Merge release-20191129.0-48-g0d02726 (automated)

13 files changed, 1496 insertions, 0 deletions

diff --git a/pkg/tcpip/link/sharedmem/pipe/pipe.go b/pkg/tcpip/link/sharedmem/pipe/pipe.go
new file mode 100755
index 000000000..74c9f0311
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/pipe/pipe.go
@@ -0,0 +1,78 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+// Package pipe implements a shared memory ring buffer on which a single reader
+// and a single writer can operate (read/write) concurrently. The ring buffer
+// allows for data of different sizes to be written, and preserves the boundary
+// of the written data.
+//
+// Example usage is as follows:
+//
+//	wb := t.Push(20)
+//	// Write data to wb.
+//	t.Flush()
+//
+//	rb := r.Pull()
+//	// Do something with data in rb.
+//	t.Flush()
+package pipe
+
+import (
+	"math"
+)
+
+const (
+	jump           uint64 = math.MaxUint32 + 1
+	offsetMask     uint64 = math.MaxUint32
+	revolutionMask uint64 = ^offsetMask
+
+	sizeOfSlotHeader        = 8 // sizeof(uint64)
+	slotFree         uint64 = 1 << 63
+	slotSizeMask     uint64 = math.MaxUint32
+)
+
+// payloadToSlotSize calculates the total size of a slot based on its payload
+// size. The  total size is the header size, plus the payload size, plus padding
+// if necessary to make the total size a multiple of sizeOfSlotHeader.
+func payloadToSlotSize(payloadSize uint64) uint64 {
+	s := sizeOfSlotHeader + payloadSize
+	return (s + sizeOfSlotHeader - 1) &^ (sizeOfSlotHeader - 1)
+}
+
+// slotToPayloadSize calculates the payload size of a slot based on the total
+// size of the slot. This is only meant to be used when creating slots that
+// don't carry information (e.g., free slots or wrap slots).
+func slotToPayloadSize(offset uint64) uint64 {
+	return offset - sizeOfSlotHeader
+}
+
+// pipe is a basic data structure used by both (transmit & receive) ends of a
+// pipe. Indices into this pipe are split into two fields: offset, which counts
+// the number of bytes from the beginning of the buffer, and revolution, which
+// counts the number of times the index has wrapped around.
+type pipe struct {
+	buffer []byte
+}
+
+// init initializes the pipe buffer such that its size is a multiple of the size
+// of the slot header.
+func (p *pipe) init(b []byte) {
+	p.buffer = b[:len(b)&^(sizeOfSlotHeader-1)]
+}
+
+// data returns a section of the buffer starting at the given index (which may
+// include revolution information) and with the given size.
+func (p *pipe) data(idx uint64, size uint64) []byte {
+	return p.buffer[(idx&offsetMask)+sizeOfSlotHeader:][:size]
+}
diff --git a/pkg/tcpip/link/sharedmem/pipe/pipe_state_autogen.go b/pkg/tcpip/link/sharedmem/pipe/pipe_state_autogen.go
new file mode 100755
index 000000000..c7c7c21b3
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/pipe/pipe_state_autogen.go
@@ -0,0 +1,4 @@
+// automatically generated by stateify.
+
+package pipe
+
diff --git a/pkg/tcpip/link/sharedmem/pipe/pipe_unsafe.go b/pkg/tcpip/link/sharedmem/pipe/pipe_unsafe.go
new file mode 100755
index 000000000..62d17029e
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/pipe/pipe_unsafe.go
@@ -0,0 +1,35 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+package pipe
+
+import (
+	"sync/atomic"
+	"unsafe"
+)
+
+func (p *pipe) write(idx uint64, v uint64) {
+	ptr := (*uint64)(unsafe.Pointer(&p.buffer[idx&offsetMask:][:8][0]))
+	*ptr = v
+}
+
+func (p *pipe) writeAtomic(idx uint64, v uint64) {
+	ptr := (*uint64)(unsafe.Pointer(&p.buffer[idx&offsetMask:][:8][0]))
+	atomic.StoreUint64(ptr, v)
+}
+
+func (p *pipe) readAtomic(idx uint64) uint64 {
+	ptr := (*uint64)(unsafe.Pointer(&p.buffer[idx&offsetMask:][:8][0]))
+	return atomic.LoadUint64(ptr)
+}
diff --git a/pkg/tcpip/link/sharedmem/pipe/rx.go b/pkg/tcpip/link/sharedmem/pipe/rx.go
new file mode 100755
index 000000000..f22e533ac
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/pipe/rx.go
@@ -0,0 +1,93 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+package pipe
+
+// Rx is the receive side of the shared memory ring buffer.
+type Rx struct {
+	p pipe
+
+	tail uint64
+	head uint64
+}
+
+// Init initializes the receive end of the pipe. In the initial state, the next
+// slot to be inspected is the very first one.
+func (r *Rx) Init(b []byte) {
+	r.p.init(b)
+	r.tail = 0xfffffffe * jump
+	r.head = r.tail
+}
+
+// Pull reads the next buffer from the pipe, returning nil if there isn't one
+// currently available.
+//
+// The returned slice is available until Flush() is next called. After that, it
+// must not be touched.
+func (r *Rx) Pull() []byte {
+	if r.head == r.tail+jump {
+		// We've already pulled the whole pipe.
+		return nil
+	}
+
+	header := r.p.readAtomic(r.head)
+	if header&slotFree != 0 {
+		// The next slot is free, we can't pull it yet.
+		return nil
+	}
+
+	payloadSize := header & slotSizeMask
+	newHead := r.head + payloadToSlotSize(payloadSize)
+	headWrap := (r.head & revolutionMask) | uint64(len(r.p.buffer))
+
+	// Check if this is a wrapping slot. If that's the case, it carries no
+	// data, so we just skip it and try again from the first slot.
+	if int64(newHead-headWrap) >= 0 {
+		if int64(newHead-headWrap) > int64(jump) || newHead&offsetMask != 0 {
+			return nil
+		}
+
+		if r.tail == r.head {
+			// If this is the first pull since the last Flush()
+			// call, we flush the state so that the sender can use
+			// this space if it needs to.
