tilegx network driver: initial support
authorChris Metcalf <cmetcalf@tilera.com>
Thu, 7 Jun 2012 10:45:02 +0000 (10:45 +0000)
committerDavid S. Miller <davem@davemloft.net>
Tue, 12 Jun 2012 00:01:26 +0000 (17:01 -0700)
This change adds support for the tilegx network driver based on the
GXIO IORPC support in the tilegx software stack, using the on-chip
mPIPE packet processing engine.

Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/tile/Kconfig
drivers/net/ethernet/tile/Makefile
drivers/net/ethernet/tile/tilegx.c [new file with mode: 0644]

index 2d9218f..098b1c4 100644 (file)
@@ -7,6 +7,8 @@ config TILE_NET
        depends on TILE
        default y
        select CRC32
+       select TILE_GXIO_MPIPE if TILEGX
+       select HIGH_RES_TIMERS if TILEGX
        ---help---
          This is a standard Linux network device driver for the
          on-chip Tilera Gigabit Ethernet and XAUI interfaces.
index f634f14..0ef9eef 100644 (file)
@@ -4,7 +4,7 @@
 
 obj-$(CONFIG_TILE_NET) += tile_net.o
 ifdef CONFIG_TILEGX
-tile_net-objs := tilegx.o mpipe.o iorpc_mpipe.o dma_queue.o
+tile_net-y := tilegx.o
 else
-tile_net-objs := tilepro.o
+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 (file)
index 0000000..83b4b38
--- /dev/null
@@ -0,0 +1,1898 @@
+/*
+ * 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/tcp.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)
+
+/* 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 {
+       struct hrtimer timer;
+       struct net_device *dev;
+};
+
+/* Info for a specific cpu. */
+struct tile_net_info {
+       /* The NAPI struct. */
+       struct napi_struct napi;
+       /* Packet queue. */
+       gxio_mpipe_iqueue_t iqueue;
+       /* Our cpu. */
+       int my_cpu;
+       /* True if iqueue is valid. */
+       bool has_iqueue;
+       /* NAPI flags. */
+       bool napi_added;
+       bool napi_enabled;
+       /* 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 hrtimer egress_timer;
+       /* True if "egress_timer" is scheduled. */
+       bool egress_timer_scheduled;
+       /* 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];
+};
+
+/* 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;
+       /* Total stats. */
+       struct net_device_stats stats;
+};
+
+/* Egress info, indexed by "priv->echannel" (lazily created as needed). */
+static 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.
+ */
+static struct net_device *tile_net_devs_for_channel[TILE_NET_CHANNELS];
+
+/* 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 "context" for all devices. */
+static gxio_mpipe_context_t context;
+
+/* Buffer sizes and mpipe enum codes for buffer stacks.
+ * See arch/tile/include/gxio/mpipe.h for the set of possible values.
+ */
+#define BUFFER_SIZE_SMALL_ENUM GXIO_MPIPE_BUFFER_SIZE_128
+#define BUFFER_SIZE_SMALL 128
+#define BUFFER_SIZE_LARGE_ENUM GXIO_MPIPE_BUFFER_SIZE_1664
+#define BUFFER_SIZE_LARGE 1664
+
+/* The small/large "buffer stacks". */
+static int small_buffer_stack = -1;
+static int large_buffer_stack = -1;
+
+/* Amount of memory allocated for each buffer stack. */
+static size_t buffer_stack_size;
+
+/* The actual memory allocated for the buffer stacks. */
+static void *small_buffer_stack_va;
+static void *large_buffer_stack_va;
+
+/* The buckets. */
+static int first_bucket = -1;
+static int num_buckets = 1;
+
+/* The ingress irq. */
+static int ingress_irq = -1;
+
+/* 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 "loopify=LINK" was specified, this is "LINK". */
+static char *loopify_link_name;
+
+/* If "tile_net.custom" was specified, this is non-NULL. */
+static char *custom_str;
+
+/* 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)
+{
+       char buf[1024];
+       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;
+       }
+
+       cpulist_scnprintf(buf, sizeof(buf), &network_cpus_map);
+       pr_info("Linux network CPUs: %s\n", buf);
+       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_str, charp, 0444);
+MODULE_PARM_DESC(custom, "indicates a (heavily) customized classifier");
+
+/* 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(bool small)
