REBASE!

2024-12-22 10:58:57 +00:00 · 2013-12-15 11:43:31 +01:00 · 2013-12-15 11:43:31 +01:00 · bfe935b5ab
commit bfe935b5ab
26 changed files with 2866 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@ -0,0 +1 @@
+syncthing
--- a/19
+++ b/19
@ -0,0 +1,19 @@
+Copyright (C) 2013 Jakob Borg
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+
+- The above copyright notice and this permission notice shall be included in
+  all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
--- a/README.md
+++ b/README.md
@ -0,0 +1,161 @@
+syncthing
+=========
+
+This is `syncthing`, an open BitTorrent Sync alternative. It is
+currently far from ready for mass consumption, but it is a usable proof
+of concept and tech demo. The following are the project goals:
+
+ 1. Define an open, secure, language neutral protocol usable for
+    efficient synchronization of a file repository between an arbitrary
+    number of nodes. This is the [Block Exchange
+    Protocol](https://github.com/calmh/syncthing/blob/master/protocol/PROTOCOL.md)
+    (BEP).
+
+ 2. Provide the reference implementation to demonstrate the usability of
+    said protocol. This is the `syncthing` utility.
+
+The two are evolving together; the protocol is not to be considered
+stable until syncthing 1.0 is released, at which point it is locked down
+for incompatible changes.
+
+Syncthing does not use the BitTorrent protocol. The reasons for this are
+1) we don't know if BitTorrent Sync does either, so there's nothing to
+be compatible with, 2) BitTorrent includes a lot of functionality for
+making sure large swarms of selfish agents behave and somehow work
+towards a common goal. Here we have a much smaller swarm of cooperative
+agents and a simpler approach will suffice.
+
+Features
+--------
+
+The following features are _currently implemented and working_:
+
+ * The formation of a cluster of nodes, certificate authenticated and
+   communicating over TLS over TCP.
+
+ * Synchronization of a single directory among the cluster nodes.
+
+ * Change detection by periodic scanning of the local repository.
+
+ * Static configuration of cluster nodes.
+
+ * Automatic discovery of cluster nodes on the local network. See
+   [discover.go](https://github.com/calmh/syncthing/blob/master/discover/discover.go)
+   for the protocol specification.
+
+ * Handling of deleted files. Deletes can be propagated or ignored per
+   client.
+
+The following features are _not yet implemented but planned_:
+
+ * Syncing multiple directories from the same syncthing instance.
+
+ * Change detection by listening to file system notifications instead of
+   periodic scanning.
+
+ * HTTP GUI.
+
+The following features are _not implemented but may be implemented_ in
+the future:
+
+ * Automatic remote node discovery using a DHT. This is not technically
+   very difficult but requires one or more globally reachable root
+   nodes. This is open for discussion -- perhaps we can piggyback on an
+   existing DHT, or root nodes need to be established in some other
+   manner.
+
+ * Automatic NAT handling via UPNP. Required for the above, not very
+   useful without it.
+
+ * Conflict resolution. Currently whichever file has the newest
+   modification time "wins". The correct behavior in the face of
+   conflicts is open for discussion.
+
+Security
+--------
+
+Security is one of the primary project goals. This means that it should
+not be possible for an attacker to join a cluster uninvited, and it
+should not be possible to extract private information from intercepted
+traffic. Currently this is implemented as follows.
+
+All traffic is protected by TLS. To prevent uninvited nodes from joining
+a cluster, the certificate fingerprint of each node is compared to a
+preset list of acceptable nodes at connection establishment. The
+fingerprint is computed as the SHA-1 hash of the certificate and
+displayed in BASE32 encoding to form a compact yet convenient string.
+Currently SHA-1 is deemed secure against preimage attacks.
+
+Usage
+=====
+
+`go get github.com/calmh/syncthing`
+
+Check out the options:
+
+```
+$ syncthing --help
+Usage:
+  syncthing [options]
+
+...
+```
+
+Run syncthing to let it create it's config directory and certificate:
+
+```
+$ syncthing
+11:34:13 tls.go:61: OK: wrote cert.pem
+11:34:13 tls.go:67: OK: wrote key.pem
+11:34:13 main.go:66: INFO: My ID: NCTBZAAHXR6ZZP3D7SL3DLYFFQERMW4Q
+11:34:13 main.go:90: FATAL: No config file
+```
+
+Take note of the "My ID: ..." line. Perform the same operation on
+another computer (or the same computer but with a different `--home` for
+testing) to create another node. Take note of that ID as well, and
+create a config file `~/.syncthing/syncthing.ini` looking something like
+this:
+
+```
+[repository]
+dir = /Users/jb/Synced
+
+[nodes]
+NCTBZAAHXR6ZZP3D7SL3DLYFFQERMW4Q = 172.16.32.1:22000 192.23.34.56:22000
+CUGAE43Y5N64CRJU26YFH6MTWPSBLSUL = dynamic
+```
+
+This assumes that the first node is reachable on either of the two
+addresses listed (perhaps one internal and one port-forwarded external)
+and that the other node is not normally reachable from the outside. Save
+this config file, identically, to both nodes. If both nodes are running
+on the same network, you can set all addresses to 'dynamic' and they
+will find each other by local node discovery.
+
+Start syncthing on both nodes. If you're running both on the same
+computer, one needs a different repository directory (in the config
+file) and listening port (set as a command line paramter). For the
+cautious, one side can be set to be read only.
+
+```
+$ syncthing --ro
+13:30:55 main.go:102: INFO: My ID: NCTBZAAHXR6ZZP3D7SL3DLYFFQERMW4Q
+13:30:55 main.go:149: INFO: Initial repository scan in progress
+13:30:59 main.go:153: INFO: Listening for incoming connections
+13:30:59 main.go:157: INFO: Attempting to connect to other nodes
+13:30:59 main.go:247: INFO: Starting local discovery
+13:30:59 main.go:165: OK: Ready to synchronize
+13:31:04 discover.go:113: INFO: Discovered node CUGAE43Y5N64CRJU26YFH6MTWPSBLSUL at 172.16.32.24:23456
+13:31:14 main.go:296: OK: Connected to node CUGAE43Y5N64CRJU26YFH6MTWPSBLSUL
+13:31:19 main.go:345: INFO: Transferred 139 KiB in (14 KiB/s), 139 KiB out (14 KiB/s)
+...
+```
+You should see the synchronization start and then finish a short while
+later. Add nodes to taste.
+
+License
+=======
+
+MIT
+
--- a/blocks.go
+++ b/blocks.go
@ -0,0 +1,67 @@
+package main
+
+import (
+	"bytes"
+	"crypto/sha256"
+	"io"
+)
+
+type BlockList []Block
+
+type Block struct {
+	Offset uint64
+	Length uint32
+	Hash   []byte
+}
+
+// Blocks returns the blockwise hash of the reader.
+func Blocks(r io.Reader, blocksize int) (BlockList, error) {
+	var blocks BlockList
+	var offset uint64
+	for {
+		lr := &io.LimitedReader{r, int64(blocksize)}
+		hf := sha256.New()
+		n, err := io.Copy(hf, lr)
+		if err != nil {
+			return nil, err
+		}
+
+		if n == 0 {
+			break
+		}
+
+		b := Block{
+			Offset: offset,
+			Length: uint32(n),
+			Hash:   hf.Sum(nil),
+		}
+		blocks = append(blocks, b)
+		offset += uint64(n)
+	}
+
+	return blocks, nil
+}
+
+// To returns the list of blocks necessary to transform src into dst.
+// Both block lists must have been created with the same block size.
