Native VPP driver for Intel Niantic family of NICs

Available only in vpp_lite platform

Change-Id: I09d112af5f7f4521ec25196ecdd8c02c20eedd5f
Signed-off-by: Damjan Marion <damarion@cisco.com>

5 files changed, 4268 insertions, 0 deletions

diff --git a/vnet/Makefile.am b/vnet/Makefile.am
index 84ee79e9178..9feffc50eae 100644
--- a/vnet/Makefile.am
+++ b/vnet/Makefile.am
@@ -636,6 +636,12 @@ libvnet_la_SOURCES +=				\
 nobase_include_HEADERS +=			\
   vnet/devices/dpdk/dpdk.h			\
   vnet/devices/dpdk/threads.h
+else
+libvnet_la_SOURCES +=				\
+  vnet/devices/nic/ixge.c                       \
+  vnet/devices/nic/ixge.h                       \
+  vnet/devices/nic/sfp.c                        \
+  vnet/devices/nic/sfp.h
 endif
 
 ########################################
diff --git a/vnet/vnet/devices/nic/ixge.c b/vnet/vnet/devices/nic/ixge.c
new file mode 100644
index 00000000000..8d2144142c6
--- /dev/null
+++ b/vnet/vnet/devices/nic/ixge.c
@@ -0,0 +1,2818 @@
+/*
+ * Copyright (c) 2016 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/*
+ *   WARNING!
+ *   This driver is not intended for production use and it is unsupported.
+ *   It is provided for educational use only.
+ *   Please use supported DPDK driver instead.
+ */
+
+#include <vppinfra/vector.h>
+
+#ifndef CLIB_HAVE_VEC128
+#warning HACK: ixge driver wont really work, missing u32x4
+typedef unsigned long long u32x4;
+#endif
+
+#include <vlib/vlib.h>
+#include <vlib/unix/unix.h>
+#include <vlib/pci/pci.h>
+#include <vnet/vnet.h>
+#include <vnet/devices/nic/ixge.h>
+#include <vnet/ethernet/ethernet.h>
+
+#define IXGE_ALWAYS_POLL 0
+
+#define EVENT_SET_FLAGS 0
+#define IXGE_HWBP_RACE_ELOG 0
+
+#define PCI_VENDOR_ID_INTEL 0x8086
+
+/* 10 GIG E (XGE) PHY IEEE 802.3 clause 45 definitions. */
+#define XGE_PHY_DEV_TYPE_PMA_PMD 1
+#define XGE_PHY_DEV_TYPE_PHY_XS 4
+#define XGE_PHY_ID1 0x2
+#define XGE_PHY_ID2 0x3
+#define XGE_PHY_CONTROL 0x0
+#define XGE_PHY_CONTROL_RESET (1 << 15)
+
+ixge_main_t ixge_main;
+static vlib_node_registration_t ixge_input_node;
+static vlib_node_registration_t ixge_process_node;
+
+static void ixge_semaphore_get (ixge_device_t * xd)
+{
+  ixge_main_t * xm = &ixge_main;
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_regs_t * r = xd->regs;
+  u32 i;
+
+  i = 0;
+  while (! (r->software_semaphore & (1 << 0)))
+    {
+      if (i > 0)
+	vlib_process_suspend (vm, 100e-6);
+      i++;
+    }
+  do {
+    r->software_semaphore |= 1 << 1;
+  } while (! (r->software_semaphore & (1 << 1)));
+}
+
+static void ixge_semaphore_release (ixge_device_t * xd)
+{
+  ixge_regs_t * r = xd->regs;
+  r->software_semaphore &= ~3;
+}
+
+static void ixge_software_firmware_sync (ixge_device_t * xd, u32 sw_mask)
+{
+  ixge_main_t * xm = &ixge_main;
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_regs_t * r = xd->regs;
+  u32 fw_mask = sw_mask << 5;
+  u32 m, done = 0;
+
+  while (! done)
+    {
+      ixge_semaphore_get (xd);
+      m = r->software_firmware_sync;
+      done = (m & fw_mask) == 0;
+      if (done)
+	r->software_firmware_sync = m | sw_mask;
+      ixge_semaphore_release (xd);
+      if (! done)
+	vlib_process_suspend (vm, 10e-3);
+    }
+}
+
+static void ixge_software_firmware_sync_release (ixge_device_t * xd, u32 sw_mask)
+{
+  ixge_regs_t * r = xd->regs;
+  ixge_semaphore_get (xd);
+  r->software_firmware_sync &= ~sw_mask;
+  ixge_semaphore_release (xd);
+}
+
+u32 ixge_read_write_phy_reg (ixge_device_t * xd, u32 dev_type, u32 reg_index, u32 v, u32 is_read)
+{
+  ixge_regs_t * r = xd->regs;
+  const u32 busy_bit = 1 << 30;
+  u32 x;
+
+  ASSERT (xd->phy_index < 2);
+  ixge_software_firmware_sync (xd, 1 << (1 + xd->phy_index));
+
+  ASSERT (reg_index < (1 << 16));
+  ASSERT (dev_type < (1 << 5));
+  if (! is_read)
+    r->xge_mac.phy_data = v;
+
+  /* Address cycle. */
+  x = reg_index | (dev_type << 16) | (xd->phys[xd->phy_index].mdio_address << 21);
+  r->xge_mac.phy_command = x | busy_bit;
+  /* Busy wait timed to take 28e-6 secs.  No suspend. */
+  while (r->xge_mac.phy_command & busy_bit)
+    ;
+
+  r->xge_mac.phy_command = x | ((is_read ? 2 : 1) << 26) | busy_bit;
+  while (r->xge_mac.phy_command & busy_bit)
+    ;
+
+  if (is_read)
+    v = r->xge_mac.phy_data >> 16;
+
+  ixge_software_firmware_sync_release (xd, 1 << (1 + xd->phy_index));
+
+  return v;
+}
+
+static u32 ixge_read_phy_reg (ixge_device_t * xd, u32 dev_type, u32 reg_index)
+{ return ixge_read_write_phy_reg (xd, dev_type, reg_index, 0, /* is_read */ 1); }
+
+static void ixge_write_phy_reg (ixge_device_t * xd, u32 dev_type, u32 reg_index, u32 v)
+{ (void) ixge_read_write_phy_reg (xd, dev_type, reg_index, v, /* is_read */ 0); }
+
+static void ixge_i2c_put_bits (i2c_bus_t * b, int scl, int sda)
+{
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, b->private_data);
+  u32 v;
+
+  v = 0;
+  v |= (sda != 0) << 3;
+  v |= (scl != 0) << 1;
+  xd->regs->i2c_control = v;
+}
+
+static void ixge_i2c_get_bits (i2c_bus_t * b, int * scl, int * sda)
+{
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, b->private_data);
+  u32 v;
+
+  v = xd->regs->i2c_control;
+  *sda = (v & (1 << 2)) != 0;
+  *scl = (v & (1 << 0)) != 0;
+}
+
+static u16 ixge_read_eeprom (ixge_device_t * xd, u32 address)
+{
+  ixge_regs_t * r = xd->regs;
+  u32 v;
+  r->eeprom_read = ((/* start bit */ (1 << 0)) | (address << 2));
+  /* Wait for done bit. */
+  while (! ((v = r->eeprom_read) & (1 << 1)))
+    ;
+  return v >> 16;
+}
+
+static void
+ixge_sfp_enable_disable_laser (ixge_device_t * xd, uword enable)
+{
+  u32 tx_disable_bit = 1 << 3;
+  if (enable)
+    xd->regs->sdp_control &= ~tx_disable_bit;
+  else
+    xd->regs->sdp_control |= tx_disable_bit;
+}
+
+static void
+ixge_sfp_enable_disable_10g (ixge_device_t * xd, uword enable)
+{
+  u32 is_10g_bit = 1 << 5;
+  if (enable)
+    xd->regs->sdp_control |= is_10g_bit;
+  else
+    xd->regs->sdp_control &= ~is_10g_bit;
+}
+
+static clib_error_t *
+ixge_sfp_phy_init_from_eeprom (ixge_device_t * xd, u16 sfp_type)
+{
+  u16 a, id, reg_values_addr = 0;
+
+  a = ixge_read_eeprom (xd, 0x2b);
+  if (a == 0 || a == 0xffff)
+    return clib_error_create ("no init sequence in eeprom");
+
+  while (1)
+    {
+      id = ixge_read_eeprom (xd, ++a);
+      if (id == 0xffff)
+	break;
+      reg_values_addr = ixge_read_eeprom (xd, ++a);
+      if (id == sfp_type)
+	break;
+    }
+  if (id != sfp_type)
+    return clib_error_create ("failed to find id 0x%x", sfp_type);
+
+  ixge_software_firmware_sync (xd, 1 << 3);
+  while (1)
+    {
+      u16 v = ixge_read_eeprom (xd, ++reg_values_addr);
+      if (v == 0xffff)
+	break;
+      xd->regs->core_analog_config = v;
+    }
+  ixge_software_firmware_sync_release (xd, 1 << 3);
+
+  /* Make sure laser is off.  We'll turn on the laser when
+     the interface is brought up. */
+  ixge_sfp_enable_disable_laser (xd, /* enable */ 0);
+  ixge_sfp_enable_disable_10g (xd, /* is_10g */ 1);
+
+  return 0;
+}
+
+static void
+ixge_sfp_device_up_down (ixge_device_t * xd, uword is_up)
+{
+  u32 v;
+
+  if (is_up)
+    {
+      /* pma/pmd 10g serial SFI. */
+      xd->regs->xge_mac.auto_negotiation_control2 &= ~(3 << 16);
+      xd->regs->xge_mac.auto_negotiation_control2 |= 2 << 16;
+
+      v = xd->regs->xge_mac.auto_negotiation_control;
+      v &= ~(7 << 13);
+      v |= (0 << 13);
+      /* Restart autoneg. */
+      v |= (1 << 12);
+      xd->regs->xge_mac.auto_negotiation_control = v;
+
+      while (! (xd->regs->xge_mac.link_partner_ability[0] & 0xf0000))
+	;
+
+      v = xd->regs->xge_mac.auto_negotiation_control;
+
+      /* link mode 10g sfi serdes */
+      v &= ~(7 << 13);
+      v |= (3 << 13);
+
+      /* Restart autoneg. */
+      v |= (1 << 12);
+      xd->regs->xge_mac.auto_negotiation_control = v;
+
+      xd->regs->xge_mac.link_status;
+    }
+
+  ixge_sfp_enable_disable_laser (xd, /* enable */ is_up);
+
+  /* Give time for link partner to notice that we're up. */
+  if (is_up &&
+      vlib_in_process_context(vlib_get_main())) {
+    vlib_process_suspend (vlib_get_main(), 300e-3);
+  }
+}
+
+always_inline ixge_dma_regs_t *
+get_dma_regs (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 qi)
+{
+  ixge_regs_t * r = xd->regs;
+  ASSERT (qi < 128);
+  if (rt == VLIB_RX)
+    return qi < 64 ? &r->rx_dma0[qi] : &r->rx_dma1[qi - 64];
+  else
+    return &r->tx_dma[qi];
+}
+
+static clib_error_t *
+ixge_interface_admin_up_down (vnet_main_t * vnm, u32 hw_if_index, u32 flags)
+{
+  vnet_hw_interface_t * hif = vnet_get_hw_interface (vnm, hw_if_index);
+  uword is_up = (flags & VNET_SW_INTERFACE_FLAG_ADMIN_UP) != 0;
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, hif->dev_instance);
+  ixge_dma_regs_t * dr = get_dma_regs (xd, VLIB_RX, 0);
+
+  if (is_up)
+    {
+      xd->regs->rx_enable |= 1;
+      xd->regs->tx_dma_control |= 1;
+      dr->control |= 1 << 25;
+      while (! (dr->control & (1 << 25)))
+        ;
+    }
+  else
+    {
+      xd->regs->rx_enable &= ~1;
+      xd->regs->tx_dma_control &= ~1;
+    }
+
+  ixge_sfp_device_up_down (xd, is_up);
+
+  return /* no error */ 0;
+}
+
+static void ixge_sfp_phy_init (ixge_device_t * xd)
+{
+  ixge_phy_t * phy = xd->phys + xd->phy_index;
+  i2c_bus_t * ib = &xd->i2c_bus;
+
+  ib->private_data = xd->device_index;
+  ib->put_bits = ixge_i2c_put_bits;
+  ib->get_bits = ixge_i2c_get_bits;
+  vlib_i2c_init (ib);
+
+  vlib_i2c_read_eeprom (ib, 0x50, 0, 128, (u8 *) &xd->sfp_eeprom);
+
+  if ( vlib_i2c_bus_timed_out(ib)  || ! sfp_eeprom_is_valid (&xd->sfp_eeprom))
+    xd->sfp_eeprom.id = SFP_ID_unknown;
+  else
+    {
+      /* FIXME 5 => SR/LR eeprom ID. */
+      clib_error_t * e = ixge_sfp_phy_init_from_eeprom (xd, 5 + xd->pci_function);
+      if (e)
+	clib_error_report (e);
+    }
+
+  phy->mdio_address = ~0;
+}
+
+static void ixge_phy_init (ixge_device_t * xd)
+{
+  ixge_main_t * xm = &ixge_main;
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_phy_t * phy = xd->phys + xd->phy_index;
+
+  switch (xd->device_id)
+    {
+    case IXGE_82599_sfp:
+    case IXGE_82599_sfp_em:
+    case IXGE_82599_sfp_fcoe:
+      /* others? */
+      return ixge_sfp_phy_init (xd);
+
+    default:
+      break;
+    }
+
+  /* Probe address of phy. */
+  {
+    u32 i, v;
+
+    phy->mdio_address = ~0;
+    for (i = 0; i < 32; i++)
+      {
+	phy->mdio_address = i;
+	v = ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PMA_PMD, XGE_PHY_ID1);
+	if (v != 0xffff && v != 0)
+	  break;
+      }
+
+    /* No PHY found? */
+    if (i >= 32)
+      return;
+  }
+
+  phy->id = ((ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PMA_PMD, XGE_PHY_ID1) << 16)
+	     | ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PMA_PMD, XGE_PHY_ID2));
+
+  {
+    ELOG_TYPE_DECLARE (e) = {
+      .function = (char *) __FUNCTION__,
+      .format = "ixge %d, phy id 0x%d mdio address %d",
+      .format_args = "i4i4i4",
+    };
+    struct { u32 instance, id, address; } * ed;
+    ed = ELOG_DATA (&vm->elog_main, e);
+    ed->instance = xd->device_index;
+    ed->id = phy->id;
+    ed->address = phy->mdio_address;
+  }
+
+  /* Reset phy. */
+  ixge_write_phy_reg (xd, XGE_PHY_DEV_TYPE_PHY_XS, XGE_PHY_CONTROL, XGE_PHY_CONTROL_RESET);
+
+  /* Wait for self-clearning reset bit to clear. */
+  do {
+    vlib_process_suspend (vm, 1e-3);
+  } while (ixge_read_phy_reg (xd, XGE_PHY_DEV_TYPE_PHY_XS, XGE_PHY_CONTROL) & XGE_PHY_CONTROL_RESET);
+}
+
+static u8 * format_ixge_rx_from_hw_descriptor (u8 * s, va_list * va)
+{
+  ixge_rx_from_hw_descriptor_t * d = va_arg (*va, ixge_rx_from_hw_descriptor_t *);
+  u32 s0 = d->status[0], s2 = d->status[2];
+  u32 is_ip4, is_ip6, is_ip, is_tcp, is_udp;
+  uword indent = format_get_indent (s);
+
+  s = format (s, "%s-owned",
+	      (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE) ? "sw" : "hw");
+  s = format (s, ", length this descriptor %d, l3 offset %d",
+	      d->n_packet_bytes_this_descriptor,
+	      IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s0));
+  if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET)
+    s = format (s, ", end-of-packet");
+
+  s = format (s, "\n%U", format_white_space, indent);
+
+  if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_ETHERNET_ERROR)
+    s = format (s, "layer2 error");
+
+  if (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_LAYER2)
+    {
+      s = format (s, "layer 2 type %d", (s0 & 0x1f));
+      return s;
+    }
+
+  if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_VLAN)
+    s = format (s, "vlan header 0x%x\n%U", d->vlan_tag,
+		format_white_space, indent);
+
+  if ((is_ip4 = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4)))
+    {
+      s = format (s, "ip4%s",
+		  (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4_EXT) ? " options" : "");
+      if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED)
+	s = format (s, " checksum %s",
+		    (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR) ? "bad" : "ok");
+    }
+  if ((is_ip6 = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6)))
+    s = format (s, "ip6%s",
+		(s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6_EXT) ? " extended" : "");
+  is_tcp = is_udp = 0;
+  if ((is_ip = (is_ip4 | is_ip6)))
+    {
+      is_tcp = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_TCP) != 0;
+      is_udp = (s0 & IXGE_RX_DESCRIPTOR_STATUS0_IS_UDP) != 0;
+      if (is_tcp)
+	s = format (s, ", tcp");
+      if (is_udp)
+	s = format (s, ", udp");
+    }
+
+  if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED)
+    s = format (s, ", tcp checksum %s",
+		(s2 & IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR) ? "bad" : "ok");
+  if (s2 & IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED)
+    s = format (s, ", udp checksum %s",
+		(s2 & IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR) ? "bad" : "ok");
+
+  return s;
+}
+
+static u8 * format_ixge_tx_descriptor (u8 * s, va_list * va)
+{
+  ixge_tx_descriptor_t * d = va_arg (*va, ixge_tx_descriptor_t *);
+  u32 s0 = d->status0, s1 = d->status1;
+  uword indent = format_get_indent (s);
+  u32 v;
+
+  s = format (s, "buffer 0x%Lx, %d packet bytes, %d bytes this buffer",
+	      d->buffer_address,
+	      s1 >> 14, d->n_bytes_this_buffer);
+
+  s = format (s, "\n%U", format_white_space, indent);
+
+  if ((v = (s0 >> 0) & 3))
+    s = format (s, "reserved 0x%x, ", v);
+
+  if ((v = (s0 >> 2) & 3))
+    s = format (s, "mac 0x%x, ", v);
+
+  if ((v = (s0 >> 4) & 0xf) != 3)
+    s = format (s, "type 0x%x, ", v);
+
+  s = format (s, "%s%s%s%s%s%s%s%s",
+	      (s0 & (1 << 8)) ? "eop, " : "",
+	      (s0 & (1 << 9)) ? "insert-fcs, " : "",
+	      (s0 & (1 << 10)) ? "reserved26, " : "",
+	      (s0 & (1 << 11)) ? "report-status, " : "",
+	      (s0 & (1 << 12)) ? "reserved28, " : "",
+	      (s0 & (1 << 13)) ? "is-advanced, " : "",
+	      (s0 & (1 << 14)) ? "vlan-enable, " : "",
+	      (s0 & (1 << 15)) ? "tx-segmentation, " : "");
+
+  if ((v = s1 & 0xf) != 0)
+    s = format (s, "status 0x%x, ", v);
+
+  if ((v = (s1 >> 4) & 0xf))
+    s = format (s, "context 0x%x, ", v);
+
+  if ((v = (s1 >> 8) & 0x3f))
+    s = format (s, "options 0x%x, ", v);
+
+  return s;
+}
+
+typedef struct {
+  ixge_descriptor_t before, after;
+
+  u32 buffer_index;
+
+  u16 device_index;
+
+  u8 queue_index;
+
+  u8 is_start_of_packet;
+
+  /* Copy of VLIB buffer; packet data stored in pre_data. */
+  vlib_buffer_t buffer;
+} ixge_rx_dma_trace_t;
+
+static u8 * format_ixge_rx_dma_trace (u8 * s, va_list * va)
+{
+  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *);
+  vlib_node_t * node = va_arg (*va, vlib_node_t *);
+  vnet_main_t * vnm = vnet_get_main();
+  ixge_rx_dma_trace_t * t = va_arg (*va, ixge_rx_dma_trace_t *);
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, t->device_index);
+  format_function_t * f;
+  uword indent = format_get_indent (s);
+
+  {
+    vnet_sw_interface_t * sw = vnet_get_sw_interface (vnm, xd->vlib_sw_if_index);
+    s = format (s, "%U rx queue %d",
+		format_vnet_sw_interface_name, vnm, sw,
+		t->queue_index);
+  }
+
+  s = format (s, "\n%Ubefore: %U",
+	      format_white_space, indent,
+	      format_ixge_rx_from_hw_descriptor, &t->before);
+  s = format (s, "\n%Uafter : head/tail address 0x%Lx/0x%Lx",
+	      format_white_space, indent,
+	      t->after.rx_to_hw.head_address,
+	      t->after.rx_to_hw.tail_address);
+
+  s = format (s, "\n%Ubuffer 0x%x: %U",
+	      format_white_space, indent,
+	      t->buffer_index,
+	      format_vlib_buffer, &t->buffer);
+
+  s = format (s, "\n%U",
+	      format_white_space, indent);
+
+  f = node->format_buffer;
+  if (! f || ! t->is_start_of_packet)
+    f = format_hex_bytes;
+  s = format (s, "%U", f, t->buffer.pre_data, sizeof (t->buffer.pre_data));
+
+  return s;
+}
+
+#define foreach_ixge_error					\
+  _ (none, "no error")						\
+  _ (tx_full_drops, "tx ring full drops")			\
+  _ (ip4_checksum_error, "ip4 checksum errors")			\
+  _ (rx_alloc_fail, "rx buf alloc from free list failed")	\
+  _ (rx_alloc_no_physmem, "rx buf alloc failed no physmem")
+
+typedef enum {
+#define _(f,s) IXGE_ERROR_##f,
+  foreach_ixge_error
+#undef _
+  IXGE_N_ERROR,
+} ixge_error_t;
+
+always_inline void
+ixge_rx_next_and_error_from_status_x1 (ixge_device_t *xd,
+                                       u32 s00, u32 s02,
+				       u8 * next0, u8 * error0, u32 * flags0)
+{
+  u8 is0_ip4, is0_ip6, n0, e0;
+  u32 f0;
+
+  e0 = IXGE_ERROR_none;
+  n0 = IXGE_RX_NEXT_ETHERNET_INPUT;
+
+  is0_ip4 = s02 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED;
+  n0 = is0_ip4 ? IXGE_RX_NEXT_IP4_INPUT : n0;
+
+  e0 = (is0_ip4 && (s02 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR)
+        ? IXGE_ERROR_ip4_checksum_error
+        : e0);
+
+  is0_ip6 = s00 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6;
+  n0 = is0_ip6 ? IXGE_RX_NEXT_IP6_INPUT : n0;
+
+  n0 = (xd->per_interface_next_index != ~0) ?