+			r.p.writeAtomic(r.head, slotFree|slotToPayloadSize(newHead-r.head))
+			r.tail = newHead
+		}
+
+		r.head = newHead
+		return r.Pull()
+	}
+
+	// Grab the buffer before updating r.head.
+	b := r.p.data(r.head, payloadSize)
+	r.head = newHead
+	return b
+}
+
+// Flush tells the transmitter that all buffers pulled since the last Flush()
+// have been used, so the transmitter is free to used their slots for further
+// transmission.
+func (r *Rx) Flush() {
+	if r.head == r.tail {
+		return
+	}
+	r.p.writeAtomic(r.tail, slotFree|slotToPayloadSize(r.head-r.tail))
+	r.tail = r.head
+}
+
+// Bytes returns the byte slice on which the pipe operates.
+func (r *Rx) Bytes() []byte {
+	return r.p.buffer
+}
diff --git a/pkg/tcpip/link/sharedmem/pipe/tx.go b/pkg/tcpip/link/sharedmem/pipe/tx.go
new file mode 100755
index 000000000..9841eb231
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/pipe/tx.go
@@ -0,0 +1,161 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+package pipe
+
+// Tx is the transmit side of the shared memory ring buffer.
+type Tx struct {
+	p              pipe
+	maxPayloadSize uint64
+
+	head uint64
+	tail uint64
+	next uint64
+
+	tailHeader uint64
+}
+
+// Init initializes the transmit end of the pipe. In the initial state, the next
+// slot to be written is the very first one, and the transmitter has the whole
+// ring buffer available to it.
+func (t *Tx) Init(b []byte) {
+	t.p.init(b)
+	// maxPayloadSize excludes the header of the payload, and the header
+	// of the wrapping message.
+	t.maxPayloadSize = uint64(len(t.p.buffer)) - 2*sizeOfSlotHeader
+	t.tail = 0xfffffffe * jump
+	t.next = t.tail
+	t.head = t.tail + jump
+	t.p.write(t.tail, slotFree)
+}
+
+// Capacity determines how many records of the given size can be written to the
+// pipe before it fills up.
+func (t *Tx) Capacity(recordSize uint64) uint64 {
+	available := uint64(len(t.p.buffer)) - sizeOfSlotHeader
+	entryLen := payloadToSlotSize(recordSize)
+	return available / entryLen
+}
+
+// Push reserves "payloadSize" bytes for transmission in the pipe. The caller
+// populates the returned slice with the data to be transferred and enventually
+// calls Flush() to make the data visible to the reader, or Abort() to make the
+// pipe forget all Push() calls since the last Flush().
+//
+// The returned slice is available until Flush() or Abort() is next called.
+// After that, it must not be touched.
+func (t *Tx) Push(payloadSize uint64) []byte {
+	// Fail request if we know we will never have enough room.
+	if payloadSize > t.maxPayloadSize {
+		return nil
+	}
+
+	totalLen := payloadToSlotSize(payloadSize)
+	newNext := t.next + totalLen
+	nextWrap := (t.next & revolutionMask) | uint64(len(t.p.buffer))
+	if int64(newNext-nextWrap) >= 0 {
+		// The new buffer would overflow the pipe, so we push a wrapping
+		// slot, then try to add the actual slot to the front of the
+		// pipe.
+		newNext = (newNext & revolutionMask) + jump
+		wrappingPayloadSize := slotToPayloadSize(newNext - t.next)
+		if !t.reclaim(newNext) {
+			return nil
+		}
+
+		oldNext := t.next
+		t.next = newNext
+		if oldNext != t.tail {
+			t.p.write(oldNext, wrappingPayloadSize)
+		} else {
+			t.tailHeader = wrappingPayloadSize
+			t.Flush()
+		}
+
+		newNext += totalLen
+	}
+
+	// Check that we have enough room for the buffer.
+	if !t.reclaim(newNext) {
+		return nil
+	}
+
+	if t.next != t.tail {
+		t.p.write(t.next, payloadSize)
+	} else {
+		t.tailHeader = payloadSize
+	}
+
+	// Grab the buffer before updating t.next.
+	b := t.p.data(t.next, payloadSize)
+	t.next = newNext
+
+	return b
+}
+
+// reclaim attempts to advance the head until at least newNext. If the head is
+// already at or beyond newNext, nothing happens and true is returned; otherwise
+// it tries to reclaim slots that have already been consumed by the receive end
+// of the pipe (they will be marked as free) and returns a boolean indicating
+// whether it was successful in reclaiming enough slots.
+func (t *Tx) reclaim(newNext uint64) bool {
+	for int64(newNext-t.head) > 0 {
+		// Can't reclaim if slot is not free.
+		header := t.p.readAtomic(t.head)
+		if header&slotFree == 0 {
+			return false
+		}
+
+		payloadSize := header & slotSizeMask
+		newHead := t.head + payloadToSlotSize(payloadSize)
+
+		// Check newHead is within bounds and valid.
+		if int64(newHead-t.tail) > int64(jump) || newHead&offsetMask >= uint64(len(t.p.buffer)) {
+			return false
+		}
+
+		t.head = newHead
+	}
+
+	return true
+}
+
+// Abort causes all Push() calls since the last Flush() to be forgotten and
+// therefore they will not be made visible to the receiver.
+func (t *Tx) Abort() {
+	t.next = t.tail
+}
+
+// Flush causes all buffers pushed since the last Flush() [or Abort(), whichever
+// is the most recent] to be made visible to the receiver.
+func (t *Tx) Flush() {
+	if t.next == t.tail {
+		// Nothing to do if there are no pushed buffers.
+		return
+	}
+
+	if t.next != t.head {
+		// The receiver will spin in t.next, so we must make sure that
+		// the slotFree bit is set.
+		t.p.write(t.next, slotFree)
+	}
+
+	t.p.writeAtomic(t.tail, t.tailHeader)
+	t.tail = t.next
+}
+
+// Bytes returns the byte slice on which the pipe operates.