+func (src BlockList) To(tgt BlockList) (have, need BlockList) {
+	if len(tgt) == 0 && len(src) != 0 {
+		return nil, nil
+	}
+
+	if len(tgt) != 0 && len(src) == 0 {
+		// Copy the entire file
+		return nil, tgt
+	}
+
+	for i := range tgt {
+		if i >= len(src) || bytes.Compare(tgt[i].Hash, src[i].Hash) != 0 {
+			// Copy differing block
+			need = append(need, tgt[i])
+		} else {
+			have = append(have, tgt[i])
+		}
+	}
+
+	return have, need
+}
--- a/blocks_test.go
+++ b/blocks_test.go
@ -0,0 +1,115 @@
+package main
+
+import (
+	"bytes"
+	"fmt"
+	"testing"
+)
+
+var blocksTestData = []struct {
+	data      []byte
+	blocksize int
+	hash      []string
+}{
+	{[]byte(""), 1024, []string{}},
+	{[]byte("contents"), 1024, []string{
+		"d1b2a59fbea7e20077af9f91b27e95e865061b270be03ff539ab3b73587882e8"}},
+	{[]byte("contents"), 9, []string{
+		"d1b2a59fbea7e20077af9f91b27e95e865061b270be03ff539ab3b73587882e8"}},
+	{[]byte("contents"), 8, []string{
+		"d1b2a59fbea7e20077af9f91b27e95e865061b270be03ff539ab3b73587882e8"}},
+	{[]byte("contents"), 7, []string{
+		"ed7002b439e9ac845f22357d822bac1444730fbdb6016d3ec9432297b9ec9f73",
+		"043a718774c572bd8a25adbeb1bfcd5c0256ae11cecf9f9c3f925d0e52beaf89"},
+	},
+	{[]byte("contents"), 3, []string{
+		"1143da2bc54c495c4be31d3868785d39ffdfd56df5668f0645d8f14d47647952",
+		"e4432baa90819aaef51d2a7f8e148bf7e679610f3173752fabb4dcb2d0f418d3",
+		"44ad63f60af0f6db6fdde6d5186ef78176367df261fa06be3079b6c80c8adba4"},
+	},
+	{[]byte("conconts"), 3, []string{
+		"1143da2bc54c495c4be31d3868785d39ffdfd56df5668f0645d8f14d47647952",
+		"1143da2bc54c495c4be31d3868785d39ffdfd56df5668f0645d8f14d47647952",
+		"44ad63f60af0f6db6fdde6d5186ef78176367df261fa06be3079b6c80c8adba4"},
+	},
+	{[]byte("contenten"), 3, []string{
+		"1143da2bc54c495c4be31d3868785d39ffdfd56df5668f0645d8f14d47647952",
+		"e4432baa90819aaef51d2a7f8e148bf7e679610f3173752fabb4dcb2d0f418d3",
+		"e4432baa90819aaef51d2a7f8e148bf7e679610f3173752fabb4dcb2d0f418d3"},
+	},
+}
+
+func TestBlocks(t *testing.T) {
+	for _, test := range blocksTestData {
+		buf := bytes.NewBuffer(test.data)
+		blocks, err := Blocks(buf, test.blocksize)
+
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		if l := len(blocks); l != len(test.hash) {
+			t.Fatalf("Incorrect number of blocks %d != %d", l, len(test.hash))
+		} else {
+			i := 0
+			for off := uint64(0); off < uint64(len(test.data)); off += uint64(test.blocksize) {
+				if blocks[i].Offset != off {
+					t.Errorf("Incorrect offset for block %d: %d != %d", i, blocks[i].Offset, off)
+				}
+
+				bs := test.blocksize
+				if rem := len(test.data) - int(off); bs > rem {
+					bs = rem
+				}
+				if int(blocks[i].Length) != bs {
+					t.Errorf("Incorrect length for block %d: %d != %d", i, blocks[i].Length, bs)
+				}
+				if h := fmt.Sprintf("%x", blocks[i].Hash); h != test.hash[i] {
+					t.Errorf("Incorrect block hash %q != %q", h, test.hash[i])
+				}
+
+				i++
+			}
+		}
+	}
+}
+
+var diffTestData = []struct {
+	a string
+	b string
+	s int
+	d []Block
+}{
+	{"contents", "contents", 1024, []Block{}},
+	{"", "", 1024, []Block{}},
+	{"contents", "contents", 3, []Block{}},
+	{"contents", "cantents", 3, []Block{{0, 3, nil}}},
+	{"contents", "contants", 3, []Block{{3, 3, nil}}},
+	{"contents", "cantants", 3, []Block{{0, 3, nil}, {3, 3, nil}}},
+	{"contents", "", 3, nil},
+	{"", "contents", 3, []Block{{0, 3, nil}, {3, 3, nil}, {6, 2, nil}}},
+	{"con", "contents", 3, []Block{{3, 3, nil}, {6, 2, nil}}},
+	{"contents", "con", 3, nil},
+	{"contents", "cont", 3, []Block{{3, 1, nil}}},
+	{"cont", "contents", 3, []Block{{3, 3, nil}, {6, 2, nil}}},
+}
+
+func TestDiff(t *testing.T) {
+	for i, test := range diffTestData {
+		a, _ := Blocks(bytes.NewBufferString(test.a), test.s)
+		b, _ := Blocks(bytes.NewBufferString(test.b), test.s)
+		_, d := a.To(b)
+		if len(d) != len(test.d) {
+			t.Fatalf("Incorrect length for diff %d; %d != %d", i, len(d), len(test.d))
+		} else {
+			for j := range test.d {
+				if d[j].Offset != test.d[j].Offset {
+					t.Errorf("Incorrect offset for diff %d block %d; %d != %d", i, j, d[j].Offset, test.d[j].Offset)
+				}
+				if d[j].Length != test.d[j].Length {
+					t.Errorf("Incorrect length for diff %d block %d; %d != %d", i, j, d[j].Length, test.d[j].Length)
+				}
+			}
+		}
+	}
+}
--- a/buffers/buffers.go
+++ b/buffers/buffers.go
@ -0,0 +1,26 @@
+package buffers
+
+var buffers = make(chan []byte, 32)
+
+func Get(size int) []byte {
+	var buf []byte
+	select {
+	case buf = <-buffers:
+	default:
+	}
+	if len(buf) < size {
+		return make([]byte, size)
+	}
+	return buf[:size]
+}
+
+func Put(buf []byte) {
+	if cap(buf) == 0 {
+		return
+	}
+	buf = buf[:cap(buf)]
+	select {
+	case buffers <- buf:
+	default:
+	}
+}
--- a/discover/discover.go
+++ b/discover/discover.go
@ -0,0 +1,121 @@
+/*
+This is the local node discovery protocol. In principle we might be better
+served by something more standardized, such as mDNS / DNS-SD. In practice, this
+was much easier and quicker to get up and running.
+
+The mode of operation is to periodically (currently once every 30 seconds)
+transmit a broadcast UDP packet to the well known port number 21025. The packet
+has the following format:
+
+     0                   1                   2                   3
+     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+    |                         Magic Number                          |
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+    |          Port Number          |        Length of NodeID       |
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+    /                                                               /
+    \                   NodeID (variable length)                    \
+    /                                                               /
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+
+The sending node's address is not encoded -- it is taken to be the source
+address of the announcement. Every time such a packet is received, a local
+table that maps NodeID to Address is updated. When the local node wants to
+connect to another node with the address specification 'dynamic', this table is
+consulted.
+*/
+package discover
+
+import (
+	"encoding/binary"
+	"fmt"
+	"net"
+	"sync"
+	"time"
+)
+
+type Discoverer struct {
+	MyID          string
+	ListenPort    int
+	BroadcastIntv time.Duration
+
+	conn         *net.UDPConn
+	registry     map[string]string
+	registryLock sync.RWMutex
+}
+
+func NewDiscoverer(id string, port int) (*Discoverer, error) {
+	local4 := &net.UDPAddr{IP: net.IP{0, 0, 0, 0}, Port: 21025}
+	conn, err := net.ListenUDP("udp4", local4)
+	if err != nil {
+		return nil, err
+	}
+
+	disc := &Discoverer{
+		MyID:          id,
+		ListenPort:    port,
+		BroadcastIntv: 30 * time.Second,
+		conn:          conn,
+		registry:      make(map[string]string),
+	}
+
+	go disc.sendAnnouncements()
+	go disc.recvAnnouncements()
+
+	return disc, nil
+}
+
+func (d *Discoverer) sendAnnouncements() {
+	remote4 := &net.UDPAddr{IP: net.IP{255, 255, 255, 255}, Port: 21025}
+
+	idbs := []byte(d.MyID)
+	buf := make([]byte, 4+4+4+len(idbs))
+
+	binary.BigEndian.PutUint32(buf, uint32(0x121025))
+	binary.BigEndian.PutUint16(buf[4:], uint16(d.ListenPort))
+	binary.BigEndian.PutUint16(buf[6:], uint16(len(idbs)))
+	copy(buf[8:], idbs)
+
+	for {
+		_, _, err := d.conn.WriteMsgUDP(buf, nil, remote4)
+		if err != nil {
+			panic(err)
+		}
+		time.Sleep(d.BroadcastIntv)
+	}
+}
+
+func (d *Discoverer) recvAnnouncements() {
+	var buf = make([]byte, 1024)
+	for {
+		_, addr, err := d.conn.ReadFromUDP(buf)
+		if err != nil {
+			panic(err)
+		}
+		magic := binary.BigEndian.Uint32(buf)
+		if magic != 0x121025 {
+			continue
+		}
+		port := binary.BigEndian.Uint16(buf[4:])
+		l := binary.BigEndian.Uint16(buf[6:])
+		idbs := buf[8 : l+8]
+		id := string(idbs)
+
+		if id != d.MyID {
+			nodeAddr := fmt.Sprintf("%s:%d", addr.IP.String(), port)
+			d.registryLock.Lock()
+			if d.registry[id] != nodeAddr {
+				d.registry[id] = nodeAddr
+			}
+			d.registryLock.Unlock()
+		}
+	}
+}
+
+func (d *Discoverer) Lookup(node string) (string, bool) {
+	d.registryLock.Lock()
+	defer d.registryLock.Unlock()
+	addr, ok := d.registry[node]
+	return addr, ok
+}
--- a/logger.go
+++ b/logger.go
@ -0,0 +1,72 @@
+package main
+
+import (
+	"fmt"
+	"log"
+	"os"
+)
+
+var debugEnabled = true
+var logger = log.New(os.Stderr, "", log.Lshortfile|log.Ltime)
+
+func debugln(vals ...interface{}) {
+	if debugEnabled {
+		s := fmt.Sprintln(vals...)