+    xd->per_interface_next_index : n0;
+
+  /* Check for error. */
+  n0 = e0 != IXGE_ERROR_none ? IXGE_RX_NEXT_DROP : n0;
+
+  f0 = ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED
+		| IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED))
+	? IP_BUFFER_L4_CHECKSUM_COMPUTED
+	: 0);
+
+  f0 |= ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR
+		 | IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR))
+	 ? 0
+	 : IP_BUFFER_L4_CHECKSUM_CORRECT);
+
+  *error0 = e0;
+  *next0 = n0;
+  *flags0 = f0;
+}
+
+always_inline void
+ixge_rx_next_and_error_from_status_x2 (ixge_device_t *xd,
+                                       u32 s00, u32 s02,
+				       u32 s10, u32 s12,
+				       u8 * next0, u8 * error0, u32 * flags0,
+				       u8 * next1, u8 * error1, u32 * flags1)
+{
+  u8 is0_ip4, is0_ip6, n0, e0;
+  u8 is1_ip4, is1_ip6, n1, e1;
+  u32 f0, f1;
+
+  e0 = e1 = IXGE_ERROR_none;
+  n0 = n1 = IXGE_RX_NEXT_IP4_INPUT;
+
+  is0_ip4 = s02 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED;
+  is1_ip4 = s12 & IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED;
+
+  n0 = is0_ip4 ? IXGE_RX_NEXT_IP4_INPUT : n0;
+  n1 = is1_ip4 ? IXGE_RX_NEXT_IP4_INPUT : n1;
+
+  e0 = (is0_ip4 && (s02 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR)
+        ? IXGE_ERROR_ip4_checksum_error
+        : e0);
+  e1 = (is1_ip4 && (s12 & IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR)
+	    ? IXGE_ERROR_ip4_checksum_error
+        : e1);
+
+  is0_ip6 = s00 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6;
+  is1_ip6 = s10 & IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6;
+
+  n0 = is0_ip6 ? IXGE_RX_NEXT_IP6_INPUT : n0;
+  n1 = is1_ip6 ? IXGE_RX_NEXT_IP6_INPUT : n1;
+
+  n0 = (xd->per_interface_next_index != ~0) ?
+    xd->per_interface_next_index : n0;
+  n1 = (xd->per_interface_next_index != ~0) ?
+    xd->per_interface_next_index : n1;
+
+  /* Check for error. */
+  n0 = e0 != IXGE_ERROR_none ? IXGE_RX_NEXT_DROP : n0;
+  n1 = e1 != IXGE_ERROR_none ? IXGE_RX_NEXT_DROP : n1;
+
+  *error0 = e0;
+  *error1 = e1;
+
+  *next0 = n0;
+  *next1 = n1;
+
+  f0 = ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED
+		| IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED))
+	? IP_BUFFER_L4_CHECKSUM_COMPUTED
+	: 0);
+  f1 = ((s12 & (IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED
+		| IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED))
+	? IP_BUFFER_L4_CHECKSUM_COMPUTED
+	: 0);
+
+  f0 |= ((s02 & (IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR
+		 | IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR))
+	 ? 0
+	 : IP_BUFFER_L4_CHECKSUM_CORRECT);
+  f1 |= ((s12 & (IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR
+		 | IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR))
+	 ? 0
+	 : IP_BUFFER_L4_CHECKSUM_CORRECT);
+
+  *flags0 = f0;
+  *flags1 = f1;
+}
+
+static void
+ixge_rx_trace (ixge_main_t * xm,
+	       ixge_device_t * xd,
+	       ixge_dma_queue_t * dq,
+	       ixge_descriptor_t * before_descriptors,
+	       u32 * before_buffers,
+	       ixge_descriptor_t * after_descriptors,
+	       uword n_descriptors)
+{
+  vlib_main_t * vm = xm->vlib_main;
+  vlib_node_runtime_t * node = dq->rx.node;
+  ixge_rx_from_hw_descriptor_t * bd;
+  ixge_rx_to_hw_descriptor_t * ad;
+  u32 * b, n_left, is_sop, next_index_sop;
+
+  n_left = n_descriptors;
+  b = before_buffers;
+  bd = &before_descriptors->rx_from_hw;
+  ad = &after_descriptors->rx_to_hw;
+  is_sop = dq->rx.is_start_of_packet;
+  next_index_sop = dq->rx.saved_start_of_packet_next_index;
+
+  while (n_left >= 2)
+    {
+      u32 bi0, bi1, flags0, flags1;
+      vlib_buffer_t * b0, * b1;
+      ixge_rx_dma_trace_t * t0, * t1;
+      u8 next0, error0, next1, error1;
+
+      bi0 = b[0];
+      bi1 = b[1];
+      n_left -= 2;
+
+      b0 = vlib_get_buffer (vm, bi0);
+      b1 = vlib_get_buffer (vm, bi1);
+
+      ixge_rx_next_and_error_from_status_x2 (xd,
+                                             bd[0].status[0], bd[0].status[2],
+					     bd[1].status[0], bd[1].status[2],
+					     &next0, &error0, &flags0,
+					     &next1, &error1, &flags1);
+
+      next_index_sop = is_sop ? next0 : next_index_sop;
+      vlib_trace_buffer (vm, node, next_index_sop, b0, /* follow_chain */ 0);
+      t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0]));
+      t0->is_start_of_packet = is_sop;
+      is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      next_index_sop = is_sop ? next1 : next_index_sop;
+      vlib_trace_buffer (vm, node, next_index_sop, b1, /* follow_chain */ 0);
+      t1 = vlib_add_trace (vm, node, b1, sizeof (t1[0]));
+      t1->is_start_of_packet = is_sop;
+      is_sop = (b1->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      t0->queue_index = dq->queue_index;
+      t1->queue_index = dq->queue_index;
+      t0->device_index = xd->device_index;
+      t1->device_index = xd->device_index;
+      t0->before.rx_from_hw = bd[0];
+      t1->before.rx_from_hw = bd[1];
+      t0->after.rx_to_hw = ad[0];
+      t1->after.rx_to_hw = ad[1];
+      t0->buffer_index = bi0;
+      t1->buffer_index = bi1;
+      memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data));
+      memcpy (&t1->buffer, b1, sizeof (b1[0]) - sizeof (b0->pre_data));
+      memcpy (t0->buffer.pre_data, b0->data + b0->current_data,
+	      sizeof (t0->buffer.pre_data));
+      memcpy (t1->buffer.pre_data, b1->data + b1->current_data,
+	      sizeof (t1->buffer.pre_data));
+
+      b += 2;
+      bd += 2;
+      ad += 2;
+    }
+
+  while (n_left >= 1)
+    {
+      u32 bi0, flags0;
+      vlib_buffer_t * b0;
+      ixge_rx_dma_trace_t * t0;
+      u8 next0, error0;
+
+      bi0 = b[0];
+      n_left -= 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      ixge_rx_next_and_error_from_status_x1 (xd,
+                                             bd[0].status[0], bd[0].status[2],
+					     &next0, &error0, &flags0);
+
+      next_index_sop = is_sop ? next0 : next_index_sop;
+      vlib_trace_buffer (vm, node, next_index_sop, b0, /* follow_chain */ 0);
+      t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0]));
+      t0->is_start_of_packet = is_sop;
+      is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      t0->queue_index = dq->queue_index;
+      t0->device_index = xd->device_index;
+      t0->before.rx_from_hw = bd[0];
+      t0->after.rx_to_hw = ad[0];
+      t0->buffer_index = bi0;
+      memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data));
+      memcpy (t0->buffer.pre_data, b0->data + b0->current_data,
+	      sizeof (t0->buffer.pre_data));
+
+      b += 1;
+      bd += 1;
+      ad += 1;
+    }
+}
+
+typedef struct {
+  ixge_tx_descriptor_t descriptor;
+
+  u32 buffer_index;
+
+  u16 device_index;
+
+  u8 queue_index;
+
+  u8 is_start_of_packet;
+
+  /* Copy of VLIB buffer; packet data stored in pre_data. */
+  vlib_buffer_t buffer;
+} ixge_tx_dma_trace_t;
+
+static u8 * format_ixge_tx_dma_trace (u8 * s, va_list * va)
+{
+  CLIB_UNUSED (vlib_main_t * vm) = va_arg (*va, vlib_main_t *);
+  CLIB_UNUSED (vlib_node_t * node) = va_arg (*va, vlib_node_t *);
+  ixge_tx_dma_trace_t * t = va_arg (*va, ixge_tx_dma_trace_t *);
+  vnet_main_t * vnm = vnet_get_main();
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, t->device_index);
+  format_function_t * f;
+  uword indent = format_get_indent (s);
+
+  {
+    vnet_sw_interface_t * sw = vnet_get_sw_interface (vnm, xd->vlib_sw_if_index);
+    s = format (s, "%U tx queue %d",
+		format_vnet_sw_interface_name, vnm, sw,
+		t->queue_index);
+  }
+
+  s = format (s, "\n%Udescriptor: %U",
+	      format_white_space, indent,
+	      format_ixge_tx_descriptor, &t->descriptor);
+
+  s = format (s, "\n%Ubuffer 0x%x: %U",
+	      format_white_space, indent,
+	      t->buffer_index,
+	      format_vlib_buffer, &t->buffer);
+
+  s = format (s, "\n%U",
+	      format_white_space, indent);
+
+  f = format_ethernet_header_with_length;
+  if (! f || ! t->is_start_of_packet)
+    f = format_hex_bytes;
+  s = format (s, "%U", f, t->buffer.pre_data, sizeof (t->buffer.pre_data));
+
+  return s;
+}
+
+typedef struct {
+  vlib_node_runtime_t * node;
+
+  u32 is_start_of_packet;
+
+  u32 n_bytes_in_packet;
+
+  ixge_tx_descriptor_t * start_of_packet_descriptor;
+} ixge_tx_state_t;
+
+static void
+ixge_tx_trace (ixge_main_t * xm,
+	       ixge_device_t * xd,
+	       ixge_dma_queue_t * dq,
+	       ixge_tx_state_t * tx_state,
+	       ixge_tx_descriptor_t * descriptors,
+	       u32 * buffers,
+	       uword n_descriptors)
+{
+  vlib_main_t * vm = xm->vlib_main;
+  vlib_node_runtime_t * node = tx_state->node;
+  ixge_tx_descriptor_t * d;
+  u32 * b, n_left, is_sop;
+
+  n_left = n_descriptors;
+  b = buffers;
+  d = descriptors;
+  is_sop = tx_state->is_start_of_packet;
+
+  while (n_left >= 2)
+    {
+      u32 bi0, bi1;
+      vlib_buffer_t * b0, * b1;
+      ixge_tx_dma_trace_t * t0, * t1;
+
+      bi0 = b[0];
+      bi1 = b[1];
+      n_left -= 2;
+
+      b0 = vlib_get_buffer (vm, bi0);
+      b1 = vlib_get_buffer (vm, bi1);
+
+      t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0]));
+      t0->is_start_of_packet = is_sop;
+      is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      t1 = vlib_add_trace (vm, node, b1, sizeof (t1[0]));
+      t1->is_start_of_packet = is_sop;
+      is_sop = (b1->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      t0->queue_index = dq->queue_index;
+      t1->queue_index = dq->queue_index;
+      t0->device_index = xd->device_index;
+      t1->device_index = xd->device_index;
+      t0->descriptor = d[0];
+      t1->descriptor = d[1];
+      t0->buffer_index = bi0;
+      t1->buffer_index = bi1;
+      memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data));
+      memcpy (&t1->buffer, b1, sizeof (b1[0]) - sizeof (b0->pre_data));
+      memcpy (t0->buffer.pre_data, b0->data + b0->current_data,
+	      sizeof (t0->buffer.pre_data));
+      memcpy (t1->buffer.pre_data, b1->data + b1->current_data,
+	      sizeof (t1->buffer.pre_data));
+
+      b += 2;
+      d += 2;
+    }
+
+  while (n_left >= 1)
+    {
+      u32 bi0;
+      vlib_buffer_t * b0;
+      ixge_tx_dma_trace_t * t0;
+
+      bi0 = b[0];
+      n_left -= 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      t0 = vlib_add_trace (vm, node, b0, sizeof (t0[0]));
+      t0->is_start_of_packet = is_sop;
+      is_sop = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      t0->queue_index = dq->queue_index;
+      t0->device_index = xd->device_index;
+      t0->descriptor = d[0];
+      t0->buffer_index = bi0;
+      memcpy (&t0->buffer, b0, sizeof (b0[0]) - sizeof (b0->pre_data));
+      memcpy (t0->buffer.pre_data, b0->data + b0->current_data,
+	      sizeof (t0->buffer.pre_data));
+
+      b += 1;
+      d += 1;
+    }
+}
+
+always_inline uword
+ixge_ring_sub (ixge_dma_queue_t * q, u32 i0, u32 i1)
+{
+  i32 d = i1 - i0;
+  ASSERT (i0 < q->n_descriptors);
+  ASSERT (i1 < q->n_descriptors);
+  return d < 0 ? q->n_descriptors + d : d;
+}
+
+always_inline uword
+ixge_ring_add (ixge_dma_queue_t * q, u32 i0, u32 i1)
+{
+  u32 d = i0 + i1;
+  ASSERT (i0 < q->n_descriptors);
+  ASSERT (i1 < q->n_descriptors);
+  d -= d >= q->n_descriptors ? q->n_descriptors : 0;
+  return d;
+}
+
+always_inline uword
+ixge_tx_descriptor_matches_template (ixge_main_t * xm, ixge_tx_descriptor_t * d)
+{
+  u32 cmp;
+
+  cmp = ((d->status0 & xm->tx_descriptor_template_mask.status0)
+	 ^ xm->tx_descriptor_template.status0);
+  if (cmp)
+    return 0;
+  cmp = ((d->status1 & xm->tx_descriptor_template_mask.status1)
+	 ^ xm->tx_descriptor_template.status1);
+  if (cmp)
+    return 0;
+
+  return 1;
+}
+
+static uword
+ixge_tx_no_wrap (ixge_main_t * xm,
+		 ixge_device_t * xd,
+		 ixge_dma_queue_t * dq,
+		 u32 * buffers,
+		 u32 start_descriptor_index,
+		 u32 n_descriptors,
+		 ixge_tx_state_t * tx_state)
+{
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_tx_descriptor_t * d, * d_sop;
+  u32 n_left = n_descriptors;
+  u32 * to_free = vec_end (xm->tx_buffers_pending_free);
+  u32 * to_tx = vec_elt_at_index (dq->descriptor_buffer_indices, start_descriptor_index);
+  u32 is_sop = tx_state->is_start_of_packet;
+  u32 len_sop = tx_state->n_bytes_in_packet;
+  u16 template_status = xm->tx_descriptor_template.status0;
+  u32 descriptor_prefetch_rotor = 0;
+
+  ASSERT (start_descriptor_index + n_descriptors <= dq->n_descriptors);
+  d = &dq->descriptors[start_descriptor_index].tx;
+  d_sop = is_sop ? d : tx_state->start_of_packet_descriptor;
+
+  while (n_left >= 4)
+    {
+      vlib_buffer_t * b0, * b1;
+      u32 bi0, fi0, len0;
+      u32 bi1, fi1, len1;
+      u8 is_eop0, is_eop1;
+
+      /* Prefetch next iteration. */
+      vlib_prefetch_buffer_with_index (vm, buffers[2], LOAD);
+      vlib_prefetch_buffer_with_index (vm, buffers[3], LOAD);
+
+      if ((descriptor_prefetch_rotor & 0x3) == 0)
+        CLIB_PREFETCH (d + 4, CLIB_CACHE_LINE_BYTES, STORE);
+
+      descriptor_prefetch_rotor += 2;
+
+      bi0 = buffers[0];
+      bi1 = buffers[1];
+
+      to_free[0] = fi0 = to_tx[0];
+      to_tx[0] = bi0;
+      to_free += fi0 != 0;
+
+      to_free[0] = fi1 = to_tx[1];
+      to_tx[1] = bi1;
+      to_free += fi1 != 0;
+
+      buffers += 2;
+      n_left -= 2;
+      to_tx += 2;
+
+      b0 = vlib_get_buffer (vm, bi0);
+      b1 = vlib_get_buffer (vm, bi1);
+
+      is_eop0 = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+      is_eop1 = (b1->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      len0 = b0->current_length;
+      len1 = b1->current_length;
+
+      ASSERT (ixge_tx_descriptor_matches_template (xm, d + 0));
+      ASSERT (ixge_tx_descriptor_matches_template (xm, d + 1));
+
+      d[0].buffer_address = vlib_get_buffer_data_physical_address (vm, bi0) + b0->current_data;
+      d[1].buffer_address = vlib_get_buffer_data_physical_address (vm, bi1) + b1->current_data;
+
+      d[0].n_bytes_this_buffer = len0;
+      d[1].n_bytes_this_buffer = len1;
+
+      d[0].status0 = template_status | (is_eop0 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET);
+      d[1].status0 = template_status | (is_eop1 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET);
+
+      len_sop = (is_sop ? 0 : len_sop) + len0;
+      d_sop[0].status1 = IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET (len_sop);
+      d += 1;
+      d_sop = is_eop0 ? d : d_sop;
+
+      is_sop = is_eop0;
+
+      len_sop = (is_sop ? 0 : len_sop) + len1;
+      d_sop[0].status1 = IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET (len_sop);
+      d += 1;
+      d_sop = is_eop1 ? d : d_sop;
+
+      is_sop = is_eop1;
+    }
+
+  while (n_left > 0)
+    {
+      vlib_buffer_t * b0;
+      u32 bi0, fi0, len0;
+      u8 is_eop0;
+
+      bi0 = buffers[0];
+
+      to_free[0] = fi0 = to_tx[0];
+      to_tx[0] = bi0;
+      to_free += fi0 != 0;
+
+      buffers += 1;
+      n_left -= 1;
+      to_tx += 1;
+
+      b0 = vlib_get_buffer (vm, bi0);
+
+      is_eop0 = (b0->flags & VLIB_BUFFER_NEXT_PRESENT) == 0;
+
+      len0 = b0->current_length;
+
+      ASSERT (ixge_tx_descriptor_matches_template (xm, d + 0));
+
+      d[0].buffer_address = vlib_get_buffer_data_physical_address (vm, bi0) + b0->current_data;
+
+      d[0].n_bytes_this_buffer = len0;
+
+      d[0].status0 = template_status | (is_eop0 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET);
+
+      len_sop = (is_sop ? 0 : len_sop) + len0;
+      d_sop[0].status1 = IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET (len_sop);
+      d += 1;
+      d_sop = is_eop0 ? d : d_sop;
+
+      is_sop = is_eop0;
+    }
+
+  if (tx_state->node->flags & VLIB_NODE_FLAG_TRACE)
+    {
+      to_tx = vec_elt_at_index (dq->descriptor_buffer_indices, start_descriptor_index);
+      ixge_tx_trace (xm, xd, dq, tx_state,
+		     &dq->descriptors[start_descriptor_index].tx,
+		     to_tx,
+		     n_descriptors);
+    }
+
+  _vec_len (xm->tx_buffers_pending_free) = to_free - xm->tx_buffers_pending_free;
+