+func (t *Tx) Bytes() []byte {
+	return t.p.buffer
+}
diff --git a/pkg/tcpip/link/sharedmem/queue/queue_state_autogen.go b/pkg/tcpip/link/sharedmem/queue/queue_state_autogen.go
new file mode 100755
index 000000000..eec17d734
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/queue/queue_state_autogen.go
@@ -0,0 +1,4 @@
+// automatically generated by stateify.
+
+package queue
+
diff --git a/pkg/tcpip/link/sharedmem/queue/rx.go b/pkg/tcpip/link/sharedmem/queue/rx.go
new file mode 100755
index 000000000..696e6c9e5
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/queue/rx.go
@@ -0,0 +1,221 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+// Package queue provides the implementation of transmit and receive queues
+// based on shared memory ring buffers.
+package queue
+
+import (
+	"encoding/binary"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/tcpip/link/sharedmem/pipe"
+)
+
+const (
+	// Offsets within a posted buffer.
+	postedOffset           = 0
+	postedSize             = 8
+	postedRemainingInGroup = 12
+	postedUserData         = 16
+	postedID               = 24
+
+	sizeOfPostedBuffer = 32
+
+	// Offsets within a received packet header.
+	consumedPacketSize     = 0
+	consumedPacketReserved = 4
+
+	sizeOfConsumedPacketHeader = 8
+
+	// Offsets within a consumed buffer.
+	consumedOffset   = 0
+	consumedSize     = 8
+	consumedUserData = 12
+	consumedID       = 20
+
+	sizeOfConsumedBuffer = 28
+
+	// The following are the allowed states of the shared data area.
+	eventFDUninitialized = 0
+	eventFDDisabled      = 1
+	eventFDEnabled       = 2
+)
+
+// RxBuffer is the descriptor of a receive buffer.
+type RxBuffer struct {
+	Offset   uint64
+	Size     uint32
+	ID       uint64
+	UserData uint64
+}
+
+// Rx is a receive queue. It is implemented with one tx and one rx pipe: the tx
+// pipe is used to "post" buffers, while the rx pipe is used to receive packets
+// whose contents have been written to previously posted buffers.
+//
+// This struct is thread-compatible.
+type Rx struct {
+	tx                 pipe.Tx
+	rx                 pipe.Rx
+	sharedEventFDState *uint32
+}
+
+// Init initializes the receive queue with the given pipes, and shared state
+// pointer -- the latter is used to enable/disable eventfd notifications.
+func (r *Rx) Init(tx, rx []byte, sharedEventFDState *uint32) {
+	r.sharedEventFDState = sharedEventFDState
+	r.tx.Init(tx)
+	r.rx.Init(rx)
+}
+
+// EnableNotification updates the shared state such that the peer will notify
+// the eventfd when there are packets to be dequeued.
+func (r *Rx) EnableNotification() {
+	atomic.StoreUint32(r.sharedEventFDState, eventFDEnabled)
+}
+
+// DisableNotification updates the shared state such that the peer will not
+// notify the eventfd.
+func (r *Rx) DisableNotification() {
+	atomic.StoreUint32(r.sharedEventFDState, eventFDDisabled)
+}
+
+// PostedBuffersLimit returns the maximum number of buffers that can be posted
+// before the tx queue fills up.
+func (r *Rx) PostedBuffersLimit() uint64 {
+	return r.tx.Capacity(sizeOfPostedBuffer)
+}
+
+// PostBuffers makes the given buffers available for receiving data from the
+// peer. Once they are posted, the peer is free to write to them and will
+// eventually post them back for consumption.
+func (r *Rx) PostBuffers(buffers []RxBuffer) bool {
+	for i := range buffers {
+		b := r.tx.Push(sizeOfPostedBuffer)
+		if b == nil {
+			r.tx.Abort()
+			return false
+		}
+
+		pb := &buffers[i]
+		binary.LittleEndian.PutUint64(b[postedOffset:], pb.Offset)
+		binary.LittleEndian.PutUint32(b[postedSize:], pb.Size)
+		binary.LittleEndian.PutUint32(b[postedRemainingInGroup:], 0)
+		binary.LittleEndian.PutUint64(b[postedUserData:], pb.UserData)
+		binary.LittleEndian.PutUint64(b[postedID:], pb.ID)
+	}
+
+	r.tx.Flush()
+
+	return true
+}
+
+// Dequeue receives buffers that have been previously posted by PostBuffers()
+// and that have been filled by the peer and posted back.
+//
+// This is similar to append() in that new buffers are appended to "bufs", with
+// reallocation only if "bufs" doesn't have enough capacity.
+func (r *Rx) Dequeue(bufs []RxBuffer) ([]RxBuffer, uint32) {
+	for {
+		outBufs := bufs
+
+		// Pull the next descriptor from the rx pipe.
+		b := r.rx.Pull()
+		if b == nil {
+			return bufs, 0
+		}
+
+		if len(b) < sizeOfConsumedPacketHeader {
+			log.Warningf("Ignoring packet header: size (%v) is less than header size (%v)", len(b), sizeOfConsumedPacketHeader)
+			r.rx.Flush()
+			continue
+		}
+
+		totalDataSize := binary.LittleEndian.Uint32(b[consumedPacketSize:])
+
+		// Calculate the number of buffer descriptors and copy them
+		// over to the output.
+		count := (len(b) - sizeOfConsumedPacketHeader) / sizeOfConsumedBuffer
+		offset := sizeOfConsumedPacketHeader
+		buffersSize := uint32(0)
+		for i := count; i > 0; i-- {
+			s := binary.LittleEndian.Uint32(b[offset+consumedSize:])
+			buffersSize += s
+			if buffersSize < s {
+				// The buffer size overflows an unsigned 32-bit
+				// integer, so break out and force it to be
+				// ignored.
+				totalDataSize = 1
+				buffersSize = 0
+				break
+			}
+
+			outBufs = append(outBufs, RxBuffer{
+				Offset: binary.LittleEndian.Uint64(b[offset+consumedOffset:]),
+				Size:   s,
+				ID:     binary.LittleEndian.Uint64(b[offset+consumedID:]),
+			})
+
+			offset += sizeOfConsumedBuffer
+		}
+
+		r.rx.Flush()
+
+		if buffersSize < totalDataSize {
+			// The descriptor is corrupted, ignore it.