+		logger.Output(2, "DEBUG: "+s)
+	}
+}
+
+func debugf(format string, vals ...interface{}) {
+	if debugEnabled {
+		s := fmt.Sprintf(format, vals...)
+		logger.Output(2, "DEBUG: "+s)
+	}
+}
+
+func infoln(vals ...interface{}) {
+	s := fmt.Sprintln(vals...)
+	logger.Output(2, "INFO: "+s)
+}
+
+func infof(format string, vals ...interface{}) {
+	s := fmt.Sprintf(format, vals...)
+	logger.Output(2, "INFO: "+s)
+}
+
+func okln(vals ...interface{}) {
+	s := fmt.Sprintln(vals...)
+	logger.Output(2, "OK: "+s)
+}
+
+func okf(format string, vals ...interface{}) {
+	s := fmt.Sprintf(format, vals...)
+	logger.Output(2, "OK: "+s)
+}
+
+func warnln(vals ...interface{}) {
+	s := fmt.Sprintln(vals...)
+	logger.Output(2, "WARNING: "+s)
+}
+
+func warnf(format string, vals ...interface{}) {
+	s := fmt.Sprintf(format, vals...)
+	logger.Output(2, "WARNING: "+s)
+}
+
+func fatalln(vals ...interface{}) {
+	s := fmt.Sprintln(vals...)
+	logger.Output(2, "FATAL: "+s)
+	os.Exit(3)
+}
+
+func fatalf(format string, vals ...interface{}) {
+	s := fmt.Sprintf(format, vals...)
+	logger.Output(2, "FATAL: "+s)
+	os.Exit(3)
+}
+
+func fatalErr(err error) {
+	if err != nil {
+		fatalf(err.Error())
+	}
+}
--- a/main.go
+++ b/main.go
@ -0,0 +1,352 @@
+package main
+
+import (
+	"crypto/sha1"
+	"crypto/tls"
+	"fmt"
+	"log"
+	"net"
+	"net/http"
+	_ "net/http/pprof"
+	"os"
+	"path"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/calmh/ini"
+	"github.com/calmh/syncthing/discover"
+	"github.com/calmh/syncthing/protocol"
+	docopt "github.com/docopt/docopt.go"
+)
+
+const (
+	confDirName  = ".syncthing"
+	confFileName = "syncthing.ini"
+	usage        = `Usage:
+  syncthing [options]
+
+Options:
+  -l <addr>        Listening address [default: :22000]
+  -p <addr>        Enable HTTP profiler on addr
+  --home <path>    Home directory
+  --delete         Delete files that were deleted on a peer node
+  --ro             Local repository is read only
+  --scan-intv <s>  Repository scan interval, in seconds [default: 60]
+  --conn-intv <s>  Node reconnect interval, in seconds [default: 15]
+  --no-stats       Don't print transfer statistics
+
+Help Options:
+  -h, --help       Show this help
+  --version        Show version
+
+Debug Options:
+  --trace-file     Trace file operations
+  --trace-net      Trace network operations
+  --trace-idx      Trace sent indexes
+`
+)
+
+var (
+	config    ini.Config
+	nodeAddrs = make(map[string][]string)
+)
+
+// Options
+var (
+	confDir    = path.Join(getHomeDir(), confDirName)
+	addr       string
+	prof       string
+	readOnly   bool
+	scanIntv   int
+	connIntv   int
+	traceNet   bool
+	traceFile  bool
+	traceIdx   bool
+	printStats bool
+	doDelete   bool
+)
+
+func main() {
+	// Useful for debugging; to be adjusted.
+	log.SetFlags(log.Ltime | log.Lshortfile)
+
+	arguments, _ := docopt.Parse(usage, nil, true, "syncthing 0.1", false)
+
+	addr = arguments["-l"].(string)
+	prof, _ = arguments["-p"].(string)
+	readOnly, _ = arguments["--ro"].(bool)
+
+	if arguments["--home"] != nil {
+		confDir, _ = arguments["--home"].(string)
+	}
+
+	scanIntv, _ = strconv.Atoi(arguments["--scan-intv"].(string))
+	if scanIntv == 0 {
+		fatalln("Invalid --scan-intv")
+	}
+
+	connIntv, _ = strconv.Atoi(arguments["--conn-intv"].(string))
+	if connIntv == 0 {
+		fatalln("Invalid --conn-intv")
+	}
+
+	doDelete = arguments["--delete"].(bool)
+	traceFile = arguments["--trace-file"].(bool)
+	traceNet = arguments["--trace-net"].(bool)
+	traceIdx = arguments["--trace-idx"].(bool)
+	printStats = !arguments["--no-stats"].(bool)
+
+	// Ensure that our home directory exists and that we have a certificate and key.
+
+	ensureDir(confDir)
+	cert, err := loadCert(confDir)
+	if err != nil {
+		newCertificate(confDir)
+		cert, err = loadCert(confDir)
+		fatalErr(err)
+	}
+
+	myID := string(certId(cert.Certificate[0]))
+	infoln("My ID:", myID)
+
+	if prof != "" {
+		okln("Profiler listening on", prof)
+		go func() {
+			http.ListenAndServe(prof, nil)
+		}()
+	}
+
+	// The TLS configuration is used for both the listening socket and outgoing
+	// connections.
+
+	cfg := &tls.Config{
+		ClientAuth:         tls.RequestClientCert,
+		ServerName:         "syncthing",
+		NextProtos:         []string{"bep/1.0"},
+		InsecureSkipVerify: true,
+		Certificates:       []tls.Certificate{cert},
+	}
+
+	// Load the configuration file, if it exists.
+
+	cf, err := os.Open(path.Join(confDir, confFileName))
+	if err != nil {
+		fatalln("No config file")
+		config = ini.Config{}
+	}
+	config = ini.Parse(cf)
+	cf.Close()
+
+	var dir = config.Get("repository", "dir")
+
+	// Create a map of desired node connections based on the configuration file
+	// directives.
+
+	for nodeID, addrs := range config.OptionMap("nodes") {
+		addrs := strings.Fields(addrs)
+		nodeAddrs[nodeID] = addrs
+	}
+
+	m := NewModel(dir)
+
+	// Walk the repository and update the local model before establishing any
+	// connections to other nodes.
+
+	infoln("Iniial repository scan in progress")
+	loadIndex(m)
+	updateLocalModel(m)
+
+	// Routine to listen for incoming connections
+	infoln("Listening for incoming connections")
+	go listen(myID, addr, m, cfg)
+
+	// Routine to connect out to configured nodes
+	infoln("Attempting to connect to other nodes")
+	go connect(myID, addr, nodeAddrs, m, cfg)
+
+	// Routine to pull blocks from other nodes to synchronize the local
+	// repository. Does not run when we are in read only (publish only) mode.
+	if !readOnly {
+		infoln("Cleaning out incomplete synchronizations")
+		CleanTempFiles(dir)
+		okln("Ready to synchronize")
+		m.Start()
+	}
+
+	// Periodically scan the repository and update the local model.
+	// XXX: Should use some fsnotify mechanism.
