pass to the upper level networking layer.
+ *
+ */
+
+
+#include <linux/config.h>
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/pci.h>
+#include <linux/init.h>
+#include <linux/ioport.h>
+#include <linux/netdevice.h>
+#include <linux/etherdevice.h>
+#include <linux/delay.h>
+#include <linux/ethtool.h>
+#include <linux/mii.h>
+#include <linux/smp.h>
+#include <linux/string.h>
+#include <asm/sn/bte.h>
+#include <asm/sn/io.h>
+#include <asm/sn/sn_sal.h>
+#include <asm/types.h>
+#include <asm/atomic.h>
+#include <asm/sn/xp.h>
+
+
+/*
+ * The message payload transferred by XPC.
+ *
+ * buf_pa is the physical address where the DMA should pull from.
+ *
+ * NOTE: for performance reasons, buf_pa should _ALWAYS_ begin on a
+ * cacheline boundary.  To accomplish this, we record the number of
+ * bytes from the beginning of the first cacheline to the first useful
+ * byte of the skb (leadin_ignore) and the number of bytes from the
+ * last useful byte of the skb to the end of the last cacheline
+ * (tailout_ignore).
+ *
+ * size is the number of bytes to transfer which includes the skb->len
+ * (useful bytes of the senders skb) plus the leadin and tailout
+ */
+struct xpnet_message {
+       u16 version;            /* Version for this message */
+       u16 embedded_bytes;     /* #of bytes embedded in XPC message */
+       u32 magic;              /* Special number indicating this is xpnet */
+       u64 buf_pa;             /* phys address of buffer to retrieve */
+       u32 size;               /* #of bytes in buffer */
+       u8 leadin_ignore;       /* #of bytes to ignore at the beginning */
+       u8 tailout_ignore;      /* #of bytes to ignore at the end */
+       unsigned char data;     /* body of small packets */
+};
+
+/*
+ * Determine the size of our message, the cacheline aligned size,
+ * and then the number of message will request from XPC.
+ *
+ * XPC expects each message to exist in an individual cacheline.
+ */
+#define XPNET_MSG_SIZE         (L1_CACHE_BYTES - XPC_MSG_PAYLOAD_OFFSET)
+#define XPNET_MSG_DATA_MAX     \
+               (XPNET_MSG_SIZE - (u64)(&((struct xpnet_message *)0)->data))
+#define XPNET_MSG_ALIGNED_SIZE (L1_CACHE_ALIGN(XPNET_MSG_SIZE))
+#define XPNET_MSG_NENTRIES     (PAGE_SIZE / XPNET_MSG_ALIGNED_SIZE)
+
+
+#define XPNET_MAX_KTHREADS     (XPNET_MSG_NENTRIES + 1)
+#define XPNET_MAX_IDLE_KTHREADS        (XPNET_MSG_NENTRIES + 1)
+
+/*
+ * Version number of XPNET implementation. XPNET can always talk to versions
+ * with same major #, and never talk to versions with a different version.
+ */
+#define _XPNET_VERSION(_major, _minor) (((_major) << 4) | (_minor))
+#define XPNET_VERSION_MAJOR(_v)                ((_v) >> 4)
+#define XPNET_VERSION_MINOR(_v)                ((_v) & 0xf)
+
+#define        XPNET_VERSION _XPNET_VERSION(1,0)               /* version 1.0 */
+#define        XPNET_VERSION_EMBED _XPNET_VERSION(1,1)         /* version 1.1 */
+#define XPNET_MAGIC    0x88786984 /* "XNET" */
+
+#define XPNET_VALID_MSG(_m)                                                 \
+   ((XPNET_VERSION_MAJOR(_m->version) == XPNET_VERSION_MAJOR(XPNET_VERSION)) \
+    && (msg->magic == XPNET_MAGIC))
+
+#define XPNET_DEVICE_NAME              "xp0"
+
+
+/*
+ * When messages are queued with xpc_send_notify, a kmalloc'd buffer
+ * of the following type is passed as a notification cookie.  When the
+ * notification function is called, we use the cookie to decide
+ * whether all outstanding message sends have completed.  The skb can
+ * then be released.