+{
+       int stack = small ? small_buffer_stack : large_buffer_stack;
+       const unsigned long buffer_alignment = 128;
+       struct sk_buff *skb;
+       int len;
+
+       len = sizeof(struct sk_buff **) + buffer_alignment;
+       len += (small ? BUFFER_SIZE_SMALL : BUFFER_SIZE_LARGE);
+       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(&context, stack,
+                              (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 stack)
+{
+       for (;;) {
+               tile_io_addr_t addr =
+                       (tile_io_addr_t)gxio_mpipe_pop_buffer(&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 = &__get_cpu_var(per_cpu_info);
+
+       while (info->num_needed_small_buffers != 0) {
+               if (!tile_net_provide_buffer(true))
+                       goto oops;
+               info->num_needed_small_buffers--;
+       }
+
+       while (info->num_needed_large_buffers != 0) {
+               if (!tile_net_provide_buffer(false))
+                       goto oops;
+               info->num_needed_large_buffers--;
+       }
+
+       return;
+
+oops:
+       /* Add a description to the page allocation failure dump. */
+       pr_notice("Tile %d still needs some buffers\n", info->my_cpu);
+}
+
+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) &&
+           compare_ether_addr(dev->dev_addr, buf) != 0)
+               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 = &__get_cpu_var(per_cpu_info);
+       struct tile_net_priv *priv = netdev_priv(dev);
+
+       /* 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;
+
+       netif_receive_skb(skb);
+
+       /* Update stats. */
+       tile_net_stats_add(1, &priv->stats.rx_packets);
+       tile_net_stats_add(len, &priv->stats.rx_bytes);
+
+       /* Need a new buffer. */
+       if (idesc->size == BUFFER_SIZE_SMALL_ENUM)
+               info->num_needed_small_buffers++;
+       else
+               info->num_needed_large_buffers++;
+}
+
+/* Handle a packet.  Return true if "processed", false if "filtered". */
+static bool tile_net_handle_packet(gxio_mpipe_idesc_t *idesc)
+{
+       struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
+       struct net_device *dev = 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.
+        * NOTE: This happens under heavy load.
+        */
+       if (idesc->be) {
+               struct tile_net_priv *priv = netdev_priv(dev);
+               tile_net_stats_add(1, &priv->stats.rx_dropped);
+               gxio_mpipe_iqueue_consume(&info->iqueue, idesc);
+               if (net_ratelimit())
+                       pr_info("Dropping packet (insufficient buffers).\n");
+               return false;
+       }
+
+       /* Get the "l2_offset", if allowed. */
+       l2_offset = custom_str ? 0 : gxio_mpipe_idesc_get_l2_offset(idesc);
+
+       /* Get the raw buffer VA (includes "headroom"). */
+       va = tile_io_addr_to_va((unsigned long)(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) {
+               gxio_mpipe_iqueue_drop(&info->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->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 = &__get_cpu_var(per_cpu_info);
+       unsigned int work = 0;
+       gxio_mpipe_idesc_t *idesc;
+       int i, n;
+
+       /* Process packets. */
+       while ((n = gxio_mpipe_iqueue_try_peek(&info->iqueue, &idesc)) > 0) {
+               for (i = 0; i < n; i++) {
+                       if (i == TILE_NET_BATCH)
+                               goto done;
+                       if (tile_net_handle_packet(idesc + i)) {
+                               if (++work >= budget)
+                                       goto done;
+                       }
+               }
+       }
+
+       /* There are no packets left. */
+       napi_complete(&info->napi);
+
+       /* Re-enable hypervisor interrupts. */
+       gxio_mpipe_enable_notif_ring_interrupt(&context, info->iqueue.ring);
+
+       /* HACK: Avoid the "rotting packet" problem. */
+       if (gxio_mpipe_iqueue_try_peek(&info->iqueue, &idesc) > 0)
+               napi_schedule(&info->napi);
+