+  /* When we are done d_sop can point to end of ring.  Wrap it if so. */
+  {
+    ixge_tx_descriptor_t * d_start = &dq->descriptors[0].tx;
+
+    ASSERT (d_sop - d_start <= dq->n_descriptors);
+    d_sop = d_sop - d_start == dq->n_descriptors ? d_start : d_sop;
+  }
+
+  tx_state->is_start_of_packet = is_sop;
+  tx_state->start_of_packet_descriptor = d_sop;
+  tx_state->n_bytes_in_packet = len_sop;
+
+  return n_descriptors;
+}
+
+static uword
+ixge_interface_tx (vlib_main_t * vm,
+		   vlib_node_runtime_t * node,
+		   vlib_frame_t * f)
+{
+  ixge_main_t * xm = &ixge_main;
+  vnet_interface_output_runtime_t * rd = (void *) node->runtime_data;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, rd->dev_instance);
+  ixge_dma_queue_t * dq;
+  u32 * from, n_left_tx, n_descriptors_to_tx, n_tail_drop;
+  u32 queue_index = 0;		/* fixme parameter */
+  ixge_tx_state_t tx_state;
+
+  tx_state.node = node;
+  tx_state.is_start_of_packet = 1;
+  tx_state.start_of_packet_descriptor = 0;
+  tx_state.n_bytes_in_packet = 0;
+
+  from = vlib_frame_vector_args (f);
+
+  dq = vec_elt_at_index (xd->dma_queues[VLIB_TX], queue_index);
+
+  dq->head_index = dq->tx.head_index_write_back[0];
+
+  /* Since head == tail means ring is empty we can send up to dq->n_descriptors - 1. */
+  n_left_tx = dq->n_descriptors - 1;
+  n_left_tx -= ixge_ring_sub (dq, dq->head_index, dq->tail_index);
+
+  _vec_len (xm->tx_buffers_pending_free) = 0;
+
+  n_descriptors_to_tx = f->n_vectors;
+  n_tail_drop = 0;
+  if (PREDICT_FALSE (n_descriptors_to_tx > n_left_tx))
+    {
+      i32 i, n_ok, i_eop, i_sop;
+
+      i_sop = i_eop = ~0;
+      for (i = n_left_tx - 1; i >= 0; i--)
+	{
+	  vlib_buffer_t * b = vlib_get_buffer (vm, from[i]);
+	  if (! (b->flags & VLIB_BUFFER_NEXT_PRESENT))
+	    {
+	      if (i_sop != ~0 && i_eop != ~0)
+		break;
+	      i_eop = i;
+	      i_sop = i + 1;
+	    }
+	}
+      if (i == 0)
+	n_ok = 0;
+      else
+	n_ok = i_eop + 1;
+
+      {
+	ELOG_TYPE_DECLARE (e) = {
+	  .function = (char *) __FUNCTION__,
+	  .format = "ixge %d, ring full to tx %d head %d tail %d",
+	  .format_args = "i2i2i2i2",
+	};
+	struct { u16 instance, to_tx, head, tail; } * ed;
+	ed = ELOG_DATA (&vm->elog_main, e);
+	ed->instance = xd->device_index;
+	ed->to_tx = n_descriptors_to_tx;
+	ed->head = dq->head_index;
+	ed->tail = dq->tail_index;
+      }
+
+      if (n_ok < n_descriptors_to_tx)
+	{
+	  n_tail_drop = n_descriptors_to_tx - n_ok;
+	  vec_add (xm->tx_buffers_pending_free, from + n_ok, n_tail_drop);
+	  vlib_error_count (vm, ixge_input_node.index, IXGE_ERROR_tx_full_drops, n_tail_drop);
+	}
+
+      n_descriptors_to_tx = n_ok;
+    }
+
+  dq->tx.n_buffers_on_ring += n_descriptors_to_tx;
+
+  /* Process from tail to end of descriptor ring. */
+  if (n_descriptors_to_tx > 0 && dq->tail_index < dq->n_descriptors)
+    {
+      u32 n = clib_min (dq->n_descriptors - dq->tail_index, n_descriptors_to_tx);
+      n = ixge_tx_no_wrap (xm, xd, dq, from, dq->tail_index, n, &tx_state);
+      from += n;
+      n_descriptors_to_tx -= n;
+      dq->tail_index += n;
+      ASSERT (dq->tail_index <= dq->n_descriptors);
+      if (dq->tail_index == dq->n_descriptors)
+	dq->tail_index = 0;
+    }
+
+  if (n_descriptors_to_tx > 0)
+    {
+      u32 n = ixge_tx_no_wrap (xm, xd, dq, from, 0, n_descriptors_to_tx, &tx_state);
+      from += n;
+      ASSERT (n == n_descriptors_to_tx);
+      dq->tail_index += n;
+      ASSERT (dq->tail_index <= dq->n_descriptors);
+      if (dq->tail_index == dq->n_descriptors)
+	dq->tail_index = 0;
+    }
+
+  /* We should only get full packets. */
+  ASSERT (tx_state.is_start_of_packet);
+
+  /* Report status when last descriptor is done. */
+  {
+    u32 i = dq->tail_index == 0 ? dq->n_descriptors - 1 : dq->tail_index - 1;
+    ixge_tx_descriptor_t * d = &dq->descriptors[i].tx;
+    d->status0 |= IXGE_TX_DESCRIPTOR_STATUS0_REPORT_STATUS;
+  }
+
+  /* Give new descriptors to hardware. */
+  {
+    ixge_dma_regs_t * dr = get_dma_regs (xd, VLIB_TX, queue_index);
+
+    CLIB_MEMORY_BARRIER ();
+
+    dr->tail_index = dq->tail_index;
+  }
+
+  /* Free any buffers that are done. */
+  {
+    u32 n = _vec_len (xm->tx_buffers_pending_free);
+    if (n > 0)
+      {
+	vlib_buffer_free_no_next (vm, xm->tx_buffers_pending_free, n);
+	_vec_len (xm->tx_buffers_pending_free) = 0;
+	ASSERT (dq->tx.n_buffers_on_ring >= n);
+	dq->tx.n_buffers_on_ring -= (n - n_tail_drop);
+      }
+  }
+
+  return f->n_vectors;
+}
+
+static uword
+ixge_rx_queue_no_wrap (ixge_main_t * xm,
+		       ixge_device_t * xd,
+		       ixge_dma_queue_t * dq,
+		       u32 start_descriptor_index,
+		       u32 n_descriptors)
+{
+  vlib_main_t * vm = xm->vlib_main;
+  vlib_node_runtime_t * node = dq->rx.node;
+  ixge_descriptor_t * d;
+  static ixge_descriptor_t * d_trace_save;
+  static u32 * d_trace_buffers;
+  u32 n_descriptors_left = n_descriptors;
+  u32 * to_rx = vec_elt_at_index (dq->descriptor_buffer_indices, start_descriptor_index);
+  u32 * to_add;
+  u32 bi_sop = dq->rx.saved_start_of_packet_buffer_index;
+  u32 bi_last = dq->rx.saved_last_buffer_index;
+  u32 next_index_sop = dq->rx.saved_start_of_packet_next_index;
+  u32 is_sop = dq->rx.is_start_of_packet;
+  u32 next_index, n_left_to_next, * to_next;
+  u32 n_packets = 0;
+  u32 n_bytes = 0;
+  u32 n_trace = vlib_get_trace_count (vm, node);
+  vlib_buffer_t * b_last, b_dummy;
+
+  ASSERT (start_descriptor_index + n_descriptors <= dq->n_descriptors);
+  d = &dq->descriptors[start_descriptor_index];
+
+  b_last = bi_last != ~0 ? vlib_get_buffer (vm, bi_last) : &b_dummy;
+  next_index = dq->rx.next_index;
+
+  if (n_trace > 0)
+    {
+      u32 n = clib_min (n_trace, n_descriptors);
+      if (d_trace_save)
+	{
+	  _vec_len (d_trace_save) = 0;
+	  _vec_len (d_trace_buffers) = 0;
+	}
+      vec_add (d_trace_save, (ixge_descriptor_t *) d, n);
+      vec_add (d_trace_buffers, to_rx, n);
+    }
+
+  {
+    uword l = vec_len (xm->rx_buffers_to_add);
+
+    if (l < n_descriptors_left)
+      {
+	u32 n_to_alloc = 2 * dq->n_descriptors - l;
+	u32 n_allocated;
+
+	vec_resize (xm->rx_buffers_to_add, n_to_alloc);
+
+	_vec_len (xm->rx_buffers_to_add) = l;
+	n_allocated = vlib_buffer_alloc_from_free_list
+	  (vm, xm->rx_buffers_to_add + l, n_to_alloc,
+	   xm->vlib_buffer_free_list_index);
+	_vec_len (xm->rx_buffers_to_add) += n_allocated;
+
+        /* Handle transient allocation failure */
+	if (PREDICT_FALSE(l + n_allocated <= n_descriptors_left))
+	  {
+	    if (n_allocated == 0)
+	      vlib_error_count (vm, ixge_input_node.index,
+				IXGE_ERROR_rx_alloc_no_physmem, 1);
+	    else
+	      vlib_error_count (vm, ixge_input_node.index,
+				IXGE_ERROR_rx_alloc_fail, 1);
+
+	    n_descriptors_left = l + n_allocated;
+	  }
+        n_descriptors = n_descriptors_left;
+      }
+
+    /* Add buffers from end of vector going backwards. */
+    to_add = vec_end (xm->rx_buffers_to_add) - 1;
+  }
+
+  while (n_descriptors_left > 0)
+    {
+      vlib_get_next_frame (vm, node, next_index,
+			   to_next, n_left_to_next);
+
+      while (n_descriptors_left >= 4 && n_left_to_next >= 2)
+	{
+	  vlib_buffer_t * b0, * b1;
+	  u32 bi0, fi0, len0, l3_offset0, s20, s00, flags0;
+	  u32 bi1, fi1, len1, l3_offset1, s21, s01, flags1;
+	  u8 is_eop0, error0, next0;
+	  u8 is_eop1, error1, next1;
+          ixge_descriptor_t d0, d1;
+
+	  vlib_prefetch_buffer_with_index (vm, to_rx[2], STORE);
+	  vlib_prefetch_buffer_with_index (vm, to_rx[3], STORE);
+
+          CLIB_PREFETCH (d + 2, 32, STORE);
+
+          d0.as_u32x4 = d[0].as_u32x4;
+          d1.as_u32x4 = d[1].as_u32x4;
+
+	  s20 = d0.rx_from_hw.status[2];
+	  s21 = d1.rx_from_hw.status[2];
+
+	  s00 = d0.rx_from_hw.status[0];
+	  s01 = d1.rx_from_hw.status[0];
+
+	  if (! ((s20 & s21) & IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE))
+	    goto found_hw_owned_descriptor_x2;
+
+	  bi0 = to_rx[0];
+	  bi1 = to_rx[1];
+
+	  ASSERT (to_add - 1 >= xm->rx_buffers_to_add);
+	  fi0 = to_add[0];
+	  fi1 = to_add[-1];
+
+	  to_rx[0] = fi0;
+	  to_rx[1] = fi1;
+	  to_rx += 2;
+	  to_add -= 2;
+
+	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == vlib_buffer_is_known (vm, bi0));
+	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == vlib_buffer_is_known (vm, bi1));
+	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == vlib_buffer_is_known (vm, fi0));
+	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == vlib_buffer_is_known (vm, fi1));
+
+	  b0 = vlib_get_buffer (vm, bi0);
+	  b1 = vlib_get_buffer (vm, bi1);
+
+          /*
+           * Turn this on if you run into
+           * "bad monkey" contexts, and you want to know exactly
+           * which nodes they've visited... See main.c...
+           */
+          VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b0);
+          VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b1);
+
+	  CLIB_PREFETCH (b0->data, CLIB_CACHE_LINE_BYTES, LOAD);
+	  CLIB_PREFETCH (b1->data, CLIB_CACHE_LINE_BYTES, LOAD);
+
+	  is_eop0 = (s20 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) != 0;
+	  is_eop1 = (s21 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) != 0;
+
+	  ixge_rx_next_and_error_from_status_x2 (xd, s00, s20, s01, s21,
+						 &next0, &error0, &flags0,
+						 &next1, &error1, &flags1);
+
+	  next0 = is_sop ? next0 : next_index_sop;
+	  next1 = is_eop0 ? next1 : next0;
+	  next_index_sop = next1;
+
+	  b0->flags |= flags0 | (!is_eop0 << VLIB_BUFFER_LOG2_NEXT_PRESENT);
+	  b1->flags |= flags1 | (!is_eop1 << VLIB_BUFFER_LOG2_NEXT_PRESENT);
+
+	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index;
+	  vnet_buffer (b1)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index;
+	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32)~0;
+	  vnet_buffer (b1)->sw_if_index[VLIB_TX] = (u32)~0;
+
+	  b0->error = node->errors[error0];
+	  b1->error = node->errors[error1];
+
+	  len0 = d0.rx_from_hw.n_packet_bytes_this_descriptor;
+	  len1 = d1.rx_from_hw.n_packet_bytes_this_descriptor;
+	  n_bytes += len0 + len1;
+	  n_packets += is_eop0 + is_eop1;
+
+	  /* Give new buffers to hardware. */
+	  d0.rx_to_hw.tail_address =
+              vlib_get_buffer_data_physical_address (vm, fi0);
+	  d1.rx_to_hw.tail_address =
+              vlib_get_buffer_data_physical_address (vm, fi1);
+	  d0.rx_to_hw.head_address = d[0].rx_to_hw.tail_address;
+	  d1.rx_to_hw.head_address = d[1].rx_to_hw.tail_address;
+          d[0].as_u32x4 = d0.as_u32x4;
+          d[1].as_u32x4 = d1.as_u32x4;
+
+	  d += 2;
+	  n_descriptors_left -= 2;
+
+	  /* Point to either l2 or l3 header depending on next. */
+	  l3_offset0 = (is_sop && (next0 != IXGE_RX_NEXT_ETHERNET_INPUT))
+              ? IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s00)
+              : 0;
+	  l3_offset1 = (is_eop0 && (next1 != IXGE_RX_NEXT_ETHERNET_INPUT))
+              ? IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s01)
+              : 0;
+
+	  b0->current_length = len0 - l3_offset0;
+	  b1->current_length = len1 - l3_offset1;
+	  b0->current_data = l3_offset0;
+	  b1->current_data = l3_offset1;
+
+	  b_last->next_buffer = is_sop ? ~0 : bi0;
+	  b0->next_buffer = is_eop0 ? ~0 : bi1;
+	  bi_last = bi1;
+	  b_last = b1;
+
+	  if (CLIB_DEBUG > 0)
+	    {
+	      u32 bi_sop0 = is_sop ? bi0 : bi_sop;
+	      u32 bi_sop1 = is_eop0 ? bi1 : bi_sop0;
+
+	      if (is_eop0)
+		{
+		  u8 * msg = vlib_validate_buffer (vm, bi_sop0, /* follow_buffer_next */ 1);
+		  ASSERT (! msg);
+		}
+	      if (is_eop1)
+		{
+		  u8 * msg = vlib_validate_buffer (vm, bi_sop1, /* follow_buffer_next */ 1);
+		  ASSERT (! msg);
+		}
+	    }
+          if (0) /* "Dave" version */
+            {
+              u32 bi_sop0 = is_sop ? bi0 : bi_sop;
+              u32 bi_sop1 = is_eop0 ? bi1 : bi_sop0;
+
+              if (is_eop0)
+                {
+                  to_next[0] = bi_sop0;
+                  to_next++;
+                  n_left_to_next--;
+
+                  vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
+                                                   to_next, n_left_to_next,
+                                                   bi_sop0, next0);
+                }
+              if (is_eop1)
+                {
+                  to_next[0] = bi_sop1;
+                  to_next++;
+                  n_left_to_next--;
+
+                  vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
+                                                   to_next, n_left_to_next,
+                                                   bi_sop1, next1);
+                }
+              is_sop = is_eop1;
+              bi_sop = bi_sop1;
+            }
+          if (1) /* "Eliot" version */
+            {
+              /* Speculatively enqueue to cached next. */
+              u8 saved_is_sop = is_sop;
+              u32 bi_sop_save = bi_sop;
+
+              bi_sop = saved_is_sop ? bi0 : bi_sop;
+              to_next[0] = bi_sop;
+              to_next += is_eop0;
+              n_left_to_next -= is_eop0;
+
+              bi_sop = is_eop0 ? bi1 : bi_sop;
+              to_next[0] = bi_sop;
+              to_next += is_eop1;
+              n_left_to_next -= is_eop1;
+
+              is_sop = is_eop1;
+
+              if (PREDICT_FALSE (! (next0 == next_index && next1 == next_index)))
+                {
+                  /* Undo speculation. */
+                  to_next -= is_eop0 + is_eop1;
+                  n_left_to_next += is_eop0 + is_eop1;
+
+                  /* Re-do both descriptors being careful about where we enqueue. */
+                  bi_sop = saved_is_sop ? bi0 : bi_sop_save;
+                  if (is_eop0)
+                    {
+                      if (next0 != next_index)
+                        vlib_set_next_frame_buffer (vm, node, next0, bi_sop);
+                      else
+                        {
+                          to_next[0] = bi_sop;
+                          to_next += 1;
+                          n_left_to_next -= 1;
+                        }
+                    }
+
+                  bi_sop = is_eop0 ? bi1 : bi_sop;
+                  if (is_eop1)
+                    {
+                      if (next1 != next_index)
+                        vlib_set_next_frame_buffer (vm, node, next1, bi_sop);
+                      else
+                        {
+                          to_next[0] = bi_sop;
+                          to_next += 1;
+                          n_left_to_next -= 1;
+                        }
+                    }
+
+                  /* Switch cached next index when next for both packets is the same. */
+                  if (is_eop0 && is_eop1 && next0 == next1)
+                    {
+                      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+                      next_index = next0;
+                      vlib_get_next_frame (vm, node, next_index,
+                                           to_next, n_left_to_next);
+                    }
+                }
+            }
+	}
+
+    /* Bail out of dual loop and proceed with single loop. */
+    found_hw_owned_descriptor_x2:
+
+      while (n_descriptors_left > 0 && n_left_to_next > 0)
+	{
+	  vlib_buffer_t * b0;
+	  u32 bi0, fi0, len0, l3_offset0, s20, s00, flags0;
+	  u8 is_eop0, error0, next0;
+          ixge_descriptor_t d0;
+
+          d0.as_u32x4 = d[0].as_u32x4;
+
+	  s20 = d0.rx_from_hw.status[2];
+	  s00 = d0.rx_from_hw.status[0];
+
+	  if (! (s20 & IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE))
+	    goto found_hw_owned_descriptor_x1;
+
+	  bi0 = to_rx[0];
+	  ASSERT (to_add >= xm->rx_buffers_to_add);
+	  fi0 = to_add[0];
+
+	  to_rx[0] = fi0;
+	  to_rx += 1;
+	  to_add -= 1;
+
+	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == vlib_buffer_is_known (vm, bi0));
+	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED == vlib_buffer_is_known (vm, fi0));
+
+	  b0 = vlib_get_buffer (vm, bi0);
+
+          /*
+           * Turn this on if you run into
+           * "bad monkey" contexts, and you want to know exactly
+           * which nodes they've visited...
