Initial import

4 files changed, 4732 insertions, 0 deletions

diff --git a/drivers/net/ethernet/tile/Kconfig b/drivers/net/ethernet/tile/Kconfig
new file mode 100644
index 000000000..f59a6c265
--- /dev/null
+++ b/drivers/net/ethernet/tile/Kconfig
@@ -0,0 +1,18 @@
+#
+# Tilera network device configuration
+#
+
+config TILE_NET
+	tristate "Tilera GBE/XGBE network driver support"
+	depends on TILE
+	default y
+	select CRC32
+	select TILE_GXIO_MPIPE if TILEGX
+	select HIGH_RES_TIMERS if TILEGX
+	select PTP_1588_CLOCK if TILEGX
+	---help---
+	  This is a standard Linux network device driver for the
+	  on-chip Tilera Gigabit Ethernet and XAUI interfaces.
+
+	  To compile this driver as a module, choose M here: the module
+	  will be called tile_net.
diff --git a/drivers/net/ethernet/tile/Makefile b/drivers/net/ethernet/tile/Makefile
new file mode 100644
index 000000000..0ef9eefd3
--- /dev/null
+++ b/drivers/net/ethernet/tile/Makefile
@@ -0,0 +1,10 @@
+#
+# Makefile for the TILE on-chip networking support.
+#
+
+obj-$(CONFIG_TILE_NET) += tile_net.o
+ifdef CONFIG_TILEGX
+tile_net-y := tilegx.o
+else
+tile_net-y := tilepro.o
+endif
diff --git a/drivers/net/ethernet/tile/tilegx.c b/drivers/net/ethernet/tile/tilegx.c
new file mode 100644
index 000000000..a3f761000
--- /dev/null
+++ b/drivers/net/ethernet/tile/tilegx.c
@@ -0,0 +1,2281 @@
+/*
+ * Copyright 2012 Tilera Corporation. All Rights Reserved.
+ *
+ *   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, version 2.
+ *
+ *   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, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/moduleparam.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>      /* printk() */
+#include <linux/slab.h>        /* kmalloc() */
+#include <linux/errno.h>       /* error codes */
+#include <linux/types.h>       /* size_t */
+#include <linux/interrupt.h>
+#include <linux/in.h>
+#include <linux/irq.h>
+#include <linux/netdevice.h>   /* struct device, and other headers */
+#include <linux/etherdevice.h> /* eth_type_trans */
+#include <linux/skbuff.h>
+#include <linux/ioctl.h>
+#include <linux/cdev.h>
+#include <linux/hugetlb.h>
+#include <linux/in6.h>
+#include <linux/timer.h>
+#include <linux/hrtimer.h>
+#include <linux/ktime.h>
+#include <linux/io.h>
+#include <linux/ctype.h>
+#include <linux/ip.h>
+#include <linux/ipv6.h>
+#include <linux/tcp.h>
+#include <linux/net_tstamp.h>
+#include <linux/ptp_clock_kernel.h>
+
+#include <asm/checksum.h>
+#include <asm/homecache.h>
+#include <gxio/mpipe.h>
+#include <arch/sim.h>
+
+/* Default transmit lockup timeout period, in jiffies. */
+#define TILE_NET_TIMEOUT (5 * HZ)
+
+/* The maximum number of distinct channels (idesc.channel is 5 bits). */
+#define TILE_NET_CHANNELS 32
+
+/* Maximum number of idescs to handle per "poll". */
+#define TILE_NET_BATCH 128
+
+/* Maximum number of packets to handle per "poll". */
+#define TILE_NET_WEIGHT 64
+
+/* Number of entries in each iqueue. */
+#define IQUEUE_ENTRIES 512
+
+/* Number of entries in each equeue. */
+#define EQUEUE_ENTRIES 2048
+
+/* Total header bytes per equeue slot.  Must be big enough for 2 bytes
+ * of NET_IP_ALIGN alignment, plus 14 bytes (?) of L2 header, plus up to
+ * 60 bytes of actual TCP header.  We round up to align to cache lines.
+ */
+#define HEADER_BYTES 128
+
+/* Maximum completions per cpu per device (must be a power of two).
+ * ISSUE: What is the right number here?  If this is too small, then
+ * egress might block waiting for free space in a completions array.
+ * ISSUE: At the least, allocate these only for initialized echannels.
+ */
+#define TILE_NET_MAX_COMPS 64
+
+#define MAX_FRAGS (MAX_SKB_FRAGS + 1)
+
+/* The "kinds" of buffer stacks (small/large/jumbo). */
+#define MAX_KINDS 3
+
+/* Size of completions data to allocate.
+ * ISSUE: Probably more than needed since we don't use all the channels.
+ */
+#define COMPS_SIZE (TILE_NET_CHANNELS * sizeof(struct tile_net_comps))
+
+/* Size of NotifRing data to allocate. */
+#define NOTIF_RING_SIZE (IQUEUE_ENTRIES * sizeof(gxio_mpipe_idesc_t))
+
+/* Timeout to wake the per-device TX timer after we stop the queue.
+ * We don't want the timeout too short (adds overhead, and might end
+ * up causing stop/wake/stop/wake cycles) or too long (affects performance).
+ * For the 10 Gb NIC, 30 usec means roughly 30+ 1500-byte packets.
+ */
+#define TX_TIMER_DELAY_USEC 30
+
+/* Timeout to wake the per-cpu egress timer to free completions. */
+#define EGRESS_TIMER_DELAY_USEC 1000
+
+MODULE_AUTHOR("Tilera Corporation");
+MODULE_LICENSE("GPL");
+
+/* A "packet fragment" (a chunk of memory). */
+struct frag {
+	void *buf;
+	size_t length;
+};
+
+/* A single completion. */
+struct tile_net_comp {
+	/* The "complete_count" when the completion will be complete. */
+	s64 when;
+	/* The buffer to be freed when the completion is complete. */
+	struct sk_buff *skb;
+};
+
+/* The completions for a given cpu and echannel. */
+struct tile_net_comps {
+	/* The completions. */
+	struct tile_net_comp comp_queue[TILE_NET_MAX_COMPS];
+	/* The number of completions used. */
+	unsigned long comp_next;
+	/* The number of completions freed. */
+	unsigned long comp_last;
+};
+
+/* The transmit wake timer for a given cpu and echannel. */
+struct tile_net_tx_wake {
+	int tx_queue_idx;
+	struct hrtimer timer;
+	struct net_device *dev;
+};
+
+/* Info for a specific cpu. */
+struct tile_net_info {
+	/* Our cpu. */
+	int my_cpu;
+	/* A timer for handling egress completions. */
+	struct hrtimer egress_timer;
+	/* True if "egress_timer" is scheduled. */
+	bool egress_timer_scheduled;
+	struct info_mpipe {
+		/* Packet queue. */
+		gxio_mpipe_iqueue_t iqueue;
+		/* The NAPI struct. */
+		struct napi_struct napi;
+		/* Number of buffers (by kind) which must still be provided. */
+		unsigned int num_needed_buffers[MAX_KINDS];
+		/* instance id. */
+		int instance;
+		/* True if iqueue is valid. */
+		bool has_iqueue;
+		/* NAPI flags. */
+		bool napi_added;
+		bool napi_enabled;
+		/* Comps for each egress channel. */
+		struct tile_net_comps *comps_for_echannel[TILE_NET_CHANNELS];
+		/* Transmit wake timer for each egress channel. */
+		struct tile_net_tx_wake tx_wake[TILE_NET_CHANNELS];
+	} mpipe[NR_MPIPE_MAX];
+};
+
+/* Info for egress on a particular egress channel. */
+struct tile_net_egress {
+	/* The "equeue". */
+	gxio_mpipe_equeue_t *equeue;
+	/* The headers for TSO. */
+	unsigned char *headers;
+};
+
+/* Info for a specific device. */
+struct tile_net_priv {
+	/* Our network device. */
+	struct net_device *dev;
+	/* The primary link. */
+	gxio_mpipe_link_t link;
+	/* The primary channel, if open, else -1. */
+	int channel;
+	/* The "loopify" egress link, if needed. */
+	gxio_mpipe_link_t loopify_link;
+	/* The "loopify" egress channel, if open, else -1. */
+	int loopify_channel;
+	/* The egress channel (channel or loopify_channel). */
+	int echannel;
+	/* mPIPE instance, 0 or 1. */
+	int instance;
+	/* The timestamp config. */
+	struct hwtstamp_config stamp_cfg;
+};
+
+static struct mpipe_data {
+	/* The ingress irq. */
+	int ingress_irq;
+
+	/* The "context" for all devices. */
+	gxio_mpipe_context_t context;
+
+	/* Egress info, indexed by "priv->echannel"
+	 * (lazily created as needed).
+	 */
+	struct tile_net_egress
+	egress_for_echannel[TILE_NET_CHANNELS];
+
+	/* Devices currently associated with each channel.
+	 * NOTE: The array entry can become NULL after ifconfig down, but
+	 * we do not free the underlying net_device structures, so it is
+	 * safe to use a pointer after reading it from this array.
+	 */
+	struct net_device
+	*tile_net_devs_for_channel[TILE_NET_CHANNELS];
+
+	/* The actual memory allocated for the buffer stacks. */
+	void *buffer_stack_vas[MAX_KINDS];
+
+	/* The amount of memory allocated for each buffer stack. */
+	size_t buffer_stack_bytes[MAX_KINDS];
+
+	/* The first buffer stack index
+	 * (small = +0, large = +1, jumbo = +2).
+	 */
+	int first_buffer_stack;
+
+	/* The buckets. */
+	int first_bucket;
+	int num_buckets;
+
+	/* PTP-specific data. */
+	struct ptp_clock *ptp_clock;
+	struct ptp_clock_info caps;
+
+	/* Lock for ptp accessors. */
+	struct mutex ptp_lock;
+
+} mpipe_data[NR_MPIPE_MAX] = {
+	[0 ... (NR_MPIPE_MAX - 1)] {
+		.ingress_irq = -1,
+		.first_buffer_stack = -1,
+		.first_bucket = -1,
+		.num_buckets = 1
+	}
+};
+
+/* A mutex for "tile_net_devs_for_channel". */
+static DEFINE_MUTEX(tile_net_devs_for_channel_mutex);
+
+/* The per-cpu info. */
+static DEFINE_PER_CPU(struct tile_net_info, per_cpu_info);
+
+
+/* The buffer size enums for each buffer stack.
+ * See arch/tile/include/gxio/mpipe.h for the set of possible values.
+ * We avoid the "10384" size because it can induce "false chaining"
+ * on "cut-through" jumbo packets.
+ */
+static gxio_mpipe_buffer_size_enum_t buffer_size_enums[MAX_KINDS] = {
+	GXIO_MPIPE_BUFFER_SIZE_128,
+	GXIO_MPIPE_BUFFER_SIZE_1664,
+	GXIO_MPIPE_BUFFER_SIZE_16384
+};
+
+/* Text value of tile_net.cpus if passed as a module parameter. */
+static char *network_cpus_string;
+
+/* The actual cpus in "network_cpus". */
+static struct cpumask network_cpus_map;
+
+/* If "tile_net.loopify=LINK" was specified, this is "LINK". */
+static char *loopify_link_name;
+
+/* If "tile_net.custom" was specified, this is true. */
+static bool custom_flag;
+
+/* If "tile_net.jumbo=NUM" was specified, this is "NUM". */
+static uint jumbo_num;
+
+/* Obtain mpipe instance from struct tile_net_priv given struct net_device. */
+static inline int mpipe_instance(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	return priv->instance;
+}
+
+/* The "tile_net.cpus" argument specifies the cpus that are dedicated
+ * to handle ingress packets.
+ *
+ * The parameter should be in the form "tile_net.cpus=m-n[,x-y]", where
+ * m, n, x, y are integer numbers that represent the cpus that can be
+ * neither a dedicated cpu nor a dataplane cpu.
+ */
+static bool network_cpus_init(void)
+{
+	int rc;
+
+	if (network_cpus_string == NULL)
+		return false;
+
+	rc = cpulist_parse_crop(network_cpus_string, &network_cpus_map);
+	if (rc != 0) {
+		pr_warn("tile_net.cpus=%s: malformed cpu list\n",
+			network_cpus_string);
+		return false;
+	}
+
+	/* Remove dedicated cpus. */
+	cpumask_and(&network_cpus_map, &network_cpus_map, cpu_possible_mask);
+
+	if (cpumask_empty(&network_cpus_map)) {
+		pr_warn("Ignoring empty tile_net.cpus='%s'.\n",
+			network_cpus_string);
+		return false;
+	}
+
+	pr_info("Linux network CPUs: %*pbl\n",
+		cpumask_pr_args(&network_cpus_map));
+	return true;
+}
+
+module_param_named(cpus, network_cpus_string, charp, 0444);
+MODULE_PARM_DESC(cpus, "cpulist of cores that handle network interrupts");
+
+/* The "tile_net.loopify=LINK" argument causes the named device to
+ * actually use "loop0" for ingress, and "loop1" for egress.  This
+ * allows an app to sit between the actual link and linux, passing
+ * (some) packets along to linux, and forwarding (some) packets sent
+ * out by linux.
+ */
+module_param_named(loopify, loopify_link_name, charp, 0444);
+MODULE_PARM_DESC(loopify, "name the device to use loop0/1 for ingress/egress");
+
+/* The "tile_net.custom" argument causes us to ignore the "conventional"
+ * classifier metadata, in particular, the "l2_offset".
+ */
+module_param_named(custom, custom_flag, bool, 0444);
+MODULE_PARM_DESC(custom, "indicates a (heavily) customized classifier");
+
+/* The "tile_net.jumbo" argument causes us to support "jumbo" packets,
+ * and to allocate the given number of "jumbo" buffers.
+ */
+module_param_named(jumbo, jumbo_num, uint, 0444);
+MODULE_PARM_DESC(jumbo, "the number of buffers to support jumbo packets");
+
+/* Atomically update a statistics field.
+ * Note that on TILE-Gx, this operation is fire-and-forget on the
+ * issuing core (single-cycle dispatch) and takes only a few cycles
+ * longer than a regular store when the request reaches the home cache.
+ * No expensive bus management overhead is required.
+ */
+static void tile_net_stats_add(unsigned long value, unsigned long *field)
+{
+	BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(unsigned long));
+	atomic_long_add(value, (atomic_long_t *)field);
+}
+
+/* Allocate and push a buffer. */
+static bool tile_net_provide_buffer(int instance, int kind)
+{
+	struct mpipe_data *md = &mpipe_data[instance];
+	gxio_mpipe_buffer_size_enum_t bse = buffer_size_enums[kind];
+	size_t bs = gxio_mpipe_buffer_size_enum_to_buffer_size(bse);
+	const unsigned long buffer_alignment = 128;
+	struct sk_buff *skb;
+	int len;
+
+	len = sizeof(struct sk_buff **) + buffer_alignment + bs;
+	skb = dev_alloc_skb(len);
+	if (skb == NULL)
+		return false;
+
+	/* Make room for a back-pointer to 'skb' and guarantee alignment. */
+	skb_reserve(skb, sizeof(struct sk_buff **));
+	skb_reserve(skb, -(long)skb->data & (buffer_alignment - 1));
+
+	/* Save a back-pointer to 'skb'. */
+	*(struct sk_buff **)(skb->data - sizeof(struct sk_buff **)) = skb;
+
+	/* Make sure "skb" and the back-pointer have been flushed. */
+	wmb();
+
+	gxio_mpipe_push_buffer(&md->context, md->first_buffer_stack + kind,
+			       (void *)va_to_tile_io_addr(skb->data));
+
+	return true;
+}
+
+/* Convert a raw mpipe buffer to its matching skb pointer. */
+static struct sk_buff *mpipe_buf_to_skb(void *va)
+{
+	/* Acquire the associated "skb". */
+	struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+	struct sk_buff *skb = *skb_ptr;
+
+	/* Paranoia. */
+	if (skb->data != va) {
+		/* Panic here since there's a reasonable chance
+		 * that corrupt buffers means generic memory
+		 * corruption, with unpredictable system effects.
+		 */
+		panic("Corrupt linux buffer! va=%p, skb=%p, skb->data=%p",
+		      va, skb, skb->data);
+	}
+
+	return skb;
+}
+
+static void tile_net_pop_all_buffers(int instance, int stack)
+{
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	for (;;) {
+		tile_io_addr_t addr =
+			(tile_io_addr_t)gxio_mpipe_pop_buffer(&md->context,
+							      stack);
+		if (addr == 0)
+			break;
+		dev_kfree_skb_irq(mpipe_buf_to_skb(tile_io_addr_to_va(addr)));
+	}
+}
+
+/* Provide linux buffers to mPIPE. */
+static void tile_net_provide_needed_buffers(void)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	int instance, kind;
+	for (instance = 0; instance < NR_MPIPE_MAX &&
+		     info->mpipe[instance].has_iqueue; instance++)	{
+		for (kind = 0; kind < MAX_KINDS; kind++) {
+			while (info->mpipe[instance].num_needed_buffers[kind]
+			       != 0) {
+				if (!tile_net_provide_buffer(instance, kind)) {
+					pr_notice("Tile %d still needs"
+						  " some buffers\n",
+						  info->my_cpu);
+					return;
+				}
+				info->mpipe[instance].
+					num_needed_buffers[kind]--;
+			}
+		}
+	}
+}
+
+/* Get RX timestamp, and store it in the skb. */
+static void tile_rx_timestamp(struct tile_net_priv *priv, struct sk_buff *skb,
+			      gxio_mpipe_idesc_t *idesc)
+{
+	if (unlikely(priv->stamp_cfg.rx_filter != HWTSTAMP_FILTER_NONE)) {
+		struct skb_shared_hwtstamps *shhwtstamps = skb_hwtstamps(skb);
+		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
+		shhwtstamps->hwtstamp = ktime_set(idesc->time_stamp_sec,
+						  idesc->time_stamp_ns);
+	}
+}
+
+/* Get TX timestamp, and store it in the skb. */
+static void tile_tx_timestamp(struct sk_buff *skb, int instance)
+{
+	struct skb_shared_info *shtx = skb_shinfo(skb);
+	if (unlikely((shtx->tx_flags & SKBTX_HW_TSTAMP) != 0)) {
+		struct mpipe_data *md = &mpipe_data[instance];
+		struct skb_shared_hwtstamps shhwtstamps;
+		struct timespec ts;
+
+		shtx->tx_flags |= SKBTX_IN_PROGRESS;
+		gxio_mpipe_get_timestamp(&md->context, &ts);
+		memset(&shhwtstamps, 0, sizeof(shhwtstamps));
+		shhwtstamps.hwtstamp = ktime_set(ts.tv_sec, ts.tv_nsec);
+		skb_tstamp_tx(skb, &shhwtstamps);
+	}
+}
+
+/* Use ioctl() to enable or disable TX or RX timestamping. */
+static int tile_hwtstamp_set(struct net_device *dev, struct ifreq *rq)
+{
+	struct hwtstamp_config config;
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	if (copy_from_user(&config, rq->ifr_data, sizeof(config)))
+		return -EFAULT;
+
+	if (config.flags)  /* reserved for future extensions */
+		return -EINVAL;
+
+	switch (config.tx_type) {
+	case HWTSTAMP_TX_OFF:
+	case HWTSTAMP_TX_ON:
+		break;
+	default:
+		return -ERANGE;
+	}
+
+	switch (config.rx_filter) {
+	case HWTSTAMP_FILTER_NONE:
+		break;
+	case HWTSTAMP_FILTER_ALL:
+	case HWTSTAMP_FILTER_SOME:
+	case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
+	case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
+	case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
+	case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
+	case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
+	case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
+	case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
+	case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
+	case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
+	case HWTSTAMP_FILTER_PTP_V2_EVENT:
+	case HWTSTAMP_FILTER_PTP_V2_SYNC:
+	case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
+		config.rx_filter = HWTSTAMP_FILTER_ALL;
+		break;
+	default:
+		return -ERANGE;
+	}
+
+	if (copy_to_user(rq->ifr_data, &config, sizeof(config)))
+		return -EFAULT;
+
+	priv->stamp_cfg = config;
+	return 0;
+}
+
+static int tile_hwtstamp_get(struct net_device *dev, struct ifreq *rq)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	if (copy_to_user(rq->ifr_data, &priv->stamp_cfg,
+			 sizeof(priv->stamp_cfg)))
+		return -EFAULT;
+
+	return 0;
+}
+
+static inline bool filter_packet(struct net_device *dev, void *buf)
+{
+	/* Filter packets received before we're up. */
+	if (dev == NULL || !(dev->flags & IFF_UP))
+		return true;
+
+	/* Filter out packets that aren't for us. */
+	if (!(dev->flags & IFF_PROMISC) &&
+	    !is_multicast_ether_addr(buf) &&
+	    !ether_addr_equal(dev->dev_addr, buf))
+		return true;
+
+	return false;
+}
+
+static void tile_net_receive_skb(struct net_device *dev, struct sk_buff *skb,
+				 gxio_mpipe_idesc_t *idesc, unsigned long len)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int instance = priv->instance;
+
+	/* Encode the actual packet length. */
+	skb_put(skb, len);
+
+	skb->protocol = eth_type_trans(skb, dev);
+
+	/* Acknowledge "good" hardware checksums. */
+	if (idesc->cs && idesc->csum_seed_val == 0xFFFF)
+		skb->ip_summed = CHECKSUM_UNNECESSARY;
+
+	/* Get RX timestamp from idesc. */
+	tile_rx_timestamp(priv, skb, idesc);
+
+	napi_gro_receive(&info->mpipe[instance].napi, skb);
+
+	/* Update stats. */
+	tile_net_stats_add(1, &dev->stats.rx_packets);
+	tile_net_stats_add(len, &dev->stats.rx_bytes);
+
+	/* Need a new buffer. */
+	if (idesc->size == buffer_size_enums[0])
+		info->mpipe[instance].num_needed_buffers[0]++;
+	else if (idesc->size == buffer_size_enums[1])
+		info->mpipe[instance].num_needed_buffers[1]++;
+	else
+		info->mpipe[instance].num_needed_buffers[2]++;
+}
+
+/* Handle a packet.  Return true if "processed", false if "filtered". */
+static bool tile_net_handle_packet(int instance, gxio_mpipe_idesc_t *idesc)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	struct mpipe_data *md = &mpipe_data[instance];
+	struct net_device *dev = md->tile_net_devs_for_channel[idesc->channel];
+	uint8_t l2_offset;
+	void *va;
+	void *buf;
+	unsigned long len;
+	bool filter;
+
+	/* Drop packets for which no buffer was available (which can
+	 * happen under heavy load), or for which the me/tr/ce flags
+	 * are set (which can happen for jumbo cut-through packets,
+	 * or with a customized classifier).