+       /* ISSUE: Handle completions? */
+
+done:
+       tile_net_provide_needed_buffers();
+
+       return work;
+}
+
+/* Handle an ingress interrupt on the current cpu. */
+static irqreturn_t tile_net_handle_ingress_irq(int irq, void *unused)
+{
+       struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
+       napi_schedule(&info->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)
+{
+       struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
+       struct tile_net_priv *priv = netdev_priv(dev);
+
+       hrtimer_start(&info->tx_wake[priv->echannel].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, smp_processor_id());
+       return HRTIMER_NORESTART;
+}
+
+/* Make sure the egress timer is scheduled. */
+static void tile_net_schedule_egress_timer(void)
+{
+       struct tile_net_info *info = &__get_cpu_var(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 = &__get_cpu_var(per_cpu_info);
+       unsigned long irqflags;
+       bool pending = false;
+       int i;
+
+       local_irq_save(irqflags);
+
+       /* The timer is no longer scheduled. */
+       info->egress_timer_scheduled = false;
+
+       /* Free all possible comps for this tile. */
+       for (i = 0; i < TILE_NET_CHANNELS; i++) {
+               struct tile_net_egress *egress = &egress_for_echannel[i];
+               struct tile_net_comps *comps = info->comps_for_echannel[i];
+               if (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;
+}
+
+/* Helper function for "tile_net_update()".
+ * "dev" (i.e. arg) is the device being brought up or down,
+ * or NULL if all devices are now down.
+ */
+static void tile_net_update_cpu(void *arg)
+{
+       struct tile_net_info *info = &__get_cpu_var(per_cpu_info);
+       struct net_device *dev = arg;
+
+       if (!info->has_iqueue)
+               return;
+
+       if (dev != NULL) {
+               if (!info->napi_added) {
+                       netif_napi_add(dev, &info->napi,
+                                      tile_net_poll, TILE_NET_WEIGHT);
+                       info->napi_added = true;
+               }
+               if (!info->napi_enabled) {
+                       napi_enable(&info->napi);
+                       info->napi_enabled = true;
+               }
+               enable_percpu_irq(ingress_irq, 0);
+       } else {
+               disable_percpu_irq(ingress_irq);
+               if (info->napi_enabled) {
+                       napi_disable(&info->napi);
+                       info->napi_enabled = false;
+               }
+               /* FIXME: Drain the iqueue. */
+       }
+}
+
+/* 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 channel;
+       int rc;
+       int cpu;
+
+       gxio_mpipe_rules_init(&rules, &context);
+
+       for (channel = 0; channel < TILE_NET_CHANNELS; channel++) {
+               if (tile_net_devs_for_channel[channel] == NULL)
+                       continue;
+               if (!saw_channel) {
+                       saw_channel = true;
+                       gxio_mpipe_rules_begin(&rules, first_bucket,
+                                              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 failed: %d\n", rc);
+               return -EIO;
+       }
+
+       /* Update all cpus, sequentially (to protect "netif_napi_add()"). */
+       for_each_online_cpu(cpu)
+               smp_call_function_single(cpu, tile_net_update_cpu,
+                                        (saw_channel ? dev : NULL), 1);
+
+       /* HACK: Allow packets to flow in the simulator. */
+       if (saw_channel)
+               sim_enable_mpipe_links(0, -1);
+
+       return 0;
+}
+
+/* Allocate and initialize mpipe buffer stacks, and register them in
+ * the mPIPE TLBs, for both small and large packet sizes.
+ * This routine supports tile_net_init_mpipe(), below.
+ */
+static int init_buffer_stacks(struct net_device *dev, int num_buffers)
+{
+       pte_t hash_pte = pte_set_home((pte_t) { 0 }, PAGE_HOME_HASH);
+       int rc;
+
+       /* Compute stack bytes; we round up to 64KB and then use
+        * alloc_pages() so we get the required 64KB alignment as well.