+ */
+struct xpnet_pending_msg {
+       struct list_head free_list;
+       struct sk_buff *skb;
+       atomic_t use_count;
+};
+
+/* driver specific structure pointed to by the device structure */
+struct xpnet_dev_private {
+       struct net_device_stats stats;
+};
+
+struct net_device *xpnet_device;
+
+/*
+ * When we are notified of other partitions activating, we add them to
+ * our bitmask of partitions to which we broadcast.
+ */
+static u64 xpnet_broadcast_partitions;
+/* protect above */
+static spinlock_t xpnet_broadcast_lock = SPIN_LOCK_UNLOCKED;
+
+/*
+ * Since the Block Transfer Engine (BTE) is being used for the transfer
+ * and it relies upon cache-line size transfers, we need to reserve at
+ * least one cache-line for head and tail alignment.  The BTE is
+ * limited to 8MB transfers.
+ *
+ * Testing has shown that changing MTU to greater than 64KB has no effect
+ * on TCP as the two sides negotiate a Max Segment Size that is limited
+ * to 64K.  Other protocols May use packets greater than this, but for
+ * now, the default is 64KB.
+ */
+#define XPNET_MAX_MTU (0x800000UL - L1_CACHE_BYTES)
+/* 32KB has been determined to be the ideal */
+#define XPNET_DEF_MTU (0x8000UL)
+
+
+/*
+ * The partition id is encapsulated in the MAC address.  The following
+ * define locates the octet the partid is in.
+ */
+#define XPNET_PARTID_OCTET     1
+#define XPNET_LICENSE_OCTET    2
+
+
+/*
+ * Define the XPNET debug device structure that is to be used with dev_dbg(),
+ * dev_err(), dev_warn(), and dev_info().
+ */
+struct device_driver xpnet_dbg_name = {
+       .name = "xpnet"
+};
+
+struct device xpnet_dbg_subname = {
+       .bus_id = {0},                  /* set to "" */
+       .driver = &xpnet_dbg_name
+};
+
+struct device *xpnet = &xpnet_dbg_subname;
+
+/*
+ * Packet was recevied by XPC and forwarded to us.
+ */
+static void
+xpnet_receive(partid_t partid, int channel, struct xpnet_message *msg)
+{
+       struct sk_buff *skb;
+       bte_result_t bret;
+       struct xpnet_dev_private *priv =
+               (struct xpnet_dev_private *) xpnet_device->priv;
+
+
+       if (!XPNET_VALID_MSG(msg)) {
+               /*
+                * Packet with a different XPC version.  Ignore.
+                */
+               xpc_received(partid, channel, (void *) msg);
+
+               priv->stats.rx_errors++;
+
+               return;
+       }
+       dev_dbg(xpnet, "received 0x%lx, %d, %d, %d\n", msg->buf_pa, msg->size,
+               msg->leadin_ignore, msg->tailout_ignore);
+
+
+       /* reserve an extra cache line */
+       skb = dev_alloc_skb(msg->size + L1_CACHE_BYTES);
+       if (!skb) {
+               dev_err(xpnet, "failed on dev_alloc_skb(%d)\n",
+                       msg->size + L1_CACHE_BYTES);
+
+               xpc_received(partid, channel, (void *) msg);
+
+               priv->stats.rx_errors++;
+
+               return;
+       }
+
+       /*
+        * The allocated skb has some reserved space.
+        * In order to use bte_copy, we need to get the
+        * skb->data pointer moved forward.
+        */
+       skb_reserve(skb, (L1_CACHE_BYTES - ((u64)skb->data &
+                                           (L1_CACHE_BYTES - 1)) +
+                         msg->leadin_ignore));
+
+       /*
+        * Update the tail pointer to indicate data actually
+        * transferred.
+        */
+       skb_put(skb, (msg->size - msg->leadin_ignore - msg->tailout_ignore));
+
+       /*
+        * Move the data over from the the other side.