+           */
+          VLIB_BUFFER_TRACE_TRAJECTORY_INIT(b0);
+
+	  is_eop0 = (s20 & IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET) != 0;
+	  ixge_rx_next_and_error_from_status_x1
+            (xd, s00, s20, &next0, &error0, &flags0);
+
+	  next0 = is_sop ? next0 : next_index_sop;
+	  next_index_sop = next0;
+
+	  b0->flags |= flags0 | (!is_eop0 << VLIB_BUFFER_LOG2_NEXT_PRESENT);
+
+	  vnet_buffer (b0)->sw_if_index[VLIB_RX] = xd->vlib_sw_if_index;
+	  vnet_buffer (b0)->sw_if_index[VLIB_TX] = (u32)~0;
+
+	  b0->error = node->errors[error0];
+
+	  len0 = d0.rx_from_hw.n_packet_bytes_this_descriptor;
+	  n_bytes += len0;
+	  n_packets += is_eop0;
+
+	  /* Give new buffer to hardware. */
+	  d0.rx_to_hw.tail_address =
+              vlib_get_buffer_data_physical_address (vm, fi0);
+	  d0.rx_to_hw.head_address = d0.rx_to_hw.tail_address;
+          d[0].as_u32x4 = d0.as_u32x4;
+
+	  d += 1;
+	  n_descriptors_left -= 1;
+
+	  /* Point to either l2 or l3 header depending on next. */
+	  l3_offset0 = (is_sop && (next0 != IXGE_RX_NEXT_ETHERNET_INPUT))
+              ? IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET (s00)
+              : 0;
+	  b0->current_length = len0 - l3_offset0;
+	  b0->current_data = l3_offset0;
+
+	  b_last->next_buffer = is_sop ? ~0 : bi0;
+	  bi_last = bi0;
+	  b_last = b0;
+
+	  bi_sop = is_sop ? bi0 : bi_sop;
+
+	  if (CLIB_DEBUG > 0 && is_eop0)
+	    {
+	      u8 * msg = vlib_validate_buffer (vm, bi_sop, /* follow_buffer_next */ 1);
+	      ASSERT (! msg);
+	    }
+
+          if (0) /* "Dave" version */
+            {
+              if (is_eop0)
+                {
+                  to_next[0] = bi_sop;
+                  to_next++;
+                  n_left_to_next--;
+
+                  vlib_validate_buffer_enqueue_x1 (vm, node, next_index,
+                                                   to_next, n_left_to_next,
+                                                   bi_sop, next0);
+                }
+            }
+          if (1) /* "Eliot" version */
+            {
+              if (PREDICT_TRUE (next0 == next_index))
+                {
+                  to_next[0] = bi_sop;
+                  to_next += is_eop0;
+                  n_left_to_next -= is_eop0;
+                }
+              else
+                {
+                  if (next0 != next_index && is_eop0)
+                    vlib_set_next_frame_buffer (vm, node, next0, bi_sop);
+
+                  vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+                  next_index = next0;
+                  vlib_get_next_frame (vm, node, next_index,
+                                       to_next, n_left_to_next);
+                }
+            }
+          is_sop = is_eop0;
+	}
+      vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+    }
+
+ found_hw_owned_descriptor_x1:
+  if (n_descriptors_left > 0)
+    vlib_put_next_frame (vm, node, next_index, n_left_to_next);
+
+  _vec_len (xm->rx_buffers_to_add) = (to_add + 1) - xm->rx_buffers_to_add;
+
+  {
+    u32 n_done = n_descriptors - n_descriptors_left;
+
+    if (n_trace > 0 && n_done > 0)
+      {
+	u32 n = clib_min (n_trace, n_done);
+	ixge_rx_trace (xm, xd, dq,
+		       d_trace_save,
+		       d_trace_buffers,
+		       &dq->descriptors[start_descriptor_index],
+		       n);
+	vlib_set_trace_count (vm, node, n_trace - n);
+      }
+    if (d_trace_save)
+      {
+	_vec_len (d_trace_save) = 0;
+	_vec_len (d_trace_buffers) = 0;
+      }
+
+    /* Don't keep a reference to b_last if we don't have to.
+       Otherwise we can over-write a next_buffer pointer after already haven
+       enqueued a packet. */
+    if (is_sop)
+      {
+        b_last->next_buffer = ~0;
+        bi_last = ~0;
+      }
+
+    dq->rx.n_descriptors_done_this_call = n_done;
+    dq->rx.n_descriptors_done_total += n_done;
+    dq->rx.is_start_of_packet = is_sop;
+    dq->rx.saved_start_of_packet_buffer_index = bi_sop;
+    dq->rx.saved_last_buffer_index = bi_last;
+    dq->rx.saved_start_of_packet_next_index = next_index_sop;
+    dq->rx.next_index = next_index;
+    dq->rx.n_bytes += n_bytes;
+
+    return n_packets;
+  }
+}
+
+static uword
+ixge_rx_queue (ixge_main_t * xm,
+	       ixge_device_t * xd,
+	       vlib_node_runtime_t * node,
+	       u32 queue_index)
+{
+  ixge_dma_queue_t * dq = vec_elt_at_index (xd->dma_queues[VLIB_RX], queue_index);
+  ixge_dma_regs_t * dr = get_dma_regs (xd, VLIB_RX, dq->queue_index);
+  uword n_packets = 0;
+  u32 hw_head_index, sw_head_index;
+
+  /* One time initialization. */
+  if (! dq->rx.node)
+    {
+      dq->rx.node = node;
+      dq->rx.is_start_of_packet = 1;
+      dq->rx.saved_start_of_packet_buffer_index = ~0;
+      dq->rx.saved_last_buffer_index = ~0;
+    }
+
+  dq->rx.next_index = node->cached_next_index;
+
+  dq->rx.n_descriptors_done_total = 0;
+  dq->rx.n_descriptors_done_this_call = 0;
+  dq->rx.n_bytes = 0;
+
+  /* Fetch head from hardware and compare to where we think we are. */
+  hw_head_index = dr->head_index;
+  sw_head_index = dq->head_index;
+
+  if (hw_head_index == sw_head_index)
+    goto done;
+
+  if (hw_head_index < sw_head_index)
+    {
+      u32 n_tried = dq->n_descriptors - sw_head_index;
+      n_packets += ixge_rx_queue_no_wrap (xm, xd, dq, sw_head_index, n_tried);
+      sw_head_index = ixge_ring_add (dq, sw_head_index, dq->rx.n_descriptors_done_this_call);
+
+      if (dq->rx.n_descriptors_done_this_call != n_tried)
+	goto done;
+    }
+  if (hw_head_index >= sw_head_index)
+    {
+      u32 n_tried = hw_head_index - sw_head_index;
+      n_packets += ixge_rx_queue_no_wrap (xm, xd, dq, sw_head_index, n_tried);
+      sw_head_index = ixge_ring_add (dq, sw_head_index, dq->rx.n_descriptors_done_this_call);
+    }
+
+ done:
+  dq->head_index = sw_head_index;
+  dq->tail_index = ixge_ring_add (dq, dq->tail_index, dq->rx.n_descriptors_done_total);
+
+  /* Give tail back to hardware. */
+  CLIB_MEMORY_BARRIER ();
+
+  dr->tail_index = dq->tail_index;
+
+  vlib_increment_combined_counter (vnet_main.interface_main.combined_sw_if_counters
+				   + VNET_INTERFACE_COUNTER_RX,
+                                   0 /* cpu_index */,
+				   xd->vlib_sw_if_index,
+				   n_packets,
+				   dq->rx.n_bytes);
+
+  return n_packets;
+}
+
+static void ixge_interrupt (ixge_main_t * xm, ixge_device_t * xd, u32 i)
+{
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_regs_t * r = xd->regs;
+
+  if (i != 20)
+    {
+      ELOG_TYPE_DECLARE (e) = {
+	.function = (char *) __FUNCTION__,
+	.format = "ixge %d, %s",
+	.format_args = "i1t1",
+	.n_enum_strings = 16,
+	.enum_strings = {
+	  "flow director",
+	  "rx miss",
+	  "pci exception",
+	  "mailbox",
+	  "link status change",
+	  "linksec key exchange",
+	  "manageability event",
+	  "reserved23",
+	  "sdp0",
+	  "sdp1",
+	  "sdp2",
+	  "sdp3",
+	  "ecc",
+	  "descriptor handler error",
+	  "tcp timer",
+	  "other",
+	},
+      };
+      struct { u8 instance; u8 index; } * ed;
+      ed = ELOG_DATA (&vm->elog_main, e);
+      ed->instance = xd->device_index;
+      ed->index = i - 16;
+    }
+  else
+    {
+      u32 v = r->xge_mac.link_status;
+      uword is_up = (v & (1 << 30)) != 0;
+
+      ELOG_TYPE_DECLARE (e) = {
+	.function = (char *) __FUNCTION__,
+	.format = "ixge %d, link status change 0x%x",
+	.format_args = "i4i4",
+      };
+      struct { u32 instance, link_status; } * ed;
+      ed = ELOG_DATA (&vm->elog_main, e);
+      ed->instance = xd->device_index;
+      ed->link_status = v;
+      xd->link_status_at_last_link_change = v;
+
+      vlib_process_signal_event (vm, ixge_process_node.index,
+                                 EVENT_SET_FLAGS,
+                                 ((is_up<<31) | xd->vlib_hw_if_index));
+    }
+}
+
+always_inline u32
+clean_block (u32 * b, u32 * t, u32 n_left)
+{
+  u32 * t0 = t;
+
+  while (n_left >= 4)
+    {
+      u32 bi0, bi1, bi2, bi3;
+
+      t[0] = bi0 = b[0];
+      b[0] = 0;
+      t += bi0 != 0;
+
+      t[0] = bi1 = b[1];
+      b[1] = 0;
+      t += bi1 != 0;
+
+      t[0] = bi2 = b[2];
+      b[2] = 0;
+      t += bi2 != 0;
+
+      t[0] = bi3 = b[3];
+      b[3] = 0;
+      t += bi3 != 0;
+
+      b += 4;
+      n_left -= 4;
+    }
+
+  while (n_left > 0)
+    {
+      u32 bi0;
+
+      t[0] = bi0 = b[0];
+      b[0] = 0;
+      t += bi0 != 0;
+      b += 1;
+      n_left -= 1;
+    }
+
+  return t - t0;
+}
+
+static void
+ixge_tx_queue (ixge_main_t * xm, ixge_device_t * xd, u32 queue_index)
+{
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_dma_queue_t * dq = vec_elt_at_index (xd->dma_queues[VLIB_TX], queue_index);
+  u32 n_clean, * b, * t, * t0;
+  i32 n_hw_owned_descriptors;
+  i32 first_to_clean, last_to_clean;
+  u64 hwbp_race = 0;
+
+  /* Handle case where head write back pointer update
+   * arrives after the interrupt during high PCI bus loads.
+   */
+  while ((dq->head_index == dq->tx.head_index_write_back[0]) &&
+	 dq->tx.n_buffers_on_ring && (dq->head_index != dq->tail_index))
+    {
+      hwbp_race++;
+      if (IXGE_HWBP_RACE_ELOG && (hwbp_race == 1))
+	{
+	  ELOG_TYPE_DECLARE (e) = {
+	    .function = (char *) __FUNCTION__,
+	    .format = "ixge %d tx head index race: head %4d, tail %4d, buffs %4d",
+	    .format_args = "i4i4i4i4",
+	  };
+	  struct { u32 instance, head_index, tail_index, n_buffers_on_ring; } * ed;
+	  ed = ELOG_DATA (&vm->elog_main, e);
+	  ed->instance = xd->device_index;
+	  ed->head_index = dq->head_index;
+	  ed->tail_index = dq->tail_index;
+	  ed->n_buffers_on_ring = dq->tx.n_buffers_on_ring;
+	}
+    }
+
+  dq->head_index = dq->tx.head_index_write_back[0];
+  n_hw_owned_descriptors = ixge_ring_sub (dq, dq->head_index, dq->tail_index);
+  ASSERT(dq->tx.n_buffers_on_ring >= n_hw_owned_descriptors);
+  n_clean = dq->tx.n_buffers_on_ring - n_hw_owned_descriptors;
+
+  if (IXGE_HWBP_RACE_ELOG && hwbp_race)
+    {
+	  ELOG_TYPE_DECLARE (e) = {
+	    .function = (char *) __FUNCTION__,
+	    .format = "ixge %d tx head index race: head %4d, hw_owned %4d, n_clean %4d, retries %d",
+	    .format_args = "i4i4i4i4i4",
+	  };
+	  struct { u32 instance, head_index, n_hw_owned_descriptors, n_clean, retries; } * ed;
+	  ed = ELOG_DATA (&vm->elog_main, e);
+	  ed->instance = xd->device_index;
+	  ed->head_index = dq->head_index;
+	  ed->n_hw_owned_descriptors = n_hw_owned_descriptors;
+	  ed->n_clean = n_clean;
+	  ed->retries = hwbp_race;
+    }
+
+  /*
+   * This function used to wait until hardware owned zero descriptors.
+   * At high PPS rates, that doesn't happen until the TX ring is
+   * completely full of descriptors which need to be cleaned up.
+   * That, in turn, causes TX ring-full drops and/or long RX service
+   * interruptions.
+   */
+  if (n_clean == 0)
+    return;
+
+  /* Clean the n_clean descriptors prior to the reported hardware head */
+  last_to_clean = dq->head_index - 1;
+  last_to_clean = (last_to_clean < 0) ? last_to_clean + dq->n_descriptors :
+      last_to_clean;
+
+  first_to_clean = (last_to_clean) - (n_clean - 1);
+  first_to_clean = (first_to_clean < 0) ? first_to_clean + dq->n_descriptors :
+      first_to_clean;
+
+  vec_resize (xm->tx_buffers_pending_free, dq->n_descriptors - 1);
+  t0 = t = xm->tx_buffers_pending_free;
+  b = dq->descriptor_buffer_indices + first_to_clean;
+
+  /* Wrap case: clean from first to end, then start to last */
+  if (first_to_clean > last_to_clean)
+    {
+      t += clean_block (b, t, (dq->n_descriptors - 1) - first_to_clean);
+      first_to_clean = 0;
+      b = dq->descriptor_buffer_indices;
+    }
+
+  /* Typical case: clean from first to last */
+  if (first_to_clean <= last_to_clean)
+      t += clean_block (b, t, (last_to_clean - first_to_clean) + 1);
+
+  if (t > t0)
+    {
+      u32 n = t - t0;
+      vlib_buffer_free_no_next (vm, t0, n);
+      ASSERT (dq->tx.n_buffers_on_ring >= n);
+      dq->tx.n_buffers_on_ring -= n;
+      _vec_len (xm->tx_buffers_pending_free) = 0;
+    }
+}
+
+/* RX queue interrupts 0 thru 7; TX 8 thru 15. */
+always_inline uword ixge_interrupt_is_rx_queue (uword i)
+{ return i < 8; }
+
+always_inline uword ixge_interrupt_is_tx_queue (uword i)
+{ return i >= 8 && i < 16; }
+
+always_inline uword ixge_tx_queue_to_interrupt (uword i)
+{ return 8 + i; }
+
+always_inline uword ixge_rx_queue_to_interrupt (uword i)
+{ return 0 + i; }
+
+always_inline uword ixge_interrupt_rx_queue (uword i)
+{
+  ASSERT (ixge_interrupt_is_rx_queue (i));
+  return i - 0;
+}
+
+always_inline uword ixge_interrupt_tx_queue (uword i)
+{
+  ASSERT (ixge_interrupt_is_tx_queue (i));
+  return i - 8;
+}
+
+static uword
+ixge_device_input (ixge_main_t * xm,
+		   ixge_device_t * xd,
+		   vlib_node_runtime_t * node)
+{
+  ixge_regs_t * r = xd->regs;
+  u32 i, s;
+  uword n_rx_packets = 0;
+
+  s = r->interrupt.status_write_1_to_set;
+  if (s)
+    r->interrupt.status_write_1_to_clear = s;
+
+  foreach_set_bit (i, s, ({
+    if (ixge_interrupt_is_rx_queue (i))
+      n_rx_packets += ixge_rx_queue (xm, xd, node, ixge_interrupt_rx_queue (i));
+
+    else if (ixge_interrupt_is_tx_queue (i))
+      ixge_tx_queue (xm, xd, ixge_interrupt_tx_queue (i));
+
+    else
+      ixge_interrupt (xm, xd, i);
+  }));
+
+  return n_rx_packets;
+}
+
+static uword
+ixge_input (vlib_main_t * vm,
+	    vlib_node_runtime_t * node,
+	    vlib_frame_t * f)
+{
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd;
+  uword n_rx_packets = 0;
+
+  if (node->state == VLIB_NODE_STATE_INTERRUPT)
+    {
+      uword i;
+
+      /* Loop over devices with interrupts. */
+      foreach_set_bit (i, node->runtime_data[0], ({
+	xd = vec_elt_at_index (xm->devices, i);
+	n_rx_packets += ixge_device_input (xm, xd, node);
+
+	/* Re-enable interrupts since we're going to stay in interrupt mode. */
+	if (! (node->flags & VLIB_NODE_FLAG_SWITCH_FROM_INTERRUPT_TO_POLLING_MODE))
+	  xd->regs->interrupt.enable_write_1_to_set = ~0;
+      }));
+
+      /* Clear mask of devices with pending interrupts. */
+      node->runtime_data[0] = 0;
+    }
+  else
+    {
+      /* Poll all devices for input/interrupts. */
+      vec_foreach (xd, xm->devices)
+	{
+	  n_rx_packets += ixge_device_input (xm, xd, node);
+
+	  /* Re-enable interrupts when switching out of polling mode. */
+	  if (node->flags &
+	      VLIB_NODE_FLAG_SWITCH_FROM_POLLING_TO_INTERRUPT_MODE)
+	    xd->regs->interrupt.enable_write_1_to_set = ~0;
+	}
+    }
+
+  return n_rx_packets;
+}
+
+static char * ixge_error_strings[] = {
+#define _(n,s) s,
+    foreach_ixge_error
+#undef _
+};
+
+static vlib_node_registration_t ixge_input_node = {
+  .function = ixge_input,
+  .type = VLIB_NODE_TYPE_INPUT,
+  .name = "ixge-input",