+	 */
+	if (idesc->be || idesc->me || idesc->tr || idesc->ce) {
+		if (dev)
+			tile_net_stats_add(1, &dev->stats.rx_errors);
+		goto drop;
+	}
+
+	/* Get the "l2_offset", if allowed. */
+	l2_offset = custom_flag ? 0 : gxio_mpipe_idesc_get_l2_offset(idesc);
+
+	/* Get the VA (including NET_IP_ALIGN bytes of "headroom"). */
+	va = tile_io_addr_to_va((unsigned long)idesc->va);
+
+	/* Get the actual packet start/length. */
+	buf = va + l2_offset;
+	len = idesc->l2_size - l2_offset;
+
+	/* Point "va" at the raw buffer. */
+	va -= NET_IP_ALIGN;
+
+	filter = filter_packet(dev, buf);
+	if (filter) {
+		if (dev)
+			tile_net_stats_add(1, &dev->stats.rx_dropped);
+drop:
+		gxio_mpipe_iqueue_drop(&info->mpipe[instance].iqueue, idesc);
+	} else {
+		struct sk_buff *skb = mpipe_buf_to_skb(va);
+
+		/* Skip headroom, and any custom header. */
+		skb_reserve(skb, NET_IP_ALIGN + l2_offset);
+
+		tile_net_receive_skb(dev, skb, idesc, len);
+	}
+
+	gxio_mpipe_iqueue_consume(&info->mpipe[instance].iqueue, idesc);
+	return !filter;
+}
+
+/* Handle some packets for the current CPU.
+ *
+ * This function handles up to TILE_NET_BATCH idescs per call.
+ *
+ * ISSUE: Since we do not provide new buffers until this function is
+ * complete, we must initially provide enough buffers for each network
+ * cpu to fill its iqueue and also its batched idescs.
+ *
+ * ISSUE: The "rotting packet" race condition occurs if a packet
+ * arrives after the queue appears to be empty, and before the
+ * hypervisor interrupt is re-enabled.
+ */
+static int tile_net_poll(struct napi_struct *napi, int budget)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	unsigned int work = 0;
+	gxio_mpipe_idesc_t *idesc;
+	int instance, i, n;
+	struct mpipe_data *md;
+	struct info_mpipe *info_mpipe =
+		container_of(napi, struct info_mpipe, napi);
+
+	if (budget <= 0)
+		goto done;
+
+	instance = info_mpipe->instance;
+	while ((n = gxio_mpipe_iqueue_try_peek(
+			&info_mpipe->iqueue,
+			&idesc)) > 0) {
+		for (i = 0; i < n; i++) {
+			if (i == TILE_NET_BATCH)
+				goto done;
+			if (tile_net_handle_packet(instance,
+						   idesc + i)) {
+				if (++work >= budget)
+					goto done;
+			}
+		}
+	}
+
+	/* There are no packets left. */
+	napi_complete(&info_mpipe->napi);
+
+	md = &mpipe_data[instance];
+	/* Re-enable hypervisor interrupts. */
+	gxio_mpipe_enable_notif_ring_interrupt(
+		&md->context, info->mpipe[instance].iqueue.ring);
+
+	/* HACK: Avoid the "rotting packet" problem. */
+	if (gxio_mpipe_iqueue_try_peek(&info_mpipe->iqueue, &idesc) > 0)
+		napi_schedule(&info_mpipe->napi);
+
+	/* ISSUE: Handle completions? */
+
+done:
+	tile_net_provide_needed_buffers();
+
+	return work;
+}
+
+/* Handle an ingress interrupt from an instance on the current cpu. */
+static irqreturn_t tile_net_handle_ingress_irq(int irq, void *id)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	napi_schedule(&info->mpipe[(uint64_t)id].napi);
+	return IRQ_HANDLED;
+}
+
+/* Free some completions.  This must be called with interrupts blocked. */
+static int tile_net_free_comps(gxio_mpipe_equeue_t *equeue,
+				struct tile_net_comps *comps,
+				int limit, bool force_update)
+{
+	int n = 0;
+	while (comps->comp_last < comps->comp_next) {
+		unsigned int cid = comps->comp_last % TILE_NET_MAX_COMPS;
+		struct tile_net_comp *comp = &comps->comp_queue[cid];
+		if (!gxio_mpipe_equeue_is_complete(equeue, comp->when,
+						   force_update || n == 0))
+			break;
+		dev_kfree_skb_irq(comp->skb);
+		comps->comp_last++;
+		if (++n == limit)
+			break;
+	}
+	return n;
+}
+
+/* Add a completion.  This must be called with interrupts blocked.
+ * tile_net_equeue_try_reserve() will have ensured a free completion entry.
+ */
+static void add_comp(gxio_mpipe_equeue_t *equeue,
+		     struct tile_net_comps *comps,
+		     uint64_t when, struct sk_buff *skb)
+{
+	int cid = comps->comp_next % TILE_NET_MAX_COMPS;
+	comps->comp_queue[cid].when = when;
+	comps->comp_queue[cid].skb = skb;
+	comps->comp_next++;
+}
+
+static void tile_net_schedule_tx_wake_timer(struct net_device *dev,
+                                            int tx_queue_idx)
+{
+	struct tile_net_info *info = &per_cpu(per_cpu_info, tx_queue_idx);
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int instance = priv->instance;
+	struct tile_net_tx_wake *tx_wake =
+		&info->mpipe[instance].tx_wake[priv->echannel];
+
+	hrtimer_start(&tx_wake->timer,
+		      ktime_set(0, TX_TIMER_DELAY_USEC * 1000UL),
+		      HRTIMER_MODE_REL_PINNED);
+}
+
+static enum hrtimer_restart tile_net_handle_tx_wake_timer(struct hrtimer *t)
+{
+	struct tile_net_tx_wake *tx_wake =
+		container_of(t, struct tile_net_tx_wake, timer);
+	netif_wake_subqueue(tx_wake->dev, tx_wake->tx_queue_idx);
+	return HRTIMER_NORESTART;
+}
+
+/* Make sure the egress timer is scheduled. */
+static void tile_net_schedule_egress_timer(void)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+
+	if (!info->egress_timer_scheduled) {
+		hrtimer_start(&info->egress_timer,
+			      ktime_set(0, EGRESS_TIMER_DELAY_USEC * 1000UL),
+			      HRTIMER_MODE_REL_PINNED);
+		info->egress_timer_scheduled = true;
+	}
+}
+
+/* The "function" for "info->egress_timer".
+ *
+ * This timer will reschedule itself as long as there are any pending
+ * completions expected for this tile.
+ */
+static enum hrtimer_restart tile_net_handle_egress_timer(struct hrtimer *t)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	unsigned long irqflags;
+	bool pending = false;
+	int i, instance;
+
+	local_irq_save(irqflags);
+
+	/* The timer is no longer scheduled. */
+	info->egress_timer_scheduled = false;
+
+	/* Free all possible comps for this tile. */
+	for (instance = 0; instance < NR_MPIPE_MAX &&
+		     info->mpipe[instance].has_iqueue; instance++) {
+		for (i = 0; i < TILE_NET_CHANNELS; i++) {
+			struct tile_net_egress *egress =
+				&mpipe_data[instance].egress_for_echannel[i];
+			struct tile_net_comps *comps =
+				info->mpipe[instance].comps_for_echannel[i];
+			if (!egress || comps->comp_last >= comps->comp_next)
+				continue;
+			tile_net_free_comps(egress->equeue, comps, -1, true);
+			pending = pending ||
+				(comps->comp_last < comps->comp_next);
+		}
+	}
+
+	/* Reschedule timer if needed. */
+	if (pending)
+		tile_net_schedule_egress_timer();
+
+	local_irq_restore(irqflags);
+
+	return HRTIMER_NORESTART;
+}
+
+/* PTP clock operations. */
+
+static int ptp_mpipe_adjfreq(struct ptp_clock_info *ptp, s32 ppb)
+{
+	int ret = 0;
+	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
+	mutex_lock(&md->ptp_lock);
+	if (gxio_mpipe_adjust_timestamp_freq(&md->context, ppb))
+		ret = -EINVAL;
+	mutex_unlock(&md->ptp_lock);
+	return ret;
+}
+
+static int ptp_mpipe_adjtime(struct ptp_clock_info *ptp, s64 delta)
+{
+	int ret = 0;
+	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
+	mutex_lock(&md->ptp_lock);
+	if (gxio_mpipe_adjust_timestamp(&md->context, delta))
+		ret = -EBUSY;
+	mutex_unlock(&md->ptp_lock);
+	return ret;
+}
+
+static int ptp_mpipe_gettime(struct ptp_clock_info *ptp,
+			     struct timespec64 *ts)
+{
+	int ret = 0;
+	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
+	mutex_lock(&md->ptp_lock);
+	if (gxio_mpipe_get_timestamp(&md->context, ts))
+		ret = -EBUSY;
+	mutex_unlock(&md->ptp_lock);
+	return ret;
+}
+
+static int ptp_mpipe_settime(struct ptp_clock_info *ptp,
+			     const struct timespec64 *ts)
+{
+	int ret = 0;
+	struct mpipe_data *md = container_of(ptp, struct mpipe_data, caps);
+	mutex_lock(&md->ptp_lock);
+	if (gxio_mpipe_set_timestamp(&md->context, ts))
+		ret = -EBUSY;
+	mutex_unlock(&md->ptp_lock);
+	return ret;
+}
+
+static int ptp_mpipe_enable(struct ptp_clock_info *ptp,
+			    struct ptp_clock_request *request, int on)
+{
+	return -EOPNOTSUPP;
+}
+
+static struct ptp_clock_info ptp_mpipe_caps = {
+	.owner		= THIS_MODULE,
+	.name		= "mPIPE clock",
+	.max_adj	= 999999999,
+	.n_ext_ts	= 0,
+	.n_pins		= 0,
+	.pps		= 0,
+	.adjfreq	= ptp_mpipe_adjfreq,
+	.adjtime	= ptp_mpipe_adjtime,
+	.gettime64	= ptp_mpipe_gettime,
+	.settime64	= ptp_mpipe_settime,
+	.enable		= ptp_mpipe_enable,
+};
+
+/* Sync mPIPE's timestamp up with Linux system time and register PTP clock. */
+static void register_ptp_clock(struct net_device *dev, struct mpipe_data *md)
+{
+	struct timespec ts;
+
+	getnstimeofday(&ts);
+	gxio_mpipe_set_timestamp(&md->context, &ts);
+
+	mutex_init(&md->ptp_lock);
+	md->caps = ptp_mpipe_caps;
+	md->ptp_clock = ptp_clock_register(&md->caps, NULL);
+	if (IS_ERR(md->ptp_clock))
+		netdev_err(dev, "ptp_clock_register failed %ld\n",
+			   PTR_ERR(md->ptp_clock));
+}
+
+/* Initialize PTP fields in a new device. */
+static void init_ptp_dev(struct tile_net_priv *priv)
+{
+	priv->stamp_cfg.rx_filter = HWTSTAMP_FILTER_NONE;
+	priv->stamp_cfg.tx_type = HWTSTAMP_TX_OFF;
+}
+
+/* Helper functions for "tile_net_update()". */
+static void enable_ingress_irq(void *irq)
+{
+	enable_percpu_irq((long)irq, 0);
+}
+
+static void disable_ingress_irq(void *irq)
+{
+	disable_percpu_irq((long)irq);
+}
+
+/* Helper function for tile_net_open() and tile_net_stop().
+ * Always called under tile_net_devs_for_channel_mutex.
+ */
+static int tile_net_update(struct net_device *dev)
+{
+	static gxio_mpipe_rules_t rules;  /* too big to fit on the stack */
+	bool saw_channel = false;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+	int channel;
+	int rc;
+	int cpu;
+
+	saw_channel = false;
+	gxio_mpipe_rules_init(&rules, &md->context);
+
+	for (channel = 0; channel < TILE_NET_CHANNELS; channel++) {
+		if (md->tile_net_devs_for_channel[channel] == NULL)
+			continue;
+		if (!saw_channel) {
+			saw_channel = true;
+			gxio_mpipe_rules_begin(&rules, md->first_bucket,
+					       md->num_buckets, NULL);
+			gxio_mpipe_rules_set_headroom(&rules, NET_IP_ALIGN);
+		}
+		gxio_mpipe_rules_add_channel(&rules, channel);
+	}
+
+	/* NOTE: This can fail if there is no classifier.
+	 * ISSUE: Can anything else cause it to fail?
+	 */
+	rc = gxio_mpipe_rules_commit(&rules);
+	if (rc != 0) {
+		netdev_warn(dev, "gxio_mpipe_rules_commit: mpipe[%d] %d\n",
+			    instance, rc);
+		return -EIO;
+	}
+
+	/* Update all cpus, sequentially (to protect "netif_napi_add()").
+	 * We use on_each_cpu to handle the IPI mask or unmask.
+	 */
+	if (!saw_channel)
+		on_each_cpu(disable_ingress_irq,
+			    (void *)(long)(md->ingress_irq), 1);
+	for_each_online_cpu(cpu) {
+		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+
+		if (!info->mpipe[instance].has_iqueue)
+			continue;
+		if (saw_channel) {
+			if (!info->mpipe[instance].napi_added) {
+				netif_napi_add(dev, &info->mpipe[instance].napi,
+					       tile_net_poll, TILE_NET_WEIGHT);
+				info->mpipe[instance].napi_added = true;
+			}
+			if (!info->mpipe[instance].napi_enabled) {
+				napi_enable(&info->mpipe[instance].napi);
+				info->mpipe[instance].napi_enabled = true;
+			}
+		} else {
+			if (info->mpipe[instance].napi_enabled) {
+				napi_disable(&info->mpipe[instance].napi);
+				info->mpipe[instance].napi_enabled = false;
+			}
+			/* FIXME: Drain the iqueue. */
+		}
+	}
+	if (saw_channel)
+		on_each_cpu(enable_ingress_irq,
+			    (void *)(long)(md->ingress_irq), 1);
+
+	/* HACK: Allow packets to flow in the simulator. */
+	if (saw_channel)
+		sim_enable_mpipe_links(instance, -1);
+
+	return 0;
+}
+
+/* Initialize a buffer stack. */
+static int create_buffer_stack(struct net_device *dev,
+			       int kind, size_t num_buffers)
+{
+	pte_t hash_pte = pte_set_home((pte_t) { 0 }, PAGE_HOME_HASH);
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+	size_t needed = gxio_mpipe_calc_buffer_stack_bytes(num_buffers);
+	int stack_idx = md->first_buffer_stack + kind;
+	void *va;
+	int i, rc;
+
+	/* Round up to 64KB and then use alloc_pages() so we get the
+	 * required 64KB alignment.
+	 */
+	md->buffer_stack_bytes[kind] =
+		ALIGN(needed, 64 * 1024);
+
+	va = alloc_pages_exact(md->buffer_stack_bytes[kind], GFP_KERNEL);
+	if (va == NULL) {
+		netdev_err(dev,
+			   "Could not alloc %zd bytes for buffer stack %d\n",
+			   md->buffer_stack_bytes[kind], kind);
+		return -ENOMEM;
+	}
+
+	/* Initialize the buffer stack. */
+	rc = gxio_mpipe_init_buffer_stack(&md->context, stack_idx,
+					  buffer_size_enums[kind],  va,
+					  md->buffer_stack_bytes[kind], 0);
+	if (rc != 0) {
+		netdev_err(dev, "gxio_mpipe_init_buffer_stack: mpipe[%d] %d\n",
+			   instance, rc);
+		free_pages_exact(va, md->buffer_stack_bytes[kind]);
+		return rc;
+	}
+
+	md->buffer_stack_vas[kind] = va;
+
+	rc = gxio_mpipe_register_client_memory(&md->context, stack_idx,
+					       hash_pte, 0);
+	if (rc != 0) {
+		netdev_err(dev,
+			   "gxio_mpipe_register_client_memory: mpipe[%d] %d\n",
+			   instance, rc);
+		return rc;
+	}
+
+	/* Provide initial buffers. */
+	for (i = 0; i < num_buffers; i++) {
+		if (!tile_net_provide_buffer(instance, kind)) {
+			netdev_err(dev, "Cannot allocate initial sk_bufs!\n");
+			return -ENOMEM;
+		}
+	}
+
+	return 0;
+}
+
+/* Allocate and initialize mpipe buffer stacks, and register them in
+ * the mPIPE TLBs, for small, large, and (possibly) jumbo packet sizes.
+ * This routine supports tile_net_init_mpipe(), below.
+ */
+static int init_buffer_stacks(struct net_device *dev,
+			      int network_cpus_count)
+{
+	int num_kinds = MAX_KINDS - (jumbo_num == 0);
+	size_t num_buffers;
+	int rc;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	/* Allocate the buffer stacks. */
+	rc = gxio_mpipe_alloc_buffer_stacks(&md->context, num_kinds, 0, 0);
+	if (rc < 0) {
+		netdev_err(dev,
+			   "gxio_mpipe_alloc_buffer_stacks: mpipe[%d] %d\n",
+			   instance, rc);
+		return rc;
+	}
+	md->first_buffer_stack = rc;
+
+	/* Enough small/large buffers to (normally) avoid buffer errors. */
+	num_buffers =
+		network_cpus_count * (IQUEUE_ENTRIES + TILE_NET_BATCH);
+
+	/* Allocate the small memory stack. */
+	if (rc >= 0)
+		rc = create_buffer_stack(dev, 0, num_buffers);
+
+	/* Allocate the large buffer stack. */
+	if (rc >= 0)
+		rc = create_buffer_stack(dev, 1, num_buffers);
+
+	/* Allocate the jumbo buffer stack if needed. */
+	if (rc >= 0 && jumbo_num != 0)
+		rc = create_buffer_stack(dev, 2, jumbo_num);
+
+	return rc;
+}
+
+/* Allocate per-cpu resources (memory for completions and idescs).