+	go func() {
+		for {
+			time.Sleep(time.Duration(scanIntv) * time.Second)
+			updateLocalModel(m)
+		}
+	}()
+
+	select {}
+}
+
+func listen(myID string, addr string, m *Model, cfg *tls.Config) {
+	l, err := tls.Listen("tcp", addr, cfg)
+	fatalErr(err)
+
+listen:
+	for {
+		conn, err := l.Accept()
+		if err != nil {
+			warnln(err)
+			continue
+		}
+
+		if traceNet {
+			debugln("NET: Connect from", conn.RemoteAddr())
+		}
+
+		tc := conn.(*tls.Conn)
+		err = tc.Handshake()
+		if err != nil {
+			warnln(err)
+			tc.Close()
+			continue
+		}
+
+		remoteID := certId(tc.ConnectionState().PeerCertificates[0].Raw)
+
+		if remoteID == myID {
+			warnf("Connect from myself (%s) - should not happen", remoteID)
+			conn.Close()
+			continue
+		}
+
+		if m.ConnectedTo(remoteID) {
+			warnf("Connect from connected node (%s)", remoteID)
+		}
+
+		for nodeID := range nodeAddrs {
+			if nodeID == remoteID {
+				nc := protocol.NewConnection(remoteID, conn, conn, m)
+				m.AddNode(nc)
+				okln("Connected to nodeID", remoteID, "(in)")
+				continue listen
+			}
+		}
+
+		warnln("Connect from unknown node", remoteID)
+		conn.Close()
+	}
+}
+
+func connect(myID string, addr string, nodeAddrs map[string][]string, m *Model, cfg *tls.Config) {
+	_, portstr, err := net.SplitHostPort(addr)
+	fatalErr(err)
+	port, _ := strconv.Atoi(portstr)
+
+	infoln("Starting local discovery")
+	disc, err := discover.NewDiscoverer(myID, port)
+	if err != nil {
+		warnln("No local discovery possible")
+	}
+
+	for {
+	nextNode:
+		for nodeID, addrs := range nodeAddrs {
+			if nodeID == myID {
+				continue
+			}
+			if m.ConnectedTo(nodeID) {
+				continue
+			}
+			for _, addr := range addrs {
+				if addr == "dynamic" {
+					var ok bool
+					if disc != nil {
+						addr, ok = disc.Lookup(nodeID)
+					}
+					if !ok {
+						continue
+					}
+				}
+
+				if traceNet {
+					debugln("NET: Dial", nodeID, addr)
+				}
+				conn, err := tls.Dial("tcp", addr, cfg)
+				if err != nil {
+					if traceNet {
+						debugln("NET:", err)
+					}
+					continue
+				}
+
+				remoteID := certId(conn.ConnectionState().PeerCertificates[0].Raw)
+				if remoteID != nodeID {
+					warnln("Unexpected nodeID", remoteID, "!=", nodeID)
+					conn.Close()
+					continue
+				}
+
+				nc := protocol.NewConnection(nodeID, conn, conn, m)
+				okln("Connected to node", remoteID, "(out)")
+				m.AddNode(nc)
+				if traceNet {
+					t0 := time.Now()
+					nc.Ping()
+					timing("NET: Ping reply", t0)
+				}
+				continue nextNode
+			}
+		}
+
+		time.Sleep(time.Duration(connIntv) * time.Second)
+	}
+}
+
+func updateLocalModel(m *Model) {
+	files := Walk(m.Dir(), m)
+	m.ReplaceLocal(files)
+	saveIndex(m)
+}
+
+func saveIndex(m *Model) {
+	fname := fmt.Sprintf("%x.idx", sha1.Sum([]byte(m.Dir())))
+	idxf, err := os.Create(path.Join(confDir, fname))
+	if err != nil {
+		return
+	}
+	protocol.WriteIndex(idxf, m.ProtocolIndex())
+	idxf.Close()
+}
+
+func loadIndex(m *Model) {
+	fname := fmt.Sprintf("%x.idx", sha1.Sum([]byte(m.Dir())))
+	idxf, err := os.Open(path.Join(confDir, fname))
+	if err != nil {
+		return
+	}
+	defer idxf.Close()
+
+	idx, err := protocol.ReadIndex(idxf)
+	if err != nil {
+		return
+	}
+	m.SeedIndex(idx)
+}
+
+func ensureDir(dir string) {
+	fi, err := os.Stat(dir)
+	if os.IsNotExist(err) {
+		err := os.MkdirAll(dir, 0700)
+		fatalErr(err)
+	} else if fi.Mode()&0077 != 0 {
+		err := os.Chmod(dir, 0700)
+		fatalErr(err)
+	}
+}
+
+func getHomeDir() string {
+	home := os.Getenv("HOME")
+	if home == "" {
+		fatalln("No home directory?")
+	}
+	return home
+}
--- a/model.go
+++ b/model.go
@ -0,0 +1,373 @@
+package main
+
+/*
+
+Locking
+=======
+
+The model has read and write locks. These must be acquired as appropriate by
+public methods. To prevent deadlock situations, private methods should never
+acquire locks, but document what locks they require.
+
+TODO(jb): Keep global and per node transfer and performance statistics.
+
+*/
+
+import (
+	"os"
+	"path"
+	"sync"
+	"time"
+
+	"github.com/calmh/syncthing/buffers"
+	"github.com/calmh/syncthing/protocol"
+)
+
+type Model struct {
+	sync.RWMutex
+	dir     string
+	updated int64
+	global  map[string]File // the latest version of each file as it exists in the cluster
+	local   map[string]File // the files we currently have locally on disk
+	remote  map[string]map[string]File
+	need    map[string]bool // the files we need to update
+	nodes   map[string]*protocol.Connection
+}
+
+const (
+	RemoteFetchers = 4
+	FlagDeleted    = 1 << 12
+)
+
+func NewModel(dir string) *Model {
+	m := &Model{
+		dir:    dir,
+		global: make(map[string]File),
+		local:  make(map[string]File),
+		remote: make(map[string]map[string]File),
+		need:   make(map[string]bool),
+		nodes:  make(map[string]*protocol.Connection),
+	}
+
+	return m
+}
+
+func (m *Model) Start() {
+	go m.puller()
+}
+
+func (m *Model) Index(nodeID string, fs []protocol.FileInfo) {
+	m.Lock()
+	defer m.Unlock()
+
+	if traceNet {
+		debugf("NET IDX(in): %s: %d files", nodeID, len(fs))
+	}
+
+	m.remote[nodeID] = make(map[string]File)
+	for _, f := range fs {
+		if f.Flags&FlagDeleted != 0 && !doDelete {
+			// Files marked as deleted do not even enter the model
+			continue
+		}
+		mf := File{
+			Name:     f.Name,
+			Flags:    f.Flags,
+			Modified: int64(f.Modified),
+		}
+		var offset uint64
+		for _, b := range f.Blocks {
+			mf.Blocks = append(mf.Blocks, Block{
+				Offset: offset,
+				Length: b.Length,
+				Hash:   b.Hash,
+			})
+			offset += uint64(b.Length)
+		}
+		m.remote[nodeID][f.Name] = mf
+	}
+
+	m.recomputeGlobal()
+	m.recomputeNeed()
+}
+
+func (m *Model) SeedIndex(fs []protocol.FileInfo) {
+	m.Lock()
+	defer m.Unlock()
+
+	m.local = make(map[string]File)
+	for _, f := range fs {
+		mf := File{
+			Name:     f.Name,
+			Flags:    f.Flags,
+			Modified: int64(f.Modified),
+		}
+		var offset uint64
+		for _, b := range f.Blocks {
+			mf.Blocks = append(mf.Blocks, Block{
+				Offset: offset,
+				Length: b.Length,
+				Hash:   b.Hash,
+			})
+			offset += uint64(b.Length)
+		}
+		m.local[f.Name] = mf
+	}
+
+	m.recomputeGlobal()
+	m.recomputeNeed()
+}
+
+func (m *Model) Close(node string) {
+	m.Lock()
+	defer m.Unlock()
+
+	if traceNet {
+		debugf("NET CLOSE: %s", node)
+	}
+
+	delete(m.remote, node)
+	delete(m.nodes, node)
+
+	m.recomputeGlobal()
+	m.recomputeNeed()
+}
+
+func (m *Model) Request(nodeID, name string, offset uint64, size uint32, hash []byte) ([]byte, error) {
+	if traceNet && nodeID != "<local>" {
+		debugf("NET REQ(in): %s: %q o=%d s=%d h=%x", nodeID, name, offset, size, hash)
+	}
+	fn := path.Join(m.dir, name)
+	fd, err := os.Open(fn) // XXX: Inefficient, should cache fd?
+	if err != nil {
+		return nil, err
+	}
+	defer fd.Close()
+
+	buf := buffers.Get(int(size))
+	_, err = fd.ReadAt(buf, int64(offset))
+	if err != nil {
+		return nil, err
+	}
+
+	return buf, nil
+}
+
+func (m *Model) RequestGlobal(nodeID, name string, offset uint64, size uint32, hash []byte) ([]byte, error) {
+	m.RLock()
+	nc := m.nodes[nodeID]
+	m.RUnlock()
+
+	if traceNet {
+		debugf("NET REQ(out): %s: %q o=%d s=%d h=%x", nodeID, name, offset, size, hash)
+	}
+
+	return nc.Request(name, offset, size, hash)
+}
+
+func (m *Model) ReplaceLocal(fs []File) {
+	m.Lock()
+	defer m.Unlock()
+
+	var updated bool
+	var newLocal = make(map[string]File)
+
+	for _, f := range fs {
+		newLocal[f.Name] = f
+		if ef := m.local[f.Name]; ef.Modified != f.Modified {
+			updated = true
+		}
+	}
+
+	if m.markDeletedLocals(newLocal) {
+		updated = true
+	}
+
+	if len(newLocal) != len(m.local) {
+		updated = true
+	}
+
+	if updated {
+		m.local = newLocal
+		m.recomputeGlobal()
+		m.recomputeNeed()
+		m.updated = time.Now().Unix()
+		go m.broadcastIndex()
+	}
+}
+
+// Must be called with the read lock held.
+func (m *Model) broadcastIndex() {
+	idx := m.protocolIndex()
+	for _, node := range m.nodes {
+		if traceNet {
+			debugf("NET IDX(out): %s: %d files", node.ID, len(idx))
+		}
+		node.Index(idx)
+	}
+}
+
+// markDeletedLocals sets the deleted flag on files that have gone missing locally.
+// Must be called with the write lock held.
+func (m *Model) markDeletedLocals(newLocal map[string]File) bool {
+	// For every file in the existing local table, check if they are also
+	// present in the new local table. If they are not, check that we already
+	// had the newest version available according to the global table and if so
+	// note the file as having been deleted.