+        */
+       buffer_stack_size =
+               ALIGN(gxio_mpipe_calc_buffer_stack_bytes(num_buffers),
+                     64 * 1024);
+
+       /* Allocate two buffer stack indices. */
+       rc = gxio_mpipe_alloc_buffer_stacks(&context, 2, 0, 0);
+       if (rc < 0) {
+               netdev_err(dev, "gxio_mpipe_alloc_buffer_stacks failed: %d\n",
+                          rc);
+               return rc;
+       }
+       small_buffer_stack = rc;
+       large_buffer_stack = rc + 1;
+
+       /* Allocate the small memory stack. */
+       small_buffer_stack_va =
+               alloc_pages_exact(buffer_stack_size, GFP_KERNEL);
+       if (small_buffer_stack_va == NULL) {
+               netdev_err(dev,
+                          "Could not alloc %zd bytes for buffer stacks\n",
+                          buffer_stack_size);
+               return -ENOMEM;
+       }
+       rc = gxio_mpipe_init_buffer_stack(&context, small_buffer_stack,
+                                         BUFFER_SIZE_SMALL_ENUM,
+                                         small_buffer_stack_va,
+                                         buffer_stack_size, 0);
+       if (rc != 0) {
+               netdev_err(dev, "gxio_mpipe_init_buffer_stack: %d\n", rc);
+               return rc;
+       }
+       rc = gxio_mpipe_register_client_memory(&context, small_buffer_stack,
+                                              hash_pte, 0);
+       if (rc != 0) {
+               netdev_err(dev,
+                          "gxio_mpipe_register_buffer_memory failed: %d\n",
+                          rc);
+               return rc;
+       }
+
+       /* Allocate the large buffer stack. */
+       large_buffer_stack_va =
+               alloc_pages_exact(buffer_stack_size, GFP_KERNEL);
+       if (large_buffer_stack_va == NULL) {
+               netdev_err(dev,
+                          "Could not alloc %zd bytes for buffer stacks\n",
+                          buffer_stack_size);
+               return -ENOMEM;
+       }
+       rc = gxio_mpipe_init_buffer_stack(&context, large_buffer_stack,
+                                         BUFFER_SIZE_LARGE_ENUM,
+                                         large_buffer_stack_va,
+                                         buffer_stack_size, 0);
+       if (rc != 0) {
+               netdev_err(dev, "gxio_mpipe_init_buffer_stack failed: %d\n",
+                          rc);
+               return rc;
+       }
+       rc = gxio_mpipe_register_client_memory(&context, large_buffer_stack,
+                                              hash_pte, 0);
+       if (rc != 0) {
+               netdev_err(dev,
+                          "gxio_mpipe_register_buffer_memory failed: %d\n",
+                          rc);
+               return rc;
+       }
+
+       return 0;
+}
+
+/* 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;
+       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->comps_for_echannel[i] =
+                       addr + i * sizeof(struct tile_net_comps);
+
+       /* If this is a network cpu, create an iqueue. */
+       if (cpu_isset(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->iqueue, &context, ring++,
+                                           addr, NOTIF_RING_SIZE, 0);
+               if (rc < 0) {
+                       netdev_err(dev,
+                                  "gxio_mpipe_iqueue_init failed: %d\n", rc);
+                       return rc;
+               }
+               info->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;
+
+       /* Allocate one NotifGroup. */
+       rc = gxio_mpipe_alloc_notif_groups(&context, 1, 0, 0);
+       if (rc < 0) {
+               netdev_err(dev, "gxio_mpipe_alloc_notif_groups failed: %d\n",
+                          rc);
+               return rc;
+       }
+       group = rc;
+
+       /* Initialize global num_buckets value. */
+       if (network_cpus_count > 4)
+               num_buckets = 256;
+       else if (network_cpus_count > 1)
+               num_buckets = 16;
+
+       /* Allocate some buckets, and set global first_bucket value. */
+       rc = gxio_mpipe_alloc_buckets(&context, num_buckets, 0, 0);
+       if (rc < 0) {
+               netdev_err(dev, "gxio_mpipe_alloc_buckets failed: %d\n", rc);
+               return rc;
+       }
+       first_bucket = rc;
+
+       /* Init group and buckets. */
+       rc = gxio_mpipe_init_notif_group_and_buckets(