+ * This routine supports tile_net_init_mpipe(), below.
+ */
+static int alloc_percpu_mpipe_resources(struct net_device *dev,
+					int cpu, int ring)
+{
+	struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+	int order, i, rc;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+	struct page *page;
+	void *addr;
+
+	/* Allocate the "comps". */
+	order = get_order(COMPS_SIZE);
+	page = homecache_alloc_pages(GFP_KERNEL, order, cpu);
+	if (page == NULL) {
+		netdev_err(dev, "Failed to alloc %zd bytes comps memory\n",
+			   COMPS_SIZE);
+		return -ENOMEM;
+	}
+	addr = pfn_to_kaddr(page_to_pfn(page));
+	memset(addr, 0, COMPS_SIZE);
+	for (i = 0; i < TILE_NET_CHANNELS; i++)
+		info->mpipe[instance].comps_for_echannel[i] =
+			addr + i * sizeof(struct tile_net_comps);
+
+	/* If this is a network cpu, create an iqueue. */
+	if (cpumask_test_cpu(cpu, &network_cpus_map)) {
+		order = get_order(NOTIF_RING_SIZE);
+		page = homecache_alloc_pages(GFP_KERNEL, order, cpu);
+		if (page == NULL) {
+			netdev_err(dev,
+				   "Failed to alloc %zd bytes iqueue memory\n",
+				   NOTIF_RING_SIZE);
+			return -ENOMEM;
+		}
+		addr = pfn_to_kaddr(page_to_pfn(page));
+		rc = gxio_mpipe_iqueue_init(&info->mpipe[instance].iqueue,
+					    &md->context, ring++, addr,
+					    NOTIF_RING_SIZE, 0);
+		if (rc < 0) {
+			netdev_err(dev,
+				   "gxio_mpipe_iqueue_init failed: %d\n", rc);
+			return rc;
+		}
+		info->mpipe[instance].has_iqueue = true;
+	}
+
+	return ring;
+}
+
+/* Initialize NotifGroup and buckets.
+ * This routine supports tile_net_init_mpipe(), below.
+ */
+static int init_notif_group_and_buckets(struct net_device *dev,
+					int ring, int network_cpus_count)
+{
+	int group, rc;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	/* Allocate one NotifGroup. */
+	rc = gxio_mpipe_alloc_notif_groups(&md->context, 1, 0, 0);
+	if (rc < 0) {
+		netdev_err(dev, "gxio_mpipe_alloc_notif_groups: mpipe[%d] %d\n",
+			   instance, rc);
+		return rc;
+	}
+	group = rc;
+
+	/* Initialize global num_buckets value. */
+	if (network_cpus_count > 4)
+		md->num_buckets = 256;
+	else if (network_cpus_count > 1)
+		md->num_buckets = 16;
+
+	/* Allocate some buckets, and set global first_bucket value. */
+	rc = gxio_mpipe_alloc_buckets(&md->context, md->num_buckets, 0, 0);
+	if (rc < 0) {
+		netdev_err(dev, "gxio_mpipe_alloc_buckets: mpipe[%d] %d\n",
+			   instance, rc);
+		return rc;
+	}
+	md->first_bucket = rc;
+
+	/* Init group and buckets. */
+	rc = gxio_mpipe_init_notif_group_and_buckets(
+		&md->context, group, ring, network_cpus_count,
+		md->first_bucket, md->num_buckets,
+		GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY);
+	if (rc != 0) {
+		netdev_err(dev,	"gxio_mpipe_init_notif_group_and_buckets: "
+			   "mpipe[%d] %d\n", instance, rc);
+		return rc;
+	}
+
+	return 0;
+}
+
+/* Create an irq and register it, then activate the irq and request
+ * interrupts on all cores.  Note that "ingress_irq" being initialized
+ * is how we know not to call tile_net_init_mpipe() again.
+ * This routine supports tile_net_init_mpipe(), below.
+ */
+static int tile_net_setup_interrupts(struct net_device *dev)
+{
+	int cpu, rc, irq;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	irq = md->ingress_irq;
+	if (irq < 0) {
+		irq = irq_alloc_hwirq(-1);
+		if (!irq) {
+			netdev_err(dev,
+				   "create_irq failed: mpipe[%d] %d\n",
+				   instance, irq);
+			return irq;
+		}
+		tile_irq_activate(irq, TILE_IRQ_PERCPU);
+
+		rc = request_irq(irq, tile_net_handle_ingress_irq,
+				 0, "tile_net", (void *)((uint64_t)instance));
+
+		if (rc != 0) {
+			netdev_err(dev, "request_irq failed: mpipe[%d] %d\n",
+				   instance, rc);
+			irq_free_hwirq(irq);
+			return rc;
+		}
+		md->ingress_irq = irq;
+	}
+
+	for_each_online_cpu(cpu) {
+		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+		if (info->mpipe[instance].has_iqueue) {
+			gxio_mpipe_request_notif_ring_interrupt(&md->context,
+				cpu_x(cpu), cpu_y(cpu), KERNEL_PL, irq,
+				info->mpipe[instance].iqueue.ring);
+		}
+	}
+
+	return 0;
+}
+
+/* Undo any state set up partially by a failed call to tile_net_init_mpipe. */
+static void tile_net_init_mpipe_fail(int instance)
+{
+	int kind, cpu;
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	/* Do cleanups that require the mpipe context first. */
+	for (kind = 0; kind < MAX_KINDS; kind++) {
+		if (md->buffer_stack_vas[kind] != NULL) {
+			tile_net_pop_all_buffers(instance,
+						 md->first_buffer_stack +
+						 kind);
+		}
+	}
+
+	/* Destroy mpipe context so the hardware no longer owns any memory. */
+	gxio_mpipe_destroy(&md->context);
+
+	for_each_online_cpu(cpu) {
+		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+		free_pages(
+			(unsigned long)(
+				info->mpipe[instance].comps_for_echannel[0]),
+			get_order(COMPS_SIZE));
+		info->mpipe[instance].comps_for_echannel[0] = NULL;
+		free_pages((unsigned long)(info->mpipe[instance].iqueue.idescs),
+			   get_order(NOTIF_RING_SIZE));
+		info->mpipe[instance].iqueue.idescs = NULL;
+	}
+
+	for (kind = 0; kind < MAX_KINDS; kind++) {
+		if (md->buffer_stack_vas[kind] != NULL) {
+			free_pages_exact(md->buffer_stack_vas[kind],
+					 md->buffer_stack_bytes[kind]);
+			md->buffer_stack_vas[kind] = NULL;
+		}
+	}
+
+	md->first_buffer_stack = -1;
+	md->first_bucket = -1;
+}
+
+/* The first time any tilegx network device is opened, we initialize
+ * the global mpipe state.  If this step fails, we fail to open the
+ * device, but if it succeeds, we never need to do it again, and since
+ * tile_net can't be unloaded, we never undo it.
+ *
+ * Note that some resources in this path (buffer stack indices,
+ * bindings from init_buffer_stack, etc.) are hypervisor resources
+ * that are freed implicitly by gxio_mpipe_destroy().
+ */
+static int tile_net_init_mpipe(struct net_device *dev)
+{
+	int rc;
+	int cpu;
+	int first_ring, ring;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+	int network_cpus_count = cpumask_weight(&network_cpus_map);
+
+	if (!hash_default) {
+		netdev_err(dev, "Networking requires hash_default!\n");
+		return -EIO;
+	}
+
+	rc = gxio_mpipe_init(&md->context, instance);
+	if (rc != 0) {
+		netdev_err(dev, "gxio_mpipe_init: mpipe[%d] %d\n",
+			   instance, rc);
+		return -EIO;
+	}
+
+	/* Set up the buffer stacks. */
+	rc = init_buffer_stacks(dev, network_cpus_count);
+	if (rc != 0)
+		goto fail;
+
+	/* Allocate one NotifRing for each network cpu. */
+	rc = gxio_mpipe_alloc_notif_rings(&md->context,
+					  network_cpus_count, 0, 0);
+	if (rc < 0) {
+		netdev_err(dev, "gxio_mpipe_alloc_notif_rings failed %d\n",
+			   rc);
+		goto fail;
+	}
+
+	/* Init NotifRings per-cpu. */
+	first_ring = rc;
+	ring = first_ring;
+	for_each_online_cpu(cpu) {
+		rc = alloc_percpu_mpipe_resources(dev, cpu, ring);
+		if (rc < 0)
+			goto fail;
+		ring = rc;
+	}
+
+	/* Initialize NotifGroup and buckets. */
+	rc = init_notif_group_and_buckets(dev, first_ring, network_cpus_count);
+	if (rc != 0)
+		goto fail;
+
+	/* Create and enable interrupts. */
+	rc = tile_net_setup_interrupts(dev);
+	if (rc != 0)
+		goto fail;
+
+	/* Register PTP clock and set mPIPE timestamp, if configured. */
+	register_ptp_clock(dev, md);
+
+	return 0;
+
+fail:
+	tile_net_init_mpipe_fail(instance);
+	return rc;
+}
+
+/* Create persistent egress info for a given egress channel.
+ * Note that this may be shared between, say, "gbe0" and "xgbe0".
+ * ISSUE: Defer header allocation until TSO is actually needed?
+ */
+static int tile_net_init_egress(struct net_device *dev, int echannel)
+{
+	static int ering = -1;
+	struct page *headers_page, *edescs_page, *equeue_page;
+	gxio_mpipe_edesc_t *edescs;
+	gxio_mpipe_equeue_t *equeue;
+	unsigned char *headers;
+	int headers_order, edescs_order, equeue_order;
+	size_t edescs_size;
+	int rc = -ENOMEM;
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	/* Only initialize once. */
+	if (md->egress_for_echannel[echannel].equeue != NULL)
+		return 0;
+
+	/* Allocate memory for the "headers". */
+	headers_order = get_order(EQUEUE_ENTRIES * HEADER_BYTES);
+	headers_page = alloc_pages(GFP_KERNEL, headers_order);
+	if (headers_page == NULL) {
+		netdev_warn(dev,
+			    "Could not alloc %zd bytes for TSO headers.\n",
+			    PAGE_SIZE << headers_order);
+		goto fail;
+	}
+	headers = pfn_to_kaddr(page_to_pfn(headers_page));
+
+	/* Allocate memory for the "edescs". */
+	edescs_size = EQUEUE_ENTRIES * sizeof(*edescs);
+	edescs_order = get_order(edescs_size);
+	edescs_page = alloc_pages(GFP_KERNEL, edescs_order);
+	if (edescs_page == NULL) {
+		netdev_warn(dev,
+			    "Could not alloc %zd bytes for eDMA ring.\n",
+			    edescs_size);
+		goto fail_headers;
+	}
+	edescs = pfn_to_kaddr(page_to_pfn(edescs_page));
+
+	/* Allocate memory for the "equeue". */
+	equeue_order = get_order(sizeof(*equeue));
+	equeue_page = alloc_pages(GFP_KERNEL, equeue_order);
+	if (equeue_page == NULL) {
+		netdev_warn(dev,
+			    "Could not alloc %zd bytes for equeue info.\n",
+			    PAGE_SIZE << equeue_order);
+		goto fail_edescs;
+	}
+	equeue = pfn_to_kaddr(page_to_pfn(equeue_page));
+
+	/* Allocate an edma ring (using a one entry "free list"). */
+	if (ering < 0) {
+		rc = gxio_mpipe_alloc_edma_rings(&md->context, 1, 0, 0);
+		if (rc < 0) {
+			netdev_warn(dev, "gxio_mpipe_alloc_edma_rings: "
+				    "mpipe[%d] %d\n", instance, rc);
+			goto fail_equeue;
+		}
+		ering = rc;
+	}
+
+	/* Initialize the equeue. */
+	rc = gxio_mpipe_equeue_init(equeue, &md->context, ering, echannel,
+				    edescs, edescs_size, 0);
+	if (rc != 0) {
+		netdev_err(dev, "gxio_mpipe_equeue_init: mpipe[%d] %d\n",
+			   instance, rc);
+		goto fail_equeue;
+	}
+
+	/* Don't reuse the ering later. */
+	ering = -1;
+
+	if (jumbo_num != 0) {
+		/* Make sure "jumbo" packets can be egressed safely. */
+		if (gxio_mpipe_equeue_set_snf_size(equeue, 10368) < 0) {
+			/* ISSUE: There is no "gxio_mpipe_equeue_destroy()". */
+			netdev_warn(dev, "Jumbo packets may not be egressed"
+				    " properly on channel %d\n", echannel);
+		}
+	}
+
+	/* Done. */
+	md->egress_for_echannel[echannel].equeue = equeue;
+	md->egress_for_echannel[echannel].headers = headers;
+	return 0;
+
+fail_equeue:
+	__free_pages(equeue_page, equeue_order);
+
+fail_edescs:
+	__free_pages(edescs_page, edescs_order);
+
+fail_headers:
+	__free_pages(headers_page, headers_order);
+
+fail:
+	return rc;
+}
+
+/* Return channel number for a newly-opened link. */
+static int tile_net_link_open(struct net_device *dev, gxio_mpipe_link_t *link,
+			      const char *link_name)
+{
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+	int rc = gxio_mpipe_link_open(link, &md->context, link_name, 0);
+	if (rc < 0) {
+		netdev_err(dev, "Failed to open '%s', mpipe[%d], %d\n",
+			   link_name, instance, rc);
+		return rc;
+	}
+	if (jumbo_num != 0) {
+		u32 attr = GXIO_MPIPE_LINK_RECEIVE_JUMBO;
+		rc = gxio_mpipe_link_set_attr(link, attr, 1);
+		if (rc != 0) {
+			netdev_err(dev,
+				   "Cannot receive jumbo packets on '%s'\n",
+				   link_name);
+			gxio_mpipe_link_close(link);
+			return rc;
+		}
+	}
+	rc = gxio_mpipe_link_channel(link);
+	if (rc < 0 || rc >= TILE_NET_CHANNELS) {
+		netdev_err(dev, "gxio_mpipe_link_channel bad value: %d\n", rc);
+		gxio_mpipe_link_close(link);
+		return -EINVAL;
+	}
+	return rc;
+}
+
+/* Help the kernel activate the given network interface. */
+static int tile_net_open(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int cpu, rc, instance;
+
+	mutex_lock(&tile_net_devs_for_channel_mutex);
+
+	/* Get the instance info. */
+	rc = gxio_mpipe_link_instance(dev->name);
+	if (rc < 0 || rc >= NR_MPIPE_MAX) {
+		mutex_unlock(&tile_net_devs_for_channel_mutex);
+		return -EIO;
+	}
+
+	priv->instance = rc;
+	instance = rc;
+	if (!mpipe_data[rc].context.mmio_fast_base) {
+		/* Do one-time initialization per instance the first time
+		 * any device is opened.
+		 */
+		rc = tile_net_init_mpipe(dev);
+		if (rc != 0)
+			goto fail;
+	}
+
+	/* Determine if this is the "loopify" device. */
+	if (unlikely((loopify_link_name != NULL) &&
+		     !strcmp(dev->name, loopify_link_name))) {
+		rc = tile_net_link_open(dev, &priv->link, "loop0");
+		if (rc < 0)
+			goto fail;
+		priv->channel = rc;
+		rc = tile_net_link_open(dev, &priv->loopify_link, "loop1");
+		if (rc < 0)
+			goto fail;
+		priv->loopify_channel = rc;
+		priv->echannel = rc;
+	} else {
+		rc = tile_net_link_open(dev, &priv->link, dev->name);
+		if (rc < 0)
+			goto fail;
+		priv->channel = rc;
+		priv->echannel = rc;
+	}
+
+	/* Initialize egress info (if needed).  Once ever, per echannel. */
+	rc = tile_net_init_egress(dev, priv->echannel);
+	if (rc != 0)
+		goto fail;
+
+	mpipe_data[instance].tile_net_devs_for_channel[priv->channel] = dev;
+
+	rc = tile_net_update(dev);
+	if (rc != 0)
+		goto fail;
+
+	mutex_unlock(&tile_net_devs_for_channel_mutex);
+
+	/* Initialize the transmit wake timer for this device for each cpu. */
+	for_each_online_cpu(cpu) {
+		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+		struct tile_net_tx_wake *tx_wake =
+			&info->mpipe[instance].tx_wake[priv->echannel];
+
+		hrtimer_init(&tx_wake->timer, CLOCK_MONOTONIC,
+			     HRTIMER_MODE_REL);
+		tx_wake->tx_queue_idx = cpu;
+		tx_wake->timer.function = tile_net_handle_tx_wake_timer;
+		tx_wake->dev = dev;
+	}
+
+	for_each_online_cpu(cpu)
+		netif_start_subqueue(dev, cpu);
+	netif_carrier_on(dev);
+	return 0;
+
+fail:
+	if (priv->loopify_channel >= 0) {
+		if (gxio_mpipe_link_close(&priv->loopify_link) != 0)
+			netdev_warn(dev, "Failed to close loopify link!\n");
+		priv->loopify_channel = -1;
+	}
+	if (priv->channel >= 0) {
+		if (gxio_mpipe_link_close(&priv->link) != 0)
+			netdev_warn(dev, "Failed to close link!\n");
+		priv->channel = -1;
+	}
+	priv->echannel = -1;
+	mpipe_data[instance].tile_net_devs_for_channel[priv->channel] =	NULL;
+	mutex_unlock(&tile_net_devs_for_channel_mutex);
+
+	/* Don't return raw gxio error codes to generic Linux. */
+	return (rc > -512) ? rc : -EIO;
+}
+
+/* Help the kernel deactivate the given network interface. */
+static int tile_net_stop(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int cpu;
+	int instance = priv->instance;
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	for_each_online_cpu(cpu) {
+		struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+		struct tile_net_tx_wake *tx_wake =
+			&info->mpipe[instance].tx_wake[priv->echannel];
+
+		hrtimer_cancel(&tx_wake->timer);
+		netif_stop_subqueue(dev, cpu);
+	}
+
+	mutex_lock(&tile_net_devs_for_channel_mutex);
+	md->tile_net_devs_for_channel[priv->channel] = NULL;
+	(void)tile_net_update(dev);
+	if (priv->loopify_channel >= 0) {
+		if (gxio_mpipe_link_close(&priv->loopify_link) != 0)
+			netdev_warn(dev, "Failed to close loopify link!\n");
+		priv->loopify_channel = -1;
+	}
+	if (priv->channel >= 0) {
+		if (gxio_mpipe_link_close(&priv->link) != 0)
+			netdev_warn(dev, "Failed to close link!\n");
+		priv->channel = -1;
+	}
+	priv->echannel = -1;
+	mutex_unlock(&tile_net_devs_for_channel_mutex);
+
+	return 0;
+}
+
+/* Determine the VA for a fragment. */
+static inline void *tile_net_frag_buf(skb_frag_t *f)
+{
+	unsigned long pfn = page_to_pfn(skb_frag_page(f));
+	return pfn_to_kaddr(pfn) + f->page_offset;
+}
+
+/* Acquire a completion entry and an egress slot, or if we can't,
+ * stop the queue and schedule the tx_wake timer.
+ */
+static s64 tile_net_equeue_try_reserve(struct net_device *dev,
+				       int tx_queue_idx,
+				       struct tile_net_comps *comps,
+				       gxio_mpipe_equeue_t *equeue,
+				       int num_edescs)
+{
+	/* Try to acquire a completion entry. */
+	if (comps->comp_next - comps->comp_last < TILE_NET_MAX_COMPS - 1 ||
+	    tile_net_free_comps(equeue, comps, 32, false) != 0) {
+
+		/* Try to acquire an egress slot. */
+		s64 slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs);
+		if (slot >= 0)
+			return slot;
+
+		/* Freeing some completions gives the equeue time to drain. */
+		tile_net_free_comps(equeue, comps, TILE_NET_MAX_COMPS, false);
+
+		slot = gxio_mpipe_equeue_try_reserve(equeue, num_edescs);
+		if (slot >= 0)
+			return slot;
+	}
+
+	/* Still nothing; give up and stop the queue for a short while. */
+	netif_stop_subqueue(dev, tx_queue_idx);
+	tile_net_schedule_tx_wake_timer(dev, tx_queue_idx);
+	return -1;
+}
+
+/* Determine how many edesc's are needed for TSO.