+	var updated bool
+	for n, f := range m.local {
+		if _, ok := newLocal[n]; !ok {
+			if gf := m.global[n]; gf.Modified <= f.Modified {
+				if f.Flags&FlagDeleted == 0 {
+					f.Flags = FlagDeleted
+					f.Modified = f.Modified + 1
+					f.Blocks = nil
+					updated = true
+				}
+				newLocal[n] = f
+			}
+		}
+	}
+	return updated
+}
+
+func (m *Model) UpdateLocal(f File) {
+	m.Lock()
+	defer m.Unlock()
+
+	if ef, ok := m.local[f.Name]; !ok || ef.Modified != f.Modified {
+		m.local[f.Name] = f
+		m.recomputeGlobal()
+		m.recomputeNeed()
+		m.updated = time.Now().Unix()
+		go m.broadcastIndex()
+	}
+}
+
+func (m *Model) Dir() string {
+	m.RLock()
+	defer m.RUnlock()
+	return m.dir
+}
+
+func (m *Model) HaveFiles() []File {
+	m.RLock()
+	defer m.RUnlock()
+	var files []File
+	for _, file := range m.local {
+		files = append(files, file)
+	}
+	return files
+}
+
+func (m *Model) LocalFile(name string) (File, bool) {
+	m.RLock()
+	defer m.RUnlock()
+	f, ok := m.local[name]
+	return f, ok
+}
+
+func (m *Model) GlobalFile(name string) (File, bool) {
+	m.RLock()
+	defer m.RUnlock()
+	f, ok := m.global[name]
+	return f, ok
+}
+
+// Must be called with the write lock held.
+func (m *Model) recomputeGlobal() {
+	var newGlobal = make(map[string]File)
+
+	for n, f := range m.local {
+		newGlobal[n] = f
+	}
+
+	for _, fs := range m.remote {
+		for n, f := range fs {
+			if cf, ok := newGlobal[n]; !ok || cf.Modified < f.Modified {
+				newGlobal[n] = f
+			}
+		}
+	}
+
+	m.global = newGlobal
+}
+
+// Must be called with the write lock held.
+func (m *Model) recomputeNeed() {
+	m.need = make(map[string]bool)
+	for n, f := range m.global {
+		hf, ok := m.local[n]
+		if !ok || f.Modified > hf.Modified {
+			m.need[n] = true
+		}
+	}
+}
+
+// Must be called with the read lock held.
+func (m *Model) whoHas(name string) []string {
+	var remote []string
+
+	gf := m.global[name]
+	for node, files := range m.remote {
+		if file, ok := files[name]; ok && file.Modified == gf.Modified {
+			remote = append(remote, node)
+		}
+	}
+
+	return remote
+}
+
+func (m *Model) ConnectedTo(nodeID string) bool {
+	m.RLock()
+	defer m.RUnlock()
+	_, ok := m.nodes[nodeID]
+	return ok
+}
+
+func (m *Model) ProtocolIndex() []protocol.FileInfo {
+	m.RLock()
+	defer m.RUnlock()
+	return m.protocolIndex()
+}
+
+// Must be called with the read lock held.
+func (m *Model) protocolIndex() []protocol.FileInfo {
+	var index []protocol.FileInfo
+	for _, f := range m.local {
+		mf := protocol.FileInfo{
+			Name:     f.Name,
+			Flags:    f.Flags,
+			Modified: int64(f.Modified),
+		}
+		for _, b := range f.Blocks {
+			mf.Blocks = append(mf.Blocks, protocol.BlockInfo{
+				Length: b.Length,
+				Hash:   b.Hash,
+			})
+		}
+		if traceIdx {
+			var flagComment string
+			if mf.Flags&FlagDeleted != 0 {
+				flagComment = " (deleted)"
+			}
+			debugf("IDX: %q m=%d f=%o%s (%d blocks)", mf.Name, mf.Modified, mf.Flags, flagComment, len(mf.Blocks))
+		}
+		index = append(index, mf)
+	}
+	return index
+}
+
+func (m *Model) AddNode(node *protocol.Connection) {
+	m.Lock()
+	m.nodes[node.ID] = node
+	m.Unlock()
+	m.RLock()
+	idx := m.protocolIndex()
+	m.RUnlock()
+
+	if traceNet {
+		debugf("NET IDX(out): %s: %d files", node.ID, len(idx))
+	}
+	node.Index(idx)
+}
--- a/model_puller.go
+++ b/model_puller.go
@ -0,0 +1,212 @@
+package main
+
+/*
+
+Locking
+=======
+
+These methods are never called from the outside so don't follow the locking
+policy in model.go. Instead, appropriate locks are acquired when needed and
+held for as short a time as possible.
+
+TODO(jb): Refactor this into smaller and cleaner pieces.
+
+TODO(jb): Some kind of coalescing / rate limiting of index sending, so we don't
+send hundreds of index updates in a short period if time when deleting files
+etc.
+
+*/
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"os"
+	"path"
+	"sync"
+	"time"
+
+	"github.com/calmh/syncthing/buffers"
+)
+
+func (m *Model) pullFile(name string) error {
+	m.RLock()
+	var localFile = m.local[name]
+	var globalFile = m.global[name]
+	m.RUnlock()
+
+	filename := path.Join(m.dir, name)
+	sdir := path.Dir(filename)
+
+	_, err := os.Stat(sdir)
+	if err != nil && os.IsNotExist(err) {
+		os.MkdirAll(sdir, 0777)
+	}
+
+	tmpFilename := tempName(filename, globalFile.Modified)
+	tmpFile, err := os.Create(tmpFilename)
+	if err != nil {
+		return err
+	}
+	defer tmpFile.Close()
+
+	contentChan := make(chan content, 32)
+	var applyDone sync.WaitGroup
+	applyDone.Add(1)
+	go func() {
+		applyContent(contentChan, tmpFile)
+		applyDone.Done()
+	}()
+
+	local, remote := localFile.Blocks.To(globalFile.Blocks)
+	var fetchDone sync.WaitGroup
+
+	// One local copy routing
+
+	fetchDone.Add(1)
+	go func() {
+		for _, block := range local {
+			data, err := m.Request("<local>", name, block.Offset, block.Length, block.Hash)
+			if err != nil {
+				break
+			}
+			contentChan <- content{
+				offset: int64(block.Offset),
+				data:   data,
+			}
+		}
+		fetchDone.Done()
+	}()
+
+	// N remote copy routines
+
+	m.RLock()
+	var nodeIDs = m.whoHas(name)
+	m.RUnlock()
+	var remoteBlocksChan = make(chan Block)
+	go func() {
+		for _, block := range remote {
+			remoteBlocksChan <- block
+		}
+		close(remoteBlocksChan)
+	}()
+
+	// XXX: This should be rewritten into something nicer that takes differing
+	// peer performance into account.