+
+  /* Will be enabled if/when hardware is detected. */
+  .state = VLIB_NODE_STATE_DISABLED,
+
+  .format_buffer = format_ethernet_header_with_length,
+  .format_trace = format_ixge_rx_dma_trace,
+
+  .n_errors = IXGE_N_ERROR,
+  .error_strings = ixge_error_strings,
+
+  .n_next_nodes = IXGE_RX_N_NEXT,
+  .next_nodes = {
+    [IXGE_RX_NEXT_DROP] = "error-drop",
+    [IXGE_RX_NEXT_ETHERNET_INPUT] = "ethernet-input",
+    [IXGE_RX_NEXT_IP4_INPUT] = "ip4-input",
+    [IXGE_RX_NEXT_IP6_INPUT] = "ip6-input",
+  },
+};
+
+VLIB_NODE_FUNCTION_MULTIARCH_CLONE (ixge_input)
+CLIB_MULTIARCH_SELECT_FN (ixge_input)
+
+static u8 * format_ixge_device_name (u8 * s, va_list * args)
+{
+  u32 i = va_arg (*args, u32);
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, i);
+  return format (s, "TenGigabitEthernet%U",
+		 format_vlib_pci_handle, &xd->pci_device.bus_address);
+}
+
+#define IXGE_COUNTER_IS_64_BIT (1 << 0)
+#define IXGE_COUNTER_NOT_CLEAR_ON_READ (1 << 1)
+
+static u8 ixge_counter_flags[] = {
+#define _(a,f) 0,
+#define _64(a,f) IXGE_COUNTER_IS_64_BIT,
+  foreach_ixge_counter
+#undef _
+#undef _64
+};
+
+static void ixge_update_counters (ixge_device_t * xd)
+{
+  /* Byte offset for counter registers. */
+  static u32 reg_offsets[] = {
+#define _(a,f) (a) / sizeof (u32),
+#define _64(a,f) _(a,f)
+    foreach_ixge_counter
+#undef _
+#undef _64
+  };
+  volatile u32 * r = (volatile u32 *) xd->regs;
+  int i;
+
+  for (i = 0; i < ARRAY_LEN (xd->counters); i++)
+    {
+      u32 o = reg_offsets[i];
+      xd->counters[i] += r[o];
+      if (ixge_counter_flags[i] & IXGE_COUNTER_NOT_CLEAR_ON_READ)
+	r[o] = 0;
+      if (ixge_counter_flags[i] & IXGE_COUNTER_IS_64_BIT)
+	xd->counters[i] += (u64) r[o+1] << (u64) 32;
+    }
+}
+
+static u8 * format_ixge_device_id (u8 * s, va_list * args)
+{
+  u32 device_id = va_arg (*args, u32);
+  char * t = 0;
+  switch (device_id)
+    {
+#define _(f,n) case n: t = #f; break;
+      foreach_ixge_pci_device_id;
+#undef _
+    default:
+      t = 0;
+      break;
+    }
+  if (t == 0)
+    s = format (s, "unknown 0x%x", device_id);
+  else
+    s = format (s, "%s", t);
+  return s;
+}
+
+static u8 * format_ixge_link_status (u8 * s, va_list * args)
+{
+  ixge_device_t * xd = va_arg (*args, ixge_device_t *);
+  u32 v = xd->link_status_at_last_link_change;
+
+  s = format (s, "%s", (v & (1 << 30)) ? "up" : "down");
+
+  {
+    char * modes[] = {
+      "1g", "10g parallel", "10g serial", "autoneg",
+    };
+    char * speeds[] = {
+      "unknown", "100m", "1g", "10g",
+    };
+    s = format (s, ", mode %s, speed %s",
+		modes[(v >> 26) & 3],
+		speeds[(v >> 28) & 3]);
+  }
+
+  return s;
+}
+
+static u8 * format_ixge_device (u8 * s, va_list * args)
+{
+  u32 dev_instance = va_arg (*args, u32);
+  CLIB_UNUSED (int verbose) = va_arg (*args, int);
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, dev_instance);
+  ixge_phy_t * phy = xd->phys + xd->phy_index;
+  uword indent = format_get_indent (s);
+
+  ixge_update_counters (xd);
+  xd->link_status_at_last_link_change = xd->regs->xge_mac.link_status;
+
+  s = format (s, "Intel 8259X: id %U\n%Ulink %U",
+	      format_ixge_device_id, xd->device_id,
+	      format_white_space, indent + 2,
+	      format_ixge_link_status, xd);
+
+  {
+
+    s = format (s, "\n%UPCIe %U", format_white_space, indent + 2,
+		format_vlib_pci_link_speed, &xd->pci_device);
+  }
+
+  s = format (s, "\n%U", format_white_space, indent + 2);
+  if (phy->mdio_address != ~0)
+    s = format (s, "PHY address %d, id 0x%x", phy->mdio_address, phy->id);
+  else if (xd->sfp_eeprom.id == SFP_ID_sfp)
+    s = format (s, "SFP %U", format_sfp_eeprom, &xd->sfp_eeprom);
+  else
+    s = format (s, "PHY not found");
+
+  /* FIXME */
+  {
+    ixge_dma_queue_t * dq = vec_elt_at_index (xd->dma_queues[VLIB_RX], 0);
+    ixge_dma_regs_t * dr = get_dma_regs (xd, VLIB_RX, 0);
+    u32 hw_head_index = dr->head_index;
+    u32 sw_head_index = dq->head_index;
+    u32 nitems;
+
+    nitems = ixge_ring_sub (dq, hw_head_index, sw_head_index);
+    s = format (s, "\n%U%d unprocessed, %d total buffers on rx queue 0 ring",
+                format_white_space, indent + 2, nitems, dq->n_descriptors);
+
+    s = format (s, "\n%U%d buffers in driver rx cache",
+                format_white_space, indent + 2, vec_len(xm->rx_buffers_to_add));
+
+    s = format (s, "\n%U%d buffers on tx queue 0 ring",
+                format_white_space, indent + 2,
+                xd->dma_queues[VLIB_TX][0].tx.n_buffers_on_ring);
+  }
+  {
+    u32 i;
+    u64 v;
+    static char * names[] = {
+#define _(a,f) #f,
+#define _64(a,f) _(a,f)
+    foreach_ixge_counter
+#undef _
+#undef _64
+    };
+
+    for (i = 0; i < ARRAY_LEN (names); i++)
+      {
+	v = xd->counters[i] - xd->counters_last_clear[i];
+	if (v != 0)
+	  s = format (s, "\n%U%-40U%16Ld",
+		      format_white_space, indent + 2,
+		      format_c_identifier, names[i],
+		      v);
+      }
+  }
+
+  return s;
+}
+
+static void ixge_clear_hw_interface_counters (u32 instance)
+{
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, instance);
+  ixge_update_counters (xd);
+  memcpy (xd->counters_last_clear, xd->counters, sizeof (xd->counters));
+}
+
+/*
+ * Dynamically redirect all pkts from a specific interface
+ * to the specified node
+ */
+static void ixge_set_interface_next_node (vnet_main_t *vnm, u32 hw_if_index,
+                                          u32 node_index)
+{
+  ixge_main_t * xm = &ixge_main;
+  vnet_hw_interface_t *hw = vnet_get_hw_interface (vnm, hw_if_index);
+  ixge_device_t * xd = vec_elt_at_index (xm->devices, hw->dev_instance);
+
+  /* Shut off redirection */
+  if (node_index == ~0)
+    {
+      xd->per_interface_next_index = node_index;
+      return;
+    }
+
+  xd->per_interface_next_index =
+    vlib_node_add_next (xm->vlib_main, ixge_input_node.index, node_index);
+}
+
+
+VNET_DEVICE_CLASS (ixge_device_class) = {
+  .name = "ixge",
+  .tx_function = ixge_interface_tx,
+  .format_device_name = format_ixge_device_name,
+  .format_device = format_ixge_device,
+  .format_tx_trace = format_ixge_tx_dma_trace,
+  .clear_counters = ixge_clear_hw_interface_counters,
+  .admin_up_down_function = ixge_interface_admin_up_down,
+  .rx_redirect_to_node = ixge_set_interface_next_node,
+};
+
+#define IXGE_N_BYTES_IN_RX_BUFFER  (2048)  // DAW-HACK: Set Rx buffer size so all packets < ETH_MTU_SIZE fit in the buffer (i.e. sop & eop for all descriptors).
+
+static clib_error_t *
+ixge_dma_init (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 queue_index)
+{
+  ixge_main_t * xm = &ixge_main;
+  vlib_main_t * vm = xm->vlib_main;
+  ixge_dma_queue_t * dq;
+  clib_error_t * error = 0;
+
+  vec_validate (xd->dma_queues[rt], queue_index);
+  dq = vec_elt_at_index (xd->dma_queues[rt], queue_index);
+
+  if (! xm->n_descriptors_per_cache_line)
+    xm->n_descriptors_per_cache_line = CLIB_CACHE_LINE_BYTES / sizeof (dq->descriptors[0]);
+
+  if (! xm->n_bytes_in_rx_buffer)
+    xm->n_bytes_in_rx_buffer = IXGE_N_BYTES_IN_RX_BUFFER;
+  xm->n_bytes_in_rx_buffer = round_pow2 (xm->n_bytes_in_rx_buffer, 1024);
+  if (! xm->vlib_buffer_free_list_index)
+    {
+      xm->vlib_buffer_free_list_index = vlib_buffer_get_or_create_free_list (vm, xm->n_bytes_in_rx_buffer, "ixge rx");
+      ASSERT (xm->vlib_buffer_free_list_index != 0);
+    }
+
+  if (! xm->n_descriptors[rt])
+    xm->n_descriptors[rt] = 4 * VLIB_FRAME_SIZE;
+
+  dq->queue_index = queue_index;
+  dq->n_descriptors = round_pow2 (xm->n_descriptors[rt], xm->n_descriptors_per_cache_line);
+  dq->head_index = dq->tail_index = 0;
+
+  dq->descriptors = vlib_physmem_alloc_aligned (vm, &error,
+						dq->n_descriptors * sizeof (dq->descriptors[0]),
+						128 /* per chip spec */);
+  if (error)
+    return error;
+
+  memset (dq->descriptors, 0, dq->n_descriptors * sizeof (dq->descriptors[0]));
+  vec_resize (dq->descriptor_buffer_indices, dq->n_descriptors);
+
+  if (rt == VLIB_RX)
+    {
+      u32 n_alloc, i;
+
+      n_alloc = vlib_buffer_alloc_from_free_list
+	(vm, dq->descriptor_buffer_indices, vec_len (dq->descriptor_buffer_indices),
+	 xm->vlib_buffer_free_list_index);
+      ASSERT (n_alloc == vec_len (dq->descriptor_buffer_indices));
+      for (i = 0; i < n_alloc; i++)
+	{
+	  vlib_buffer_t * b = vlib_get_buffer (vm, dq->descriptor_buffer_indices[i]);
+	  dq->descriptors[i].rx_to_hw.tail_address = vlib_physmem_virtual_to_physical (vm, b->data);
+	}
+    }
+  else
+    {
+      u32 i;
+
+      dq->tx.head_index_write_back = vlib_physmem_alloc (vm, &error, CLIB_CACHE_LINE_BYTES);
+
+      for (i = 0; i < dq->n_descriptors; i++)
+	dq->descriptors[i].tx = xm->tx_descriptor_template;
+
+      vec_validate (xm->tx_buffers_pending_free, dq->n_descriptors - 1);
+    }
+
+  {
+    ixge_dma_regs_t * dr = get_dma_regs (xd, rt, queue_index);
+    u64 a;
+
+    a = vlib_physmem_virtual_to_physical (vm, dq->descriptors);
+    dr->descriptor_address[0] = a & 0xFFFFFFFF;
+    dr->descriptor_address[1] = a >> (u64) 32;
+    dr->n_descriptor_bytes = dq->n_descriptors * sizeof (dq->descriptors[0]);
+    dq->head_index = dq->tail_index = 0;
+
+    if (rt == VLIB_RX)
+      {
+	ASSERT ((xm->n_bytes_in_rx_buffer / 1024) < 32);
+	dr->rx_split_control =
+	  (/* buffer size */ ((xm->n_bytes_in_rx_buffer / 1024) << 0)
+	   | (/* lo free descriptor threshold (units of 64 descriptors) */
+	      (1 << 22))
+	   | (/* descriptor type: advanced one buffer */
+	      (1 << 25))
+	   | (/* drop if no descriptors available */
+	      (1 << 28)));
+
+	/* Give hardware all but last 16 cache lines' worth of descriptors. */
+	dq->tail_index = dq->n_descriptors -
+          16*xm->n_descriptors_per_cache_line;
+      }
+    else
+      {
+	/* Make sure its initialized before hardware can get to it. */
+	dq->tx.head_index_write_back[0] = dq->head_index;
+
+	a = vlib_physmem_virtual_to_physical (vm, dq->tx.head_index_write_back);
+	dr->tx.head_index_write_back_address[0] = /* enable bit */ 1 | a;
+	dr->tx.head_index_write_back_address[1] = (u64) a >> (u64) 32;
+      }
+
+    /* DMA on 82599 does not work with [13] rx data write relaxed ordering
+       and [12] undocumented set. */
+    if (rt == VLIB_RX)
+      dr->dca_control &= ~((1 << 13) | (1 << 12));
+
+    CLIB_MEMORY_BARRIER ();
+
+    if (rt == VLIB_TX)
+      {
+        xd->regs->tx_dma_control |= (1 << 0);
+        dr->control |= ((32 << 0) /* prefetch threshold */
+                        | (64 << 8) /* host threshold */
+                        | (0 << 16) /* writeback threshold*/);
+      }
+
+    /* Enable this queue and wait for hardware to initialize
+       before adding to tail. */
+    if (rt == VLIB_TX)
+      {
+        dr->control |= 1 << 25;
+        while (! (dr->control & (1 << 25)))
+          ;
+      }
+
+    /* Set head/tail indices and enable DMA. */
+    dr->head_index = dq->head_index;
+    dr->tail_index = dq->tail_index;
+  }
+
+  return error;
+}
+
+static u32 ixge_flag_change (vnet_main_t * vnm,
+                             vnet_hw_interface_t * hw,
+                             u32 flags)
+{
+  ixge_device_t * xd;
+  ixge_regs_t * r;
+  u32 old;
+  ixge_main_t * xm = &ixge_main;
+
+  xd = vec_elt_at_index (xm->devices, hw->dev_instance);
+  r = xd->regs;
+
+  old = r->filter_control;
+
+  if (flags & ETHERNET_INTERFACE_FLAG_ACCEPT_ALL)
+    r->filter_control = old |(1 << 9) /* unicast promiscuous */;
+  else
+    r->filter_control = old & ~(1 << 9);
+
+  return old;
+}
+
+static void ixge_device_init (ixge_main_t * xm)
+{
+  vnet_main_t * vnm = vnet_get_main();
+  ixge_device_t * xd;
+
+  /* Reset chip(s). */
+  vec_foreach (xd, xm->devices)
+    {
+      ixge_regs_t * r = xd->regs;
+      const u32 reset_bit = (1 << 26) | (1 << 3);
+
+      r->control |= reset_bit;
+
+      /* No need to suspend.  Timed to take ~1e-6 secs */
+      while (r->control & reset_bit)
+	;
+
+      /* Software loaded. */
+      r->extended_control |= (1 << 28);
+
+      ixge_phy_init (xd);
+
+      /* Register ethernet interface. */
+      {
+	u8 addr8[6];
+	u32 i, addr32[2];
+	clib_error_t * error;
+
+	addr32[0] = r->rx_ethernet_address0[0][0];
+	addr32[1] = r->rx_ethernet_address0[0][1];
+	for (i = 0; i < 6; i++)
+	  addr8[i] = addr32[i / 4] >> ((i % 4) * 8);
+
+	error = ethernet_register_interface
+	  (vnm,
+	   ixge_device_class.index,
+	   xd->device_index,
+	   /* ethernet address */ addr8,
+	   &xd->vlib_hw_if_index,
+           ixge_flag_change);
+	if (error)
+	  clib_error_report (error);
+      }
+
+      {
+	vnet_sw_interface_t * sw = vnet_get_hw_sw_interface (vnm, xd->vlib_hw_if_index);
+	xd->vlib_sw_if_index = sw->sw_if_index;
+      }
+
+      ixge_dma_init (xd, VLIB_RX, /* queue_index */ 0);
+
+      xm->n_descriptors[VLIB_TX] = 20 * VLIB_FRAME_SIZE;
+
+      ixge_dma_init (xd, VLIB_TX, /* queue_index */ 0);
+
+      /* RX/TX queue 0 gets mapped to interrupt bits 0 & 8. */
+      r->interrupt.queue_mapping[0] =
+	((/* valid bit */ (1 << 7) |
+	  ixge_rx_queue_to_interrupt (0)) << 0);
+
+      r->interrupt.queue_mapping[0] |=
+	((/* valid bit */ (1 << 7) |
+	  ixge_tx_queue_to_interrupt (0)) << 8);
+
+      /* No use in getting too many interrupts.
+	 Limit them to one every 3/4 ring size at line rate
+	 min sized packets.
+	 No need for this since kernel/vlib main loop provides adequate interrupt
+	 limiting scheme. */
+      if (0)
+	{
+	  f64 line_rate_max_pps = 10e9 / (8 * (64 + /* interframe padding */ 20));
+	  ixge_throttle_queue_interrupt (r, 0, .75 * xm->n_descriptors[VLIB_RX] / line_rate_max_pps);
+	}
+
+      /* Accept all multicast and broadcast packets. Should really add them
+	 to the dst_ethernet_address register array. */
+      r->filter_control |= (1 << 10) | (1 << 8);
+
+      /* Enable frames up to size in mac frame size register. */
+      r->xge_mac.control |= 1 << 2;
+      r->xge_mac.rx_max_frame_size = (9216 + 14) << 16;
+
+      /* Enable all interrupts. */
+      if (! IXGE_ALWAYS_POLL)
+	r->interrupt.enable_write_1_to_set = ~0;
+    }
+}
+
+static uword
+ixge_process (vlib_main_t * vm,
+	      vlib_node_runtime_t * rt,
+	      vlib_frame_t * f)
+{
+  vnet_main_t * vnm = vnet_get_main();
+  ixge_main_t * xm = &ixge_main;
+  ixge_device_t * xd;
+  uword event_type, * event_data = 0;
+  f64 timeout, link_debounce_deadline;
+
+  ixge_device_init (xm);
+
+  /* Clear all counters. */
+  vec_foreach (xd, xm->devices)
+    {
+      ixge_update_counters (xd);
+      memset (xd->counters, 0, sizeof (xd->counters));
+    }
+
+  timeout = 30.0;
+  link_debounce_deadline = 1e70;
+
+  while (1)
+    {
+      /* 36 bit stat counters could overflow in ~50 secs.