+        */
+       if ((XPNET_VERSION_MINOR(msg->version) == 1) &&
+                                               (msg->embedded_bytes != 0)) {
+               dev_dbg(xpnet, "copying embedded message. memcpy(0x%p, 0x%p, "
+                       "%lu)\n", skb->data, &msg->data,
+                       (size_t) msg->embedded_bytes);
+
+               memcpy(skb->data, &msg->data, (size_t) msg->embedded_bytes);
+       } else {
+               dev_dbg(xpnet, "transferring buffer to the skb->data area;\n\t"
+                       "bte_copy(0x%p, 0x%p, %hu)\n", (void *)msg->buf_pa,
+                       (void *)__pa((u64)skb->data & ~(L1_CACHE_BYTES - 1)),
+                       msg->size);
+
+               bret = bte_copy(msg->buf_pa,
+                               __pa((u64)skb->data & ~(L1_CACHE_BYTES - 1)),
+                               msg->size, (BTE_NOTIFY | BTE_WACQUIRE), NULL);
+
+               if (bret != BTE_SUCCESS) {
+                       // >>> Need better way of cleaning skb.  Currently skb
+                       // >>> appears in_use and we can't just call
+                       // >>> dev_kfree_skb.
+                       dev_err(xpnet, "bte_copy(0x%p, 0x%p, 0x%hx) returned "
+                               "error=0x%x\n", (void *)msg->buf_pa,
+                               (void *)__pa((u64)skb->data &
+                                                       ~(L1_CACHE_BYTES - 1)),
+                               msg->size, bret);
+
+                       xpc_received(partid, channel, (void *) msg);
+
+                       priv->stats.rx_errors++;
+
+                       return;
+               }
+       }
+
+       dev_dbg(xpnet, "<skb->head=0x%p skb->data=0x%p skb->tail=0x%p "
+               "skb->end=0x%p skb->len=%d\n", (void *) skb->head,
+               (void *) skb->data, (void *) skb->tail, (void *) skb->end,
+               skb->len);
+
+       skb->dev = xpnet_device;
+       skb->protocol = eth_type_trans(skb, xpnet_device);
+       skb->ip_summed = CHECKSUM_UNNECESSARY;
+
+       dev_dbg(xpnet, "passing skb to network layer; \n\tskb->head=0x%p "
+               "skb->data=0x%p skb->tail=0x%p skb->end=0x%p skb->len=%d\n",
+               (void *) skb->head, (void *) skb->data, (void *) skb->tail,
+               (void *) skb->end, skb->len);
+
+
+       xpnet_device->last_rx = jiffies;
+       priv->stats.rx_packets++;
+       priv->stats.rx_bytes += skb->len + ETH_HLEN;
+
+       netif_rx_ni(skb);
+       xpc_received(partid, channel, (void *) msg);
+}
+
+
+/*
+ * This is the handler which XPC calls during any sort of change in
+ * state or message reception on a connection.
+ */
+static void
+xpnet_connection_activity(enum xpc_retval reason, partid_t partid, int channel,
+                         void *data, void *key)
+{
+       long bp;
+
+
+       DBUG_ON(partid <= 0 || partid >= XP_MAX_PARTITIONS);
+       DBUG_ON(channel != XPC_NET_CHANNEL);
+
+       switch(reason) {
+       case xpcMsgReceived:    /* message received */
+               DBUG_ON(data == NULL);
+
+               xpnet_receive(partid, channel, (struct xpnet_message *) data);
+               break;
+
+       case xpcConnected:      /* connection completed to a partition */
+               spin_lock_bh(&xpnet_broadcast_lock);
+               xpnet_broadcast_partitions |= 1UL << (partid -1 );
+               bp = xpnet_broadcast_partitions;
+               spin_unlock_bh(&xpnet_broadcast_lock);
+
+               netif_carrier_on(xpnet_device);
+
+               dev_dbg(xpnet, "%s connection created to partition %d; "
+                       "xpnet_broadcast_partitions=0x%lx\n",
+                       xpnet_device->name, partid, bp);
+               break;
+
+       default:
+               spin_lock_bh(&xpnet_broadcast_lock);
+               xpnet_broadcast_partitions &= ~(1UL << (partid -1 ));
+               bp = xpnet_broadcast_partitions;
+               spin_unlock_bh(&xpnet_broadcast_lock);
+
+               if (bp == 0) {
+                       netif_carrier_off(xpnet_device);
+               }
+
+               dev_dbg(xpnet, "%s disconnected from partition %d; "
+                       "xpnet_broadcast_partitions=0x%lx\n",
+                       xpnet_device->name, partid, bp);