+
+	for i := 0; i < RemoteFetchers; i++ {
+		for _, nodeID := range nodeIDs {
+			fetchDone.Add(1)
+			go func(nodeID string) {
+				for block := range remoteBlocksChan {
+					data, err := m.RequestGlobal(nodeID, name, block.Offset, block.Length, block.Hash)
+					if err != nil {
+						break
+					}
+					contentChan <- content{
+						offset: int64(block.Offset),
+						data:   data,
+					}
+				}
+				fetchDone.Done()
+			}(nodeID)
+		}
+	}
+
+	fetchDone.Wait()
+	close(contentChan)
+	applyDone.Wait()
+
+	rf, err := os.Open(tmpFilename)
+	if err != nil {
+		return err
+	}
+	defer rf.Close()
+
+	writtenBlocks, err := Blocks(rf, BlockSize)
+	if err != nil {
+		return err
+	}
+	if len(writtenBlocks) != len(globalFile.Blocks) {
+		return fmt.Errorf("%s: incorrect number of blocks after sync", tmpFilename)
+	}
+	for i := range writtenBlocks {
+		if bytes.Compare(writtenBlocks[i].Hash, globalFile.Blocks[i].Hash) != 0 {
+			return fmt.Errorf("%s: hash mismatch after sync\n  %v\n  %v", tmpFilename, writtenBlocks[i], globalFile.Blocks[i])
+		}
+	}
+
+	err = os.Chtimes(tmpFilename, time.Unix(globalFile.Modified, 0), time.Unix(globalFile.Modified, 0))
+	if err != nil {
+		return err
+	}
+
+	err = os.Rename(tmpFilename, filename)
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
+func (m *Model) puller() {
+	for {
+		for {
+			var n string
+			var f File
+
+			m.RLock()
+			for n = range m.need {
+				break // just pick first name
+			}
+			if len(n) != 0 {
+				f = m.global[n]
+			}
+			m.RUnlock()
+
+			if len(n) == 0 {
+				// we got nothing
+				break
+			}
+
+			var err error
+			if f.Flags&FlagDeleted == 0 {
+				if traceFile {
+					debugf("FILE: Pull %q", n)
+				}
+				err = m.pullFile(n)
+			} else {
+				if traceFile {
+					debugf("FILE: Remove %q", n)
+				}
+				// Cheerfully ignore errors here
+				_ = os.Remove(path.Join(m.dir, n))
+			}
+			if err == nil {
+				m.UpdateLocal(f)
+			} else {
+				warnln(err)
+			}
+		}
+		time.Sleep(time.Second)
+	}
+}
+
+type content struct {
+	offset int64
+	data   []byte
+}
+
+func applyContent(cc <-chan content, dst io.WriterAt) error {
+	var err error
+
+	for c := range cc {
+		_, err = dst.WriteAt(c.data, c.offset)
+		if err != nil {
+			return err
+		}
+		buffers.Put(c.data)
+	}
+
+	return nil
+}
--- a/model_test.go
+++ b/model_test.go
@ -0,0 +1,275 @@
+package main
+
+import (
+	"reflect"
+	"testing"
+	"time"
+
+	"github.com/calmh/syncthing/protocol"
+)
+
+func TestNewModel(t *testing.T) {
+	m := NewModel("foo")
+
+	if m == nil {
+		t.Fatalf("NewModel returned nil")
+	}
+
+	if len(m.need) > 0 {
+		t.Errorf("New model should have no Need")
+	}
+
+	if len(m.local) > 0 {
+		t.Errorf("New model should have no Have")
+	}
+}
+
+var testDataExpected = map[string]File{
+	"foo": File{
+		Name:     "foo",
+		Flags:    0644,
+		Modified: 1384244572,
+		Blocks:   []Block{{Offset: 0x0, Length: 0x7, Hash: []uint8{0xae, 0xc0, 0x70, 0x64, 0x5f, 0xe5, 0x3e, 0xe3, 0xb3, 0x76, 0x30, 0x59, 0x37, 0x61, 0x34, 0xf0, 0x58, 0xcc, 0x33, 0x72, 0x47, 0xc9, 0x78, 0xad, 0xd1, 0x78, 0xb6, 0xcc, 0xdf, 0xb0, 0x1, 0x9f}}},
+	},
+	"bar": File{
+		Name:     "bar",
+		Flags:    0644,
+		Modified: 1384244579,
+		Blocks:   []Block{{Offset: 0x0, Length: 0xa, Hash: []uint8{0x2f, 0x72, 0xcc, 0x11, 0xa6, 0xfc, 0xd0, 0x27, 0x1e, 0xce, 0xf8, 0xc6, 0x10, 0x56, 0xee, 0x1e, 0xb1, 0x24, 0x3b, 0xe3, 0x80, 0x5b, 0xf9, 0xa9, 0xdf, 0x98, 0xf9, 0x2f, 0x76, 0x36, 0xb0, 0x5c}}},
+	},
+	"baz/quux": File{
+		Name:     "baz/quux",
+		Flags:    0644,
+		Modified: 1384244676,
+		Blocks:   []Block{{Offset: 0x0, Length: 0x9, Hash: []uint8{0xc1, 0x54, 0xd9, 0x4e, 0x94, 0xba, 0x72, 0x98, 0xa6, 0xad, 0xb0, 0x52, 0x3a, 0xfe, 0x34, 0xd1, 0xb6, 0xa5, 0x81, 0xd6, 0xb8, 0x93, 0xa7, 0x63, 0xd4, 0x5d, 0xdc, 0x5e, 0x20, 0x9d, 0xcb, 0x83}}},
+	},
+}
+
+func TestUpdateLocal(t *testing.T) {
+	m := NewModel("foo")
+	fs := Walk("testdata", m)
+	m.ReplaceLocal(fs)
+
+	if len(m.need) > 0 {
+		t.Fatalf("Model with only local data should have no need")
+	}
+
+	if l1, l2 := len(m.local), len(testDataExpected); l1 != l2 {
+		t.Fatalf("Model len(local) incorrect, %d != %d", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(testDataExpected); l1 != l2 {
+		t.Fatalf("Model len(global) incorrect, %d != %d", l1, l2)
+	}
+	for name, file := range testDataExpected {
+		if f, ok := m.local[name]; ok {
+			if !reflect.DeepEqual(f, file) {
+				t.Errorf("Incorrect local\n%v !=\n%v\nfor file %q", f, file, name)
+			}
+		} else {
+			t.Errorf("Missing file %q in local table", name)
+		}
+		if f, ok := m.global[name]; ok {
+			if !reflect.DeepEqual(f, file) {
+				t.Errorf("Incorrect global\n%v !=\n%v\nfor file %q", f, file, name)
+			}
+		} else {
+			t.Errorf("Missing file %q in global table", name)
+		}
+	}
+
+	for _, f := range fs {
+		if hf, ok := m.local[f.Name]; !ok || hf.Modified != f.Modified {
+			t.Fatalf("Incorrect local for %q", f.Name)
+		}
+		if cf, ok := m.global[f.Name]; !ok || cf.Modified != f.Modified {
+			t.Fatalf("Incorrect global for %q", f.Name)
+		}
+	}
+}
+
+func TestRemoteUpdateExisting(t *testing.T) {
+	m := NewModel("foo")
+	fs := Walk("testdata", m)
+	m.ReplaceLocal(fs)
+
+	newFile := protocol.FileInfo{
+		Name:     "foo",
+		Modified: time.Now().Unix(),
+		Blocks:   []protocol.BlockInfo{{100, []byte("some hash bytes")}},
+	}
+	m.Index(string("42"), []protocol.FileInfo{newFile})
+
+	if l := len(m.need); l != 1 {
+		t.Errorf("Model missing Need for one file (%d != 1)", l)
+	}
+}
+
+func TestRemoteAddNew(t *testing.T) {
+	m := NewModel("foo")
+	fs := Walk("testdata", m)
+	m.ReplaceLocal(fs)
+
+	newFile := protocol.FileInfo{
+		Name:     "a new file",
+		Modified: time.Now().Unix(),
+		Blocks:   []protocol.BlockInfo{{100, []byte("some hash bytes")}},
+	}
+	m.Index(string("42"), []protocol.FileInfo{newFile})
+
+	if l1, l2 := len(m.need), 1; l1 != l2 {
+		t.Errorf("Model len(m.need) incorrect (%d != %d)", l1, l2)
+	}
+}
+
+func TestRemoteUpdateOld(t *testing.T) {
+	m := NewModel("foo")
+	fs := Walk("testdata", m)
+	m.ReplaceLocal(fs)
+
+	oldTimeStamp := int64(1234)
+	newFile := protocol.FileInfo{
+		Name:     "foo",
+		Modified: oldTimeStamp,
+		Blocks:   []protocol.BlockInfo{{100, []byte("some hash bytes")}},
+	}
+	m.Index(string("42"), []protocol.FileInfo{newFile})
+
+	if l1, l2 := len(m.need), 0; l1 != l2 {
+		t.Errorf("Model len(need) incorrect (%d != %d)", l1, l2)
+	}
+}
+
+func TestDelete(t *testing.T) {
+	m := NewModel("foo")
+	fs := Walk("testdata", m)
+	m.ReplaceLocal(fs)
+
+	if l1, l2 := len(m.local), len(fs); l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs); l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+
+	ot := time.Now().Unix()
+	newFile := File{
+		Name:     "a new file",
+		Modified: ot,
+		Blocks:   []Block{{0, 100, []byte("some hash bytes")}},
+	}
+	m.UpdateLocal(newFile)
+
+	if l1, l2 := len(m.local), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+
+	// The deleted file is kept in the local and global tables and marked as deleted.