+	 We poll every 30 secs to be conservative. */
+      vlib_process_wait_for_event_or_clock (vm, timeout);
+
+      event_type = vlib_process_get_events (vm, &event_data);
+
+      switch (event_type) {
+      case EVENT_SET_FLAGS:
+        /* 1 ms */
+        link_debounce_deadline = vlib_time_now(vm) + 1e-3;
+        timeout = 1e-3;
+        break;
+
+      case ~0:
+	/* No events found: timer expired. */
+        if (vlib_time_now(vm) > link_debounce_deadline)
+          {
+            vec_foreach (xd, xm->devices)
+              {
+                ixge_regs_t * r = xd->regs;
+                u32 v = r->xge_mac.link_status;
+                uword is_up = (v & (1 << 30)) != 0;
+
+                vnet_hw_interface_set_flags
+                  (vnm, xd->vlib_hw_if_index,
+                   is_up ? VNET_HW_INTERFACE_FLAG_LINK_UP : 0);
+              }
+            link_debounce_deadline = 1e70;
+            timeout = 30.0;
+          }
+	break;
+
+      default:
+	ASSERT (0);
+      }
+
+      if (event_data)
+	_vec_len (event_data) = 0;
+
+      /* Query stats every 30 secs. */
+      {
+	f64 now = vlib_time_now (vm);
+	if (now - xm->time_last_stats_update > 30)
+	  {
+	    xm->time_last_stats_update = now;
+	    vec_foreach (xd, xm->devices)
+	      ixge_update_counters (xd);
+	  }
+      }
+    }
+
+  return 0;
+}
+
+static vlib_node_registration_t ixge_process_node = {
+    .function = ixge_process,
+    .type = VLIB_NODE_TYPE_PROCESS,
+    .name = "ixge-process",
+};
+
+clib_error_t * ixge_init (vlib_main_t * vm)
+{
+  ixge_main_t * xm = &ixge_main;
+  clib_error_t * error;
+
+  xm->vlib_main = vm;
+  memset (&xm->tx_descriptor_template, 0, sizeof (xm->tx_descriptor_template));
+  memset (&xm->tx_descriptor_template_mask, 0, sizeof (xm->tx_descriptor_template_mask));
+  xm->tx_descriptor_template.status0 =
+    (IXGE_TX_DESCRIPTOR_STATUS0_ADVANCED
+     | IXGE_TX_DESCRIPTOR_STATUS0_IS_ADVANCED
+     | IXGE_TX_DESCRIPTOR_STATUS0_INSERT_FCS);
+  xm->tx_descriptor_template_mask.status0 = 0xffff;
+  xm->tx_descriptor_template_mask.status1 = 0x00003fff;
+
+  xm->tx_descriptor_template_mask.status0 &=
+    ~(IXGE_TX_DESCRIPTOR_STATUS0_IS_END_OF_PACKET
+      | IXGE_TX_DESCRIPTOR_STATUS0_REPORT_STATUS);
+  xm->tx_descriptor_template_mask.status1 &=
+    ~(IXGE_TX_DESCRIPTOR_STATUS1_DONE);
+
+  error = vlib_call_init_function (vm, pci_bus_init);
+
+  return error;
+}
+
+VLIB_INIT_FUNCTION (ixge_init);
+
+
+static void
+ixge_pci_intr_handler(vlib_pci_device_t * dev)
+{
+  ixge_main_t * xm = &ixge_main;
+  vlib_main_t * vm = xm->vlib_main;
+
+  vlib_node_set_interrupt_pending (vm, ixge_input_node.index);
+
+  /* Let node know which device is interrupting. */
+  {
+    vlib_node_runtime_t * rt = vlib_node_get_runtime (vm, ixge_input_node.index);
+    rt->runtime_data[0] |= 1 << dev->private_data;
+  }
+}
+
+static clib_error_t *
+ixge_pci_init (vlib_main_t * vm, vlib_pci_device_t * dev)
+{
+  ixge_main_t * xm = &ixge_main;
+  clib_error_t * error;
+  void * r;
+  ixge_device_t * xd;
+
+  /* Device found: make sure we have dma memory. */
+  if (unix_physmem_is_fake (vm))
+    return clib_error_return (0, "no physical memory available");
+
+  error = vlib_pci_map_resource (dev, 0, &r);
+  if (error)
+    return error;
+
+  vec_add2 (xm->devices, xd, 1);
+
+  if (vec_len (xm->devices) == 1)
+    {
+      ixge_input_node.function = ixge_input_multiarch_select();
+      vlib_register_node (vm, &ixge_input_node);
+    }
+
+  xd->pci_device = dev[0];
+  xd->device_id = xd->pci_device.config0.header.device_id;
+  xd->regs = r;
+  xd->device_index = xd - xm->devices;
+  xd->pci_function = dev->bus_address.function;
+  xd->per_interface_next_index = ~0;
+
+
+  /* Chip found so enable node. */
+  {
+    vlib_node_set_state (vm, ixge_input_node.index,
+			 (IXGE_ALWAYS_POLL
+			  ? VLIB_NODE_STATE_POLLING
+			  : VLIB_NODE_STATE_INTERRUPT));
+
+    dev->private_data = xd->device_index;
+  }
+
+  if (vec_len (xm->devices) == 1)
+    {
+      vlib_register_node (vm, &ixge_process_node);
+      xm->process_node_index = ixge_process_node.index;
+    }
+
+  error = vlib_pci_bus_master_enable(dev);
+
+  if (error)
+    return error;
+
+  return vlib_pci_intr_enable(dev);
+}
+
+PCI_REGISTER_DEVICE (ixge_pci_device_registration,static) = {
+  .init_function = ixge_pci_init,
+  .interrupt_handler = ixge_pci_intr_handler,
+  .supported_devices = {
+#define _(t,i) { .vendor_id = PCI_VENDOR_ID_INTEL, .device_id = i, },
+    foreach_ixge_pci_device_id
+#undef _
+    { 0 },
+  },
+};
+
+void ixge_set_next_node (ixge_rx_next_t next, char *name)
+{
+  vlib_node_registration_t *r = &ixge_input_node;
+
+  switch (next)
+    {
+    case IXGE_RX_NEXT_IP4_INPUT:
+    case IXGE_RX_NEXT_IP6_INPUT:
+    case IXGE_RX_NEXT_ETHERNET_INPUT:
+      r->next_nodes[next] = name;
+      break;
+
+    default:
+      clib_warning ("%s: illegal next %d\n", __FUNCTION__, next);
+      break;
+    }
+}
diff --git a/vnet/vnet/devices/nic/ixge.h b/vnet/vnet/devices/nic/ixge.h
new file mode 100644
index 00000000000..8062b8c8a97
--- /dev/null
+++ b/vnet/vnet/devices/nic/ixge.h
@@ -0,0 +1,1241 @@
+/*
+ * Copyright (c) 2016 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef included_ixge_h
+#define included_ixge_h
+
+#include <vnet/vnet.h>
+#include <vlib/pci/pci.h>
+#include <vlib/i2c.h>
+#include <vnet/devices/nic/sfp.h>
+#include <vnet/ip/ip4_packet.h>
+#include <vnet/ip/ip6_packet.h>
+
+typedef volatile struct {
+  /* [31:7] 128 byte aligned. */
+  u32 descriptor_address[2];
+  u32 n_descriptor_bytes;
+
+  /* [5] rx/tx descriptor dca enable
+     [6] rx packet head dca enable
+     [7] rx packet tail dca enable
+     [9] rx/tx descriptor relaxed order
+     [11] rx/tx descriptor write back relaxed order
+     [13] rx/tx data write/read relaxed order
+     [15] rx head data write relaxed order
+     [31:24] apic id for cpu's cache. */
+  u32 dca_control;
+
+  u32 head_index;
+
+  /* [4:0] tail buffer size (in 1k byte units)
+     [13:8] head buffer size (in 64 byte units)
+     [24:22] lo free descriptors threshold (units of 64 descriptors)
+     [27:25] descriptor type 0 = legacy, 1 = advanced one buffer (e.g. tail),
+     2 = advanced header splitting (head + tail), 5 = advanced header
+     splitting (head only).
+     [28] drop if no descriptors available. */
+  u32 rx_split_control;
+
+  u32 tail_index;
+  CLIB_PAD_FROM_TO (0x1c, 0x28);
+
+  /* [7:0] rx/tx prefetch threshold
+     [15:8] rx/tx host threshold
+     [24:16] rx/tx write back threshold
+     [25] rx/tx enable
+     [26] tx descriptor writeback flush
+     [30] rx strip vlan enable */
+  u32 control;
+
+  u32 rx_coallesce_control;
+
+  union {
+    struct {
+      /* packets bytes lo hi */
+      u32 stats[3];
+
+      u32 unused;
+    } rx;
+
+    struct {
+      u32 unused[2];
+
+      /* [0] enables head write back. */
+      u32 head_index_write_back_address[2];
+    } tx;
+  };
+} ixge_dma_regs_t;
+
+/* Only advanced descriptors are supported. */
+typedef struct {
+  u64 tail_address;
+  u64 head_address;
+} ixge_rx_to_hw_descriptor_t;
+
+typedef struct {
+  u32 status[3];
+  u16 n_packet_bytes_this_descriptor;
+  u16 vlan_tag;
+} ixge_rx_from_hw_descriptor_t;
+
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_LAYER2 (1 << (4 + 11))
+/* Valid if not layer2. */
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4 (1 << (4 + 0))
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP4_EXT (1 << (4 + 1))
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6 (1 << (4 + 2))
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_IP6_EXT (1 << (4 + 3))
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_TCP (1 << (4 + 4))
+#define IXGE_RX_DESCRIPTOR_STATUS0_IS_UDP (1 << (4 + 5))
+#define IXGE_RX_DESCRIPTOR_STATUS0_L3_OFFSET(s) (((s) >> 21) & 0x3ff)
+
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_OWNED_BY_SOFTWARE (1 << (0 + 0))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_END_OF_PACKET (1 << (0 + 1))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_VLAN (1 << (0 + 3))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_UDP_CHECKSUMMED (1 << (0 + 4))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_TCP_CHECKSUMMED (1 << (0 + 5))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_IP4_CHECKSUMMED (1 << (0 + 6))
+#define IXGE_RX_DESCRIPTOR_STATUS2_NOT_UNICAST (1 << (0 + 7))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IS_DOUBLE_VLAN (1 << (0 + 9))
+#define IXGE_RX_DESCRIPTOR_STATUS2_UDP_CHECKSUM_ERROR (1 << (0 + 10))
+#define IXGE_RX_DESCRIPTOR_STATUS2_ETHERNET_ERROR (1 << (20 + 9))
+#define IXGE_RX_DESCRIPTOR_STATUS2_TCP_CHECKSUM_ERROR (1 << (20 + 10))
+#define IXGE_RX_DESCRIPTOR_STATUS2_IP4_CHECKSUM_ERROR (1 << (20 + 11))
+
+/* For layer2 packets stats0 bottom 3 bits give ether type index from filter. */
+#define IXGE_RX_DESCRIPTOR_STATUS0_LAYER2_ETHERNET_TYPE(s) ((s) & 7)
+
+typedef struct {
+  u64 buffer_address;
+  u16 n_bytes_this_buffer;
+  u16 status0;
+  u32 status1;
+#define IXGE_TX_DESCRIPTOR_STATUS0_ADVANCED (3 << 4)
+#define IXGE_TX_DESCRIPTOR_STATUS0_IS_ADVANCED (1 << (8 + 5))
+#define IXGE_TX_DESCRIPTOR_STATUS0_LOG2_REPORT_STATUS (8 + 3)
+#define IXGE_TX_DESCRIPTOR_STATUS0_REPORT_STATUS (1 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_REPORT_STATUS)
+#define IXGE_TX_DESCRIPTOR_STATUS0_INSERT_FCS (1 << (8 + 1))
+#define IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET (8 + 0)
+#define IXGE_TX_DESCRIPTOR_STATUS0_IS_END_OF_PACKET (1 << IXGE_TX_DESCRIPTOR_STATUS0_LOG2_IS_END_OF_PACKET)
+#define IXGE_TX_DESCRIPTOR_STATUS1_DONE (1 << 0)
+#define IXGE_TX_DESCRIPTOR_STATUS1_CONTEXT(i) (/* valid */ (1 << 7) | ((i) << 4))
+#define IXGE_TX_DESCRIPTOR_STATUS1_IPSEC_OFFLOAD (1 << (8 + 2))
+#define IXGE_TX_DESCRIPTOR_STATUS1_INSERT_TCP_UDP_CHECKSUM (1 << (8 + 1))
+#define IXGE_TX_DESCRIPTOR_STATUS1_INSERT_IP4_CHECKSUM (1 << (8 + 0))
+#define IXGE_TX_DESCRIPTOR_STATUS0_N_BYTES_THIS_BUFFER(l) ((l) << 0)
+#define IXGE_TX_DESCRIPTOR_STATUS1_N_BYTES_IN_PACKET(l) ((l) << 14)
+} ixge_tx_descriptor_t;
+
+typedef struct {
+  struct {
+    u8 checksum_start_offset;
+    u8 checksum_insert_offset;
+    u16 checksum_end_offset;
+  } ip, tcp;
+  u32 status0;
+
+  u8 status1;
+
+  /* Byte offset after UDP/TCP header. */
+  u8 payload_offset;
+
+  u16 max_tcp_segment_size;
+} __attribute__ ((packed)) ixge_tx_context_descriptor_t;
+
+typedef union {
+  ixge_rx_to_hw_descriptor_t rx_to_hw;
+  ixge_rx_from_hw_descriptor_t rx_from_hw;
+  ixge_tx_descriptor_t tx;
+  u32x4 as_u32x4;
+} ixge_descriptor_t;
+
+typedef volatile struct {
+  /* [2] pcie master disable
+     [3] mac reset
+     [26] global device reset */
+  u32 control;
+  u32 control_alias;
+  /* [3:2] device id (0 or 1 for dual port chips)
+     [7] link is up
+     [17:10] num vfs
+     [18] io active
+     [19] pcie master enable status */
+  u32 status_read_only;
+  CLIB_PAD_FROM_TO (0xc, 0x18);
+  /* [14] pf reset done
+     [17] relaxed ordering disable
+     [26] extended vlan enable
+     [28] driver loaded */
+  u32 extended_control;
+  CLIB_PAD_FROM_TO (0x1c, 0x20);
+
+  /* software definable pins.
+     sdp_data [7:0]
+     sdp_is_output [15:8]
+     sdp_is_native [23:16]
+     sdp_function [31:24].
+  */
+  u32 sdp_control;
+  CLIB_PAD_FROM_TO (0x24, 0x28);
+
+  /* [0] i2c clock in
+     [1] i2c clock out
+     [2] i2c data in
+     [3] i2c data out */
+  u32 i2c_control;
+  CLIB_PAD_FROM_TO (0x2c, 0x4c);
+  u32 tcp_timer;
+
+  CLIB_PAD_FROM_TO (0x50, 0x200);
+
+  u32 led_control;
+
+  CLIB_PAD_FROM_TO (0x204, 0x600);
+  u32 core_spare;
+  CLIB_PAD_FROM_TO (0x604, 0x700);
+
+  struct {
+    u32 vflr_events_clear[4];
+    u32 mailbox_interrupt_status[4];
+    u32 mailbox_interrupt_enable[4];
+    CLIB_PAD_FROM_TO (0x730, 0x800);
+  } pf_foo;
+
+  struct {
+    u32 status_write_1_to_clear;
+    CLIB_PAD_FROM_TO (0x804, 0x808);
+    u32 status_write_1_to_set;
+    CLIB_PAD_FROM_TO (0x80c, 0x810);
+    u32 status_auto_clear_enable;
+    CLIB_PAD_FROM_TO (0x814, 0x820);
+
+    /* [11:3] minimum inter-interrupt interval
+         (2e-6 units; 20e-6 units for fast ethernet).
+       [15] low-latency interrupt moderation enable
+       [20:16] low-latency interrupt credit
+       [27:21] interval counter
+       [31] write disable for credit and counter (write only). */
+    u32 throttle0[24];
+
+    u32 enable_write_1_to_set;
+    CLIB_PAD_FROM_TO (0x884, 0x888);
+    u32 enable_write_1_to_clear;
+    CLIB_PAD_FROM_TO (0x88c, 0x890);
+    u32 enable_auto_clear;
+    u32 msi_to_eitr_select;
+    /* [3:0] spd 0-3 interrupt detection enable
+       [4] msi-x enable
+       [5] other clear disable (makes other bits in status not clear on read)
+       etc. */
+    u32 control;
+    CLIB_PAD_FROM_TO (0x89c, 0x900);
+
+    /* Defines interrupt mapping for 128 rx + 128 tx queues.
+       64 x 4 8 bit entries.
+       For register [i]:
+         [5:0] bit in interrupt status for rx queue 2*i + 0
+	 [7] valid bit
+	 [13:8] bit for tx queue 2*i + 0
+	 [15] valid bit
+	 similar for rx 2*i + 1 and tx 2*i + 1. */
+    u32 queue_mapping[64];
+
+    /* tcp timer [7:0] and other interrupts [15:8] */
+    u32 misc_mapping;
+    CLIB_PAD_FROM_TO (0xa04, 0xa90);
+
+    /* 64 interrupts determined by mappings. */
+    u32 status1_write_1_to_clear[4];
+    u32 enable1_write_1_to_set[4];
+    u32 enable1_write_1_to_clear[4];
+    CLIB_PAD_FROM_TO (0xac0, 0xad0);
+    u32 status1_enable_auto_clear[4];
+    CLIB_PAD_FROM_TO (0xae0, 0x1000);
+  } interrupt;
+
+  ixge_dma_regs_t rx_dma0[64];
+
+  CLIB_PAD_FROM_TO (0x2000, 0x2140);
+  u32 dcb_rx_packet_plane_t4_config[8];
+  u32 dcb_rx_packet_plane_t4_status[8];
+  CLIB_PAD_FROM_TO (0x2180, 0x2300);
+
+  /* reg i defines mapping for 4 rx queues starting at 4*i + 0. */
+  u32 rx_queue_stats_mapping[32];
+  u32 rx_queue_stats_control;
+
+  CLIB_PAD_FROM_TO (0x2384, 0x2410);
+  u32 fc_user_descriptor_ptr[2];
+  u32 fc_buffer_control;
+  CLIB_PAD_FROM_TO (0x241c, 0x2420);
+  u32 fc_rx_dma;
+  CLIB_PAD_FROM_TO (0x2424, 0x2430);
+  u32 dcb_packet_plane_control;
+  CLIB_PAD_FROM_TO (0x2434, 0x2f00);
+
+  u32 rx_dma_control;
+  u32 pf_queue_drop_enable;
+  CLIB_PAD_FROM_TO (0x2f08, 0x2f20);
+  u32 rx_dma_descriptor_cache_config;
+  CLIB_PAD_FROM_TO (0x2f24, 0x3000);
+
+  /* 1 bit. */
+  u32 rx_enable;
+  CLIB_PAD_FROM_TO (0x3004, 0x3008);
+  /* [15:0] ether type (little endian)
+     [31:16] opcode (big endian) */
+  u32 flow_control_control;
+  CLIB_PAD_FROM_TO (0x300c, 0x3020);
+  /* 3 bit traffic class for each of 8 priorities. */
+  u32 rx_priority_to_traffic_class;
+  CLIB_PAD_FROM_TO (0x3024, 0x3028);
+  u32 rx_coallesce_data_buffer_control;
+  CLIB_PAD_FROM_TO (0x302c, 0x3190);
+  u32 rx_packet_buffer_flush_detect;
+  CLIB_PAD_FROM_TO (0x3194, 0x3200);
+  u32 flow_control_tx_timers[4];		/* 2 timer values */
+  CLIB_PAD_FROM_TO (0x3210, 0x3220);
+  u32 flow_control_rx_threshold_lo[8];
+  CLIB_PAD_FROM_TO (0x3240, 0x3260);
+  u32 flow_control_rx_threshold_hi[8];
+  CLIB_PAD_FROM_TO (0x3280, 0x32a0);
+  u32 flow_control_refresh_threshold;
+  CLIB_PAD_FROM_TO (0x32a4, 0x3c00);
+  /* For each of 8 traffic classes (units of bytes). */
+  u32 rx_packet_buffer_size[8];
+  CLIB_PAD_FROM_TO (0x3c20, 0x3d00);
+  u32 flow_control_config;
+  CLIB_PAD_FROM_TO (0x3d04, 0x4200);
+
+  struct {
+    u32 pcs_config;
+    CLIB_PAD_FROM_TO (0x4204, 0x4208);
+    u32 link_control;
+    u32 link_status;
+    u32 pcs_debug[2];
+    u32 auto_negotiation;
+    u32 link_partner_ability;
+    u32 auto_negotiation_tx_next_page;
+    u32 auto_negotiation_link_partner_next_page;
+    CLIB_PAD_FROM_TO (0x4228, 0x4240);
+  } gige_mac;
+
+  struct {
+    /* [0] tx crc enable
+       [2] enable frames up to max frame size register [31:16]
+       [10] pad frames < 64 bytes if specified by user
+       [15] loopback enable
+       [16] mdc hi speed
+       [17] turn off mdc between mdio packets */
+    u32 control;
+
+    /* [5] rx symbol error (all bits clear on read)
+       [6] rx illegal symbol
+       [7] rx idle error
+       [8] rx local fault
+       [9] rx remote fault */
+    u32 status;
+
+    u32 pause_and_pace_control;
+    CLIB_PAD_FROM_TO (0x424c, 0x425c);
+    u32 phy_command;
+    u32 phy_data;
+    CLIB_PAD_FROM_TO (0x4264, 0x4268);
+
+    /* [31:16] max frame size in bytes. */
+    u32 rx_max_frame_size;
+    CLIB_PAD_FROM_TO (0x426c, 0x4288);
+
+    /* [0]
+         [2] pcs receive link up? (latch lo)
+	 [7] local fault
+       [1]
+         [0] pcs 10g base r capable
+         [1] pcs 10g base x capable
+         [2] pcs 10g base w capable
+	 [10] rx local fault
+	 [11] tx local fault
+	 [15:14] 2 => device present at this address (else not present) */
+    u32 xgxs_status[2];
+
+    u32 base_x_pcs_status;
+
+    /* [0] pass unrecognized flow control frames
+       [1] discard pause frames
+       [2] rx priority flow control enable (only in dcb mode)
+       [3] rx flow control enable. */
+    u32 flow_control;
+
+    /* [3:0] tx lanes change polarity
+       [7:4] rx lanes change polarity
+       [11:8] swizzle tx lanes
+       [15:12] swizzle rx lanes
+       4 x 2 bit tx lane swap
+       4 x 2 bit rx lane swap. */
+    u32 serdes_control;
+
+    u32 fifo_control;
+
+    /* [0] force link up
+       [1] autoneg ack2 bit to transmit
+       [6:2] autoneg selector field to transmit
+       [8:7] 10g pma/pmd type 0 => xaui, 1 kx4, 2 cx4
+       [9] 1g pma/pmd type 0 => sfi, 1 => kx/bx
+       [10] disable 10g on without main power
+       [11] restart autoneg on transition to dx power state
+       [12] restart autoneg
+       [15:13] link mode:
+         0 => 1g no autoneg
+	 1 => 10g kx4 parallel link no autoneg
+	 2 => 1g bx autoneg
+	 3 => 10g sfi serdes
+	 4 => kx4/kx/kr
+	 5 => xgmii 1g/100m
+	 6 => kx4/kx/kr 1g an
+	 7 kx4/kx/kr sgmii.