+ *
+ * Sometimes, if "sendfile()" requires copying, we will be called with
+ * "data" containing the header and payload, with "frags" being empty.
+ * Sometimes, for example when using NFS over TCP, a single segment can
+ * span 3 fragments.  This requires special care.
+ */
+static int tso_count_edescs(struct sk_buff *skb)
+{
+	struct skb_shared_info *sh = skb_shinfo(skb);
+	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	unsigned int data_len = skb->len - sh_len;
+	unsigned int p_len = sh->gso_size;
+	long f_id = -1;    /* id of the current fragment */
+	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
+	long f_used = 0;  /* bytes used from the current fragment */
+	long n;            /* size of the current piece of payload */
+	int num_edescs = 0;
+	int segment;
+
+	for (segment = 0; segment < sh->gso_segs; segment++) {
+
+		unsigned int p_used = 0;
+
+		/* One edesc for header and for each piece of the payload. */
+		for (num_edescs++; p_used < p_len; num_edescs++) {
+
+			/* Advance as needed. */
+			while (f_used >= f_size) {
+				f_id++;
+				f_size = skb_frag_size(&sh->frags[f_id]);
+				f_used = 0;
+			}
+
+			/* Use bytes from the current fragment. */
+			n = p_len - p_used;
+			if (n > f_size - f_used)
+				n = f_size - f_used;
+			f_used += n;
+			p_used += n;
+		}
+
+		/* The last segment may be less than gso_size. */
+		data_len -= p_len;
+		if (data_len < p_len)
+			p_len = data_len;
+	}
+
+	return num_edescs;
+}
+
+/* Prepare modified copies of the skbuff headers. */
+static void tso_headers_prepare(struct sk_buff *skb, unsigned char *headers,
+				s64 slot)
+{
+	struct skb_shared_info *sh = skb_shinfo(skb);
+	struct iphdr *ih;
+	struct ipv6hdr *ih6;
+	struct tcphdr *th;
+	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	unsigned int data_len = skb->len - sh_len;
+	unsigned char *data = skb->data;
+	unsigned int ih_off, th_off, p_len;
+	unsigned int isum_seed, tsum_seed, seq;
+	unsigned int uninitialized_var(id);
+	int is_ipv6;
+	long f_id = -1;    /* id of the current fragment */
+	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
+	long f_used = 0;  /* bytes used from the current fragment */
+	long n;            /* size of the current piece of payload */
+	int segment;
+
+	/* Locate original headers and compute various lengths. */
+	is_ipv6 = skb_is_gso_v6(skb);
+	if (is_ipv6) {
+		ih6 = ipv6_hdr(skb);
+		ih_off = skb_network_offset(skb);
+	} else {
+		ih = ip_hdr(skb);
+		ih_off = skb_network_offset(skb);
+		isum_seed = ((0xFFFF - ih->check) +
+			     (0xFFFF - ih->tot_len) +
+			     (0xFFFF - ih->id));
+		id = ntohs(ih->id);
+	}
+
+	th = tcp_hdr(skb);
+	th_off = skb_transport_offset(skb);
+	p_len = sh->gso_size;
+
+	tsum_seed = th->check + (0xFFFF ^ htons(skb->len));
+	seq = ntohl(th->seq);
+
+	/* Prepare all the headers. */
+	for (segment = 0; segment < sh->gso_segs; segment++) {
+		unsigned char *buf;
+		unsigned int p_used = 0;
+
+		/* Copy to the header memory for this segment. */
+		buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES +
+			NET_IP_ALIGN;
+		memcpy(buf, data, sh_len);
+
+		/* Update copied ip header. */
+		if (is_ipv6) {
+			ih6 = (struct ipv6hdr *)(buf + ih_off);
+			ih6->payload_len = htons(sh_len + p_len - ih_off -
+						 sizeof(*ih6));
+		} else {
+			ih = (struct iphdr *)(buf + ih_off);
+			ih->tot_len = htons(sh_len + p_len - ih_off);
+			ih->id = htons(id++);
+			ih->check = csum_long(isum_seed + ih->tot_len +
+					      ih->id) ^ 0xffff;
+		}
+
+		/* Update copied tcp header. */
+		th = (struct tcphdr *)(buf + th_off);
+		th->seq = htonl(seq);
+		th->check = csum_long(tsum_seed + htons(sh_len + p_len));
+		if (segment != sh->gso_segs - 1) {
+			th->fin = 0;
+			th->psh = 0;
+		}
+
+		/* Skip past the header. */
+		slot++;
+
+		/* Skip past the payload. */
+		while (p_used < p_len) {
+
+			/* Advance as needed. */
+			while (f_used >= f_size) {
+				f_id++;
+				f_size = skb_frag_size(&sh->frags[f_id]);
+				f_used = 0;
+			}
+
+			/* Use bytes from the current fragment. */
+			n = p_len - p_used;
+			if (n > f_size - f_used)
+				n = f_size - f_used;
+			f_used += n;
+			p_used += n;
+
+			slot++;
+		}
+
+		seq += p_len;
+
+		/* The last segment may be less than gso_size. */
+		data_len -= p_len;
+		if (data_len < p_len)
+			p_len = data_len;
+	}
+
+	/* Flush the headers so they are ready for hardware DMA. */
+	wmb();
+}
+
+/* Pass all the data to mpipe for egress. */
+static void tso_egress(struct net_device *dev, gxio_mpipe_equeue_t *equeue,
+		       struct sk_buff *skb, unsigned char *headers, s64 slot)
+{
+	struct skb_shared_info *sh = skb_shinfo(skb);
+	int instance = mpipe_instance(dev);
+	struct mpipe_data *md = &mpipe_data[instance];
+	unsigned int sh_len = skb_transport_offset(skb) + tcp_hdrlen(skb);
+	unsigned int data_len = skb->len - sh_len;
+	unsigned int p_len = sh->gso_size;
+	gxio_mpipe_edesc_t edesc_head = { { 0 } };
+	gxio_mpipe_edesc_t edesc_body = { { 0 } };
+	long f_id = -1;    /* id of the current fragment */
+	long f_size = skb_headlen(skb) - sh_len;  /* current fragment size */
+	long f_used = 0;  /* bytes used from the current fragment */
+	void *f_data = skb->data + sh_len;
+	long n;            /* size of the current piece of payload */
+	unsigned long tx_packets = 0, tx_bytes = 0;
+	unsigned int csum_start;
+	int segment;
+
+	/* Prepare to egress the headers: set up header edesc. */
+	csum_start = skb_checksum_start_offset(skb);
+	edesc_head.csum = 1;
+	edesc_head.csum_start = csum_start;
+	edesc_head.csum_dest = csum_start + skb->csum_offset;
+	edesc_head.xfer_size = sh_len;
+
+	/* This is only used to specify the TLB. */
+	edesc_head.stack_idx = md->first_buffer_stack;
+	edesc_body.stack_idx = md->first_buffer_stack;
+
+	/* Egress all the edescs. */
+	for (segment = 0; segment < sh->gso_segs; segment++) {
+		unsigned char *buf;
+		unsigned int p_used = 0;
+
+		/* Egress the header. */
+		buf = headers + (slot % EQUEUE_ENTRIES) * HEADER_BYTES +
+			NET_IP_ALIGN;
+		edesc_head.va = va_to_tile_io_addr(buf);
+		gxio_mpipe_equeue_put_at(equeue, edesc_head, slot);
+		slot++;
+
+		/* Egress the payload. */
+		while (p_used < p_len) {
+			void *va;
+
+			/* Advance as needed. */
+			while (f_used >= f_size) {
+				f_id++;
+				f_size = skb_frag_size(&sh->frags[f_id]);
+				f_data = tile_net_frag_buf(&sh->frags[f_id]);
+				f_used = 0;
+			}
+
+			va = f_data + f_used;
+
+			/* Use bytes from the current fragment. */
+			n = p_len - p_used;
+			if (n > f_size - f_used)
+				n = f_size - f_used;
+			f_used += n;
+			p_used += n;
+
+			/* Egress a piece of the payload. */
+			edesc_body.va = va_to_tile_io_addr(va);
+			edesc_body.xfer_size = n;
+			edesc_body.bound = !(p_used < p_len);
+			gxio_mpipe_equeue_put_at(equeue, edesc_body, slot);
+			slot++;
+		}
+
+		tx_packets++;
+		tx_bytes += sh_len + p_len;
+
+		/* The last segment may be less than gso_size. */
+		data_len -= p_len;
+		if (data_len < p_len)
+			p_len = data_len;
+	}
+
+	/* Update stats. */
+	tile_net_stats_add(tx_packets, &dev->stats.tx_packets);
+	tile_net_stats_add(tx_bytes, &dev->stats.tx_bytes);
+}
+
+/* Do "TSO" handling for egress.
+ *
+ * Normally drivers set NETIF_F_TSO only to support hardware TSO;
+ * otherwise the stack uses scatter-gather to implement GSO in software.
+ * On our testing, enabling GSO support (via NETIF_F_SG) drops network
+ * performance down to around 7.5 Gbps on the 10G interfaces, although
+ * also dropping cpu utilization way down, to under 8%.  But
+ * implementing "TSO" in the driver brings performance back up to line
+ * rate, while dropping cpu usage even further, to less than 4%.  In
+ * practice, profiling of GSO shows that skb_segment() is what causes
+ * the performance overheads; we benefit in the driver from using
+ * preallocated memory to duplicate the TCP/IP headers.
+ */
+static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int channel = priv->echannel;
+	int instance = priv->instance;
+	struct mpipe_data *md = &mpipe_data[instance];
+	struct tile_net_egress *egress = &md->egress_for_echannel[channel];
+	struct tile_net_comps *comps =
+		info->mpipe[instance].comps_for_echannel[channel];
+	gxio_mpipe_equeue_t *equeue = egress->equeue;
+	unsigned long irqflags;
+	int num_edescs;
+	s64 slot;
+
+	/* Determine how many mpipe edesc's are needed. */
+	num_edescs = tso_count_edescs(skb);
+
+	local_irq_save(irqflags);
+
+	/* Try to acquire a completion entry and an egress slot. */
+	slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps,
+					   equeue, num_edescs);
+	if (slot < 0) {
+		local_irq_restore(irqflags);
+		return NETDEV_TX_BUSY;
+	}
+
+	/* Set up copies of header data properly. */
+	tso_headers_prepare(skb, egress->headers, slot);
+
+	/* Actually pass the data to the network hardware. */
+	tso_egress(dev, equeue, skb, egress->headers, slot);
+
+	/* Add a completion record. */
+	add_comp(equeue, comps, slot + num_edescs - 1, skb);
+
+	local_irq_restore(irqflags);
+
+	/* Make sure the egress timer is scheduled. */
+	tile_net_schedule_egress_timer();
+
+	return NETDEV_TX_OK;
+}
+
+/* Analyze the body and frags for a transmit request. */
+static unsigned int tile_net_tx_frags(struct frag *frags,
+				       struct sk_buff *skb,
+				       void *b_data, unsigned int b_len)
+{
+	unsigned int i, n = 0;
+
+	struct skb_shared_info *sh = skb_shinfo(skb);
+
+	if (b_len != 0) {
+		frags[n].buf = b_data;
+		frags[n++].length = b_len;
+	}
+
+	for (i = 0; i < sh->nr_frags; i++) {
+		skb_frag_t *f = &sh->frags[i];
+		frags[n].buf = tile_net_frag_buf(f);
+		frags[n++].length = skb_frag_size(f);
+	}
+
+	return n;
+}
+
+/* Help the kernel transmit a packet. */
+static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int instance = priv->instance;
+	struct mpipe_data *md = &mpipe_data[instance];
+	struct tile_net_egress *egress =
+		&md->egress_for_echannel[priv->echannel];
+	gxio_mpipe_equeue_t *equeue = egress->equeue;
+	struct tile_net_comps *comps =
+		info->mpipe[instance].comps_for_echannel[priv->echannel];
+	unsigned int len = skb->len;
+	unsigned char *data = skb->data;
+	unsigned int num_edescs;
+	struct frag frags[MAX_FRAGS];
+	gxio_mpipe_edesc_t edescs[MAX_FRAGS];
+	unsigned long irqflags;
+	gxio_mpipe_edesc_t edesc = { { 0 } };
+	unsigned int i;
+	s64 slot;
+
+	if (skb_is_gso(skb))
+		return tile_net_tx_tso(skb, dev);
+
+	num_edescs = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
+
+	/* This is only used to specify the TLB. */
+	edesc.stack_idx = md->first_buffer_stack;
+
+	/* Prepare the edescs. */
+	for (i = 0; i < num_edescs; i++) {
+		edesc.xfer_size = frags[i].length;
+		edesc.va = va_to_tile_io_addr(frags[i].buf);
+		edescs[i] = edesc;
+	}
+
+	/* Mark the final edesc. */
+	edescs[num_edescs - 1].bound = 1;
+
+	/* Add checksum info to the initial edesc, if needed. */
+	if (skb->ip_summed == CHECKSUM_PARTIAL) {
+		unsigned int csum_start = skb_checksum_start_offset(skb);
+		edescs[0].csum = 1;
+		edescs[0].csum_start = csum_start;
+		edescs[0].csum_dest = csum_start + skb->csum_offset;
+	}
+
+	local_irq_save(irqflags);
+
+	/* Try to acquire a completion entry and an egress slot. */
+	slot = tile_net_equeue_try_reserve(dev, skb->queue_mapping, comps,
+					   equeue, num_edescs);
+	if (slot < 0) {
+		local_irq_restore(irqflags);
+		return NETDEV_TX_BUSY;
+	}
+
+	for (i = 0; i < num_edescs; i++)
+		gxio_mpipe_equeue_put_at(equeue, edescs[i], slot++);
+
+	/* Store TX timestamp if needed. */
+	tile_tx_timestamp(skb, instance);
+
+	/* Add a completion record. */
+	add_comp(equeue, comps, slot - 1, skb);
+
+	/* NOTE: Use ETH_ZLEN for short packets (e.g. 42 < 60). */
+	tile_net_stats_add(1, &dev->stats.tx_packets);
+	tile_net_stats_add(max_t(unsigned int, len, ETH_ZLEN),
+			   &dev->stats.tx_bytes);
+
+	local_irq_restore(irqflags);
+
+	/* Make sure the egress timer is scheduled. */
+	tile_net_schedule_egress_timer();
+
+	return NETDEV_TX_OK;
+}
+
+/* Return subqueue id on this core (one per core). */
+static u16 tile_net_select_queue(struct net_device *dev, struct sk_buff *skb,
+				 void *accel_priv, select_queue_fallback_t fallback)
+{
+	return smp_processor_id();
+}
+
+/* Deal with a transmit timeout. */
+static void tile_net_tx_timeout(struct net_device *dev)
+{
+	int cpu;
+
+	for_each_online_cpu(cpu)
+		netif_wake_subqueue(dev, cpu);
+}
+
+/* Ioctl commands. */
+static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
+{
+	if (cmd == SIOCSHWTSTAMP)
+		return tile_hwtstamp_set(dev, rq);
+	if (cmd == SIOCGHWTSTAMP)
+		return tile_hwtstamp_get(dev, rq);
+
+	return -EOPNOTSUPP;
+}
+
+/* Change the MTU. */
+static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
+{
+	if (new_mtu < 68)
+		return -EINVAL;
+	if (new_mtu > ((jumbo_num != 0) ? 9000 : 1500))
+		return -EINVAL;
+	dev->mtu = new_mtu;
+	return 0;
+}
+
+/* Change the Ethernet address of the NIC.
+ *
+ * The hypervisor driver does not support changing MAC address.  However,
+ * the hardware does not do anything with the MAC address, so the address
+ * which gets used on outgoing packets, and which is accepted on incoming
+ * packets, is completely up to us.
+ *
+ * Returns 0 on success, negative on failure.
+ */
+static int tile_net_set_mac_address(struct net_device *dev, void *p)
+{
+	struct sockaddr *addr = p;
+
+	if (!is_valid_ether_addr(addr->sa_data))
+		return -EINVAL;
+	memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
+	return 0;
+}
+
+#ifdef CONFIG_NET_POLL_CONTROLLER
+/* Polling 'interrupt' - used by things like netconsole to send skbs
+ * without having to re-enable interrupts. It's not called while
+ * the interrupt routine is executing.
+ */
+static void tile_net_netpoll(struct net_device *dev)
+{
+	int instance = mpipe_instance(dev);
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	struct mpipe_data *md = &mpipe_data[instance];
+
+	disable_percpu_irq(md->ingress_irq);
+	napi_schedule(&info->mpipe[instance].napi);
+	enable_percpu_irq(md->ingress_irq, 0);
+}
+#endif
+
+static const struct net_device_ops tile_net_ops = {
+	.ndo_open = tile_net_open,
+	.ndo_stop = tile_net_stop,
+	.ndo_start_xmit = tile_net_tx,
+	.ndo_select_queue = tile_net_select_queue,
+	.ndo_do_ioctl = tile_net_ioctl,
+	.ndo_change_mtu = tile_net_change_mtu,
+	.ndo_tx_timeout = tile_net_tx_timeout,
+	.ndo_set_mac_address = tile_net_set_mac_address,
+#ifdef CONFIG_NET_POLL_CONTROLLER
+	.ndo_poll_controller = tile_net_netpoll,
+#endif
+};
+
+/* The setup function.
+ *
+ * This uses ether_setup() to assign various fields in dev, including
+ * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
+ */
+static void tile_net_setup(struct net_device *dev)
+{
+	netdev_features_t features = 0;
+
+	ether_setup(dev);
+	dev->netdev_ops = &tile_net_ops;
+	dev->watchdog_timeo = TILE_NET_TIMEOUT;
+	dev->mtu = 1500;
+
+	features |= NETIF_F_HW_CSUM;
+	features |= NETIF_F_SG;
+	features |= NETIF_F_TSO;
+	features |= NETIF_F_TSO6;
+
+	dev->hw_features   |= features;
+	dev->vlan_features |= features;
+	dev->features      |= features;
+}
+
+/* Allocate the device structure, register the device, and obtain the
+ * MAC address from the hypervisor.
+ */
+static void tile_net_dev_init(const char *name, const uint8_t *mac)
+{
+	int ret;
+	struct net_device *dev;
+	struct tile_net_priv *priv;
+
+	/* HACK: Ignore "loop" links. */
+	if (strncmp(name, "loop", 4) == 0)
+		return;
+
+	/* Allocate the device structure.  Normally, "name" is a
+	 * template, instantiated by register_netdev(), but not for us.
+	 */
+	dev = alloc_netdev_mqs(sizeof(*priv), name, NET_NAME_UNKNOWN,
+			       tile_net_setup, NR_CPUS, 1);
+	if (!dev) {
+		pr_err("alloc_netdev_mqs(%s) failed\n", name);
+		return;
+	}
+
+	/* Initialize "priv". */
+	priv = netdev_priv(dev);
+	priv->dev = dev;
+	priv->channel = -1;
+	priv->loopify_channel = -1;
+	priv->echannel = -1;
+	init_ptp_dev(priv);
+
+	/* Get the MAC address and set it in the device struct; this must
+	 * be done before the device is opened.  If the MAC is all zeroes,
+	 * we use a random address, since we're probably on the simulator.