+
+	m.ReplaceLocal(fs)
+
+	if l1, l2 := len(m.local), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+
+	if m.local["a new file"].Flags&(1<<12) == 0 {
+		t.Error("Unexpected deleted flag = 0 in local table")
+	}
+	if len(m.local["a new file"].Blocks) != 0 {
+		t.Error("Unexpected non-zero blocks for deleted file in local")
+	}
+	if ft := m.local["a new file"].Modified; ft != ot+1 {
+		t.Errorf("Unexpected time %d != %d for deleted file in local", ft, ot+1)
+	}
+
+	if m.global["a new file"].Flags&(1<<12) == 0 {
+		t.Error("Unexpected deleted flag = 0 in global table")
+	}
+	if len(m.global["a new file"].Blocks) != 0 {
+		t.Error("Unexpected non-zero blocks for deleted file in global")
+	}
+	if ft := m.local["a new file"].Modified; ft != ot+1 {
+		t.Errorf("Unexpected time %d != %d for deleted file in local", ft, ot+1)
+	}
+
+	// Another update should change nothing
+
+	m.ReplaceLocal(fs)
+
+	if l1, l2 := len(m.local), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+
+	if m.local["a new file"].Flags&(1<<12) == 0 {
+		t.Error("Unexpected deleted flag = 0 in local table")
+	}
+	if len(m.local["a new file"].Blocks) != 0 {
+		t.Error("Unexpected non-zero blocks for deleted file in local")
+	}
+	if ft := m.local["a new file"].Modified; ft != ot+1 {
+		t.Errorf("Unexpected time %d != %d for deleted file in local", ft, ot+1)
+	}
+
+	if m.global["a new file"].Flags&(1<<12) == 0 {
+		t.Error("Unexpected deleted flag = 0 in global table")
+	}
+	if len(m.global["a new file"].Blocks) != 0 {
+		t.Error("Unexpected non-zero blocks for deleted file in global")
+	}
+	if ft := m.local["a new file"].Modified; ft != ot+1 {
+		t.Errorf("Unexpected time %d != %d for deleted file in local", ft, ot+1)
+	}
+}
+
+func TestForgetNode(t *testing.T) {
+	m := NewModel("foo")
+	fs := Walk("testdata", m)
+	m.ReplaceLocal(fs)
+
+	if l1, l2 := len(m.local), len(fs); l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs); l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.need), 0; l1 != l2 {
+		t.Errorf("Model len(need) incorrect (%d != %d)", l1, l2)
+	}
+
+	newFile := protocol.FileInfo{
+		Name:     "new file",
+		Modified: time.Now().Unix(),
+		Blocks:   []protocol.BlockInfo{{100, []byte("some hash bytes")}},
+	}
+	m.Index(string("42"), []protocol.FileInfo{newFile})
+
+	if l1, l2 := len(m.local), len(fs); l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs)+1; l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.need), 1; l1 != l2 {
+		t.Errorf("Model len(need) incorrect (%d != %d)", l1, l2)
+	}
+
+	m.Close(string("42"))
+
+	if l1, l2 := len(m.local), len(fs); l1 != l2 {
+		t.Errorf("Model len(local) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.global), len(fs); l1 != l2 {
+		t.Errorf("Model len(global) incorrect (%d != %d)", l1, l2)
+	}
+	if l1, l2 := len(m.need), 0; l1 != l2 {
+		t.Errorf("Model len(need) incorrect (%d != %d)", l1, l2)
+	}
+}
--- a/protocol/PROTOCOL.md
+++ b/protocol/PROTOCOL.md
@ -0,0 +1,208 @@
+Block Exchange Protocol v1.0
+============================
+
+Introduction and Definitions
+----------------------------
+
+The BEP is used between two or more _nodes_ thus forming a _cluster_.
+Each node has a _repository_ of files described by the _local model_,
+containing modifications times and block hashes. The local model is sent
+to the other nodes in the cluster. The union of all files in the local
+models, with files selected for most recent modification time, forms the
+_global model_. Each node strives to get it's repository in synch with
+the global model by requesting missing blocks from the other nodes.
+
+Transport and Authentication
+----------------------------
+
+The BEP itself does not provide retransmissions, compression, encryption
+nor authentication. It is expected that this is performed at lower
+layers of the networking stack. A typical deployment stack should be
+similar to the following:
+
+    |-----------------------------|
+    |   Block Exchange Protocol   |
+    |-----------------------------|
+    |   Compression (RFC 1951)    |
+    |-----------------------------|
+    | Encryption & Auth (TLS 1.0) |
+    |-----------------------------|
+    |             TCP             |
+    |-----------------------------|
+    v                             v
+
+The exact nature of the authentication is up to the application.
+Possibilities include certificates signed by a common trusted CA,
+preshared certificates, preshared certificate fingerprints or
+certificate pinning combined with some out of band first verification.
+
+There is no required order or synchronization among BEP messages - any
+message type may be sent at any time and the sender need not await a
+response to one message before sending another. Responses must however
+be sent in the same order as the requests are received.
+
+Compression is started directly after a successfull TLS handshake,
+before the first message is sent. The compression is flushed at each
+message boundary.
+
+Messages
+--------
+
+Every message starts with one 32 bit word indicating the message version
+and type. For BEP v1.0 the Version field is set to zero. Future versions
+with incompatible message formats will increment the Version field. The
+reserved bits must be set to zero.
+
+     0                   1                   2                   3
+     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+    | Ver=0 |       Message ID      |     Type      |    Reserved   |
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+
+All data following the message header is in XDR (RFC 1014) encoding.
+The actual data types in use by BEP, in XDR naming convention, are:
+
+ - unsigned int   -- unsigned 32 bit integer
+ - hyper          -- signed 64 bit integer
+ - unsigned hyper -- signed 64 bit integer
+ - opaque<>       -- variable length opaque data
+ - string<>       -- variable length string
+
+The encoding of opaque<> and string<> are identical, the distinction is
+solely in interpretation. Opaque data should not be interpreted as such,
+but can be compared bytewise to other opaque data. All strings use the
+UTF-8 encoding.
+
+### Index (Type = 1)
+
+The Index message defines the contents of the senders repository. A Index
+message is sent by each peer immediately upon connection and whenever the
+local repository contents changes. However, if a peer has no data to
+advertise (the repository is empty, or it is set to only import data) it
+is allowed but not required to send an empty Index message (a file list of
+zero length). If the repository contents change from non-empty to empty,
+an empty Index message must be sent. There is no response to the Index
+message.
+
+    struct IndexMessage {
+        FileInfo Files<>;
+    }
+
+    struct FileInfo {
+        string Name<>;
+        unsigned int Flags;
+        hyper Modified;
+        BlockInfo Blocks<>;
+    }
+
+    struct BlockInfo {
+        unsigned int Length;
+        opaque Hash<>
+    }
+
+The file name is the part relative to the repository root. The
+modification time is expressed as the number of seconds since the Unix
+Epoch. The hash algorithm is implied by the hash length. Currently, the
+hash must be 32 bytes long and computed by SHA256.
+
+The flags field is made up of the following single bit flags:
+
+     0                   1                   2                   3
+     0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+    |              Reserved               |D|   Unix Perm. & Mode   |
+    +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+
+ - The lower 12 bits hold the common Unix permission and mode bits.
+
+ - Bit 19 ("D") is set when the file has been deleted. The block list
+   shall contain zero blocks and the modification time indicates the
+   time of deletion or, if deletion time is not reliably determinable,
+   one second past the last know modification time.
+
+ - Bit 0 through 18 are reserved for future use and shall be set to
+   zero.
+
+### Request (Type = 2)
+
+The Request message expresses the desire to receive a data block
+corresponding to a part of a certain file in the peer's repository.
+
+The requested block must correspond exactly to one block seen in the
+peer's Index message. The hash field must be set to the expected value by
+the sender. The receiver may validate that this is actually the case
+before transmitting data. Each Request message must be met with a Response
+message.
+
+    struct RequestMessage {
+        string Name<>;
+        unsigned hyper Offset;
+        unsigned int Length;
+        opaque Hash<>;
+    }
+
+The hash algorithm is implied by the hash length. Currently, the hash
+must be 32 bytes long and computed by SHA256.
+
+The Message ID in the header must set to a unique value to be able to
+correlate the request with the response message.
+
+### Response (Type = 3)
+
+The Response message is sent in response to a Request message. In case the
+requested data was not available (an outdated block was requested, or
+the file has been deleted), the Data field is empty.
+
+    struct ResponseMessage {
+        opaque Data<>
+    }
+
+The Message ID in the header is used to correlate requests and
+responses.
+
+### Ping (Type = 4)
+
+The Ping message is used to determine that a connection is alive, and to
+keep connections alive through state tracking network elements such as
+firewalls and NAT gateways. The Ping message has no contents.
+
+    struct PingMessage {
+    }
+
+### Pong (Type = 5)
+
+The Pong message is sent in response to a Ping. The Pong message has no
+contents, but copies the Message ID from the Ping.
+
+    struct PongMessage {
+    }
+
+Example Exchange
+----------------
+
+          A            B
+     1. Index->      <-Index
+     2. Request->
+     3. Request->
+     4. Request->
+     5. Request->
+     6.         <-Response
+     7.         <-Response
+     8.         <-Response
+     9.         <-Response
+    10. Index->
+        ...
+    11. Ping->
+    12.            <-Pong
+
+The connection is established and at 1. both peers send Index records.
+The Index records are received and both peers recompute their knowledge
+of the data in the cluster. In this example, peer A has four missing or
+outdated blocks. At 2 through 5 peer A sends requests for these blocks.
+The requests are received by peer B, who retrieves the data from the
+repository and transmits Response records (6 through 9). Node A updates
+their repository contents and transmits an updated Index message (10).
+Both peers enter idle state after 10. At some later time 11, peer A
+determines that it has not seen data from B for some time and sends a
+Ping request. A response is sent at 12.