+			log.Warningf("Ignoring packet: actual data size (%v) less than expected size (%v)", buffersSize, totalDataSize)
+			continue
+		}
+
+		return outBufs, totalDataSize
+	}
+}
+
+// Bytes returns the byte slices on which the queue operates.
+func (r *Rx) Bytes() (tx, rx []byte) {
+	return r.tx.Bytes(), r.rx.Bytes()
+}
+
+// DecodeRxBufferHeader decodes the header of a buffer posted on an rx queue.
+func DecodeRxBufferHeader(b []byte) RxBuffer {
+	return RxBuffer{
+		Offset:   binary.LittleEndian.Uint64(b[postedOffset:]),
+		Size:     binary.LittleEndian.Uint32(b[postedSize:]),
+		ID:       binary.LittleEndian.Uint64(b[postedID:]),
+		UserData: binary.LittleEndian.Uint64(b[postedUserData:]),
+	}
+}
+
+// RxCompletionSize returns the number of bytes needed to encode an rx
+// completion containing "count" buffers.
+func RxCompletionSize(count int) uint64 {
+	return sizeOfConsumedPacketHeader + uint64(count)*sizeOfConsumedBuffer
+}
+
+// EncodeRxCompletion encodes an rx completion header.
+func EncodeRxCompletion(b []byte, size, reserved uint32) {
+	binary.LittleEndian.PutUint32(b[consumedPacketSize:], size)
+	binary.LittleEndian.PutUint32(b[consumedPacketReserved:], reserved)
+}
+
+// EncodeRxCompletionBuffer encodes the i-th rx completion buffer header.
+func EncodeRxCompletionBuffer(b []byte, i int, rxb RxBuffer) {
+	b = b[RxCompletionSize(i):]
+	binary.LittleEndian.PutUint64(b[consumedOffset:], rxb.Offset)
+	binary.LittleEndian.PutUint32(b[consumedSize:], rxb.Size)
+	binary.LittleEndian.PutUint64(b[consumedUserData:], rxb.UserData)
+	binary.LittleEndian.PutUint64(b[consumedID:], rxb.ID)
+}
diff --git a/pkg/tcpip/link/sharedmem/queue/tx.go b/pkg/tcpip/link/sharedmem/queue/tx.go
new file mode 100755
index 000000000..beffe807b
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/queue/tx.go
@@ -0,0 +1,151 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+package queue
+
+import (
+	"encoding/binary"
+
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/tcpip/link/sharedmem/pipe"
+)
+
+const (
+	// Offsets within a packet header.
+	packetID       = 0
+	packetSize     = 8
+	packetReserved = 12
+
+	sizeOfPacketHeader = 16
+
+	// Offsets with a buffer descriptor
+	bufferOffset = 0
+	bufferSize   = 8
+
+	sizeOfBufferDescriptor = 12
+)
+
+// TxBuffer is the descriptor of a transmit buffer.
+type TxBuffer struct {
+	Next   *TxBuffer
+	Offset uint64
+	Size   uint32
+}
+
+// Tx is a transmit queue. It is implemented with one tx and one rx pipe: the
+// tx pipe is used to request the transmission of packets, while the rx pipe
+// is used to receive which transmissions have completed.
+//
+// This struct is thread-compatible.
+type Tx struct {
+	tx pipe.Tx
+	rx pipe.Rx
+}
+
+// Init initializes the transmit queue with the given pipes.
+func (t *Tx) Init(tx, rx []byte) {
+	t.tx.Init(tx)
+	t.rx.Init(rx)
+}
+
+// Enqueue queues the given linked list of buffers for transmission as one
+// packet. While it is queued, the caller must not modify them.
+func (t *Tx) Enqueue(id uint64, totalDataLen, bufferCount uint32, buffer *TxBuffer) bool {
+	// Reserve room in the tx pipe.
+	totalLen := sizeOfPacketHeader + uint64(bufferCount)*sizeOfBufferDescriptor
+
+	b := t.tx.Push(totalLen)
+	if b == nil {
+		return false
+	}
+
+	// Initialize the packet and buffer descriptors.
+	binary.LittleEndian.PutUint64(b[packetID:], id)
+	binary.LittleEndian.PutUint32(b[packetSize:], totalDataLen)
+	binary.LittleEndian.PutUint32(b[packetReserved:], 0)
+
+	offset := sizeOfPacketHeader
+	for i := bufferCount; i != 0; i-- {
+		binary.LittleEndian.PutUint64(b[offset+bufferOffset:], buffer.Offset)
+		binary.LittleEndian.PutUint32(b[offset+bufferSize:], buffer.Size)
+		offset += sizeOfBufferDescriptor
+		buffer = buffer.Next
+	}
+
+	t.tx.Flush()
+
+	return true
+}
+
+// CompletedPacket returns the id of the last completed transmission. The
+// returned id, if any, refers to a value passed on a previous call to
+// Enqueue().
+func (t *Tx) CompletedPacket() (id uint64, ok bool) {
+	for {
+		b := t.rx.Pull()
+		if b == nil {
+			return 0, false
+		}
+
+		if len(b) != 8 {
+			t.rx.Flush()
+			log.Warningf("Ignoring completed packet: size (%v) is less than expected (%v)", len(b), 8)
+			continue
+		}
+
+		v := binary.LittleEndian.Uint64(b)
+
+		t.rx.Flush()
+
+		return v, true
+	}
+}
+
+// Bytes returns the byte slices on which the queue operates.
+func (t *Tx) Bytes() (tx, rx []byte) {
+	return t.tx.Bytes(), t.rx.Bytes()
+}
+
+// TxPacketInfo holds information about a packet sent on a tx queue.
+type TxPacketInfo struct {
+	ID          uint64
+	Size        uint32
+	Reserved    uint32
+	BufferCount int
+}
+
+// DecodeTxPacketHeader decodes the header of a packet sent over a tx queue.