+	 */
+	if (!is_zero_ether_addr(mac))
+		ether_addr_copy(dev->dev_addr, mac);
+	else
+		eth_hw_addr_random(dev);
+
+	/* Register the network device. */
+	ret = register_netdev(dev);
+	if (ret) {
+		netdev_err(dev, "register_netdev failed %d\n", ret);
+		free_netdev(dev);
+		return;
+	}
+}
+
+/* Per-cpu module initialization. */
+static void tile_net_init_module_percpu(void *unused)
+{
+	struct tile_net_info *info = this_cpu_ptr(&per_cpu_info);
+	int my_cpu = smp_processor_id();
+	int instance;
+
+	for (instance = 0; instance < NR_MPIPE_MAX; instance++) {
+		info->mpipe[instance].has_iqueue = false;
+		info->mpipe[instance].instance = instance;
+	}
+	info->my_cpu = my_cpu;
+
+	/* Initialize the egress timer. */
+	hrtimer_init(&info->egress_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+	info->egress_timer.function = tile_net_handle_egress_timer;
+}
+
+/* Module initialization. */
+static int __init tile_net_init_module(void)
+{
+	int i;
+	char name[GXIO_MPIPE_LINK_NAME_LEN];
+	uint8_t mac[6];
+
+	pr_info("Tilera Network Driver\n");
+
+	BUILD_BUG_ON(NR_MPIPE_MAX != 2);
+
+	mutex_init(&tile_net_devs_for_channel_mutex);
+
+	/* Initialize each CPU. */
+	on_each_cpu(tile_net_init_module_percpu, NULL, 1);
+
+	/* Find out what devices we have, and initialize them. */
+	for (i = 0; gxio_mpipe_link_enumerate_mac(i, name, mac) >= 0; i++)
+		tile_net_dev_init(name, mac);
+
+	if (!network_cpus_init())
+		network_cpus_map = *cpu_online_mask;
+
+	return 0;
+}
+
+module_init(tile_net_init_module);
diff --git a/drivers/net/ethernet/tile/tilepro.c b/drivers/net/ethernet/tile/tilepro.c
new file mode 100644
index 000000000..3d8f60d96
--- /dev/null
+++ b/drivers/net/ethernet/tile/tilepro.c
@@ -0,0 +1,2423 @@
+/*
+ * Copyright 2011 Tilera Corporation. All Rights Reserved.
+ *
+ *   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, version 2.
+ *
+ *   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, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/moduleparam.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>      /* printk() */
+#include <linux/slab.h>        /* kmalloc() */
+#include <linux/errno.h>       /* error codes */
+#include <linux/types.h>       /* size_t */
+#include <linux/interrupt.h>
+#include <linux/in.h>
+#include <linux/netdevice.h>   /* struct device, and other headers */
+#include <linux/etherdevice.h> /* eth_type_trans */
+#include <linux/skbuff.h>
+#include <linux/ioctl.h>
+#include <linux/cdev.h>
+#include <linux/hugetlb.h>
+#include <linux/in6.h>
+#include <linux/timer.h>
+#include <linux/io.h>
+#include <linux/u64_stats_sync.h>
+#include <asm/checksum.h>
+#include <asm/homecache.h>
+
+#include <hv/drv_xgbe_intf.h>
+#include <hv/drv_xgbe_impl.h>
+#include <hv/hypervisor.h>
+#include <hv/netio_intf.h>
+
+/* For TSO */
+#include <linux/ip.h>
+#include <linux/tcp.h>
+
+
+/*
+ * First, "tile_net_init_module()" initializes all four "devices" which
+ * can be used by linux.
+ *
+ * Then, "ifconfig DEVICE up" calls "tile_net_open()", which analyzes
+ * the network cpus, then uses "tile_net_open_aux()" to initialize
+ * LIPP/LEPP, and then uses "tile_net_open_inner()" to register all
+ * the tiles, provide buffers to LIPP, allow ingress to start, and
+ * turn on hypervisor interrupt handling (and NAPI) on all tiles.
+ *
+ * If registration fails due to the link being down, then "retry_work"
+ * is used to keep calling "tile_net_open_inner()" until it succeeds.
+ *
+ * If "ifconfig DEVICE down" is called, it uses "tile_net_stop()" to
+ * stop egress, drain the LIPP buffers, unregister all the tiles, stop
+ * LIPP/LEPP, and wipe the LEPP queue.
+ *
+ * We start out with the ingress interrupt enabled on each CPU.  When
+ * this interrupt fires, we disable it, and call "napi_schedule()".
+ * This will cause "tile_net_poll()" to be called, which will pull
+ * packets from the netio queue, filtering them out, or passing them
+ * to "netif_receive_skb()".  If our budget is exhausted, we will
+ * return, knowing we will be called again later.  Otherwise, we
+ * reenable the ingress interrupt, and call "napi_complete()".
+ *
+ * HACK: Since disabling the ingress interrupt is not reliable, we
+ * ignore the interrupt if the global "active" flag is false.
+ *
+ *
+ * NOTE: The use of "native_driver" ensures that EPP exists, and that
+ * we are using "LIPP" and "LEPP".
+ *
+ * NOTE: Failing to free completions for an arbitrarily long time
+ * (which is defined to be illegal) does in fact cause bizarre
+ * problems.  The "egress_timer" helps prevent this from happening.
+ */
+
+
+/* HACK: Allow use of "jumbo" packets. */
+/* This should be 1500 if "jumbo" is not set in LIPP. */
+/* This should be at most 10226 (10240 - 14) if "jumbo" is set in LIPP. */
+/* ISSUE: This has not been thoroughly tested (except at 1500). */
+#define TILE_NET_MTU 1500
+
+/* HACK: Define this to verify incoming packets. */
+/* #define TILE_NET_VERIFY_INGRESS */
+
+/* Use 3000 to enable the Linux Traffic Control (QoS) layer, else 0. */
+#define TILE_NET_TX_QUEUE_LEN 0
+
+/* Define to dump packets (prints out the whole packet on tx and rx). */
+/* #define TILE_NET_DUMP_PACKETS */
+
+/* Define to enable debug spew (all PDEBUG's are enabled). */
+/* #define TILE_NET_DEBUG */
+
+
+/* Define to activate paranoia checks. */
+/* #define TILE_NET_PARANOIA */
+
+/* Default transmit lockup timeout period, in jiffies. */
+#define TILE_NET_TIMEOUT (5 * HZ)
+
+/* Default retry interval for bringing up the NetIO interface, in jiffies. */
+#define TILE_NET_RETRY_INTERVAL (5 * HZ)
+
+/* Number of ports (xgbe0, xgbe1, gbe0, gbe1). */
+#define TILE_NET_DEVS 4
+
+
+
+/* Paranoia. */
+#if NET_IP_ALIGN != LIPP_PACKET_PADDING
+#error "NET_IP_ALIGN must match LIPP_PACKET_PADDING."
+#endif
+
+
+/* Debug print. */
+#ifdef TILE_NET_DEBUG
+#define PDEBUG(fmt, args...) net_printk(fmt, ## args)
+#else
+#define PDEBUG(fmt, args...)
+#endif
+
+
+MODULE_AUTHOR("Tilera");
+MODULE_LICENSE("GPL");
+
+
+/*
+ * Queue of incoming packets for a specific cpu and device.
+ *
+ * Includes a pointer to the "system" data, and the actual "user" data.
+ */
+struct tile_netio_queue {
+	netio_queue_impl_t *__system_part;
+	netio_queue_user_impl_t __user_part;
+
+};
+
+
+/*
+ * Statistics counters for a specific cpu and device.
+ */
+struct tile_net_stats_t {
+	struct u64_stats_sync syncp;
+	u64 rx_packets;		/* total packets received	*/
+	u64 tx_packets;		/* total packets transmitted	*/
+	u64 rx_bytes;		/* total bytes received 	*/
+	u64 tx_bytes;		/* total bytes transmitted	*/
+	u64 rx_errors;		/* packets truncated or marked bad by hw */
+	u64 rx_dropped;		/* packets not for us or intf not up */
+};
+
+
+/*
+ * Info for a specific cpu and device.
+ *
+ * ISSUE: There is a "dev" pointer in "napi" as well.
+ */
+struct tile_net_cpu {
+	/* The NAPI struct. */
+	struct napi_struct napi;
+	/* Packet queue. */
+	struct tile_netio_queue queue;
+	/* Statistics. */
+	struct tile_net_stats_t stats;
+	/* True iff NAPI is enabled. */
+	bool napi_enabled;
+	/* True if this tile has successfully registered with the IPP. */
+	bool registered;
+	/* True if the link was down last time we tried to register. */
+	bool link_down;
+	/* True if "egress_timer" is scheduled. */
+	bool egress_timer_scheduled;
+	/* Number of small sk_buffs which must still be provided. */
+	unsigned int num_needed_small_buffers;
+	/* Number of large sk_buffs which must still be provided. */
+	unsigned int num_needed_large_buffers;
+	/* A timer for handling egress completions. */
+	struct timer_list egress_timer;
+};
+
+
+/*
+ * Info for a specific device.
+ */
+struct tile_net_priv {
+	/* Our network device. */
+	struct net_device *dev;
+	/* Pages making up the egress queue. */
+	struct page *eq_pages;
+	/* Address of the actual egress queue. */
+	lepp_queue_t *eq;
+	/* Protects "eq". */
+	spinlock_t eq_lock;
+	/* The hypervisor handle for this interface. */
+	int hv_devhdl;
+	/* The intr bit mask that IDs this device. */
+	u32 intr_id;
+	/* True iff "tile_net_open_aux()" has succeeded. */
+	bool partly_opened;
+	/* True iff the device is "active". */
+	bool active;
+	/* Effective network cpus. */
+	struct cpumask network_cpus_map;
+	/* Number of network cpus. */
+	int network_cpus_count;
+	/* Credits per network cpu. */
+	int network_cpus_credits;
+	/* For NetIO bringup retries. */
+	struct delayed_work retry_work;
+	/* Quick access to per cpu data. */
+	struct tile_net_cpu *cpu[NR_CPUS];
+};
+
+/* Log2 of the number of small pages needed for the egress queue. */
+#define EQ_ORDER  get_order(sizeof(lepp_queue_t))
+/* Size of the egress queue's pages. */
+#define EQ_SIZE   (1 << (PAGE_SHIFT + EQ_ORDER))
+
+/*
+ * The actual devices (xgbe0, xgbe1, gbe0, gbe1).
+ */
+static struct net_device *tile_net_devs[TILE_NET_DEVS];
+
+/*
+ * The "tile_net_cpu" structures for each device.
+ */
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe0);
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_xgbe1);
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe0);
+static DEFINE_PER_CPU(struct tile_net_cpu, hv_gbe1);
+
+
+/*
+ * True if "network_cpus" was specified.
+ */
+static bool network_cpus_used;
+
+/*
+ * The actual cpus in "network_cpus".
+ */
+static struct cpumask network_cpus_map;
+
+
+
+#ifdef TILE_NET_DEBUG
+/*
+ * printk with extra stuff.
+ *
+ * We print the CPU we're running in brackets.
+ */
+static void net_printk(char *fmt, ...)
+{
+	int i;
+	int len;
+	va_list args;
+	static char buf[256];
+
+	len = sprintf(buf, "tile_net[%2.2d]: ", smp_processor_id());
+	va_start(args, fmt);
+	i = vscnprintf(buf + len, sizeof(buf) - len - 1, fmt, args);
+	va_end(args);
+	buf[255] = '\0';
+	pr_notice(buf);
+}
+#endif
+
+
+#ifdef TILE_NET_DUMP_PACKETS
+/*
+ * Dump a packet.
+ */
+static void dump_packet(unsigned char *data, unsigned long length, char *s)
+{
+	int my_cpu = smp_processor_id();
+
+	unsigned long i;
+	char buf[128];
+
+	static unsigned int count;
+
+	pr_info("dump_packet(data %p, length 0x%lx s %s count 0x%x)\n",
+	       data, length, s, count++);
+
+	pr_info("\n");
+
+	for (i = 0; i < length; i++) {
+		if ((i & 0xf) == 0)
+			sprintf(buf, "[%02d] %8.8lx:", my_cpu, i);
+		sprintf(buf + strlen(buf), " %2.2x", data[i]);
+		if ((i & 0xf) == 0xf || i == length - 1) {
+			strcat(buf, "\n");
+			pr_info("%s", buf);
+		}
+	}
+}
+#endif
+
+
+/*
+ * Provide support for the __netio_fastio1() swint
+ * (see <hv/drv_xgbe_intf.h> for how it is used).
+ *
+ * The fastio swint2 call may clobber all the caller-saved registers.
+ * It rarely clobbers memory, but we allow for the possibility in
+ * the signature just to be on the safe side.
+ *
+ * Also, gcc doesn't seem to allow an input operand to be
+ * clobbered, so we fake it with dummy outputs.
+ *
+ * This function can't be static because of the way it is declared
+ * in the netio header.
+ */
+inline int __netio_fastio1(u32 fastio_index, u32 arg0)
+{
+	long result, clobber_r1, clobber_r10;
+	asm volatile("swint2"
+		     : "=R00" (result),
+		       "=R01" (clobber_r1), "=R10" (clobber_r10)
+		     : "R10" (fastio_index), "R01" (arg0)
+		     : "memory", "r2", "r3", "r4",
+		       "r5", "r6", "r7", "r8", "r9",
+		       "r11", "r12", "r13", "r14",
+		       "r15", "r16", "r17", "r18", "r19",
+		       "r20", "r21", "r22", "r23", "r24",
+		       "r25", "r26", "r27", "r28", "r29");
+	return result;
+}
+
+
+static void tile_net_return_credit(struct tile_net_cpu *info)
+{
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	/* Return four credits after every fourth packet. */
+	if (--qup->__receive_credit_remaining == 0) {
+		u32 interval = qup->__receive_credit_interval;
+		qup->__receive_credit_remaining = interval;
+		__netio_fastio_return_credits(qup->__fastio_index, interval);
+	}
+}
+
+
+
+/*
+ * Provide a linux buffer to LIPP.
+ */
+static void tile_net_provide_linux_buffer(struct tile_net_cpu *info,
+					  void *va, bool small)
+{
+	struct tile_netio_queue *queue = &info->queue;
+
+	/* Convert "va" and "small" to "linux_buffer_t". */
+	unsigned int buffer = ((unsigned int)(__pa(va) >> 7) << 1) + small;
+
+	__netio_fastio_free_buffer(queue->__user_part.__fastio_index, buffer);
+}
+
+
+/*
+ * Provide a linux buffer for LIPP.
+ *
+ * Note that the ACTUAL allocation for each buffer is a "struct sk_buff",
+ * plus a chunk of memory that includes not only the requested bytes, but
+ * also NET_SKB_PAD bytes of initial padding, and a "struct skb_shared_info".
+ *
+ * Note that "struct skb_shared_info" is 88 bytes with 64K pages and
+ * 268 bytes with 4K pages (since the frags[] array needs 18 entries).
+ *
+ * Without jumbo packets, the maximum packet size will be 1536 bytes,
+ * and we use 2 bytes (NET_IP_ALIGN) of padding.  ISSUE: If we told
+ * the hardware to clip at 1518 bytes instead of 1536 bytes, then we
+ * could save an entire cache line, but in practice, we don't need it.
+ *
+ * Since CPAs are 38 bits, and we can only encode the high 31 bits in
+ * a "linux_buffer_t", the low 7 bits must be zero, and thus, we must
+ * align the actual "va" mod 128.
+ *
+ * We assume that the underlying "head" will be aligned mod 64.  Note
+ * that in practice, we have seen "head" NOT aligned mod 128 even when
+ * using 2048 byte allocations, which is surprising.
+ *
+ * If "head" WAS always aligned mod 128, we could change LIPP to
+ * assume that the low SIX bits are zero, and the 7th bit is one, that
+ * is, align the actual "va" mod 128 plus 64, which would be "free".
+ *
+ * For now, the actual "head" pointer points at NET_SKB_PAD bytes of
+ * padding, plus 28 or 92 bytes of extra padding, plus the sk_buff
+ * pointer, plus the NET_IP_ALIGN padding, plus 126 or 1536 bytes for
+ * the actual packet, plus 62 bytes of empty padding, plus some
+ * padding and the "struct skb_shared_info".
+ *
+ * With 64K pages, a large buffer thus needs 32+92+4+2+1536+62+88
+ * bytes, or 1816 bytes, which fits comfortably into 2048 bytes.
+ *
+ * With 64K pages, a small buffer thus needs 32+92+4+2+126+88
+ * bytes, or 344 bytes, which means we are wasting 64+ bytes, and
+ * could presumably increase the size of small buffers.
+ *
+ * With 4K pages, a large buffer thus needs 32+92+4+2+1536+62+268
+ * bytes, or 1996 bytes, which fits comfortably into 2048 bytes.
+ *
+ * With 4K pages, a small buffer thus needs 32+92+4+2+126+268
+ * bytes, or 524 bytes, which is annoyingly wasteful.
+ *
+ * Maybe we should increase LIPP_SMALL_PACKET_SIZE to 192?
+ *
+ * ISSUE: Maybe we should increase "NET_SKB_PAD" to 64?
+ */
+static bool tile_net_provide_needed_buffer(struct tile_net_cpu *info,
+					   bool small)
+{
+#if TILE_NET_MTU <= 1536
+	/* Without "jumbo", 2 + 1536 should be sufficient. */
+	unsigned int large_size = NET_IP_ALIGN + 1536;
+#else
+	/* ISSUE: This has not been tested. */
+	unsigned int large_size = NET_IP_ALIGN + TILE_NET_MTU + 100;
+#endif
+
+	/* Avoid "false sharing" with last cache line. */
+	/* ISSUE: This is already done by "netdev_alloc_skb()". */
+	unsigned int len =
+		 (((small ? LIPP_SMALL_PACKET_SIZE : large_size) +
+		   CHIP_L2_LINE_SIZE() - 1) & -CHIP_L2_LINE_SIZE());
+
+	unsigned int padding = 128 - NET_SKB_PAD;
+	unsigned int align;
+
+	struct sk_buff *skb;
+	void *va;
+
+	struct sk_buff **skb_ptr;
+
+	/* Request 96 extra bytes for alignment purposes. */
+	skb = netdev_alloc_skb(info->napi.dev, len + padding);
+	if (skb == NULL)
+		return false;
+
+	/* Skip 32 or 96 bytes to align "data" mod 128. */
+	align = -(long)skb->data & (128 - 1);
+	BUG_ON(align > padding);
+	skb_reserve(skb, align);
+
+	/* This address is given to IPP. */
+	va = skb->data;
+
+	/* Buffers must not span a huge page. */
+	BUG_ON(((((long)va & ~HPAGE_MASK) + len) & HPAGE_MASK) != 0);
+
+#ifdef TILE_NET_PARANOIA
+#if CHIP_HAS_CBOX_HOME_MAP()
+	if (hash_default) {
+		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)va);
+		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
+			panic("Non-HFH ingress buffer! VA=%p Mode=%d PTE=%llx",
+			      va, hv_pte_get_mode(pte), hv_pte_val(pte));
+	}
+#endif
+#endif
+
+	/* Invalidate the packet buffer. */
+	if (!hash_default)
+		__inv_buffer(va, len);
+
+	/* Skip two bytes to satisfy LIPP assumptions. */
+	/* Note that this aligns IP on a 16 byte boundary. */
+	/* ISSUE: Do this when the packet arrives? */
+	skb_reserve(skb, NET_IP_ALIGN);
+
+	/* Save a back-pointer to 'skb'. */
+	skb_ptr = va - sizeof(*skb_ptr);
+	*skb_ptr = skb;
+
+	/* Make sure "skb_ptr" has been flushed. */
+	__insn_mf();
+
+	/* Provide the new buffer. */
+	tile_net_provide_linux_buffer(info, va, small);
+
+	return true;
+}
+
+
+/*
+ * Provide linux buffers for LIPP.
+ */
+static void tile_net_provide_needed_buffers(struct tile_net_cpu *info)
+{
+	while (info->num_needed_small_buffers != 0) {
+		if (!tile_net_provide_needed_buffer(info, true))
+			goto oops;
+		info->num_needed_small_buffers--;
+	}
+
+	while (info->num_needed_large_buffers != 0) {
+		if (!tile_net_provide_needed_buffer(info, false))
+			goto oops;
+		info->num_needed_large_buffers--;
+	}
+
+	return;
+
+oops:
+
+	/* Add a description to the page allocation failure dump. */
+	pr_notice("Could not provide a linux buffer to LIPP.\n");
+}
+
+
+/*
+ * Grab some LEPP completions, and store them in "comps", of size
+ * "comps_size", and return the number of completions which were
+ * stored, so the caller can free them.
+ */
+static unsigned int tile_net_lepp_grab_comps(lepp_queue_t *eq,
+					     struct sk_buff *comps[],
+					     unsigned int comps_size,
+					     unsigned int min_size)
+{
+	unsigned int n = 0;
+
+	unsigned int comp_head = eq->comp_head;
+	unsigned int comp_busy = eq->comp_busy;
+
+	while (comp_head != comp_busy && n < comps_size) {
+		comps[n++] = eq->comps[comp_head];
+		LEPP_QINC(comp_head);
+	}
+
+	if (n < min_size)
+		return 0;
+
+	eq->comp_head = comp_head;
+
+	return n;
+}
+
+
+/*
+ * Free some comps, and return true iff there are still some pending.