+               break;
+
+       }
+}
+
+
+static int
+xpnet_dev_open(struct net_device *dev)
+{
+       enum xpc_retval ret;
+
+
+       dev_dbg(xpnet, "calling xpc_connect(%d, 0x%p, NULL, %ld, %ld, %d, "
+               "%d)\n", XPC_NET_CHANNEL, xpnet_connection_activity,
+               XPNET_MSG_SIZE, XPNET_MSG_NENTRIES, XPNET_MAX_KTHREADS,
+               XPNET_MAX_IDLE_KTHREADS);
+
+       ret = xpc_connect(XPC_NET_CHANNEL, xpnet_connection_activity, NULL,
+                         XPNET_MSG_SIZE, XPNET_MSG_NENTRIES,
+                         XPNET_MAX_KTHREADS, XPNET_MAX_IDLE_KTHREADS);
+       if (ret != xpcSuccess) {
+               dev_err(xpnet, "ifconfig up of %s failed on XPC connect, "
+                       "ret=%d\n", dev->name, ret);
+
+               return -ENOMEM;
+       }
+
+       dev_dbg(xpnet, "ifconfig up of %s; XPC connected\n", dev->name);
+
+       return 0;
+}
+
+
+static int
+xpnet_dev_stop(struct net_device *dev)
+{
+       xpc_disconnect(XPC_NET_CHANNEL);
+
+       dev_dbg(xpnet, "ifconfig down of %s; XPC disconnected\n", dev->name);
+
+       return 0;
+}
+
+
+static int
+xpnet_dev_change_mtu(struct net_device *dev, int new_mtu)
+{
+       /* 68 comes from min TCP+IP+MAC header */
+       if ((new_mtu < 68) || (new_mtu > XPNET_MAX_MTU)) {
+               dev_err(xpnet, "ifconfig %s mtu %d failed; value must be "
+                       "between 68 and %ld\n", dev->name, new_mtu,
+                       XPNET_MAX_MTU);
+               return -EINVAL;
+       }
+
+       dev->mtu = new_mtu;
+       dev_dbg(xpnet, "ifconfig %s mtu set to %d\n", dev->name, new_mtu);
+       return 0;
+}
+
+
+/*
+ * Required for the net_device structure.
+ */
+static int
+xpnet_dev_set_config(struct net_device *dev, struct ifmap *new_map)
+{
+       return 0;
+}
+
+
+/*
+ * Return statistics to the caller.
+ */
+static struct net_device_stats *
+xpnet_dev_get_stats(struct net_device *dev)
+{
+       struct xpnet_dev_private *priv;
+
+
+       priv = (struct xpnet_dev_private *) dev->priv;
+
+       return &priv->stats;
+}
+
+
+/*
+ * Notification that the other end has received the message and
+ * DMA'd the skb information.  At this point, they are done with
+ * our side.  When all recipients are done processing, we
+ * release the skb and then release our pending message structure.
+ */
+static void
+xpnet_send_completed(enum xpc_retval reason, partid_t partid, int channel,
+                       void *__qm)
+{
+       struct xpnet_pending_msg *queued_msg =
+               (struct xpnet_pending_msg *) __qm;
+
+
+       DBUG_ON(queued_msg == NULL);
+
+       dev_dbg(xpnet, "message to %d notified with reason %d\n",
+               partid, reason);
+
+       if (atomic_dec_return(&queued_msg->use_count) == 0) {
+               dev_dbg(xpnet, "all acks for skb->head=-x%p\n",
+                       (void *) queued_msg->skb->head);
+
+               dev_kfree_skb_any(queued_msg->skb);
+               kfree(queued_msg);
+       }
+}
+
+
+/*
+ * Network layer has formatted a packet (skb) and is ready to place it
+ * "on the wire".  Prepare and send an xpnet_message to all partitions
+ * which have connected with us and are targets of this packet.
+ *
+ * MAC-NOTE:  For the XPNET driver, the MAC address contains the
+ * destination partition_id.  If the destination partition id word
+ * is 0xff, this packet is to broadcast to all partitions.
+ */
+static int
+xpnet_dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev)
+{
+       struct xpnet_pending_msg *queued_msg;
+       enum xpc_retval ret;
+       struct xpnet_message *msg;
+       u64 start_addr, end_addr;
+       long dp;
+       u8 second_mac_octet;
+       partid_t dest_partid;
+       struct xpnet_dev_private *priv;
+       u16 embedded_bytes;
+
+
+       priv = (struct xpnet_dev_private *) dev->priv;
+
+
+       dev_dbg(xpnet, ">skb->head=0x%p skb->data=0x%p skb->tail=0x%p "
+               "skb->end=0x%p skb->len=%d\n", (void *) skb->head,
+               (void *) skb->data, (void *) skb->tail, (void *) skb->end,
+               skb->len);
+
+
+       /*
+        * The xpnet_pending_msg tracks how many outstanding
+        * xpc_send_notifies are relying on this skb.  When none
+        * remain, release the skb.