+func DecodeTxPacketHeader(b []byte) TxPacketInfo {
+	return TxPacketInfo{
+		ID:          binary.LittleEndian.Uint64(b[packetID:]),
+		Size:        binary.LittleEndian.Uint32(b[packetSize:]),
+		Reserved:    binary.LittleEndian.Uint32(b[packetReserved:]),
+		BufferCount: (len(b) - sizeOfPacketHeader) / sizeOfBufferDescriptor,
+	}
+}
+
+// DecodeTxBufferHeader decodes the header of the i-th buffer of a packet sent
+// over a tx queue.
+func DecodeTxBufferHeader(b []byte, i int) TxBuffer {
+	b = b[sizeOfPacketHeader+i*sizeOfBufferDescriptor:]
+	return TxBuffer{
+		Offset: binary.LittleEndian.Uint64(b[bufferOffset:]),
+		Size:   binary.LittleEndian.Uint32(b[bufferSize:]),
+	}
+}
+
+// EncodeTxCompletion encodes a tx completion header.
+func EncodeTxCompletion(b []byte, id uint64) {
+	binary.LittleEndian.PutUint64(b, id)
+}
diff --git a/pkg/tcpip/link/sharedmem/rx.go b/pkg/tcpip/link/sharedmem/rx.go
new file mode 100755
index 000000000..eec11e4cb
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/rx.go
@@ -0,0 +1,159 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+// +build linux
+
+package sharedmem
+
+import (
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/tcpip/link/rawfile"
+	"gvisor.dev/gvisor/pkg/tcpip/link/sharedmem/queue"
+)
+
+// rx holds all state associated with an rx queue.
+type rx struct {
+	data       []byte
+	sharedData []byte
+	q          queue.Rx
+	eventFD    int
+}
+
+// init initializes all state needed by the rx queue based on the information
+// provided.
+//
+// The caller always retains ownership of all file descriptors passed in. The
+// queue implementation will duplicate any that it may need in the future.
+func (r *rx) init(mtu uint32, c *QueueConfig) error {
+	// Map in all buffers.
+	txPipe, err := getBuffer(c.TxPipeFD)
+	if err != nil {
+		return err
+	}
+
+	rxPipe, err := getBuffer(c.RxPipeFD)
+	if err != nil {
+		syscall.Munmap(txPipe)
+		return err
+	}
+
+	data, err := getBuffer(c.DataFD)
+	if err != nil {
+		syscall.Munmap(txPipe)
+		syscall.Munmap(rxPipe)
+		return err
+	}
+
+	sharedData, err := getBuffer(c.SharedDataFD)
+	if err != nil {
+		syscall.Munmap(txPipe)
+		syscall.Munmap(rxPipe)
+		syscall.Munmap(data)
+		return err
+	}
+
+	// Duplicate the eventFD so that caller can close it but we can still
+	// use it.
+	efd, err := syscall.Dup(c.EventFD)
+	if err != nil {
+		syscall.Munmap(txPipe)
+		syscall.Munmap(rxPipe)
+		syscall.Munmap(data)
+		syscall.Munmap(sharedData)
+		return err
+	}
+
+	// Set the eventfd as non-blocking.
+	if err := syscall.SetNonblock(efd, true); err != nil {
+		syscall.Munmap(txPipe)
+		syscall.Munmap(rxPipe)
+		syscall.Munmap(data)
+		syscall.Munmap(sharedData)
+		syscall.Close(efd)
+		return err
+	}
+
+	// Initialize state based on buffers.
+	r.q.Init(txPipe, rxPipe, sharedDataPointer(sharedData))
+	r.data = data
+	r.eventFD = efd
+	r.sharedData = sharedData
+
+	return nil
+}
+
+// cleanup releases all resources allocated during init(). It must only be
+// called if init() has previously succeeded.
+func (r *rx) cleanup() {
+	a, b := r.q.Bytes()
+	syscall.Munmap(a)
+	syscall.Munmap(b)
+
+	syscall.Munmap(r.data)
+	syscall.Munmap(r.sharedData)
+	syscall.Close(r.eventFD)
+}
+
+// postAndReceive posts the provided buffers (if any), and then tries to read
+// from the receive queue.
+//
+// Capacity permitting, it reuses the posted buffer slice to store the buffers
+// that were read as well.
+//
+// This function will block if there aren't any available packets.
+func (r *rx) postAndReceive(b []queue.RxBuffer, stopRequested *uint32) ([]queue.RxBuffer, uint32) {
+	// Post the buffers first. If we cannot post, sleep until we can. We
+	// never post more than will fit concurrently, so it's safe to wait
+	// until enough room is available.
+	if len(b) != 0 && !r.q.PostBuffers(b) {
+		r.q.EnableNotification()
+		for !r.q.PostBuffers(b) {
+			var tmp [8]byte
+			rawfile.BlockingRead(r.eventFD, tmp[:])
+			if atomic.LoadUint32(stopRequested) != 0 {
+				r.q.DisableNotification()
+				return nil, 0
+			}
+		}
+		r.q.DisableNotification()
+	}
+
+	// Read the next set of descriptors.
+	b, n := r.q.Dequeue(b[:0])
+	if len(b) != 0 {
+		return b, n
+	}
+
+	// Data isn't immediately available. Enable eventfd notifications.
+	r.q.EnableNotification()
+	for {
+		b, n = r.q.Dequeue(b)
+		if len(b) != 0 {
+			break
+		}
+
+		// Wait for notification.
+		var tmp [8]byte
+		rawfile.BlockingRead(r.eventFD, tmp[:])
+		if atomic.LoadUint32(stopRequested) != 0 {
+			r.q.DisableNotification()
+			return nil, 0
+		}
+	}
+	r.q.DisableNotification()
+
+	return b, n
+}
diff --git a/pkg/tcpip/link/sharedmem/sharedmem.go b/pkg/tcpip/link/sharedmem/sharedmem.go
new file mode 100755
index 000000000..080f9d667
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/sharedmem.go
@@ -0,0 +1,289 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+// +build linux
+
+// Package sharedmem provides the implemention of data-link layer endpoints
+// backed by shared memory.
+//
+// Shared memory endpoints can be used in the networking stack by calling New()
+// to create a new endpoint, and then passing it as an argument to
+// Stack.CreateNIC().