+ */
+static bool tile_net_lepp_free_comps(struct net_device *dev, bool all)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	lepp_queue_t *eq = priv->eq;
+
+	struct sk_buff *olds[64];
+	unsigned int wanted = 64;
+	unsigned int i, n;
+	bool pending;
+
+	spin_lock(&priv->eq_lock);
+
+	if (all)
+		eq->comp_busy = eq->comp_tail;
+
+	n = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+
+	pending = (eq->comp_head != eq->comp_tail);
+
+	spin_unlock(&priv->eq_lock);
+
+	for (i = 0; i < n; i++)
+		kfree_skb(olds[i]);
+
+	return pending;
+}
+
+
+/*
+ * Make sure the egress timer is scheduled.
+ *
+ * Note that we use "schedule if not scheduled" logic instead of the more
+ * obvious "reschedule" logic, because "reschedule" is fairly expensive.
+ */
+static void tile_net_schedule_egress_timer(struct tile_net_cpu *info)
+{
+	if (!info->egress_timer_scheduled) {
+		mod_timer_pinned(&info->egress_timer, jiffies + 1);
+		info->egress_timer_scheduled = true;
+	}
+}
+
+
+/*
+ * The "function" for "info->egress_timer".
+ *
+ * This timer will reschedule itself as long as there are any pending
+ * completions expected (on behalf of any tile).
+ *
+ * ISSUE: Realistically, will the timer ever stop scheduling itself?
+ *
+ * ISSUE: This timer is almost never actually needed, so just use a global
+ * timer that can run on any tile.
+ *
+ * ISSUE: Maybe instead track number of expected completions, and free
+ * only that many, resetting to zero if "pending" is ever false.
+ */
+static void tile_net_handle_egress_timer(unsigned long arg)
+{
+	struct tile_net_cpu *info = (struct tile_net_cpu *)arg;
+	struct net_device *dev = info->napi.dev;
+
+	/* The timer is no longer scheduled. */
+	info->egress_timer_scheduled = false;
+
+	/* Free comps, and reschedule timer if more are pending. */
+	if (tile_net_lepp_free_comps(dev, false))
+		tile_net_schedule_egress_timer(info);
+}
+
+
+static void tile_net_discard_aux(struct tile_net_cpu *info, int index)
+{
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	int index2_aux = index + sizeof(netio_pkt_t);
+	int index2 =
+		((index2_aux ==
+		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
+		 0 : index2_aux);
+
+	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
+
+	/* Extract the "linux_buffer_t". */
+	unsigned int buffer = pkt->__packet.word;
+
+	/* Convert "linux_buffer_t" to "va". */
+	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
+
+	/* Acquire the associated "skb". */
+	struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+	struct sk_buff *skb = *skb_ptr;
+
+	kfree_skb(skb);
+
+	/* Consume this packet. */
+	qup->__packet_receive_read = index2;
+}
+
+
+/*
+ * Like "tile_net_poll()", but just discard packets.
+ */
+static void tile_net_discard_packets(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	while (qup->__packet_receive_read !=
+	       qsp->__packet_receive_queue.__packet_write) {
+		int index = qup->__packet_receive_read;
+		tile_net_discard_aux(info, index);
+	}
+}
+
+
+/*
+ * Handle the next packet.  Return true if "processed", false if "filtered".
+ */
+static bool tile_net_poll_aux(struct tile_net_cpu *info, int index)
+{
+	struct net_device *dev = info->napi.dev;
+
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+	struct tile_net_stats_t *stats = &info->stats;
+
+	int filter;
+
+	int index2_aux = index + sizeof(netio_pkt_t);
+	int index2 =
+		((index2_aux ==
+		  qsp->__packet_receive_queue.__last_packet_plus_one) ?
+		 0 : index2_aux);
+
+	netio_pkt_t *pkt = (netio_pkt_t *)((unsigned long) &qsp[1] + index);
+
+	netio_pkt_metadata_t *metadata = NETIO_PKT_METADATA(pkt);
+	netio_pkt_status_t pkt_status = NETIO_PKT_STATUS_M(metadata, pkt);
+
+	/* Extract the packet size.  FIXME: Shouldn't the second line */
+	/* get subtracted?  Mostly moot, since it should be "zero". */
+	unsigned long len =
+		(NETIO_PKT_CUSTOM_LENGTH(pkt) +
+		 NET_IP_ALIGN - NETIO_PACKET_PADDING);
+
+	/* Extract the "linux_buffer_t". */
+	unsigned int buffer = pkt->__packet.word;
+
+	/* Extract "small" (vs "large"). */
+	bool small = ((buffer & 1) != 0);
+
+	/* Convert "linux_buffer_t" to "va". */
+	void *va = __va((phys_addr_t)(buffer >> 1) << 7);
+
+	/* Extract the packet data pointer. */
+	/* Compare to "NETIO_PKT_CUSTOM_DATA(pkt)". */
+	unsigned char *buf = va + NET_IP_ALIGN;
+
+	/* Invalidate the packet buffer. */
+	if (!hash_default)
+		__inv_buffer(buf, len);
+
+	/* ISSUE: Is this needed? */
+	dev->last_rx = jiffies;
+
+#ifdef TILE_NET_DUMP_PACKETS
+	dump_packet(buf, len, "rx");
+#endif /* TILE_NET_DUMP_PACKETS */
+
+#ifdef TILE_NET_VERIFY_INGRESS
+	if (pkt_status == NETIO_PKT_STATUS_OVERSIZE && len >= 64) {
+		dump_packet(buf, len, "rx");
+		panic("Unexpected OVERSIZE.");
+	}
+#endif
+
+	filter = 0;
+
+	if (pkt_status == NETIO_PKT_STATUS_BAD) {
+		/* Handle CRC error and hardware truncation. */
+		filter = 2;
+	} else if (!(dev->flags & IFF_UP)) {
+		/* Filter packets received before we're up. */
+		filter = 1;
+	} else if (NETIO_PKT_ETHERTYPE_RECOGNIZED_M(metadata, pkt) &&
+		   pkt_status == NETIO_PKT_STATUS_UNDERSIZE) {
+		/* Filter "truncated" packets. */
+		filter = 2;
+	} else if (!(dev->flags & IFF_PROMISC)) {
+		if (!is_multicast_ether_addr(buf)) {
+			/* Filter packets not for our address. */
+			const u8 *mine = dev->dev_addr;
+			filter = !ether_addr_equal(mine, buf);
+		}
+	}
+
+	u64_stats_update_begin(&stats->syncp);
+
+	if (filter != 0) {
+
+		if (filter == 1)
+			stats->rx_dropped++;
+		else
+			stats->rx_errors++;
+
+		tile_net_provide_linux_buffer(info, va, small);
+
+	} else {
+
+		/* Acquire the associated "skb". */
+		struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+		struct sk_buff *skb = *skb_ptr;
+
+		/* Paranoia. */
+		if (skb->data != buf)
+			panic("Corrupt linux buffer from LIPP! "
+			      "VA=%p, skb=%p, skb->data=%p\n",
+			      va, skb, skb->data);
+
+		/* Encode the actual packet length. */
+		skb_put(skb, len);
+
+		/* NOTE: This call also sets "skb->dev = dev". */
+		skb->protocol = eth_type_trans(skb, dev);
+
+		/* Avoid recomputing "good" TCP/UDP checksums. */
+		if (NETIO_PKT_L4_CSUM_CORRECT_M(metadata, pkt))
+			skb->ip_summed = CHECKSUM_UNNECESSARY;
+
+		netif_receive_skb(skb);
+
+		stats->rx_packets++;
+		stats->rx_bytes += len;
+	}
+
+	u64_stats_update_end(&stats->syncp);
+
+	/* ISSUE: It would be nice to defer this until the packet has */
+	/* actually been processed. */
+	tile_net_return_credit(info);
+
+	/* Consume this packet. */
+	qup->__packet_receive_read = index2;
+
+	return !filter;
+}
+
+
+/*
+ * Handle some packets for the given device on the current CPU.
+ *
+ * If "tile_net_stop()" is called on some other tile while this
+ * function is running, we will return, hopefully before that
+ * other tile asks us to call "napi_disable()".
+ *
+ * The "rotting packet" race condition occurs if a packet arrives
+ * during the extremely narrow window between the queue appearing to
+ * be empty, and the ingress interrupt being re-enabled.  This happens
+ * a LOT under heavy network load.
+ */
+static int tile_net_poll(struct napi_struct *napi, int budget)
+{
+	struct net_device *dev = napi->dev;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+	struct tile_netio_queue *queue = &info->queue;
+	netio_queue_impl_t *qsp = queue->__system_part;
+	netio_queue_user_impl_t *qup = &queue->__user_part;
+
+	unsigned int work = 0;
+
+	if (budget <= 0)
+		goto done;
+
+	while (priv->active) {
+		int index = qup->__packet_receive_read;
+		if (index == qsp->__packet_receive_queue.__packet_write)
+			break;
+
+		if (tile_net_poll_aux(info, index)) {
+			if (++work >= budget)
+				goto done;
+		}
+	}
+
+	napi_complete(&info->napi);
+
+	if (!priv->active)
+		goto done;
+
+	/* Re-enable the ingress interrupt. */
+	enable_percpu_irq(priv->intr_id, 0);
+
+	/* HACK: Avoid the "rotting packet" problem (see above). */
+	if (qup->__packet_receive_read !=
+	    qsp->__packet_receive_queue.__packet_write) {
+		/* ISSUE: Sometimes this returns zero, presumably */
+		/* because an interrupt was handled for this tile. */
+		(void)napi_reschedule(&info->napi);
+	}
+
+done:
+
+	if (priv->active)
+		tile_net_provide_needed_buffers(info);
+
+	return work;
+}
+
+
+/*
+ * Handle an ingress interrupt for the given device on the current cpu.
+ *
+ * ISSUE: Sometimes this gets called after "disable_percpu_irq()" has
+ * been called!  This is probably due to "pending hypervisor downcalls".
+ *
+ * ISSUE: Is there any race condition between the "napi_schedule()" here
+ * and the "napi_complete()" call above?
+ */
+static irqreturn_t tile_net_handle_ingress_interrupt(int irq, void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	/* Disable the ingress interrupt. */
+	disable_percpu_irq(priv->intr_id);
+
+	/* Ignore unwanted interrupts. */
+	if (!priv->active)
+		return IRQ_HANDLED;
+
+	/* ISSUE: Sometimes "info->napi_enabled" is false here. */
+
+	napi_schedule(&info->napi);
+
+	return IRQ_HANDLED;
+}
+
+
+/*
+ * One time initialization per interface.
+ */
+static int tile_net_open_aux(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	int ret;
+	int dummy;
+	unsigned int epp_lotar;
+
+	/*
+	 * Find out where EPP memory should be homed.
+	 */
+	ret = hv_dev_pread(priv->hv_devhdl, 0,
+			   (HV_VirtAddr)&epp_lotar, sizeof(epp_lotar),
+			   NETIO_EPP_SHM_OFF);
+	if (ret < 0) {
+		pr_err("could not read epp_shm_queue lotar.\n");
+		return -EIO;
+	}
+
+	/*
+	 * Home the page on the EPP.
+	 */
+	{
+		int epp_home = hv_lotar_to_cpu(epp_lotar);
+		homecache_change_page_home(priv->eq_pages, EQ_ORDER, epp_home);
+	}
+
+	/*
+	 * Register the EPP shared memory queue.
+	 */
+	{
+		netio_ipp_address_t ea = {
+			.va = 0,
+			.pa = __pa(priv->eq),
+			.pte = hv_pte(0),
+			.size = EQ_SIZE,
+		};
+		ea.pte = hv_pte_set_lotar(ea.pte, epp_lotar);
+		ea.pte = hv_pte_set_mode(ea.pte, HV_PTE_MODE_CACHE_TILE_L3);
+		ret = hv_dev_pwrite(priv->hv_devhdl, 0,
+				    (HV_VirtAddr)&ea,
+				    sizeof(ea),
+				    NETIO_EPP_SHM_OFF);
+		if (ret < 0)
+			return -EIO;
+	}
+
+	/*
+	 * Start LIPP/LEPP.
+	 */
+	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
+			  sizeof(dummy), NETIO_IPP_START_SHIM_OFF) < 0) {
+		pr_warn("Failed to start LIPP/LEPP\n");
+		return -EIO;
+	}
+
+	return 0;
+}
+
+
+/*
+ * Register with hypervisor on the current CPU.
+ *
+ * Strangely, this function does important things even if it "fails",
+ * which is especially common if the link is not up yet.  Hopefully
+ * these things are all "harmless" if done twice!
+ */
+static void tile_net_register(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info;
+
+	struct tile_netio_queue *queue;
+
+	/* Only network cpus can receive packets. */
+	int queue_id =
+		cpumask_test_cpu(my_cpu, &priv->network_cpus_map) ? 0 : 255;
+
+	netio_input_config_t config = {
+		.flags = 0,
+		.num_receive_packets = priv->network_cpus_credits,
+		.queue_id = queue_id
+	};
+
+	int ret = 0;
+	netio_queue_impl_t *queuep;
+
+	PDEBUG("tile_net_register(queue_id %d)\n", queue_id);
+
+	if (!strcmp(dev->name, "xgbe0"))
+		info = this_cpu_ptr(&hv_xgbe0);
+	else if (!strcmp(dev->name, "xgbe1"))
+		info = this_cpu_ptr(&hv_xgbe1);
+	else if (!strcmp(dev->name, "gbe0"))
+		info = this_cpu_ptr(&hv_gbe0);
+	else if (!strcmp(dev->name, "gbe1"))
+		info = this_cpu_ptr(&hv_gbe1);
+	else
+		BUG();
+
+	/* Initialize the egress timer. */
+	init_timer(&info->egress_timer);
+	info->egress_timer.data = (long)info;
+	info->egress_timer.function = tile_net_handle_egress_timer;
+
+	u64_stats_init(&info->stats.syncp);
+
+	priv->cpu[my_cpu] = info;
+
+	/*
+	 * Register ourselves with LIPP.  This does a lot of stuff,
+	 * including invoking the LIPP registration code.
+	 */
+	ret = hv_dev_pwrite(priv->hv_devhdl, 0,
+			    (HV_VirtAddr)&config,
+			    sizeof(netio_input_config_t),
+			    NETIO_IPP_INPUT_REGISTER_OFF);
+	PDEBUG("hv_dev_pwrite(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
+	       ret);
+	if (ret < 0) {
+		if (ret != NETIO_LINK_DOWN) {
+			printk(KERN_DEBUG "hv_dev_pwrite "
+			       "NETIO_IPP_INPUT_REGISTER_OFF failure %d\n",
+			       ret);
+		}
+		info->link_down = (ret == NETIO_LINK_DOWN);
+		return;
+	}
+
+	/*
+	 * Get the pointer to our queue's system part.
+	 */
+
+	ret = hv_dev_pread(priv->hv_devhdl, 0,
+			   (HV_VirtAddr)&queuep,
+			   sizeof(netio_queue_impl_t *),
+			   NETIO_IPP_INPUT_REGISTER_OFF);
+	PDEBUG("hv_dev_pread(NETIO_IPP_INPUT_REGISTER_OFF) returned %d\n",
+	       ret);
+	PDEBUG("queuep %p\n", queuep);
+	if (ret <= 0) {
+		/* ISSUE: Shouldn't this be a fatal error? */
+		pr_err("hv_dev_pread NETIO_IPP_INPUT_REGISTER_OFF failure\n");
+		return;
+	}
+
+	queue = &info->queue;
+
+	queue->__system_part = queuep;
+
+	memset(&queue->__user_part, 0, sizeof(netio_queue_user_impl_t));
+
+	/* This is traditionally "config.num_receive_packets / 2". */
+	queue->__user_part.__receive_credit_interval = 4;
+	queue->__user_part.__receive_credit_remaining =
+		queue->__user_part.__receive_credit_interval;
+
+	/*
+	 * Get a fastio index from the hypervisor.
+	 * ISSUE: Shouldn't this check the result?
+	 */
+	ret = hv_dev_pread(priv->hv_devhdl, 0,
+			   (HV_VirtAddr)&queue->__user_part.__fastio_index,
+			   sizeof(queue->__user_part.__fastio_index),
+			   NETIO_IPP_GET_FASTIO_OFF);
+	PDEBUG("hv_dev_pread(NETIO_IPP_GET_FASTIO_OFF) returned %d\n", ret);
+
+	/* Now we are registered. */
+	info->registered = true;
+}
+
+
+/*
+ * Deregister with hypervisor on the current CPU.
+ *
+ * This simply discards all our credits, so no more packets will be
+ * delivered to this tile.  There may still be packets in our queue.
+ *
+ * Also, disable the ingress interrupt.
+ */
+static void tile_net_deregister(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	/* Disable the ingress interrupt. */
+	disable_percpu_irq(priv->intr_id);
+
+	/* Do nothing else if not registered. */
+	if (info == NULL || !info->registered)
+		return;
+
+	{
+		struct tile_netio_queue *queue = &info->queue;
+		netio_queue_user_impl_t *qup = &queue->__user_part;
+
+		/* Discard all our credits. */
+		__netio_fastio_return_credits(qup->__fastio_index, -1);
+	}
+}
+
+
+/*
+ * Unregister with hypervisor on the current CPU.
+ *
+ * Also, disable the ingress interrupt.
+ */
+static void tile_net_unregister(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	int ret;
+	int dummy = 0;
+
+	/* Disable the ingress interrupt. */
+	disable_percpu_irq(priv->intr_id);
+
+	/* Do nothing else if not registered. */
+	if (info == NULL || !info->registered)
+		return;
+
+	/* Unregister ourselves with LIPP/LEPP. */
+	ret = hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
+			    sizeof(dummy), NETIO_IPP_INPUT_UNREGISTER_OFF);
+	if (ret < 0)
+		panic("Failed to unregister with LIPP/LEPP!\n");
+
+	/* Discard all packets still in our NetIO queue. */
+	tile_net_discard_packets(dev);
+
+	/* Reset state. */
+	info->num_needed_small_buffers = 0;
+	info->num_needed_large_buffers = 0;
+
+	/* Cancel egress timer. */
+	del_timer(&info->egress_timer);
+	info->egress_timer_scheduled = false;
+}
+
+
+/*
+ * Helper function for "tile_net_stop()".
+ *
+ * Also used to handle registration failure in "tile_net_open_inner()",
+ * when the various extra steps in "tile_net_stop()" are not necessary.
+ */
+static void tile_net_stop_aux(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int i;
+
+	int dummy = 0;
+
+	/*
+	 * Unregister all tiles, so LIPP will stop delivering packets.
+	 * Also, delete all the "napi" objects (sequentially, to protect
+	 * "dev->napi_list").
+	 */
+	on_each_cpu(tile_net_unregister, (void *)dev, 1);
+	for_each_online_cpu(i) {
+		struct tile_net_cpu *info = priv->cpu[i];
+		if (info != NULL && info->registered) {
+			netif_napi_del(&info->napi);
+			info->registered = false;
+		}
+	}
+
+	/* Stop LIPP/LEPP. */
+	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
+			  sizeof(dummy), NETIO_IPP_STOP_SHIM_OFF) < 0)
+		panic("Failed to stop LIPP/LEPP!\n");
+
+	priv->partly_opened = false;
+}
+
+
+/*
+ * Disable NAPI for the given device on the current cpu.
+ */
+static void tile_net_stop_disable(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	/* Disable NAPI if needed. */
+	if (info != NULL && info->napi_enabled) {
+		napi_disable(&info->napi);
+		info->napi_enabled = false;
+	}
+}
+
+
+/*
+ * Enable NAPI and the ingress interrupt for the given device
+ * on the current cpu.
+ *
+ * ISSUE: Only do this for "network cpus"?
+ */
+static void tile_net_open_enable(void *dev_ptr)
+{
+	struct net_device *dev = (struct net_device *)dev_ptr;
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+
+	/* Enable NAPI. */
+	napi_enable(&info->napi);
+	info->napi_enabled = true;
+
+	/* Enable the ingress interrupt. */
+	enable_percpu_irq(priv->intr_id, 0);
+}
+
+
+/*
+ * tile_net_open_inner does most of the work of bringing up the interface.