+       [16] kr support
+       [17] fec requested
+       [18] fec ability
+       etc. */
+    u32 auto_negotiation_control;
+
+    /* [0] signal detect 1g/100m
+       [1] fec signal detect
+       [2] 10g serial pcs fec block lock
+       [3] 10g serial high error rate
+       [4] 10g serial pcs block lock
+       [5] kx/kx4/kr autoneg next page received
+       [6] kx/kx4/kr backplane autoneg next page received
+       [7] link status clear to read
+       [11:8] 10g signal detect (4 lanes) (for serial just lane 0)
+       [12] 10g serial signal detect
+       [16:13] 10g parallel lane sync status
+       [17] 10g parallel align status
+       [18] 1g sync status
+       [19] kx/kx4/kr backplane autoneg is idle
+       [20] 1g autoneg enabled
+       [21] 1g pcs enabled for sgmii
+       [22] 10g xgxs enabled
+       [23] 10g serial fec enabled (forward error detection)
+       [24] 10g kr pcs enabled
+       [25] sgmii enabled
+       [27:26] mac link mode
+         0 => 1g
+	 1 => 10g parallel
+	 2 => 10g serial
+	 3 => autoneg
+       [29:28] link speed
+         1 => 100m
+         2 => 1g
+         3 => 10g
+       [30] link is up
+       [31] kx/kx4/kr backplane autoneg completed successfully. */
+    u32 link_status;
+
+    /* [17:16] pma/pmd for 10g serial
+         0 => kr, 2 => sfi
+       [18] disable dme pages */
+    u32 auto_negotiation_control2;
+
+    CLIB_PAD_FROM_TO (0x42ac, 0x42b0);
+    u32 link_partner_ability[2];
+    CLIB_PAD_FROM_TO (0x42b8, 0x42d0);
+    u32 manageability_control;
+    u32 link_partner_next_page[2];
+    CLIB_PAD_FROM_TO (0x42dc, 0x42e0);
+    u32 kr_pcs_control;
+    u32 kr_pcs_status;
+    u32 fec_status[2];
+    CLIB_PAD_FROM_TO (0x42f0, 0x4314);
+    u32 sgmii_control;
+    CLIB_PAD_FROM_TO (0x4318, 0x4324);
+    u32 link_status2;
+    CLIB_PAD_FROM_TO (0x4328, 0x4900);
+  } xge_mac;
+
+  u32 tx_dcb_control;
+  u32 tx_dcb_descriptor_plane_queue_select;
+  u32 tx_dcb_descriptor_plane_t1_config;
+  u32 tx_dcb_descriptor_plane_t1_status;
+  CLIB_PAD_FROM_TO (0x4910, 0x4950);
+
+  /* For each TC in units of 1k bytes. */
+  u32 tx_packet_buffer_thresholds[8];
+  CLIB_PAD_FROM_TO (0x4970, 0x4980);
+  struct {
+    u32 mmw;
+    u32 config;
+    u32 status;
+    u32 rate_drift;
+  } dcb_tx_rate_scheduler;
+  CLIB_PAD_FROM_TO (0x4990, 0x4a80);
+  u32 tx_dma_control;
+  CLIB_PAD_FROM_TO (0x4a84, 0x4a88);
+  u32 tx_dma_tcp_flags_control[2];
+  CLIB_PAD_FROM_TO (0x4a90, 0x4b00);
+  u32 pf_mailbox[64];
+  CLIB_PAD_FROM_TO (0x4c00, 0x5000);
+
+  /* RX */
+  u32 checksum_control;
+  CLIB_PAD_FROM_TO (0x5004, 0x5008);
+  u32 rx_filter_control;
+  CLIB_PAD_FROM_TO (0x500c, 0x5010);
+  u32 management_vlan_tag[8];
+  u32 management_udp_tcp_ports[8];
+  CLIB_PAD_FROM_TO (0x5050, 0x5078);
+  /* little endian. */
+  u32 extended_vlan_ether_type;
+  CLIB_PAD_FROM_TO (0x507c, 0x5080);
+  /* [1] store/dma bad packets
+     [8] accept all multicast
+     [9] accept all unicast
+     [10] accept all broadcast. */
+  u32 filter_control;
+  CLIB_PAD_FROM_TO (0x5084, 0x5088);
+  /* [15:0] vlan ethernet type (0x8100) little endian
+     [28] cfi bit expected
+     [29] drop packets with unexpected cfi bit
+     [30] vlan filter enable. */
+  u32 vlan_control;
+  CLIB_PAD_FROM_TO (0x508c, 0x5090);
+  /* [1:0] hi bit of ethernet address for 12 bit index into multicast table
+       0 => 47, 1 => 46, 2 => 45, 3 => 43.
+     [2] enable multicast filter
+   */
+  u32 multicast_control;
+  CLIB_PAD_FROM_TO (0x5094, 0x5100);
+  u32 fcoe_rx_control;
+  CLIB_PAD_FROM_TO (0x5104, 0x5108);
+  u32 fc_flt_context;
+  CLIB_PAD_FROM_TO (0x510c, 0x5110);
+  u32 fc_filter_control;
+  CLIB_PAD_FROM_TO (0x5114, 0x5120);
+  u32 rx_message_type_lo;
+  CLIB_PAD_FROM_TO (0x5124, 0x5128);
+  /* [15:0] ethernet type (little endian)
+     [18:16] matche pri in vlan tag
+     [19] priority match enable
+     [25:20] virtualization pool
+     [26] pool enable
+     [27] is fcoe
+     [30] ieee 1588 timestamp enable
+     [31] filter enable.
+     (See ethernet_type_queue_select.) */
+  u32 ethernet_type_queue_filter[8];
+  CLIB_PAD_FROM_TO (0x5148, 0x5160);
+  /* [7:0] l2 ethernet type and
+     [15:8] l2 ethernet type or */
+  u32 management_decision_filters1[8];
+  u32 vf_vm_tx_switch_loopback_enable[2];
+  u32 rx_time_sync_control;
+  CLIB_PAD_FROM_TO (0x518c, 0x5190);
+  u32 management_ethernet_type_filters[4];
+  u32 rx_timestamp_attributes_lo;
+  u32 rx_timestamp_hi;
+  u32 rx_timestamp_attributes_hi;
+  CLIB_PAD_FROM_TO (0x51ac, 0x51b0);
+  u32 pf_virtual_control;
+  CLIB_PAD_FROM_TO (0x51b4, 0x51d8);
+  u32 fc_offset_parameter;
+  CLIB_PAD_FROM_TO (0x51dc, 0x51e0);
+  u32 vf_rx_enable[2];
+  u32 rx_timestamp_lo;
+  CLIB_PAD_FROM_TO (0x51ec, 0x5200);
+  /* 12 bits determined by multicast_control
+     lookup bits in this vector. */
+  u32 multicast_enable[128];
+
+  /* [0] ethernet address [31:0]
+     [1] [15:0] ethernet address [47:32]
+         [31] valid bit.
+     Index 0 is read from eeprom after reset. */
+  u32 rx_ethernet_address0[16][2];
+
+  CLIB_PAD_FROM_TO (0x5480, 0x5800);
+  u32 wake_up_control;
+  CLIB_PAD_FROM_TO (0x5804, 0x5808);
+  u32 wake_up_filter_control;
+  CLIB_PAD_FROM_TO (0x580c, 0x5818);
+  u32 multiple_rx_queue_command_82598;
+  CLIB_PAD_FROM_TO (0x581c, 0x5820);
+  u32 management_control;
+  u32 management_filter_control;
+  CLIB_PAD_FROM_TO (0x5828, 0x5838);
+  u32 wake_up_ip4_address_valid;
+  CLIB_PAD_FROM_TO (0x583c, 0x5840);
+  u32 wake_up_ip4_address_table[4];
+  u32 management_control_to_host;
+  CLIB_PAD_FROM_TO (0x5854, 0x5880);
+  u32 wake_up_ip6_address_table[4];
+
+  /* unicast_and broadcast_and vlan_and ip_address_and
+     etc. */
+  u32 management_decision_filters[8];
+
+  u32 management_ip4_or_ip6_address_filters[4][4];
+  CLIB_PAD_FROM_TO (0x58f0, 0x5900);
+  u32 wake_up_packet_length;
+  CLIB_PAD_FROM_TO (0x5904, 0x5910);
+  u32 management_ethernet_address_filters[4][2];
+  CLIB_PAD_FROM_TO (0x5930, 0x5a00);
+  u32 wake_up_packet_memory[32];
+  CLIB_PAD_FROM_TO (0x5a80, 0x5c00);
+  u32 redirection_table_82598[32];
+  u32 rss_random_keys_82598[10];
+  CLIB_PAD_FROM_TO (0x5ca8, 0x6000);
+
+  ixge_dma_regs_t tx_dma[128];
+
+  u32 pf_vm_vlan_insert[64];
+  u32 tx_dma_tcp_max_alloc_size_requests;
+  CLIB_PAD_FROM_TO (0x8104, 0x8110);
+  u32 vf_tx_enable[2];
+  CLIB_PAD_FROM_TO (0x8118, 0x8120);
+  /* [0] dcb mode enable
+     [1] virtualization mode enable
+     [3:2] number of tcs/qs per pool. */
+  u32 multiple_tx_queues_command;
+  CLIB_PAD_FROM_TO (0x8124, 0x8200);
+  u32 pf_vf_anti_spoof[8];
+  u32 pf_dma_tx_switch_control;
+  CLIB_PAD_FROM_TO (0x8224, 0x82e0);
+  u32 tx_strict_low_latency_queues[4];
+  CLIB_PAD_FROM_TO (0x82f0, 0x8600);
+  u32 tx_queue_stats_mapping_82599[32];
+  u32 tx_queue_packet_counts[32];
+  u32 tx_queue_byte_counts[32][2];
+
+  struct {
+    u32 control;
+    u32 status;
+    u32 buffer_almost_full;
+    CLIB_PAD_FROM_TO (0x880c, 0x8810);
+    u32 buffer_min_ifg;
+    CLIB_PAD_FROM_TO (0x8814, 0x8900);
+  } tx_security;
+
+  struct {
+    u32 index;
+    u32 salt;
+    u32 key[4];
+    CLIB_PAD_FROM_TO (0x8918, 0x8a00);
+  } tx_ipsec;
+
+  struct {
+    u32 capabilities;
+    u32 control;
+    u32 tx_sci[2];
+    u32 sa;
+    u32 sa_pn[2];
+    u32 key[2][4];
+    /* untagged packets, encrypted packets, protected packets,
+       encrypted bytes, protected bytes */
+    u32 stats[5];
+    CLIB_PAD_FROM_TO (0x8a50, 0x8c00);
+  } tx_link_security;
+
+  struct {
+    u32 control;
+    u32 timestamp_value[2];
+    u32 system_time[2];
+    u32 increment_attributes;
+    u32 time_adjustment_offset[2];
+    u32 aux_control;
+    u32 target_time[2][2];
+    CLIB_PAD_FROM_TO (0x8c34, 0x8c3c);
+    u32 aux_time_stamp[2][2];
+    CLIB_PAD_FROM_TO (0x8c4c, 0x8d00);
+  } tx_timesync;
+
+  struct {
+    u32 control;
+    u32 status;
+    CLIB_PAD_FROM_TO (0x8d08, 0x8e00);
+  } rx_security;
+
+  struct {
+    u32 index;
+    u32 ip_address[4];
+    u32 spi;
+    u32 ip_index;
+    u32 key[4];
+    u32 salt;
+    u32 mode;
+    CLIB_PAD_FROM_TO (0x8e34, 0x8f00);
+  } rx_ipsec;
+
+  struct {
+    u32 capabilities;
+    u32 control;
+    u32 sci[2];
+    u32 sa[2];
+    u32 sa_pn[2];
+    u32 key[2][4];
+    /* see datasheet */
+    u32 stats[17];
+    CLIB_PAD_FROM_TO (0x8f84, 0x9000);
+  } rx_link_security;
+
+  /* 4 wake up, 2 management, 2 wake up. */
+  u32 flexible_filters[8][16][4];
+  CLIB_PAD_FROM_TO (0x9800, 0xa000);
+
+  /* 4096 bits. */
+  u32 vlan_filter[128];
+
+  /* [0] ethernet address [31:0]
+     [1] [15:0] ethernet address [47:32]
+     [31] valid bit.
+     Index 0 is read from eeprom after reset. */
+  u32 rx_ethernet_address1[128][2];
+
+  /* select one of 64 pools for each rx address. */
+  u32 rx_ethernet_address_pool_select[128][2];
+  CLIB_PAD_FROM_TO (0xaa00, 0xc800);
+  u32 tx_priority_to_traffic_class;
+  CLIB_PAD_FROM_TO (0xc804, 0xcc00);
+
+  /* In bytes units of 1k.  Total packet buffer is 160k. */
+  u32 tx_packet_buffer_size[8];
+
+  CLIB_PAD_FROM_TO (0xcc20, 0xcd10);
+  u32 tx_manageability_tc_mapping;
+  CLIB_PAD_FROM_TO (0xcd14, 0xcd20);
+  u32 dcb_tx_packet_plane_t2_config[8];
+  u32 dcb_tx_packet_plane_t2_status[8];
+  CLIB_PAD_FROM_TO (0xcd60, 0xce00);
+
+  u32 tx_flow_control_status;
+  CLIB_PAD_FROM_TO (0xce04, 0xd000);
+
+  ixge_dma_regs_t rx_dma1[64];
+
+  struct {
+    /* Bigendian ip4 src/dst address. */
+    u32 src_address[128];
+    u32 dst_address[128];
+
+    /* TCP/UDP ports [15:0] src [31:16] dst; bigendian. */
+    u32 tcp_udp_port[128];
+
+    /* [1:0] protocol tcp, udp, sctp, other
+       [4:2] match priority (highest wins)
+       [13:8] pool
+       [25] src address match disable
+       [26] dst address match disable
+       [27] src port match disable
+       [28] dst port match disable
+       [29] protocol match disable
+       [30] pool match disable
+       [31] enable. */
+    u32 control[128];
+
+    /* [12] size bypass
+       [19:13] must be 0x80
+       [20] low-latency interrupt
+       [27:21] rx queue. */
+    u32 interrupt[128];
+  } ip4_filters;
+
+  CLIB_PAD_FROM_TO (0xea00, 0xeb00);
+  /* 4 bit rss output index indexed by 7 bit hash.
+     128 8 bit fields = 32 registers. */
+  u32 redirection_table_82599[32];
+
+  u32 rss_random_key_82599[10];
+  CLIB_PAD_FROM_TO (0xeba8, 0xec00);
+  /* [15:0] reserved
+     [22:16] rx queue index
+     [29] low-latency interrupt on match
+     [31] enable */
+  u32 ethernet_type_queue_select[8];
+  CLIB_PAD_FROM_TO (0xec20, 0xec30);
+  u32 syn_packet_queue_filter;
+  CLIB_PAD_FROM_TO (0xec34, 0xec60);
+  u32 immediate_interrupt_rx_vlan_priority;
+  CLIB_PAD_FROM_TO (0xec64, 0xec70);
+  u32 rss_queues_per_traffic_class;
+  CLIB_PAD_FROM_TO (0xec74, 0xec90);
+  u32 lli_size_threshold;
+  CLIB_PAD_FROM_TO (0xec94, 0xed00);
+
+  struct {
+    u32 control;
+    CLIB_PAD_FROM_TO (0xed04, 0xed10);
+    u32 table[8];
+    CLIB_PAD_FROM_TO (0xed30, 0xee00);
+  } fcoe_redirection;
+
+  struct {
+    /* [1:0] packet buffer allocation 0 => disabled, else 64k*2^(f-1)
+       [3] packet buffer initialization done
+       [4] perfetch match mode
+       [5] report status in rss field of rx descriptors
+       [7] report status always
+       [14:8] drop queue
+       [20:16] flex 2 byte packet offset (units of 2 bytes)
+       [27:24] max linked list length
+       [31:28] full threshold. */
+    u32 control;
+    CLIB_PAD_FROM_TO (0xee04, 0xee0c);
+
+    u32 data[8];
+
+    /* [1:0] 0 => no action, 1 => add, 2 => remove, 3 => query.
+       [2] valid filter found by query command
+       [3] filter update override
+       [4] ip6 adress table
+       [6:5] l4 protocol reserved, udp, tcp, sctp
+       [7] is ip6
+       [8] clear head/tail
+       [9] packet drop action
+       [10] matched packet generates low-latency interrupt
+       [11] last in linked list
+       [12] collision
+       [15] rx queue enable
+       [22:16] rx queue
+       [29:24] pool. */
+    u32 command;
+
+    CLIB_PAD_FROM_TO (0xee30, 0xee3c);
+    /* ip4 dst/src address, tcp ports, udp ports.
+       set bits mean bit is ignored. */
+    u32 ip4_masks[4];
+    u32 filter_length;
+    u32 usage_stats;
+    u32 failed_usage_stats;
+    u32 filters_match_stats;
+    u32 filters_miss_stats;
+    CLIB_PAD_FROM_TO (0xee60, 0xee68);
+    /* Lookup, signature. */
+    u32 hash_keys[2];
+    /* [15:0] ip6 src address 1 bit per byte
+       [31:16] ip6 dst address. */
+    u32 ip6_mask;
+    /* [0] vlan id
+       [1] vlan priority
+       [2] pool
+       [3] ip protocol
+       [4] flex
+       [5] dst ip6. */
+    u32 other_mask;
+    CLIB_PAD_FROM_TO (0xee78, 0xf000);
+  } flow_director;
+
+  struct {
+    u32 l2_control[64];
+    u32 vlan_pool_filter[64];
+    u32 vlan_pool_filter_bitmap[128];
+    u32 dst_ethernet_address[128];
+    u32 mirror_rule[4];
+    u32 mirror_rule_vlan[8];
+    u32 mirror_rule_pool[8];
+    CLIB_PAD_FROM_TO (0xf650, 0x10010);
+  } pf_bar;
+
+  u32 eeprom_flash_control;
+  /* [0] start
+     [1] done
+     [15:2] address
+     [31:16] read data. */
+  u32 eeprom_read;
+  CLIB_PAD_FROM_TO (0x10018, 0x1001c);
+  u32 flash_access;
+  CLIB_PAD_FROM_TO (0x10020, 0x10114);
+  u32 flash_data;
+  u32 flash_control;
+  u32 flash_read_data;
+  CLIB_PAD_FROM_TO (0x10120, 0x1013c);
+  u32 flash_opcode;
+  u32 software_semaphore;
+  CLIB_PAD_FROM_TO (0x10144, 0x10148);
+  u32 firmware_semaphore;
+  CLIB_PAD_FROM_TO (0x1014c, 0x10160);
+  u32 software_firmware_sync;
+  CLIB_PAD_FROM_TO (0x10164, 0x10200);
+  u32 general_rx_control;
+  CLIB_PAD_FROM_TO (0x10204, 0x11000);
+
+  struct {
+    u32 control;
+    CLIB_PAD_FROM_TO (0x11004, 0x11010);
+    /* [3:0] enable counters
+       [7:4] leaky bucket counter mode
+       [29] reset
+       [30] stop
+       [31] start. */
+    u32 counter_control;
+    /* [7:0],[15:8],[23:16],[31:24] event for counters 0-3.