+               &context, group, ring, network_cpus_count,
+               first_bucket, num_buckets,
+               GXIO_MPIPE_BUCKET_STICKY_FLOW_LOCALITY);
+       if (rc != 0) {
+               netdev_err(
+                       dev,
+                       "gxio_mpipe_init_notif_group_and_buckets failed: %d\n",
+                       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;
+
+       rc = create_irq();
+       if (rc < 0) {
+               netdev_err(dev, "create_irq failed: %d\n", rc);
+               return rc;
+       }
+       ingress_irq = rc;
+       tile_irq_activate(ingress_irq, TILE_IRQ_PERCPU);
+       rc = request_irq(ingress_irq, tile_net_handle_ingress_irq,
+                        0, NULL, NULL);
+       if (rc != 0) {
+               netdev_err(dev, "request_irq failed: %d\n", rc);
+               destroy_irq(ingress_irq);
+               ingress_irq = -1;
+               return rc;
+       }
+
+       for_each_online_cpu(cpu) {
+               struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+               if (info->has_iqueue) {
+                       gxio_mpipe_request_notif_ring_interrupt(
+                               &context, cpu_x(cpu), cpu_y(cpu),
+                               1, ingress_irq, info->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(void)
+{
+       int cpu;
+
+       /* Do cleanups that require the mpipe context first. */
+       if (small_buffer_stack >= 0)
+               tile_net_pop_all_buffers(small_buffer_stack);
+       if (large_buffer_stack >= 0)
+               tile_net_pop_all_buffers(large_buffer_stack);
+
+       /* Destroy mpipe context so the hardware no longer owns any memory. */
+       gxio_mpipe_destroy(&context);
+
+       for_each_online_cpu(cpu) {
+               struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+               free_pages((unsigned long)(info->comps_for_echannel[0]),
+                          get_order(COMPS_SIZE));
+               info->comps_for_echannel[0] = NULL;
+               free_pages((unsigned long)(info->iqueue.idescs),
+                          get_order(NOTIF_RING_SIZE));
+               info->iqueue.idescs = NULL;
+       }
+
+       if (small_buffer_stack_va)
+               free_pages_exact(small_buffer_stack_va, buffer_stack_size);
+       if (large_buffer_stack_va)
+               free_pages_exact(large_buffer_stack_va, buffer_stack_size);
+
+       small_buffer_stack_va = NULL;
+       large_buffer_stack_va = NULL;
+       large_buffer_stack = -1;
+       small_buffer_stack = -1;
+       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 i, num_buffers, rc;
+       int cpu;
+       int first_ring, ring;
+       int network_cpus_count = cpus_weight(network_cpus_map);
+
+       if (!hash_default) {
+               netdev_err(dev, "Networking requires hash_default!\n");
+               return -EIO;
+       }
+
+       rc = gxio_mpipe_init(&context, 0);
+       if (rc != 0) {
+               netdev_err(dev, "gxio_mpipe_init failed: %d\n", rc);
+               return -EIO;
+       }
+
+       /* Set up the buffer stacks. */
+       num_buffers =
+               network_cpus_count * (IQUEUE_ENTRIES + TILE_NET_BATCH);
+       rc = init_buffer_stacks(dev, num_buffers);
+       if (rc != 0)
+               goto fail;
+
+       /* Provide initial buffers. */
+       rc = -ENOMEM;
+       for (i = 0; i < num_buffers; i++) {
+               if (!tile_net_provide_buffer(true)) {
+                       netdev_err(dev, "Cannot allocate initial sk_bufs!\n");
+                       goto fail;
+               }
+       }
+       for (i = 0; i < num_buffers; i++) {
+               if (!tile_net_provide_buffer(false)) {
+                       netdev_err(dev, "Cannot allocate initial sk_bufs!\n");
+                       goto fail;
+               }
+       }
+
+       /* Allocate one NotifRing for each network cpu. */
+       rc = gxio_mpipe_alloc_notif_rings(&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;
+
+       return 0;
+
+fail:
+       tile_net_init_mpipe_fail();
+       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)
+{
+       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 edma;
+       int rc = -ENOMEM;
+
+       /* Only initialize once. */
+       if (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.  Note that in practice this can't
+        * fail, which is good, because we will leak an edma ring if so.