+ * It's called from tile_net_open(), and also from tile_net_retry_open().
+ * The return value is 0 if the interface was brought up, < 0 if
+ * tile_net_open() should return the return value as an error, and > 0 if
+ * tile_net_open() should return success and schedule a work item to
+ * periodically retry the bringup.
+ */
+static int tile_net_open_inner(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info;
+	struct tile_netio_queue *queue;
+	int result = 0;
+	int i;
+	int dummy = 0;
+
+	/*
+	 * First try to register just on the local CPU, and handle any
+	 * semi-expected "link down" failure specially.  Note that we
+	 * do NOT call "tile_net_stop_aux()", unlike below.
+	 */
+	tile_net_register(dev);
+	info = priv->cpu[my_cpu];
+	if (!info->registered) {
+		if (info->link_down)
+			return 1;
+		return -EAGAIN;
+	}
+
+	/*
+	 * Now register everywhere else.  If any registration fails,
+	 * even for "link down" (which might not be possible), we
+	 * clean up using "tile_net_stop_aux()".  Also, add all the
+	 * "napi" objects (sequentially, to protect "dev->napi_list").
+	 * ISSUE: Only use "netif_napi_add()" for "network cpus"?
+	 */
+	smp_call_function(tile_net_register, (void *)dev, 1);
+	for_each_online_cpu(i) {
+		struct tile_net_cpu *info = priv->cpu[i];
+		if (info->registered)
+			netif_napi_add(dev, &info->napi, tile_net_poll, 64);
+		else
+			result = -EAGAIN;
+	}
+	if (result != 0) {
+		tile_net_stop_aux(dev);
+		return result;
+	}
+
+	queue = &info->queue;
+
+	if (priv->intr_id == 0) {
+		unsigned int irq;
+
+		/*
+		 * Acquire the irq allocated by the hypervisor.  Every
+		 * queue gets the same irq.  The "__intr_id" field is
+		 * "1 << irq", so we use "__ffs()" to extract "irq".
+		 */
+		priv->intr_id = queue->__system_part->__intr_id;
+		BUG_ON(priv->intr_id == 0);
+		irq = __ffs(priv->intr_id);
+
+		/*
+		 * Register the ingress interrupt handler for this
+		 * device, permanently.
+		 *
+		 * We used to call "free_irq()" in "tile_net_stop()",
+		 * and then re-register the handler here every time,
+		 * but that caused DNP errors in "handle_IRQ_event()"
+		 * because "desc->action" was NULL.  See bug 9143.
+		 */
+		tile_irq_activate(irq, TILE_IRQ_PERCPU);
+		BUG_ON(request_irq(irq, tile_net_handle_ingress_interrupt,
+				   0, dev->name, (void *)dev) != 0);
+	}
+
+	{
+		/* Allocate initial buffers. */
+
+		int max_buffers =
+			priv->network_cpus_count * priv->network_cpus_credits;
+
+		info->num_needed_small_buffers =
+			min(LIPP_SMALL_BUFFERS, max_buffers);
+
+		info->num_needed_large_buffers =
+			min(LIPP_LARGE_BUFFERS, max_buffers);
+
+		tile_net_provide_needed_buffers(info);
+
+		if (info->num_needed_small_buffers != 0 ||
+		    info->num_needed_large_buffers != 0)
+			panic("Insufficient memory for buffer stack!");
+	}
+
+	/* We are about to be active. */
+	priv->active = true;
+
+	/* Make sure "active" is visible to all tiles. */
+	mb();
+
+	/* On each tile, enable NAPI and the ingress interrupt. */
+	on_each_cpu(tile_net_open_enable, (void *)dev, 1);
+
+	/* Start LIPP/LEPP and activate "ingress" at the shim. */
+	if (hv_dev_pwrite(priv->hv_devhdl, 0, (HV_VirtAddr)&dummy,
+			  sizeof(dummy), NETIO_IPP_INPUT_INIT_OFF) < 0)
+		panic("Failed to activate the LIPP Shim!\n");
+
+	/* Start our transmit queue. */
+	netif_start_queue(dev);
+
+	return 0;
+}
+
+
+/*
+ * Called periodically to retry bringing up the NetIO interface,
+ * if it doesn't come up cleanly during tile_net_open().
+ */
+static void tile_net_open_retry(struct work_struct *w)
+{
+	struct delayed_work *dw =
+		container_of(w, struct delayed_work, work);
+
+	struct tile_net_priv *priv =
+		container_of(dw, struct tile_net_priv, retry_work);
+
+	/*
+	 * Try to bring the NetIO interface up.  If it fails, reschedule
+	 * ourselves to try again later; otherwise, tell Linux we now have
+	 * a working link.  ISSUE: What if the return value is negative?
+	 */
+	if (tile_net_open_inner(priv->dev) != 0)
+		schedule_delayed_work(&priv->retry_work,
+				      TILE_NET_RETRY_INTERVAL);
+	else
+		netif_carrier_on(priv->dev);
+}
+
+
+/*
+ * Called when a network interface is made active.
+ *
+ * Returns 0 on success, negative value on failure.
+ *
+ * The open entry point is called when a network interface is made
+ * active by the system (IFF_UP).  At this point all resources needed
+ * for transmit and receive operations are allocated, the interrupt
+ * handler is registered with the OS (if needed), the watchdog timer
+ * is started, and the stack is notified that the interface is ready.
+ *
+ * If the actual link is not available yet, then we tell Linux that
+ * we have no carrier, and we keep checking until the link comes up.
+ */
+static int tile_net_open(struct net_device *dev)
+{
+	int ret = 0;
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	/*
+	 * We rely on priv->partly_opened to tell us if this is the
+	 * first time this interface is being brought up. If it is
+	 * set, the IPP was already initialized and should not be
+	 * initialized again.
+	 */
+	if (!priv->partly_opened) {
+
+		int count;
+		int credits;
+
+		/* Initialize LIPP/LEPP, and start the Shim. */
+		ret = tile_net_open_aux(dev);
+		if (ret < 0) {
+			pr_err("tile_net_open_aux failed: %d\n", ret);
+			return ret;
+		}
+
+		/* Analyze the network cpus. */
+
+		if (network_cpus_used)
+			cpumask_copy(&priv->network_cpus_map,
+				     &network_cpus_map);
+		else
+			cpumask_copy(&priv->network_cpus_map, cpu_online_mask);
+
+
+		count = cpumask_weight(&priv->network_cpus_map);
+
+		/* Limit credits to available buffers, and apply min. */
+		credits = max(16, (LIPP_LARGE_BUFFERS / count) & ~1);
+
+		/* Apply "GBE" max limit. */
+		/* ISSUE: Use higher limit for XGBE? */
+		credits = min(NETIO_MAX_RECEIVE_PKTS, credits);
+
+		priv->network_cpus_count = count;
+		priv->network_cpus_credits = credits;
+
+#ifdef TILE_NET_DEBUG
+		pr_info("Using %d network cpus, with %d credits each\n",
+		       priv->network_cpus_count, priv->network_cpus_credits);
+#endif
+
+		priv->partly_opened = true;
+
+	} else {
+		/* FIXME: Is this possible? */
+		/* printk("Already partly opened.\n"); */
+	}
+
+	/*
+	 * Attempt to bring up the link.
+	 */
+	ret = tile_net_open_inner(dev);
+	if (ret <= 0) {
+		if (ret == 0)
+			netif_carrier_on(dev);
+		return ret;
+	}
+
+	/*
+	 * We were unable to bring up the NetIO interface, but we want to
+	 * try again in a little bit.  Tell Linux that we have no carrier
+	 * so it doesn't try to use the interface before the link comes up
+	 * and then remember to try again later.
+	 */
+	netif_carrier_off(dev);
+	schedule_delayed_work(&priv->retry_work, TILE_NET_RETRY_INTERVAL);
+
+	return 0;
+}
+
+
+static int tile_net_drain_lipp_buffers(struct tile_net_priv *priv)
+{
+	int n = 0;
+
+	/* Drain all the LIPP buffers. */
+	while (true) {
+		unsigned int buffer;
+
+		/* NOTE: This should never fail. */
+		if (hv_dev_pread(priv->hv_devhdl, 0, (HV_VirtAddr)&buffer,
+				 sizeof(buffer), NETIO_IPP_DRAIN_OFF) < 0)
+			break;
+
+		/* Stop when done. */
+		if (buffer == 0)
+			break;
+
+		{
+			/* Convert "linux_buffer_t" to "va". */
+			void *va = __va((phys_addr_t)(buffer >> 1) << 7);
+
+			/* Acquire the associated "skb". */
+			struct sk_buff **skb_ptr = va - sizeof(*skb_ptr);
+			struct sk_buff *skb = *skb_ptr;
+
+			kfree_skb(skb);
+		}
+
+		n++;
+	}
+
+	return n;
+}
+
+
+/*
+ * Disables a network interface.
+ *
+ * Returns 0, this is not allowed to fail.
+ *
+ * The close entry point is called when an interface is de-activated
+ * by the OS.  The hardware is still under the drivers control, but
+ * needs to be disabled.  A global MAC reset is issued to stop the
+ * hardware, and all transmit and receive resources are freed.
+ *
+ * ISSUE: How closely does "netif_running(dev)" mirror "priv->active"?
+ *
+ * Before we are called by "__dev_close()", "netif_running()" will
+ * have been cleared, so no NEW calls to "tile_net_poll()" will be
+ * made by "netpoll_poll_dev()".
+ *
+ * Often, this can cause some tiles to still have packets in their
+ * queues, so we must call "tile_net_discard_packets()" later.
+ *
+ * Note that some other tile may still be INSIDE "tile_net_poll()",
+ * and in fact, many will be, if there is heavy network load.
+ *
+ * Calling "on_each_cpu(tile_net_stop_disable, (void *)dev, 1)" when
+ * any tile is still "napi_schedule()"'d will induce a horrible crash
+ * when "msleep()" is called.  This includes tiles which are inside
+ * "tile_net_poll()" which have not yet called "napi_complete()".
+ *
+ * So, we must first try to wait long enough for other tiles to finish
+ * with any current "tile_net_poll()" call, and, hopefully, to clear
+ * the "scheduled" flag.  ISSUE: It is unclear what happens to tiles
+ * which have called "napi_schedule()" but which had not yet tried to
+ * call "tile_net_poll()", or which exhausted their budget inside
+ * "tile_net_poll()" just before this function was called.
+ */
+static int tile_net_stop(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	PDEBUG("tile_net_stop()\n");
+
+	/* Start discarding packets. */
+	priv->active = false;
+
+	/* Make sure "active" is visible to all tiles. */
+	mb();
+
+	/*
+	 * On each tile, make sure no NEW packets get delivered, and
+	 * disable the ingress interrupt.
+	 *
+	 * Note that the ingress interrupt can fire AFTER this,
+	 * presumably due to packets which were recently delivered,
+	 * but it will have no effect.
+	 */
+	on_each_cpu(tile_net_deregister, (void *)dev, 1);
+
+	/* Optimistically drain LIPP buffers. */
+	(void)tile_net_drain_lipp_buffers(priv);
+
+	/* ISSUE: Only needed if not yet fully open. */
+	cancel_delayed_work_sync(&priv->retry_work);
+
+	/* Can't transmit any more. */
+	netif_stop_queue(dev);
+
+	/* Disable NAPI on each tile. */
+	on_each_cpu(tile_net_stop_disable, (void *)dev, 1);
+
+	/*
+	 * Drain any remaining LIPP buffers.  NOTE: This "printk()"
+	 * has never been observed, but in theory it could happen.
+	 */
+	if (tile_net_drain_lipp_buffers(priv) != 0)
+		printk("Had to drain some extra LIPP buffers!\n");
+
+	/* Stop LIPP/LEPP. */
+	tile_net_stop_aux(dev);
+
+	/*
+	 * ISSUE: It appears that, in practice anyway, by the time we
+	 * get here, there are no pending completions, but just in case,
+	 * we free (all of) them anyway.
+	 */
+	while (tile_net_lepp_free_comps(dev, true))
+		/* loop */;
+
+	/* Wipe the EPP queue, and wait till the stores hit the EPP. */
+	memset(priv->eq, 0, sizeof(lepp_queue_t));
+	mb();
+
+	return 0;
+}
+
+
+/*
+ * Prepare the "frags" info for the resulting LEPP command.
+ *
+ * If needed, flush the memory used by the frags.
+ */
+static unsigned int tile_net_tx_frags(lepp_frag_t *frags,
+				      struct sk_buff *skb,
+				      void *b_data, unsigned int b_len)
+{
+	unsigned int i, n = 0;
+
+	struct skb_shared_info *sh = skb_shinfo(skb);
+
+	phys_addr_t cpa;
+
+	if (b_len != 0) {
+
+		if (!hash_default)
+			finv_buffer_remote(b_data, b_len, 0);
+
+		cpa = __pa(b_data);
+		frags[n].cpa_lo = cpa;
+		frags[n].cpa_hi = cpa >> 32;
+		frags[n].length = b_len;
+		frags[n].hash_for_home = hash_default;
+		n++;
+	}
+
+	for (i = 0; i < sh->nr_frags; i++) {
+
+		skb_frag_t *f = &sh->frags[i];
+		unsigned long pfn = page_to_pfn(skb_frag_page(f));
+
+		/* FIXME: Compute "hash_for_home" properly. */
+		/* ISSUE: The hypervisor checks CHIP_HAS_REV1_DMA_PACKETS(). */
+		int hash_for_home = hash_default;
+
+		/* FIXME: Hmmm. */
+		if (!hash_default) {
+			void *va = pfn_to_kaddr(pfn) + f->page_offset;
+			BUG_ON(PageHighMem(skb_frag_page(f)));
+			finv_buffer_remote(va, skb_frag_size(f), 0);
+		}
+
+		cpa = ((phys_addr_t)pfn << PAGE_SHIFT) + f->page_offset;
+		frags[n].cpa_lo = cpa;
+		frags[n].cpa_hi = cpa >> 32;
+		frags[n].length = skb_frag_size(f);
+		frags[n].hash_for_home = hash_for_home;
+		n++;
+	}
+
+	return n;
+}
+
+
+/*
+ * This function takes "skb", consisting of a header template and a
+ * payload, and hands it to LEPP, to emit as one or more segments,
+ * each consisting of a possibly modified header, plus a piece of the
+ * payload, via a process known as "tcp segmentation offload".
+ *
+ * Usually, "data" will contain the header template, of size "sh_len",
+ * and "sh->frags" will contain "skb->data_len" bytes of payload, and
+ * there will be "sh->gso_segs" segments.
+ *
+ * Sometimes, if "sendfile()" requires copying, we will be called with
+ * "data" containing the header and payload, with "frags" being empty.
+ *
+ * Sometimes, for example when using NFS over TCP, a single segment can
+ * span 3 fragments, which must be handled carefully in LEPP.
+ *
+ * See "emulate_large_send_offload()" for some reference code, which
+ * does not handle checksumming.
+ *
+ * ISSUE: How do we make sure that high memory DMA does not migrate?
+ */
+static int tile_net_tx_tso(struct sk_buff *skb, struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+	struct tile_net_stats_t *stats = &info->stats;
+
+	struct skb_shared_info *sh = skb_shinfo(skb);
+
+	unsigned char *data = skb->data;
+
+	/* The ip header follows the ethernet header. */
+	struct iphdr *ih = ip_hdr(skb);
+	unsigned int ih_len = ih->ihl * 4;
+
+	/* Note that "nh == ih", by definition. */
+	unsigned char *nh = skb_network_header(skb);
+	unsigned int eh_len = nh - data;
+
+	/* The tcp header follows the ip header. */
+	struct tcphdr *th = (struct tcphdr *)(nh + ih_len);
+	unsigned int th_len = th->doff * 4;
+
+	/* The total number of header bytes. */
+	/* NOTE: This may be less than skb_headlen(skb). */
+	unsigned int sh_len = eh_len + ih_len + th_len;
+
+	/* The number of payload bytes at "skb->data + sh_len". */
+	/* This is non-zero for sendfile() without HIGHDMA. */
+	unsigned int b_len = skb_headlen(skb) - sh_len;
+
+	/* The total number of payload bytes. */
+	unsigned int d_len = b_len + skb->data_len;
+
+	/* The maximum payload size. */
+	unsigned int p_len = sh->gso_size;
+
+	/* The total number of segments. */
+	unsigned int num_segs = sh->gso_segs;
+
+	/* The temporary copy of the command. */
+	u32 cmd_body[(LEPP_MAX_CMD_SIZE + 3) / 4];
+	lepp_tso_cmd_t *cmd = (lepp_tso_cmd_t *)cmd_body;
+
+	/* Analyze the "frags". */
+	unsigned int num_frags =
+		tile_net_tx_frags(cmd->frags, skb, data + sh_len, b_len);
+
+	/* The size of the command, including frags and header. */
+	size_t cmd_size = LEPP_TSO_CMD_SIZE(num_frags, sh_len);
+
+	/* The command header. */
+	lepp_tso_cmd_t cmd_init = {
+		.tso = true,
+		.header_size = sh_len,
+		.ip_offset = eh_len,
+		.tcp_offset = eh_len + ih_len,
+		.payload_size = p_len,
+		.num_frags = num_frags,
+	};
+
+	unsigned long irqflags;
+
+	lepp_queue_t *eq = priv->eq;
+
+	struct sk_buff *olds[8];
+	unsigned int wanted = 8;
+	unsigned int i, nolds = 0;
+
+	unsigned int cmd_head, cmd_tail, cmd_next;
+	unsigned int comp_tail;
+
+
+	/* Paranoia. */
+	BUG_ON(skb->protocol != htons(ETH_P_IP));
+	BUG_ON(ih->protocol != IPPROTO_TCP);
+	BUG_ON(skb->ip_summed != CHECKSUM_PARTIAL);
+	BUG_ON(num_frags > LEPP_MAX_FRAGS);
+	/*--BUG_ON(num_segs != (d_len + (p_len - 1)) / p_len); */
+	BUG_ON(num_segs <= 1);
+
+
+	/* Finish preparing the command. */
+
+	/* Copy the command header. */
+	*cmd = cmd_init;
+
+	/* Copy the "header". */
+	memcpy(&cmd->frags[num_frags], data, sh_len);
+
+
+	/* Prefetch and wait, to minimize time spent holding the spinlock. */
+	prefetch_L1(&eq->comp_tail);
+	prefetch_L1(&eq->cmd_tail);
+	mb();
+
+
+	/* Enqueue the command. */
+
+	spin_lock_irqsave(&priv->eq_lock, irqflags);
+
+	/* Handle completions if needed to make room. */
+	/* NOTE: Return NETDEV_TX_BUSY if there is still no room. */
+	if (lepp_num_free_comp_slots(eq) == 0) {
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+		if (nolds == 0) {
+busy:
+			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
+			return NETDEV_TX_BUSY;
+		}
+	}
+
+	cmd_head = eq->cmd_head;
+	cmd_tail = eq->cmd_tail;
+
+	/* Prepare to advance, detecting full queue. */
+	/* NOTE: Return NETDEV_TX_BUSY if the queue is full. */
+	cmd_next = cmd_tail + cmd_size;
+	if (cmd_tail < cmd_head && cmd_next >= cmd_head)
+		goto busy;
+	if (cmd_next > LEPP_CMD_LIMIT) {
+		cmd_next = 0;
+		if (cmd_next == cmd_head)
+			goto busy;
+	}
+
+	/* Copy the command. */
+	memcpy(&eq->cmds[cmd_tail], cmd, cmd_size);
+
+	/* Advance. */
+	cmd_tail = cmd_next;
+
+	/* Record "skb" for eventual freeing. */
+	comp_tail = eq->comp_tail;
+	eq->comps[comp_tail] = skb;
+	LEPP_QINC(comp_tail);
+	eq->comp_tail = comp_tail;
+
+	/* Flush before allowing LEPP to handle the command. */
+	/* ISSUE: Is this the optimal location for the flush? */
+	__insn_mf();
+
+	eq->cmd_tail = cmd_tail;
+
+	/* NOTE: Using "4" here is more efficient than "0" or "2", */
+	/* and, strangely, more efficient than pre-checking the number */
+	/* of available completions, and comparing it to 4. */
+	if (nolds == 0)
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
+
+	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
+
+	/* Handle completions. */
+	for (i = 0; i < nolds; i++)
+		dev_consume_skb_any(olds[i]);
+
+	/* Update stats. */
+	u64_stats_update_begin(&stats->syncp);
+	stats->tx_packets += num_segs;
+	stats->tx_bytes += (num_segs * sh_len) + d_len;
+	u64_stats_update_end(&stats->syncp);
+
+	/* Make sure the egress timer is scheduled. */
+	tile_net_schedule_egress_timer(info);
+
+	return NETDEV_TX_OK;
+}
+
+
+/*
+ * Transmit a packet (called by the kernel via "hard_start_xmit" hook).