+       event codes:
+       0x0 bad tlp
+       0x10 reqs that reached timeout
+       etc. */
+    u32 counter_event;
+    CLIB_PAD_FROM_TO (0x11018, 0x11020);
+    u32 counters_clear_on_read[4];
+    u32 counter_config[4];
+    struct {
+      u32 address;
+      u32 data;
+    } indirect_access;
+    CLIB_PAD_FROM_TO (0x11048, 0x11050);
+    u32 extended_control;
+    CLIB_PAD_FROM_TO (0x11054, 0x11064);
+    u32 mirrored_revision_id;
+    CLIB_PAD_FROM_TO (0x11068, 0x11070);
+    u32 dca_requester_id_information;
+
+    /* [0] global disable
+       [4:1] mode: 0 => legacy, 1 => dca 1.0. */
+    u32 dca_control;
+    CLIB_PAD_FROM_TO (0x11078, 0x110b0);
+    /* [0] pci completion abort
+       [1] unsupported i/o address
+       [2] wrong byte enable
+       [3] pci timeout */
+    u32 pcie_interrupt_status;
+    CLIB_PAD_FROM_TO (0x110b4, 0x110b8);
+    u32 pcie_interrupt_enable;
+    CLIB_PAD_FROM_TO (0x110bc, 0x110c0);
+    u32 msi_x_pba_clear[8];
+    CLIB_PAD_FROM_TO (0x110e0, 0x12300);
+  } pcie;
+
+  u32 interrupt_throttle1[128-24];
+  CLIB_PAD_FROM_TO (0x124a0, 0x14f00);
+
+  u32 core_analog_config;
+  CLIB_PAD_FROM_TO (0x14f04, 0x14f10);
+  u32 core_common_config;
+  CLIB_PAD_FROM_TO (0x14f14, 0x15f14);
+
+  u32 link_sec_software_firmware_interface;
+} ixge_regs_t;
+
+typedef union {
+  struct {
+    /* Addresses bigendian. */
+    union {
+      struct {
+	ip6_address_t src_address;
+	u32 unused[1];
+      } ip6;
+      struct {
+	u32 unused[3];
+	ip4_address_t src_address, dst_address;
+      } ip4;
+    };
+
+    /* [15:0] src port (little endian).
+       [31:16] dst port. */
+    u32 tcp_udp_ports;
+
+    /* [15:0] vlan (cfi bit set to 0).
+       [31:16] flex bytes.  bigendian. */
+    u32 vlan_and_flex_word;
+
+    /* [14:0] hash
+       [15] bucket valid
+       [31:16] signature (signature filers)/sw-index (perfect match). */
+    u32 hash;
+  };
+
+  u32 as_u32[8];
+} ixge_flow_director_key_t;
+
+always_inline void
+ixge_throttle_queue_interrupt (ixge_regs_t * r,
+			       u32 queue_interrupt_index,
+			       f64 inter_interrupt_interval_in_secs)
+{
+  volatile u32 * tr =
+    (queue_interrupt_index < ARRAY_LEN (r->interrupt.throttle0)
+     ? &r->interrupt.throttle0[queue_interrupt_index]
+     : &r->interrupt_throttle1[queue_interrupt_index]);
+  ASSERT (queue_interrupt_index < 128);
+  u32 v;
+  i32 i, mask = (1 << 9) - 1;
+
+  i = flt_round_nearest (inter_interrupt_interval_in_secs / 2e-6);
+  i = i < 1 ? 1 : i;
+  i = i >= mask ? mask : i;
+
+  v = tr[0];
+  v &= ~(mask << 3);
+  v |= i << 3;
+  tr[0] = v;
+}
+
+#define foreach_ixge_counter				\
+  _ (0x40d0, rx_total_packets)				\
+  _64 (0x40c0, rx_total_bytes)				\
+  _ (0x41b0, rx_good_packets_before_filtering)		\
+  _64 (0x41b4, rx_good_bytes_before_filtering)		\
+  _ (0x2f50, rx_dma_good_packets)			\
+  _64 (0x2f54, rx_dma_good_bytes)			\
+  _ (0x2f5c, rx_dma_duplicated_good_packets)		\
+  _64 (0x2f60, rx_dma_duplicated_good_bytes)		\
+  _ (0x2f68, rx_dma_good_loopback_packets)		\
+  _64 (0x2f6c, rx_dma_good_loopback_bytes)		\
+  _ (0x2f74, rx_dma_good_duplicated_loopback_packets)	\
+  _64 (0x2f78, rx_dma_good_duplicated_loopback_bytes)	\
+  _ (0x4074, rx_good_packets)				\
+  _64 (0x4088, rx_good_bytes)				\
+  _ (0x407c, rx_multicast_packets)			\
+  _ (0x4078, rx_broadcast_packets)			\
+  _ (0x405c, rx_64_byte_packets)			\
+  _ (0x4060, rx_65_127_byte_packets)			\
+  _ (0x4064, rx_128_255_byte_packets)			\
+  _ (0x4068, rx_256_511_byte_packets)			\
+  _ (0x406c, rx_512_1023_byte_packets)			\
+  _ (0x4070, rx_gt_1023_byte_packets)			\
+  _ (0x4000, rx_crc_errors)				\
+  _ (0x4120, rx_ip_checksum_errors)			\
+  _ (0x4004, rx_illegal_symbol_errors)			\
+  _ (0x4008, rx_error_symbol_errors)			\
+  _ (0x4034, rx_mac_local_faults)			\
+  _ (0x4038, rx_mac_remote_faults)			\
+  _ (0x4040, rx_length_errors)				\
+  _ (0x41a4, rx_xons)					\
+  _ (0x41a8, rx_xoffs)					\
+  _ (0x40a4, rx_undersize_packets)			\
+  _ (0x40a8, rx_fragments)				\
+  _ (0x40ac, rx_oversize_packets)			\
+  _ (0x40b0, rx_jabbers)				\
+  _ (0x40b4, rx_management_packets)			\
+  _ (0x40b8, rx_management_drops)			\
+  _ (0x3fa0, rx_missed_packets_pool_0)			\
+  _ (0x40d4, tx_total_packets)				\
+  _ (0x4080, tx_good_packets)				\
+  _64 (0x4090, tx_good_bytes)				\
+  _ (0x40f0, tx_multicast_packets)			\
+  _ (0x40f4, tx_broadcast_packets)			\
+  _ (0x87a0, tx_dma_good_packets)			\
+  _64 (0x87a4, tx_dma_good_bytes)			\
+  _ (0x40d8, tx_64_byte_packets)			\
+  _ (0x40dc, tx_65_127_byte_packets)			\
+  _ (0x40e0, tx_128_255_byte_packets)			\
+  _ (0x40e4, tx_256_511_byte_packets)			\
+  _ (0x40e8, tx_512_1023_byte_packets)			\
+  _ (0x40ec, tx_gt_1023_byte_packets)			\
+  _ (0x4010, tx_undersize_drops)			\
+  _ (0x8780, switch_security_violation_packets)		\
+  _ (0x5118, fc_crc_errors)				\
+  _ (0x241c, fc_rx_drops)				\
+  _ (0x2424, fc_last_error_count)			\
+  _ (0x2428, fcoe_rx_packets)				\
+  _ (0x242c, fcoe_rx_dwords)				\
+  _ (0x8784, fcoe_tx_packets)				\
+  _ (0x8788, fcoe_tx_dwords)				\
+  _ (0x1030, queue_0_rx_count)				\
+  _ (0x1430, queue_0_drop_count)			\
+  _ (0x1070, queue_1_rx_count)				\
+  _ (0x1470, queue_1_drop_count)			\
+  _ (0x10b0, queue_2_rx_count)				\
+  _ (0x14b0, queue_2_drop_count)			\
+  _ (0x10f0, queue_3_rx_count)				\
+  _ (0x14f0, queue_3_drop_count)			\
+  _ (0x1130, queue_4_rx_count)				\
+  _ (0x1530, queue_4_drop_count)			\
+  _ (0x1170, queue_5_rx_count)				\
+  _ (0x1570, queue_5_drop_count)			\
+  _ (0x11b0, queue_6_rx_count)				\
+  _ (0x15b0, queue_6_drop_count)			\
+  _ (0x11f0, queue_7_rx_count)				\
+  _ (0x15f0, queue_7_drop_count)			\
+  _ (0x1230, queue_8_rx_count)				\
+  _ (0x1630, queue_8_drop_count)			\
+  _ (0x1270, queue_9_rx_count)				\
+  _ (0x1270, queue_9_drop_count)
+
+
+
+
+typedef enum {
+#define _(a,f) IXGE_COUNTER_##f,
+#define _64(a,f) _(a,f)
+  foreach_ixge_counter
+#undef _
+#undef _64
+  IXGE_N_COUNTER,
+} ixge_counter_type_t;
+
+typedef struct {
+  u32 mdio_address;
+
+  /* 32 bit ID read from ID registers. */
+  u32 id;
+} ixge_phy_t;
+
+typedef struct {
+  /* Cache aligned descriptors. */
+  ixge_descriptor_t * descriptors;
+
+  /* Number of descriptors in table. */
+  u32 n_descriptors;
+
+  /* Software head and tail pointers into descriptor ring. */
+  u32 head_index, tail_index;
+
+  /* Index into dma_queues vector. */
+  u32 queue_index;
+
+  /* Buffer indices corresponding to each active descriptor. */
+  u32 * descriptor_buffer_indices;
+
+  union {
+    struct {
+      u32 * volatile head_index_write_back;
+
+      u32 n_buffers_on_ring;
+    } tx;
+
+    struct {
+      /* Buffer indices to use to replenish each descriptor. */
+      u32 * replenish_buffer_indices;
+
+      vlib_node_runtime_t * node;
+      u32 next_index;
+
+      u32 saved_start_of_packet_buffer_index;
+
+      u32 saved_start_of_packet_next_index;
+      u32 saved_last_buffer_index;
+
+      u32 is_start_of_packet;
+
+      u32 n_descriptors_done_total;
+
+      u32 n_descriptors_done_this_call;
+
+      u32 n_bytes;
+    } rx;
+  };
+} ixge_dma_queue_t;
+
+#define foreach_ixge_pci_device_id		\
+  _ (82598, 0x10b6)				\
+  _ (82598_bx, 0x1508)				\
+  _ (82598af_dual_port, 0x10c6)			\
+  _ (82598af_single_port, 0x10c7)		\
+  _ (82598at, 0x10c8)				\
+  _ (82598at2, 0x150b)				\
+  _ (82598eb_sfp_lom, 0x10db)			\
+  _ (82598eb_cx4, 0x10dd)			\
+  _ (82598_cx4_dual_port, 0x10ec)		\
+  _ (82598_da_dual_port, 0x10f1)		\
+  _ (82598_sr_dual_port_em, 0x10e1)		\
+  _ (82598eb_xf_lr, 0x10f4)			\
+  _ (82599_kx4, 0x10f7)				\
+  _ (82599_kx4_mezz, 0x1514)			\
+  _ (82599_kr, 0x1517)				\
+  _ (82599_combo_backplane, 0x10f8)		\
+  _ (82599_cx4, 0x10f9)				\
+  _ (82599_sfp, 0x10fb)				\
+  _ (82599_backplane_fcoe, 0x152a)		\
+  _ (82599_sfp_fcoe, 0x1529)			\
+  _ (82599_sfp_em, 0x1507)			\
+  _ (82599_xaui_lom, 0x10fc)			\
+  _ (82599_t3_lom, 0x151c)			\
+  _ (x540t, 0x1528)
+
+typedef enum {
+#define _(f,n) IXGE_##f = n,
+  foreach_ixge_pci_device_id
+#undef _
+} ixge_pci_device_id_t;
+
+typedef struct {
+  /* registers */
+  ixge_regs_t * regs;
+
+  /* Specific next index when using dynamic redirection */
+  u32 per_interface_next_index;
+
+  /* PCI bus info. */
+  vlib_pci_device_t pci_device;
+
+  /* From PCI config space header. */
+  ixge_pci_device_id_t device_id;
+
+  u16 device_index;
+
+  /* 0 or 1. */
+  u16 pci_function;
+
+  /* VLIB interface for this instance. */
+  u32 vlib_hw_if_index, vlib_sw_if_index;
+
+  ixge_dma_queue_t * dma_queues[VLIB_N_RX_TX];
+
+  /* Phy index (0 or 1) and address on MDI bus. */
+  u32 phy_index;
+  ixge_phy_t phys[2];
+
+  /* Value of link_status register at last link change. */
+  u32 link_status_at_last_link_change;
+
+  i2c_bus_t i2c_bus;
+  sfp_eeprom_t sfp_eeprom;
+
+  /* Counters. */
+  u64 counters[IXGE_N_COUNTER], counters_last_clear[IXGE_N_COUNTER];
+} ixge_device_t;
+
+typedef struct {
+  vlib_main_t * vlib_main;
+
+  /* Vector of devices. */
+  ixge_device_t * devices;
+
+  /* Descriptor ring sizes. */
+  u32 n_descriptors[VLIB_N_RX_TX];
+
+  /* RX buffer size.  Must be at least 1k; will be rounded to
+     next largest 1k size. */
+  u32 n_bytes_in_rx_buffer;
+
+  u32 n_descriptors_per_cache_line;
+
+  u32 vlib_buffer_free_list_index;
+
+  u32 process_node_index;
+
+  /* Template and mask for initializing/validating TX descriptors. */
+  ixge_tx_descriptor_t tx_descriptor_template, tx_descriptor_template_mask;
+
+  /* Vector of buffers for which TX is done and can be freed. */
+  u32 * tx_buffers_pending_free;
+
+  u32 * rx_buffers_to_add;
+
+  f64 time_last_stats_update;
+} ixge_main_t;
+
+ixge_main_t ixge_main;
+vnet_device_class_t ixge_device_class;
+
+typedef enum {
+  IXGE_RX_NEXT_IP4_INPUT,
+  IXGE_RX_NEXT_IP6_INPUT,
+  IXGE_RX_NEXT_ETHERNET_INPUT,
+  IXGE_RX_NEXT_DROP,
+  IXGE_RX_N_NEXT,
+} ixge_rx_next_t;
+
+void ixge_set_next_node (ixge_rx_next_t, char *);
+
+#endif /* included_ixge_h */
diff --git a/vnet/vnet/devices/nic/sfp.c b/vnet/vnet/devices/nic/sfp.c
new file mode 100644
index 00000000000..fba94d73d95
--- /dev/null
+++ b/vnet/vnet/devices/nic/sfp.c
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <vnet/devices/nic/sfp.h>
+
+static u8 * format_space_terminated (u8 * s, va_list * args)
+{
+  u32 l = va_arg (*args, u32);
+  u8 * v = va_arg (*args, u8 *);
+  u8 * p;
+
+  for (p = v + l - 1; p >= v && p[0] == ' '; p--)
+    ;
+  vec_add (s, v, clib_min (p - v + 1, l));
+  return s;
+}
+
+static u8 * format_sfp_id (u8 * s, va_list * args)
+{
+  u32 id = va_arg (*args, u32);
+  char * t = 0;
+  switch (id)
+    {
+#define _(f) case SFP_ID_##f: t = #f; break;
+      foreach_sfp_id
+#undef _
+    default:
+      return format (s, "unknown 0x%x", id);
+    }
+  return format (s, "%s", t);
+}
+
+static u8 * format_sfp_compatibility (u8 * s, va_list * args)
+{
+  u32 c = va_arg (*args, u32);
+  char * t = 0;
+  switch (c)
+    {
+#define _(a,b,f) case SFP_COMPATIBILITY_##f: t = #f; break;
+      foreach_sfp_compatibility
+#undef _
+    default:
+      return format (s, "unknown 0x%x", c);
+    }
+  return format (s, "%s", t);
+}
+
+u32 sfp_is_comatible (sfp_eeprom_t * e, sfp_compatibility_t c)
+{
+  static struct { u8 byte, bit; } t[] = {
+#define _(a,b,f) { .byte = a, .bit = b, },
+    foreach_sfp_compatibility
+#undef _
+  };
+
+  ASSERT (c < ARRAY_LEN (t));
+  return (e->compatibility[t[c].byte] & (1 << t[c].bit)) != 0;
+}
+
+u8 * format_sfp_eeprom (u8 * s, va_list * args)
+{
+  sfp_eeprom_t * e = va_arg (*args, sfp_eeprom_t *);
+  uword indent = format_get_indent (s);
+  int i;
+
+  if (e->id != SFP_ID_sfp)
+    s = format (s, "id %U, ", format_sfp_id, e->id);
+
+  s = format (s, "compatibility:");
+  for (i = 0; i < SFP_N_COMPATIBILITY; i++)
+    if (sfp_is_comatible (e, i))
+      s = format (s, " %U", format_sfp_compatibility, i);
+
+  s = format (s, "\n%Uvendor: %U, part %U",
+	      format_white_space, indent,
+	      format_space_terminated, sizeof (e->vendor_name), e->vendor_name,
+	      format_space_terminated, sizeof (e->vendor_part_number), e->vendor_part_number);
+  s = format (s, "\n%Urevision: %U, serial: %U, date code: %U",
+	      format_white_space, indent,
+	      format_space_terminated, sizeof (e->vendor_revision), e->vendor_revision,
+	      format_space_terminated, sizeof (e->vendor_serial_number), e->vendor_serial_number,
+	      format_space_terminated, sizeof (e->vendor_date_code), e->vendor_date_code);
+
+  return s;
+}
diff --git a/vnet/vnet/devices/nic/sfp.h b/vnet/vnet/devices/nic/sfp.h
new file mode 100644
index 00000000000..9d31ad31f95
--- /dev/null
+++ b/vnet/vnet/devices/nic/sfp.h
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2016 Cisco and/or its affiliates.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at:
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef included_vnet_optics_sfp_h
+#define included_vnet_optics_sfp_h
+
+#include <vppinfra/format.h>
+
+#define foreach_sfp_id				\
+  _ (unknown)					\
+  _ (gbic)					\
+  _ (on_motherboard)				\
+  _ (sfp)
+
+typedef enum {
+#define _(f) SFP_ID_##f,
+  foreach_sfp_id
+#undef _
+} sfp_id_t;
+
+typedef struct {
+  u8 id;
+  u8 extended_id;
+  u8 connector_type;
+  u8 compatibility[8];
+  u8 encoding;
+  u8 nominal_bit_rate_100mbits_per_sec;
+  u8 reserved13;
+  u8 link_length[5];
+  u8 reserved19;
+  u8 vendor_name[16];
+  u8 reserved36;
+  u8 vendor_oui[3];
+  u8 vendor_part_number[16];
+  u8 vendor_revision[4];
+  /* 16 bit value network byte order. */
+  u8 laser_wavelength_in_nm[2];
+  u8 reserved62;
+  u8 checksum_0_to_62;
+
+  u8 options[2];
+  u8 max_bit_rate_margin_percent;
+  u8 min_bit_rate_margin_percent;
+  u8 vendor_serial_number[16];
+  u8 vendor_date_code[8];
+  u8 reserved92[3];
+  u8 checksum_63_to_94;
+  u8 vendor_specific[32];
+  u8 reserved128[384];
+
+  /* Vendor specific data follows. */
+  u8 vendor_specific1[0];
+} sfp_eeprom_t;
+
+always_inline uword
+sfp_eeprom_is_valid (sfp_eeprom_t * e)
+{
+  int i;
+  u8 sum = 0;
+  for (i = 0; i < 63; i++)
+    sum += ((u8 *) e)[i];
+  return sum == e->checksum_0_to_62;
+}
+
+/* _ (byte_index, bit_index, name) */
+#define foreach_sfp_compatibility		\
+  _ (0, 4, 10g_base_sr)				\
+  _ (0, 5, 10g_base_lr)				\
+  _ (1, 2, oc48_long_reach)			\
+  _ (1, 1, oc48_intermediate_reach)		\
+  _ (1, 0, oc48_short_reach)			\
+  _ (2, 6, oc12_long_reach)			\
+  _ (2, 5, oc12_intermediate_reach)		\
+  _ (2, 4, oc12_short_reach)			\
+  _ (2, 2, oc3_long_reach)			\
+  _ (2, 1, oc3_intermediate_reach)		\
+  _ (2, 0, oc3_short_reach)			\
+  _ (3, 3, 1g_base_t)				\
+  _ (3, 2, 1g_base_cx)				\
+  _ (3, 1, 1g_base_lx)				\
+  _ (3, 0, 1g_base_sx)
+
+typedef enum {
+#define _(a,b,f) SFP_COMPATIBILITY_##f,
+  foreach_sfp_compatibility
+#undef _
+  SFP_N_COMPATIBILITY,
+} sfp_compatibility_t;
+
+u32 sfp_is_comatible (sfp_eeprom_t * e, sfp_compatibility_t c);
+
+format_function_t format_sfp_eeprom;
+
+#endif /* included_vnet_optics_sfp_h */