+        */
+       rc = gxio_mpipe_alloc_edma_rings(&context, 1, 0, 0);
+       if (rc < 0) {
+               netdev_warn(dev, "gxio_mpipe_alloc_edma_rings failed: %d\n",
+                           rc);
+               goto fail_equeue;
+       }
+       edma = rc;
+
+       /* Initialize the equeue. */
+       rc = gxio_mpipe_equeue_init(equeue, &context, edma, echannel,
+                                   edescs, edescs_size, 0);
+       if (rc != 0) {
+               netdev_err(dev, "gxio_mpipe_equeue_init failed: %d\n", rc);
+               goto fail_equeue;
+       }
+
+       /* Done. */
+       egress_for_echannel[echannel].equeue = equeue;
+       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 rc = gxio_mpipe_link_open(link, &context, link_name, 0);
+       if (rc < 0) {
+               netdev_err(dev, "Failed to open '%s'\n", link_name);
+               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;
+
+       mutex_lock(&tile_net_devs_for_channel_mutex);
+
+       /* Do one-time initialization the first time any device is opened. */
+       if (ingress_irq < 0) {
+               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;
+
+       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->tx_wake[priv->echannel];
+
+               hrtimer_init(&tx_wake->timer, CLOCK_MONOTONIC,
+                            HRTIMER_MODE_REL);
+               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;
+       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;
+
+       for_each_online_cpu(cpu) {
+               struct tile_net_info *info = &per_cpu(per_cpu_info, cpu);
+               struct tile_net_tx_wake *tx_wake =
+                       &info->tx_wake[priv->echannel];
+
+               hrtimer_cancel(&tx_wake->timer);
+               netif_stop_subqueue(dev, cpu);
+       }
+
+       mutex_lock(&tile_net_devs_for_channel_mutex);
+       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,
+                                      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, smp_processor_id());
+       tile_net_schedule_tx_wake_timer(dev);
+       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 data_len = skb->data_len;
+       unsigned int p_len = sh->gso_size;
+       long f_id = -1;    /* id of the current fragment */
+       long f_size = -1;  /* size of the current fragment */
+       long f_used = -1;  /* 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 = sh->frags[f_id].size;
+                               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.
+ * FIXME: add support for IPv6.
+ */
+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 tcphdr *th;
+       unsigned int data_len = skb->data_len;
+       unsigned char *data = skb->data;
+       unsigned int ih_off, th_off, sh_len, p_len;
+       unsigned int isum_seed, tsum_seed, id, seq;
+       long f_id = -1;    /* id of the current fragment */
+       long f_size = -1;  /* size of the current fragment */
+       long f_used = -1;  /* bytes used from the current fragment */
+       long n;            /* size of the current piece of payload */
+       int segment;
+
+       /* Locate original headers and compute various lengths. */
+       ih = ip_hdr(skb);
+       th = tcp_hdr(skb);
+       ih_off = skb_network_offset(skb);
+       th_off = skb_transport_offset(skb);
+       sh_len = th_off + tcp_hdrlen(skb);
+       p_len = sh->gso_size;
+
+       /* Set up seed values for IP and TCP csum and initialize id and seq. */
+       isum_seed = ((0xFFFF - ih->check) +
+                    (0xFFFF - ih->tot_len) +
+                    (0xFFFF - ih->id));
+       tsum_seed = th->check + (0xFFFF ^ htons(skb->len));
+       id = ntohs(ih->id);
+       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. */
+               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 = sh->frags[f_id].size;
+                               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++;
+               }
+
+               id++;
+               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 tile_net_priv *priv = netdev_priv(dev);
+       struct skb_shared_info *sh = skb_shinfo(skb);
+       unsigned int data_len = skb->data_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 = -1;  /* size of the current fragment */
+       long f_used = -1;  /* bytes used from the current fragment */
+       long n;            /* size of the current piece of payload */