+ */
+static int tile_net_tx(struct sk_buff *skb, struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	int my_cpu = smp_processor_id();
+	struct tile_net_cpu *info = priv->cpu[my_cpu];
+	struct tile_net_stats_t *stats = &info->stats;
+
+	unsigned long irqflags;
+
+	struct skb_shared_info *sh = skb_shinfo(skb);
+
+	unsigned int len = skb->len;
+	unsigned char *data = skb->data;
+
+	unsigned int csum_start = skb_checksum_start_offset(skb);
+
+	lepp_frag_t frags[1 + MAX_SKB_FRAGS];
+
+	unsigned int num_frags;
+
+	lepp_queue_t *eq = priv->eq;
+
+	struct sk_buff *olds[8];
+	unsigned int wanted = 8;
+	unsigned int i, nolds = 0;
+
+	unsigned int cmd_size = sizeof(lepp_cmd_t);
+
+	unsigned int cmd_head, cmd_tail, cmd_next;
+	unsigned int comp_tail;
+
+	lepp_cmd_t cmds[1 + MAX_SKB_FRAGS];
+
+
+	/*
+	 * This is paranoia, since we think that if the link doesn't come
+	 * up, telling Linux we have no carrier will keep it from trying
+	 * to transmit.  If it does, though, we can't execute this routine,
+	 * since data structures we depend on aren't set up yet.
+	 */
+	if (!info->registered)
+		return NETDEV_TX_BUSY;
+
+
+	/* Save the timestamp. */
+	dev->trans_start = jiffies;
+
+
+#ifdef TILE_NET_PARANOIA
+#if CHIP_HAS_CBOX_HOME_MAP()
+	if (hash_default) {
+		HV_PTE pte = *virt_to_pte(current->mm, (unsigned long)data);
+		if (hv_pte_get_mode(pte) != HV_PTE_MODE_CACHE_HASH_L3)
+			panic("Non-HFH egress buffer! VA=%p Mode=%d PTE=%llx",
+			      data, hv_pte_get_mode(pte), hv_pte_val(pte));
+	}
+#endif
+#endif
+
+
+#ifdef TILE_NET_DUMP_PACKETS
+	/* ISSUE: Does not dump the "frags". */
+	dump_packet(data, skb_headlen(skb), "tx");
+#endif /* TILE_NET_DUMP_PACKETS */
+
+
+	if (sh->gso_size != 0)
+		return tile_net_tx_tso(skb, dev);
+
+
+	/* Prepare the commands. */
+
+	num_frags = tile_net_tx_frags(frags, skb, data, skb_headlen(skb));
+
+	for (i = 0; i < num_frags; i++) {
+
+		bool final = (i == num_frags - 1);
+
+		lepp_cmd_t cmd = {
+			.cpa_lo = frags[i].cpa_lo,
+			.cpa_hi = frags[i].cpa_hi,
+			.length = frags[i].length,
+			.hash_for_home = frags[i].hash_for_home,
+			.send_completion = final,
+			.end_of_packet = final
+		};
+
+		if (i == 0 && skb->ip_summed == CHECKSUM_PARTIAL) {
+			cmd.compute_checksum = 1;
+			cmd.checksum_data.bits.start_byte = csum_start;
+			cmd.checksum_data.bits.count = len - csum_start;
+			cmd.checksum_data.bits.destination_byte =
+				csum_start + skb->csum_offset;
+		}
+
+		cmds[i] = cmd;
+	}
+
+
+	/* Prefetch and wait, to minimize time spent holding the spinlock. */
+	prefetch_L1(&eq->comp_tail);
+	prefetch_L1(&eq->cmd_tail);
+	mb();
+
+
+	/* Enqueue the commands. */
+
+	spin_lock_irqsave(&priv->eq_lock, irqflags);
+
+	/* Handle completions if needed to make room. */
+	/* NOTE: Return NETDEV_TX_BUSY if there is still no room. */
+	if (lepp_num_free_comp_slots(eq) == 0) {
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 0);
+		if (nolds == 0) {
+busy:
+			spin_unlock_irqrestore(&priv->eq_lock, irqflags);
+			return NETDEV_TX_BUSY;
+		}
+	}
+
+	cmd_head = eq->cmd_head;
+	cmd_tail = eq->cmd_tail;
+
+	/* Copy the commands, or fail. */
+	/* NOTE: Return NETDEV_TX_BUSY if the queue is full. */
+	for (i = 0; i < num_frags; i++) {
+
+		/* Prepare to advance, detecting full queue. */
+		cmd_next = cmd_tail + cmd_size;
+		if (cmd_tail < cmd_head && cmd_next >= cmd_head)
+			goto busy;
+		if (cmd_next > LEPP_CMD_LIMIT) {
+			cmd_next = 0;
+			if (cmd_next == cmd_head)
+				goto busy;
+		}
+
+		/* Copy the command. */
+		*(lepp_cmd_t *)&eq->cmds[cmd_tail] = cmds[i];
+
+		/* Advance. */
+		cmd_tail = cmd_next;
+	}
+
+	/* Record "skb" for eventual freeing. */
+	comp_tail = eq->comp_tail;
+	eq->comps[comp_tail] = skb;
+	LEPP_QINC(comp_tail);
+	eq->comp_tail = comp_tail;
+
+	/* Flush before allowing LEPP to handle the command. */
+	/* ISSUE: Is this the optimal location for the flush? */
+	__insn_mf();
+
+	eq->cmd_tail = cmd_tail;
+
+	/* NOTE: Using "4" here is more efficient than "0" or "2", */
+	/* and, strangely, more efficient than pre-checking the number */
+	/* of available completions, and comparing it to 4. */
+	if (nolds == 0)
+		nolds = tile_net_lepp_grab_comps(eq, olds, wanted, 4);
+
+	spin_unlock_irqrestore(&priv->eq_lock, irqflags);
+
+	/* Handle completions. */
+	for (i = 0; i < nolds; i++)
+		dev_consume_skb_any(olds[i]);
+
+	/* HACK: Track "expanded" size for short packets (e.g. 42 < 60). */
+	u64_stats_update_begin(&stats->syncp);
+	stats->tx_packets++;
+	stats->tx_bytes += ((len >= ETH_ZLEN) ? len : ETH_ZLEN);
+	u64_stats_update_end(&stats->syncp);
+
+	/* Make sure the egress timer is scheduled. */
+	tile_net_schedule_egress_timer(info);
+
+	return NETDEV_TX_OK;
+}
+
+
+/*
+ * Deal with a transmit timeout.
+ */
+static void tile_net_tx_timeout(struct net_device *dev)
+{
+	PDEBUG("tile_net_tx_timeout()\n");
+	PDEBUG("Transmit timeout at %ld, latency %ld\n", jiffies,
+	       jiffies - dev->trans_start);
+
+	/* XXX: ISSUE: This doesn't seem useful for us. */
+	netif_wake_queue(dev);
+}
+
+
+/*
+ * Ioctl commands.
+ */
+static int tile_net_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
+{
+	return -EOPNOTSUPP;
+}
+
+
+/*
+ * Get System Network Statistics.
+ *
+ * Returns the address of the device statistics structure.
+ */
+static struct rtnl_link_stats64 *tile_net_get_stats64(struct net_device *dev,
+		struct rtnl_link_stats64 *stats)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	u64 rx_packets = 0, tx_packets = 0;
+	u64 rx_bytes = 0, tx_bytes = 0;
+	u64 rx_errors = 0, rx_dropped = 0;
+	int i;
+
+	for_each_online_cpu(i) {
+		struct tile_net_stats_t *cpu_stats;
+		u64 trx_packets, ttx_packets, trx_bytes, ttx_bytes;
+		u64 trx_errors, trx_dropped;
+		unsigned int start;
+
+		if (priv->cpu[i] == NULL)
+			continue;
+		cpu_stats = &priv->cpu[i]->stats;
+
+		do {
+			start = u64_stats_fetch_begin_irq(&cpu_stats->syncp);
+			trx_packets = cpu_stats->rx_packets;
+			ttx_packets = cpu_stats->tx_packets;
+			trx_bytes   = cpu_stats->rx_bytes;
+			ttx_bytes   = cpu_stats->tx_bytes;
+			trx_errors  = cpu_stats->rx_errors;
+			trx_dropped = cpu_stats->rx_dropped;
+		} while (u64_stats_fetch_retry_irq(&cpu_stats->syncp, start));
+
+		rx_packets += trx_packets;
+		tx_packets += ttx_packets;
+		rx_bytes   += trx_bytes;
+		tx_bytes   += ttx_bytes;
+		rx_errors  += trx_errors;
+		rx_dropped += trx_dropped;
+	}
+
+	stats->rx_packets = rx_packets;
+	stats->tx_packets = tx_packets;
+	stats->rx_bytes   = rx_bytes;
+	stats->tx_bytes   = tx_bytes;
+	stats->rx_errors  = rx_errors;
+	stats->rx_dropped = rx_dropped;
+
+	return stats;
+}
+
+
+/*
+ * Change the "mtu".
+ *
+ * The "change_mtu" method is usually not needed.
+ * If you need it, it must be like this.
+ */
+static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
+{
+	PDEBUG("tile_net_change_mtu()\n");
+
+	/* Check ranges. */
+	if ((new_mtu < 68) || (new_mtu > 1500))
+		return -EINVAL;
+
+	/* Accept the value. */
+	dev->mtu = new_mtu;
+
+	return 0;
+}
+
+
+/*
+ * Change the Ethernet Address of the NIC.
+ *
+ * The hypervisor driver does not support changing MAC address.  However,
+ * the IPP does not do anything with the MAC address, so the address which
+ * gets used on outgoing packets, and which is accepted on incoming packets,
+ * is completely up to the NetIO program or kernel driver which is actually
+ * handling them.
+ *
+ * Returns 0 on success, negative on failure.
+ */
+static int tile_net_set_mac_address(struct net_device *dev, void *p)
+{
+	struct sockaddr *addr = p;
+
+	if (!is_valid_ether_addr(addr->sa_data))
+		return -EADDRNOTAVAIL;
+
+	/* ISSUE: Note that "dev_addr" is now a pointer. */
+	memcpy(dev->dev_addr, addr->sa_data, dev->addr_len);
+
+	return 0;
+}
+
+
+/*
+ * Obtain the MAC address from the hypervisor.
+ * This must be done before opening the device.
+ */
+static int tile_net_get_mac(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+
+	char hv_dev_name[32];
+	int len;
+
+	__netio_getset_offset_t offset = { .word = NETIO_IPP_PARAM_OFF };
+
+	int ret;
+
+	/* For example, "xgbe0". */
+	strcpy(hv_dev_name, dev->name);
+	len = strlen(hv_dev_name);
+
+	/* For example, "xgbe/0". */
+	hv_dev_name[len] = hv_dev_name[len - 1];
+	hv_dev_name[len - 1] = '/';
+	len++;
+
+	/* For example, "xgbe/0/native_hash". */
+	strcpy(hv_dev_name + len, hash_default ? "/native_hash" : "/native");
+
+	/* Get the hypervisor handle for this device. */
+	priv->hv_devhdl = hv_dev_open((HV_VirtAddr)hv_dev_name, 0);
+	PDEBUG("hv_dev_open(%s) returned %d %p\n",
+	       hv_dev_name, priv->hv_devhdl, &priv->hv_devhdl);
+	if (priv->hv_devhdl < 0) {
+		if (priv->hv_devhdl == HV_ENODEV)
+			printk(KERN_DEBUG "Ignoring unconfigured device %s\n",
+				 hv_dev_name);
+		else
+			printk(KERN_DEBUG "hv_dev_open(%s) returned %d\n",
+				 hv_dev_name, priv->hv_devhdl);
+		return -1;
+	}
+
+	/*
+	 * Read the hardware address from the hypervisor.
+	 * ISSUE: Note that "dev_addr" is now a pointer.
+	 */
+	offset.bits.class = NETIO_PARAM;
+	offset.bits.addr = NETIO_PARAM_MAC;
+	ret = hv_dev_pread(priv->hv_devhdl, 0,
+			   (HV_VirtAddr)dev->dev_addr, dev->addr_len,
+			   offset.word);
+	PDEBUG("hv_dev_pread(NETIO_PARAM_MAC) returned %d\n", ret);
+	if (ret <= 0) {
+		printk(KERN_DEBUG "hv_dev_pread(NETIO_PARAM_MAC) %s failed\n",
+		       dev->name);
+		/*
+		 * Since the device is configured by the hypervisor but we
+		 * can't get its MAC address, we are most likely running
+		 * the simulator, so let's generate a random MAC address.
+		 */
+		eth_hw_addr_random(dev);
+	}
+
+	return 0;
+}
+
+
+#ifdef CONFIG_NET_POLL_CONTROLLER
+/*
+ * Polling 'interrupt' - used by things like netconsole to send skbs
+ * without having to re-enable interrupts. It's not called while
+ * the interrupt routine is executing.
+ */
+static void tile_net_netpoll(struct net_device *dev)
+{
+	struct tile_net_priv *priv = netdev_priv(dev);
+	disable_percpu_irq(priv->intr_id);
+	tile_net_handle_ingress_interrupt(priv->intr_id, dev);
+	enable_percpu_irq(priv->intr_id, 0);
+}
+#endif
+
+
+static const struct net_device_ops tile_net_ops = {
+	.ndo_open = tile_net_open,
+	.ndo_stop = tile_net_stop,
+	.ndo_start_xmit = tile_net_tx,
+	.ndo_do_ioctl = tile_net_ioctl,
+	.ndo_get_stats64 = tile_net_get_stats64,
+	.ndo_change_mtu = tile_net_change_mtu,
+	.ndo_tx_timeout = tile_net_tx_timeout,
+	.ndo_set_mac_address = tile_net_set_mac_address,
+#ifdef CONFIG_NET_POLL_CONTROLLER
+	.ndo_poll_controller = tile_net_netpoll,
+#endif
+};
+
+
+/*
+ * The setup function.
+ *
+ * This uses ether_setup() to assign various fields in dev, including
+ * setting IFF_BROADCAST and IFF_MULTICAST, then sets some extra fields.
+ */
+static void tile_net_setup(struct net_device *dev)
+{
+	netdev_features_t features = 0;
+
+	ether_setup(dev);
+	dev->netdev_ops = &tile_net_ops;
+	dev->watchdog_timeo = TILE_NET_TIMEOUT;
+	dev->tx_queue_len = TILE_NET_TX_QUEUE_LEN;
+	dev->mtu = TILE_NET_MTU;
+
+	features |= NETIF_F_HW_CSUM;
+	features |= NETIF_F_SG;
+
+	/* We support TSO iff the HV supports sufficient frags. */
+	if (LEPP_MAX_FRAGS >= 1 + MAX_SKB_FRAGS)
+		features |= NETIF_F_TSO;
+
+	/* We can't support HIGHDMA without hash_default, since we need
+	 * to be able to finv() with a VA if we don't have hash_default.
+	 */
+	if (hash_default)
+		features |= NETIF_F_HIGHDMA;
+
+	dev->hw_features   |= features;
+	dev->vlan_features |= features;
+	dev->features      |= features;
+}
+
+
+/*
+ * Allocate the device structure, register the device, and obtain the
+ * MAC address from the hypervisor.
+ */
+static struct net_device *tile_net_dev_init(const char *name)
+{
+	int ret;
+	struct net_device *dev;
+	struct tile_net_priv *priv;
+
+	/*
+	 * Allocate the device structure.  This allocates "priv", calls
+	 * tile_net_setup(), and saves "name".  Normally, "name" is a
+	 * template, instantiated by register_netdev(), but not for us.
+	 */
+	dev = alloc_netdev(sizeof(*priv), name, NET_NAME_UNKNOWN,
+			   tile_net_setup);
+	if (!dev) {
+		pr_err("alloc_netdev(%s) failed\n", name);
+		return NULL;
+	}
+
+	priv = netdev_priv(dev);
+
+	/* Initialize "priv". */
+
+	memset(priv, 0, sizeof(*priv));
+
+	/* Save "dev" for "tile_net_open_retry()". */
+	priv->dev = dev;
+
+	INIT_DELAYED_WORK(&priv->retry_work, tile_net_open_retry);
+
+	spin_lock_init(&priv->eq_lock);
+
+	/* Allocate "eq". */
+	priv->eq_pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, EQ_ORDER);
+	if (!priv->eq_pages) {
+		free_netdev(dev);
+		return NULL;
+	}
+	priv->eq = page_address(priv->eq_pages);
+
+	/* Register the network device. */
+	ret = register_netdev(dev);
+	if (ret) {
+		pr_err("register_netdev %s failed %d\n", dev->name, ret);
+		__free_pages(priv->eq_pages, EQ_ORDER);
+		free_netdev(dev);
+		return NULL;
+	}
+
+	/* Get the MAC address. */
+	ret = tile_net_get_mac(dev);
+	if (ret < 0) {
+		unregister_netdev(dev);
+		__free_pages(priv->eq_pages, EQ_ORDER);
+		free_netdev(dev);
+		return NULL;
+	}
+
+	return dev;
+}
+
+
+/*
+ * Module cleanup.
+ *
+ * FIXME: If compiled as a module, this module cannot be "unloaded",
+ * because the "ingress interrupt handler" is registered permanently.
+ */
+static void tile_net_cleanup(void)
+{
+	int i;
+
+	for (i = 0; i < TILE_NET_DEVS; i++) {
+		if (tile_net_devs[i]) {
+			struct net_device *dev = tile_net_devs[i];
+			struct tile_net_priv *priv = netdev_priv(dev);
+			unregister_netdev(dev);
+			finv_buffer_remote(priv->eq, EQ_SIZE, 0);
+			__free_pages(priv->eq_pages, EQ_ORDER);
+			free_netdev(dev);
+		}
+	}
+}
+
+
+/*
+ * Module initialization.
+ */
+static int tile_net_init_module(void)
+{
+	pr_info("Tilera Network Driver\n");
+
+	tile_net_devs[0] = tile_net_dev_init("xgbe0");
+	tile_net_devs[1] = tile_net_dev_init("xgbe1");
+	tile_net_devs[2] = tile_net_dev_init("gbe0");
+	tile_net_devs[3] = tile_net_dev_init("gbe1");
+
+	return 0;
+}
+
+
+module_init(tile_net_init_module);
+module_exit(tile_net_cleanup);
+
+
+#ifndef MODULE
+
+/*
+ * The "network_cpus" boot argument specifies the cpus that are dedicated
+ * to handle ingress packets.
+ *
+ * The parameter should be in the form "network_cpus=m-n[,x-y]", where
+ * m, n, x, y are integer numbers that represent the cpus that can be
+ * neither a dedicated cpu nor a dataplane cpu.
+ */
+static int __init network_cpus_setup(char *str)
+{
+	int rc = cpulist_parse_crop(str, &network_cpus_map);
+	if (rc != 0) {
+		pr_warn("network_cpus=%s: malformed cpu list\n", str);
+	} else {
+
+		/* Remove dedicated cpus. */
+		cpumask_and(&network_cpus_map, &network_cpus_map,
+			    cpu_possible_mask);
+
+
+		if (cpumask_empty(&network_cpus_map)) {
+			pr_warn("Ignoring network_cpus='%s'\n", str);
+		} else {
+			pr_info("Linux network CPUs: %*pbl\n",
+				cpumask_pr_args(&network_cpus_map));
+			network_cpus_used = true;
+		}
+	}
+
+	return 0;
+}
+__setup("network_cpus=", network_cpus_setup);
+
+#endif