+package sharedmem
+
+import (
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/tcpip"
+	"gvisor.dev/gvisor/pkg/tcpip/buffer"
+	"gvisor.dev/gvisor/pkg/tcpip/header"
+	"gvisor.dev/gvisor/pkg/tcpip/link/sharedmem/queue"
+	"gvisor.dev/gvisor/pkg/tcpip/stack"
+)
+
+// QueueConfig holds all the file descriptors needed to describe a tx or rx
+// queue over shared memory. It is used when creating new shared memory
+// endpoints to describe tx and rx queues.
+type QueueConfig struct {
+	// DataFD is a file descriptor for the file that contains the data to
+	// be transmitted via this queue. Descriptors contain offsets within
+	// this file.
+	DataFD int
+
+	// EventFD is a file descriptor for the event that is signaled when
+	// data is becomes available in this queue.
+	EventFD int
+
+	// TxPipeFD is a file descriptor for the tx pipe associated with the
+	// queue.
+	TxPipeFD int
+
+	// RxPipeFD is a file descriptor for the rx pipe associated with the
+	// queue.
+	RxPipeFD int
+
+	// SharedDataFD is a file descriptor for the file that contains shared
+	// state between the two ends of the queue. This data specifies, for
+	// example, whether EventFD signaling is enabled or disabled.
+	SharedDataFD int
+}
+
+type endpoint struct {
+	// mtu (maximum transmission unit) is the maximum size of a packet.
+	mtu uint32
+
+	// bufferSize is the size of each individual buffer.
+	bufferSize uint32
+
+	// addr is the local address of this endpoint.
+	addr tcpip.LinkAddress
+
+	// rx is the receive queue.
+	rx rx
+
+	// stopRequested is to be accessed atomically only, and determines if
+	// the worker goroutines should stop.
+	stopRequested uint32
+
+	// Wait group used to indicate that all workers have stopped.
+	completed sync.WaitGroup
+
+	// mu protects the following fields.
+	mu sync.Mutex
+
+	// tx is the transmit queue.
+	tx tx
+
+	// workerStarted specifies whether the worker goroutine was started.
+	workerStarted bool
+}
+
+// New creates a new shared-memory-based endpoint. Buffers will be broken up
+// into buffers of "bufferSize" bytes.
+func New(mtu, bufferSize uint32, addr tcpip.LinkAddress, tx, rx QueueConfig) (stack.LinkEndpoint, error) {
+	e := &endpoint{
+		mtu:        mtu,
+		bufferSize: bufferSize,
+		addr:       addr,
+	}
+
+	if err := e.tx.init(bufferSize, &tx); err != nil {
+		return nil, err
+	}
+
+	if err := e.rx.init(bufferSize, &rx); err != nil {
+		e.tx.cleanup()
+		return nil, err
+	}
+
+	return e, nil
+}
+
+// Close frees all resources associated with the endpoint.
+func (e *endpoint) Close() {
+	// Tell dispatch goroutine to stop, then write to the eventfd so that
+	// it wakes up in case it's sleeping.
+	atomic.StoreUint32(&e.stopRequested, 1)
+	syscall.Write(e.rx.eventFD, []byte{1, 0, 0, 0, 0, 0, 0, 0})
+
+	// Cleanup the queues inline if the worker hasn't started yet; we also
+	// know it won't start from now on because stopRequested is set to 1.
+	e.mu.Lock()
+	workerPresent := e.workerStarted
+	e.mu.Unlock()
+
+	if !workerPresent {
+		e.tx.cleanup()
+		e.rx.cleanup()
+	}
+}
+
+// Wait implements stack.LinkEndpoint.Wait. It waits until all workers have
+// stopped after a Close() call.
+func (e *endpoint) Wait() {
+	e.completed.Wait()
+}
+
+// Attach implements stack.LinkEndpoint.Attach. It launches the goroutine that
+// reads packets from the rx queue.
+func (e *endpoint) Attach(dispatcher stack.NetworkDispatcher) {
+	e.mu.Lock()
+	if !e.workerStarted && atomic.LoadUint32(&e.stopRequested) == 0 {
+		e.workerStarted = true
+		e.completed.Add(1)
+		// Link endpoints are not savable. When transportation endpoints
+		// are saved, they stop sending outgoing packets and all
+		// incoming packets are rejected.
+		go e.dispatchLoop(dispatcher) // S/R-SAFE: see above.
+	}
+	e.mu.Unlock()
+}
+
+// IsAttached implements stack.LinkEndpoint.IsAttached.
+func (e *endpoint) IsAttached() bool {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.workerStarted
+}
+
+// MTU implements stack.LinkEndpoint.MTU. It returns the value initialized
+// during construction.
+func (e *endpoint) MTU() uint32 {
+	return e.mtu - header.EthernetMinimumSize
+}
+
+// Capabilities implements stack.LinkEndpoint.Capabilities.
+func (*endpoint) Capabilities() stack.LinkEndpointCapabilities {
+	return 0
+}
+
+// MaxHeaderLength implements stack.LinkEndpoint.MaxHeaderLength. It returns the
+// ethernet frame header size.
+func (*endpoint) MaxHeaderLength() uint16 {
+	return header.EthernetMinimumSize
+}
+
+// LinkAddress implements stack.LinkEndpoint.LinkAddress. It returns the local
+// link address.
+func (e *endpoint) LinkAddress() tcpip.LinkAddress {
+	return e.addr
+}
+
+// WritePacket writes outbound packets to the file descriptor. If it is not
+// currently writable, the packet is dropped.
+func (e *endpoint) WritePacket(r *stack.Route, _ *stack.GSO, protocol tcpip.NetworkProtocolNumber, pkt tcpip.PacketBuffer) *tcpip.Error {
+	// Add the ethernet header here.
+	eth := header.Ethernet(pkt.Header.Prepend(header.EthernetMinimumSize))
+	pkt.LinkHeader = buffer.View(eth)
+	ethHdr := &header.EthernetFields{
+		DstAddr: r.RemoteLinkAddress,
+		Type:    protocol,
+	}
+	if r.LocalLinkAddress != "" {
+		ethHdr.SrcAddr = r.LocalLinkAddress
+	} else {
+		ethHdr.SrcAddr = e.addr
+	}
+	eth.Encode(ethHdr)
+
+	v := pkt.Data.ToView()
+	// Transmit the packet.