+        */
+       queued_msg = kmalloc(sizeof(struct xpnet_pending_msg), GFP_ATOMIC);
+       if (queued_msg == NULL) {
+               dev_warn(xpnet, "failed to kmalloc %ld bytes; dropping "
+                       "packet\n", sizeof(struct xpnet_pending_msg));
+
+               priv->stats.tx_errors++;
+
+               return -ENOMEM;
+       }
+
+
+       /* get the beginning of the first cacheline and end of last */
+       start_addr = ((u64) skb->data & ~(L1_CACHE_BYTES - 1));
+       end_addr = L1_CACHE_ALIGN((u64) skb->tail);
+
+       /* calculate how many bytes to embed in the XPC message */
+       embedded_bytes = 0;
+       if (unlikely(skb->len <= XPNET_MSG_DATA_MAX)) {
+               /* skb->data does fit so embed */
+               embedded_bytes = skb->len;
+       }
+
+
+       /*
+        * Since the send occurs asynchronously, we set the count to one
+        * and begin sending.  Any sends that happen to complete before
+        * we are done sending will not free the skb.  We will be left
+        * with that task during exit.  This also handles the case of
+        * a packet destined for a partition which is no longer up.
+        */
+       atomic_set(&queued_msg->use_count, 1);
+       queued_msg->skb = skb;
+
+
+       second_mac_octet = skb->data[XPNET_PARTID_OCTET];
+       if (second_mac_octet == 0xff) {
+               /* we are being asked to broadcast to all partitions */
+               dp = xpnet_broadcast_partitions;
+       } else if (second_mac_octet != 0) {
+               dp = xpnet_broadcast_partitions &
+                                       (1UL << (second_mac_octet - 1));
+       } else {
+               /* 0 is an invalid partid.  Ignore */
+               dp = 0;
+       }
+       dev_dbg(xpnet, "destination Partitions mask (dp) = 0x%lx\n", dp);
+
+       /*
+        * If we wanted to allow promiscous mode to work like an
+        * unswitched network, this would be a good point to OR in a
+        * mask of partitions which should be receiving all packets.
+        */
+
+       /*
+        * Main send loop.
+        */
+       for (dest_partid = 1; dp && dest_partid < XP_MAX_PARTITIONS;
+            dest_partid++) {
+
+
+               if (!(dp & (1UL << (dest_partid - 1)))) {
+                       /* not destined for this partition */
+                       continue;
+               }
+
+               /* remove this partition from the destinations mask */
+               dp &= ~(1UL << (dest_partid - 1));
+
+
+               /* found a partition to send to */
+
+               ret = xpc_allocate(dest_partid, XPC_NET_CHANNEL,
+                                  XPC_NOWAIT, (void **)&msg);
+               if (unlikely(ret != xpcSuccess)) {
+                       continue;
+               }
+
+               msg->embedded_bytes = embedded_bytes;
+               if (unlikely(embedded_bytes != 0)) {
+                       msg->version = XPNET_VERSION_EMBED;
+                       dev_dbg(xpnet, "calling memcpy(0x%p, 0x%p, 0x%lx)\n",
+                               &msg->data, skb->data, (size_t) embedded_bytes);
+                       memcpy(&msg->data, skb->data, (size_t) embedded_bytes);
+               } else {
+                       msg->version = XPNET_VERSION;
+               }
+               msg->magic = XPNET_MAGIC;
+               msg->size = end_addr - start_addr;
+               msg->leadin_ignore = (u64) skb->data - start_addr;
+               msg->tailout_ignore = end_addr - (u64) skb->tail;
+               msg->buf_pa = __pa(start_addr);
+
+               dev_dbg(xpnet, "sending XPC message to %d:%d\nmsg->buf_pa="
+                       "0x%lx, msg->size=%u, msg->leadin_ignore=%u, "
+                       "msg->tailout_ignore=%u\n", dest_partid,
+                       XPC_NET_CHANNEL, msg->buf_pa, msg->size,
+                       msg->leadin_ignore, msg->tailout_ignore);
+
+
+               atomic_inc(&queued_msg->use_count);
+
+               ret = xpc_send_notify(dest_partid, XPC_NET_CHANNEL, msg,
+                                     xpnet_send_completed, queued_msg);
+               if (unlikely(ret != xpcSuccess)) {
+                       atomic_dec(&queued_msg->use_count);
+                       continue;
+               }
+
+       }
+
+       if (atomic_dec_return(&queued_msg->use_count) == 0) {
+               dev_dbg(xpnet, "no partitions to receive packet destined for "
+                       "%d\n", dest_partid);
+
+
+               dev_kfree_skb(skb);
+               kfree(queued_msg);
+       }
+
+       priv->stats.tx_packets++;
+       priv->stats.tx_bytes += skb->len;
+
+       return 0;
+}
+
+
+/*
+ * Deal with transmit timeouts coming from the network layer.