+       unsigned long tx_packets = 0, tx_bytes = 0;
+       unsigned int csum_start, sh_len;
+       int segment;
+
+       /* Prepare to egress the headers: set up header edesc. */
+       csum_start = skb_checksum_start_offset(skb);
+       sh_len = skb_transport_offset(skb) + tcp_hdrlen(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 = large_buffer_stack;
+       edesc_body.stack_idx = large_buffer_stack;
+
+       /* Egress all the edescs. */
+       for (segment = 0; segment < sh->gso_segs; segment++) {
+               void *va;
+               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) {
+
+                       /* Advance as needed. */
+                       while (f_used >= f_size) {
+                               f_id++;
+                               f_size = sh->frags[f_id].size;
+                               f_used = 0;
+                       }
+
+                       va = tile_net_frag_buf(&sh->frags[f_id]) + 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, &priv->stats.tx_packets);
+       tile_net_stats_add(tx_bytes, &priv->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 = &__get_cpu_var(per_cpu_info);
+       struct tile_net_priv *priv = netdev_priv(dev);
+       int channel = priv->echannel;
+       struct tile_net_egress *egress = &egress_for_echannel[channel];
+       struct tile_net_comps *comps = info->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, 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 = &__get_cpu_var(per_cpu_info);
+       struct tile_net_priv *priv = netdev_priv(dev);
+       struct tile_net_egress *egress = &egress_for_echannel[priv->echannel];
+       gxio_mpipe_equeue_t *equeue = egress->equeue;
+       struct tile_net_comps *comps =
+               info->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 = large_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, 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++);
+
+       /* 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, &priv->stats.tx_packets);
+       tile_net_stats_add(max_t(unsigned int, len, ETH_ZLEN),
+                          &priv->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)
+{
+       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)
+{
+       return -EOPNOTSUPP;
+}
+
+/* Get system network statistics for device. */
+static struct net_device_stats *tile_net_get_stats(struct net_device *dev)
+{
+       struct tile_net_priv *priv = netdev_priv(dev);
+       return &priv->stats;
+}
+
+/* Change the MTU. */
+static int tile_net_change_mtu(struct net_device *dev, int new_mtu)
+{
+       if ((new_mtu < 68) || (new_mtu > 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)
+{
+       disable_percpu_irq(ingress_irq);
+       tile_net_handle_ingress_irq(ingress_irq, NULL);
+       enable_percpu_irq(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_get_stats = tile_net_get_stats,
+       .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)
+{
+       ether_setup(dev);
+       dev->netdev_ops = &tile_net_ops;
+       dev->watchdog_timeo = TILE_NET_TIMEOUT;
+       dev->features |= NETIF_F_LLTX;
+       dev->features |= NETIF_F_HW_CSUM;
+       dev->features |= NETIF_F_SG;
+       dev->features |= NETIF_F_TSO;
+       dev->mtu = 1500;
+}
+
+/* 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;
+       int i;
+       int nz_addr = 0;
+       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, tile_net_setup,
+                              NR_CPUS, 1);
+       if (!dev) {
+               pr_err("alloc_netdev_mqs(%s) failed\n", name);
+               return;
+       }
+
+       /* Initialize "priv". */
+       priv = netdev_priv(dev);
+       memset(priv, 0, sizeof(*priv));
+       priv->dev = dev;
+       priv->channel = -1;
+       priv->loopify_channel = -1;
+       priv->echannel = -1;
+
+       /* 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.
+        */
+       for (i = 0; i < 6; i++)
+               nz_addr |= mac[i];
+
+       if (nz_addr) {
+               memcpy(dev->dev_addr, mac, 6);
+               dev->addr_len = 6;
+       } else {
+               random_ether_addr(dev->dev_addr);
+       }
+
+       /* 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 = &__get_cpu_var(per_cpu_info);
+       int my_cpu = smp_processor_id();
+
+       info->has_iqueue = false;
+
+       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");
+
+       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);