+	e.mu.Lock()
+	ok := e.tx.transmit(pkt.Header.View(), v)
+	e.mu.Unlock()
+
+	if !ok {
+		return tcpip.ErrWouldBlock
+	}
+
+	return nil
+}
+
+// WritePackets implements stack.LinkEndpoint.WritePackets.
+func (e *endpoint) WritePackets(r *stack.Route, _ *stack.GSO, pkts []tcpip.PacketBuffer, protocol tcpip.NetworkProtocolNumber) (int, *tcpip.Error) {
+	panic("not implemented")
+}
+
+// WriteRawPacket implements stack.LinkEndpoint.WriteRawPacket.
+func (e *endpoint) WriteRawPacket(vv buffer.VectorisedView) *tcpip.Error {
+	v := vv.ToView()
+	// Transmit the packet.
+	e.mu.Lock()
+	ok := e.tx.transmit(v, buffer.View{})
+	e.mu.Unlock()
+
+	if !ok {
+		return tcpip.ErrWouldBlock
+	}
+
+	return nil
+}
+
+// dispatchLoop reads packets from the rx queue in a loop and dispatches them
+// to the network stack.
+func (e *endpoint) dispatchLoop(d stack.NetworkDispatcher) {
+	// Post initial set of buffers.
+	limit := e.rx.q.PostedBuffersLimit()
+	if l := uint64(len(e.rx.data)) / uint64(e.bufferSize); limit > l {
+		limit = l
+	}
+	for i := uint64(0); i < limit; i++ {
+		b := queue.RxBuffer{
+			Offset: i * uint64(e.bufferSize),
+			Size:   e.bufferSize,
+			ID:     i,
+		}
+		if !e.rx.q.PostBuffers([]queue.RxBuffer{b}) {
+			log.Warningf("Unable to post %v-th buffer", i)
+		}
+	}
+
+	// Read in a loop until a stop is requested.
+	var rxb []queue.RxBuffer
+	for atomic.LoadUint32(&e.stopRequested) == 0 {
+		var n uint32
+		rxb, n = e.rx.postAndReceive(rxb, &e.stopRequested)
+
+		// Copy data from the shared area to its own buffer, then
+		// prepare to repost the buffer.
+		b := make([]byte, n)
+		offset := uint32(0)
+		for i := range rxb {
+			copy(b[offset:], e.rx.data[rxb[i].Offset:][:rxb[i].Size])
+			offset += rxb[i].Size
+
+			rxb[i].Size = e.bufferSize
+		}
+
+		if n < header.EthernetMinimumSize {
+			continue
+		}
+
+		// Send packet up the stack.
+		eth := header.Ethernet(b[:header.EthernetMinimumSize])
+		d.DeliverNetworkPacket(e, eth.SourceAddress(), eth.DestinationAddress(), eth.Type(), tcpip.PacketBuffer{
+			Data:       buffer.View(b[header.EthernetMinimumSize:]).ToVectorisedView(),
+			LinkHeader: buffer.View(eth),
+		})
+	}
+
+	// Clean state.
+	e.tx.cleanup()
+	e.rx.cleanup()
+
+	e.completed.Done()
+}
diff --git a/pkg/tcpip/link/sharedmem/sharedmem_state_autogen.go b/pkg/tcpip/link/sharedmem/sharedmem_state_autogen.go
new file mode 100755
index 000000000..e5c542528
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/sharedmem_state_autogen.go
@@ -0,0 +1,4 @@
+// automatically generated by stateify.
+
+package sharedmem
+
diff --git a/pkg/tcpip/link/sharedmem/sharedmem_unsafe.go b/pkg/tcpip/link/sharedmem/sharedmem_unsafe.go
new file mode 100755
index 000000000..f7e816a41
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/sharedmem_unsafe.go
@@ -0,0 +1,25 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+package sharedmem
+
+import (
+	"unsafe"
+)
+
+// sharedDataPointer converts the shared data slice into a pointer so that it
+// can be used in atomic operations.
+func sharedDataPointer(sharedData []byte) *uint32 {
+	return (*uint32)(unsafe.Pointer(&sharedData[0:4][0]))
+}
diff --git a/pkg/tcpip/link/sharedmem/tx.go b/pkg/tcpip/link/sharedmem/tx.go
new file mode 100755
index 000000000..6b8d7859d
--- /dev/null
+++ b/pkg/tcpip/link/sharedmem/tx.go
@@ -0,0 +1,272 @@
+// Copyright 2018 The gVisor Authors.
+//
+// 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.
+
+package sharedmem
+
+import (
+	"math"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/tcpip/link/sharedmem/queue"
+)
+
+const (
+	nilID = math.MaxUint64
+)
+
+// tx holds all state associated with a tx queue.
+type tx struct {
+	data []byte
+	q    queue.Tx
+	ids  idManager
+	bufs bufferManager
+}
+
+// init initializes all state needed by the tx queue based on the information
+// provided.
+//
+// The caller always retains ownership of all file descriptors passed in. The
+// queue implementation will duplicate any that it may need in the future.
+func (t *tx) init(mtu uint32, c *QueueConfig) error {
+	// Map in all buffers.
+	txPipe, err := getBuffer(c.TxPipeFD)
+	if err != nil {
+		return err
+	}
+
+	rxPipe, err := getBuffer(c.RxPipeFD)
+	if err != nil {
+		syscall.Munmap(txPipe)
+		return err
+	}
+
+	data, err := getBuffer(c.DataFD)
+	if err != nil {
+		syscall.Munmap(txPipe)
+		syscall.Munmap(rxPipe)
+		return err
+	}
+
+	// Initialize state based on buffers.
+	t.q.Init(txPipe, rxPipe)
+	t.ids.init()
+	t.bufs.init(0, len(data), int(mtu))
+	t.data = data
+
+	return nil
+}
+
+// cleanup releases all resources allocated during init(). It must only be
+// called if init() has previously succeeded.