+ */
+static void
+xpnet_dev_tx_timeout (struct net_device *dev)
+{
+       struct xpnet_dev_private *priv;
+
+
+       priv = (struct xpnet_dev_private *) dev->priv;
+
+       priv->stats.tx_errors++;
+       return;
+}
+
+
+static int __init
+xpnet_init(void)
+{
+       int i;
+       u32 license_num;
+       int result = -ENOMEM;
+
+
+       dev_info(xpnet, "registering network device %s\n", XPNET_DEVICE_NAME);
+
+       /*
+        * use ether_setup() to init the majority of our device
+        * structure and then override the necessary pieces.
+        */
+       xpnet_device = alloc_netdev(sizeof(struct xpnet_dev_private),
+                                   XPNET_DEVICE_NAME, ether_setup);
+       if (xpnet_device == NULL) {
+               return -ENOMEM;
+       }
+
+       netif_carrier_off(xpnet_device);
+
+       xpnet_device->mtu = XPNET_DEF_MTU;
+       xpnet_device->change_mtu = xpnet_dev_change_mtu;
+       xpnet_device->open = xpnet_dev_open;
+       xpnet_device->get_stats = xpnet_dev_get_stats;
+       xpnet_device->stop = xpnet_dev_stop;
+       xpnet_device->hard_start_xmit = xpnet_dev_hard_start_xmit;
+       xpnet_device->tx_timeout = xpnet_dev_tx_timeout;
+       xpnet_device->set_config = xpnet_dev_set_config;
+
+       /*
+        * Multicast assumes the LSB of the first octet is set for multicast
+        * MAC addresses.  We chose the first octet of the MAC to be unlikely
+        * to collide with any vendor's officially issued MAC.
+        */
+       xpnet_device->dev_addr[0] = 0xfe;
+       xpnet_device->dev_addr[XPNET_PARTID_OCTET] = sn_partition_id;
+       license_num = sn_partition_serial_number_val();
+       for (i = 3; i >= 0; i--) {
+               xpnet_device->dev_addr[XPNET_LICENSE_OCTET + i] =
+                                                       license_num & 0xff;
+               license_num = license_num >> 8;
+       }
+
+       /*
+        * ether_setup() sets this to a multicast device.  We are
+        * really not supporting multicast at this time.
+        */
+       xpnet_device->flags &= ~IFF_MULTICAST;
+
+       /*
+        * No need to checksum as it is a DMA transfer.  The BTE will
+        * report an error if the data is not retrievable and the
+        * packet will be dropped.
+        */
+       xpnet_device->features = NETIF_F_NO_CSUM;
+
+       result = register_netdev(xpnet_device);
+       if (result != 0) {
+               free_netdev(xpnet_device);
+       }
+
+       return result;
+}
+module_init(xpnet_init);
+
+
+static void __exit
+xpnet_exit(void)
+{
+       dev_info(xpnet, "unregistering network device %s\n",
+               xpnet_device[0].name);
+
+       unregister_netdev(xpnet_device);
+
+       free_netdev(xpnet_device);
+}
+module_exit(xpnet_exit);
+
+
+MODULE_AUTHOR("Silicon Graphics, Inc.");
+MODULE_DESCRIPTION("Cross Partition Network adapter (XPNET)");
+MODULE_LICENSE("GPL");
+