+func (t *tx) cleanup() {
+	a, b := t.q.Bytes()
+	syscall.Munmap(a)
+	syscall.Munmap(b)
+	syscall.Munmap(t.data)
+}
+
+// transmit sends a packet made up of up to two buffers. Returns a boolean that
+// specifies whether the packet was successfully transmitted.
+func (t *tx) transmit(a, b []byte) bool {
+	// Pull completions from the tx queue and add their buffers back to the
+	// pool so that we can reuse them.
+	for {
+		id, ok := t.q.CompletedPacket()
+		if !ok {
+			break
+		}
+
+		if buf := t.ids.remove(id); buf != nil {
+			t.bufs.free(buf)
+		}
+	}
+
+	bSize := t.bufs.entrySize
+	total := uint32(len(a) + len(b))
+	bufCount := (total + bSize - 1) / bSize
+
+	// Allocate enough buffers to hold all the data.
+	var buf *queue.TxBuffer
+	for i := bufCount; i != 0; i-- {
+		b := t.bufs.alloc()
+		if b == nil {
+			// Failed to get all buffers. Return to the pool
+			// whatever we had managed to get.
+			if buf != nil {
+				t.bufs.free(buf)
+			}
+			return false
+		}
+		b.Next = buf
+		buf = b
+	}
+
+	// Copy data into allocated buffers.
+	nBuf := buf
+	var dBuf []byte
+	for _, data := range [][]byte{a, b} {
+		for len(data) > 0 {
+			if len(dBuf) == 0 {
+				dBuf = t.data[nBuf.Offset:][:nBuf.Size]
+				nBuf = nBuf.Next
+			}
+			n := copy(dBuf, data)
+			data = data[n:]
+			dBuf = dBuf[n:]
+		}
+	}
+
+	// Get an id for this packet and send it out.
+	id := t.ids.add(buf)
+	if !t.q.Enqueue(id, total, bufCount, buf) {
+		t.ids.remove(id)
+		t.bufs.free(buf)
+		return false
+	}
+
+	return true
+}
+
+// getBuffer returns a memory region mapped to the full contents of the given
+// file descriptor.
+func getBuffer(fd int) ([]byte, error) {
+	var s syscall.Stat_t
+	if err := syscall.Fstat(fd, &s); err != nil {
+		return nil, err
+	}
+
+	// Check that size doesn't overflow an int.
+	if s.Size > int64(^uint(0)>>1) {
+		return nil, syscall.EDOM
+	}
+
+	return syscall.Mmap(fd, 0, int(s.Size), syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_SHARED|syscall.MAP_FILE)
+}
+
+// idDescriptor is used by idManager to either point to a tx buffer (in case
+// the ID is assigned) or to the next free element (if the id is not assigned).
+type idDescriptor struct {
+	buf      *queue.TxBuffer
+	nextFree uint64
+}
+
+// idManager is a manager of tx buffer identifiers. It assigns unique IDs to
+// tx buffers that are added to it; the IDs can only be reused after they have
+// been removed.
+//
+// The ID assignments are stored so that the tx buffers can be retrieved from
+// the IDs previously assigned to them.
+type idManager struct {
+	// ids is a slice containing all tx buffers. The ID is the index into
+	// this slice.
+	ids []idDescriptor
+
+	// freeList a list of free IDs.
+	freeList uint64
+}
+
+// init initializes the id manager.
+func (m *idManager) init() {
+	m.freeList = nilID
+}
+
+// add assigns an ID to the given tx buffer.
+func (m *idManager) add(b *queue.TxBuffer) uint64 {
+	if i := m.freeList; i != nilID {
+		// There is an id available in the free list, just use it.
+		m.ids[i].buf = b
+		m.freeList = m.ids[i].nextFree
+		return i
+	}
+
+	// We need to expand the id descriptor.
+	m.ids = append(m.ids, idDescriptor{buf: b})
+	return uint64(len(m.ids) - 1)
+}
+
+// remove retrieves the tx buffer associated with the given ID, and removes the
+// ID from the assigned table so that it can be reused in the future.
+func (m *idManager) remove(i uint64) *queue.TxBuffer {
+	if i >= uint64(len(m.ids)) {
+		return nil
+	}
+
+	desc := &m.ids[i]
+	b := desc.buf
+	if b == nil {
+		// The provided id is not currently assigned.
+		return nil
+	}
+
+	desc.buf = nil
+	desc.nextFree = m.freeList
+	m.freeList = i
+
+	return b
+}
+
+// bufferManager manages a buffer region broken up into smaller, equally sized
+// buffers. Smaller buffers can be allocated and freed.
+type bufferManager struct {
+	freeList  *queue.TxBuffer
+	curOffset uint64
+	limit     uint64
+	entrySize uint32
+}
+
+// init initializes the buffer manager.
+func (b *bufferManager) init(initialOffset, size, entrySize int) {
+	b.freeList = nil
+	b.curOffset = uint64(initialOffset)
+	b.limit = uint64(initialOffset + size/entrySize*entrySize)
+	b.entrySize = uint32(entrySize)
+}
+
+// alloc allocates a buffer from the manager, if one is available.
+func (b *bufferManager) alloc() *queue.TxBuffer {
+	if b.freeList != nil {
+		// There is a descriptor ready for reuse in the free list.
+		d := b.freeList
+		b.freeList = d.Next
+		d.Next = nil
+		return d
+	}
+
+	if b.curOffset < b.limit {
+		// There is room available in the never-used range, so create
+		// a new descriptor for it.
+		d := &queue.TxBuffer{
+			Offset: b.curOffset,
+			Size:   b.entrySize,
+		}
+		b.curOffset += uint64(b.entrySize)
+		return d
+	}
+
+	return nil
+}
+
+// free returns all buffers in the list to the buffer manager so that they can
+// be reused.
+func (b *bufferManager) free(d *queue.TxBuffer) {
+	// Find the last buffer in the list.
+	last := d
+	for last.Next != nil {
+		last = last.Next
+	}
+
+	// Push list onto free list.
+	last.Next = b.freeList
+	b.freeList = d
+}