+
--- a/protocol/marshal.go
+++ b/protocol/marshal.go
@ -0,0 +1,119 @@
+package protocol
+
+import (
+	"io"
+
+	"github.com/calmh/syncthing/buffers"
+)
+
+func pad(l int) int {
+	d := l % 4
+	if d == 0 {
+		return 0
+	}
+	return 4 - d
+}
+
+var padBytes = []byte{0, 0, 0}
+
+type marshalWriter struct {
+	w   io.Writer
+	tot int
+	err error
+}
+
+func (w *marshalWriter) writeString(s string) {
+	w.writeBytes([]byte(s))
+}
+
+func (w *marshalWriter) writeBytes(bs []byte) {
+	if w.err != nil {
+		return
+	}
+	w.writeUint32(uint32(len(bs)))
+	if w.err != nil {
+		return
+	}
+	_, w.err = w.w.Write(bs)
+	if p := pad(len(bs)); p > 0 {
+		w.w.Write(padBytes[:p])
+	}
+	w.tot += len(bs) + pad(len(bs))
+}
+
+func (w *marshalWriter) writeUint32(v uint32) {
+	if w.err != nil {
+		return
+	}
+	var b [4]byte
+	b[0] = byte(v >> 24)
+	b[1] = byte(v >> 16)
+	b[2] = byte(v >> 8)
+	b[3] = byte(v)
+	_, w.err = w.w.Write(b[:])
+	w.tot += 4
+}
+
+func (w *marshalWriter) writeUint64(v uint64) {
+	if w.err != nil {
+		return
+	}
+	var b [8]byte
+	b[0] = byte(v >> 56)
+	b[1] = byte(v >> 48)
+	b[2] = byte(v >> 40)
+	b[3] = byte(v >> 32)
+	b[4] = byte(v >> 24)
+	b[5] = byte(v >> 16)
+	b[6] = byte(v >> 8)
+	b[7] = byte(v)
+	_, w.err = w.w.Write(b[:])
+	w.tot += 8
+}
+
+type marshalReader struct {
+	r   io.Reader
+	tot int
+	err error
+}
+
+func (r *marshalReader) readString() string {
+	bs := r.readBytes()
+	defer buffers.Put(bs)
+	return string(bs)
+}
+
+func (r *marshalReader) readBytes() []byte {
+	if r.err != nil {
+		return nil
+	}
+	l := int(r.readUint32())
+	if r.err != nil {
+		return nil
+	}
+	b := buffers.Get(l + pad(l))
+	_, r.err = io.ReadFull(r.r, b)
+	r.tot += int(l + pad(l))
+	return b[:l]
+}
+
+func (r *marshalReader) readUint32() uint32 {
+	if r.err != nil {
+		return 0
+	}
+	var b [4]byte
+	_, r.err = io.ReadFull(r.r, b[:])
+	r.tot += 4
+	return uint32(b[3]) | uint32(b[2])<<8 | uint32(b[1])<<16 | uint32(b[0])<<24
+}
+
+func (r *marshalReader) readUint64() uint64 {
+	if r.err != nil {
+		return 0
+	}
+	var b [8]byte
+	_, r.err = io.ReadFull(r.r, b[:])
+	r.tot += 8
+	return uint64(b[7]) | uint64(b[6])<<8 | uint64(b[5])<<16 | uint64(b[4])<<24 |
+		uint64(b[3])<<32 | uint64(b[2])<<40 | uint64(b[1])<<48 | uint64(b[0])<<56
+}
--- a/protocol/messages.go
+++ b/protocol/messages.go
@ -0,0 +1,106 @@
+package protocol
+
+import "io"
+
+type request struct {
+	name   string
+	offset uint64
+	size   uint32
+	hash   []byte
+}
+
+type header struct {
+	version int
+	msgID   int
+	msgType int
+}
+
+func encodeHeader(h header) uint32 {
+	return uint32(h.version&0xf)<<28 +
+		uint32(h.msgID&0xfff)<<16 +
+		uint32(h.msgType&0xff)<<8
+}
+
+func decodeHeader(u uint32) header {
+	return header{
+		version: int(u>>28) & 0xf,
+		msgID:   int(u>>16) & 0xfff,
+		msgType: int(u>>8) & 0xff,
+	}
+}
+
+func (w *marshalWriter) writeHeader(h header) {
+	w.writeUint32(encodeHeader(h))
+}
+
+func (w *marshalWriter) writeIndex(idx []FileInfo) {
+	w.writeUint32(uint32(len(idx)))
+	for _, f := range idx {
+		w.writeString(f.Name)
+		w.writeUint32(f.Flags)
+		w.writeUint64(uint64(f.Modified))
+		w.writeUint32(uint32(len(f.Blocks)))
+		for _, b := range f.Blocks {
+			w.writeUint32(b.Length)
+			w.writeBytes(b.Hash)
+		}
+	}
+}
+
+func WriteIndex(w io.Writer, idx []FileInfo) (int, error) {
+	mw := marshalWriter{w, 0, nil}
+	mw.writeIndex(idx)
+	return mw.tot, mw.err
+}
+
+func (w *marshalWriter) writeRequest(r request) {
+	w.writeString(r.name)
+	w.writeUint64(r.offset)
+	w.writeUint32(r.size)
+	w.writeBytes(r.hash)
+}
+
+func (w *marshalWriter) writeResponse(data []byte) {
+	w.writeBytes(data)
+}
+
+func (r *marshalReader) readHeader() header {
+	return decodeHeader(r.readUint32())
+}
+
+func (r *marshalReader) readIndex() []FileInfo {
+	nfiles := r.readUint32()
+	files := make([]FileInfo, nfiles)
+	for i := range files {
+		files[i].Name = r.readString()
+		files[i].Flags = r.readUint32()
+		files[i].Modified = int64(r.readUint64())
+		nblocks := r.readUint32()
+		blocks := make([]BlockInfo, nblocks)
+		for j := range blocks {
+			blocks[j].Length = r.readUint32()
+			blocks[j].Hash = r.readBytes()
+		}
+		files[i].Blocks = blocks
+	}
+	return files
+}
+
+func ReadIndex(r io.Reader) ([]FileInfo, error) {
+	mr := marshalReader{r, 0, nil}
+	idx := mr.readIndex()
+	return idx, mr.err
+}
+
+func (r *marshalReader) readRequest() request {
+	var req request
+	req.name = r.readString()
+	req.offset = r.readUint64()
+	req.size = r.readUint32()
+	req.hash = r.readBytes()
+	return req
+}
+
+func (r *marshalReader) readResponse() []byte {
+	return r.readBytes()
+}
--- a/protocol/messages_test.go
+++ b/protocol/messages_test.go
@ -0,0 +1,115 @@
+package protocol
+
+import (
+	"bytes"
+	"io/ioutil"
+	"reflect"
+	"testing"
+	"testing/quick"
+)
+
+func TestIndex(t *testing.T) {
+	idx := []FileInfo{
+		{
+			"Foo",
+			0755,
+			1234567890,
+			[]BlockInfo{
+				{12345678, []byte("hash hash hash")},
+				{23456781, []byte("ash hash hashh")},
+				{34567812, []byte("sh hash hashha")},
+			},
+		}, {
+			"Quux/Quux",
+			0644,
+			2345678901,
+			[]BlockInfo{
+				{45678123, []byte("4321 hash hash hash")},
+				{56781234, []byte("3214 ash hash hashh")},
+				{67812345, []byte("2143 sh hash hashha")},
+			},
+		},
+	}
+
+	var buf = new(bytes.Buffer)
+	var wr = marshalWriter{buf, 0, nil}
+	wr.writeIndex(idx)
+
+	var rd = marshalReader{buf, 0, nil}
+	var idx2 = rd.readIndex()
+
+	if !reflect.DeepEqual(idx, idx2) {
+		t.Errorf("Index marshal error:\n%#v\n%#v\n", idx, idx2)
+	}
+}
+
+func TestRequest(t *testing.T) {
+	f := func(name string, offset uint64, size uint32, hash []byte) bool {
+		var buf = new(bytes.Buffer)
+		var req = request{name, offset, size, hash}
+		var wr = marshalWriter{buf, 0, nil}
+		wr.writeRequest(req)
+		var rd = marshalReader{buf, 0, nil}
+		var req2 = rd.readRequest()
+		return req.name == req2.name &&
+			req.offset == req2.offset &&
+			req.size == req2.size &&
+			bytes.Compare(req.hash, req2.hash) == 0
+	}
+	if err := quick.Check(f, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestResponse(t *testing.T) {
+	f := func(data []byte) bool {
+		var buf = new(bytes.Buffer)
+		var wr = marshalWriter{buf, 0, nil}
+		wr.writeResponse(data)
+		var rd = marshalReader{buf, 0, nil}
+		var read = rd.readResponse()
+		return bytes.Compare(read, data) == 0
+	}
+	if err := quick.Check(f, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func BenchmarkWriteIndex(b *testing.B) {
+	idx := []FileInfo{
+		{
+			"Foo",
+			0777,
+			1234567890,
+			[]BlockInfo{
+				{12345678, []byte("hash hash hash")},
+				{23456781, []byte("ash hash hashh")},
+				{34567812, []byte("sh hash hashha")},
+			},
+		}, {
+			"Quux/Quux",
+			0644,
+			2345678901,
+			[]BlockInfo{
+				{45678123, []byte("4321 hash hash hash")},
+				{56781234, []byte("3214 ash hash hashh")},
+				{67812345, []byte("2143 sh hash hashha")},
+			},
+		},
+	}
+
+	var wr = marshalWriter{ioutil.Discard, 0, nil}
+
+	for i := 0; i < b.N; i++ {
+		wr.writeIndex(idx)
+	}
+}
+
+func BenchmarkWriteRequest(b *testing.B) {
+	var req = request{"blah blah", 1231323, 13123123, []byte("hash hash hash")}
+	var wr = marshalWriter{ioutil.Discard, 0, nil}
+
+	for i := 0; i < b.N; i++ {
+		wr.writeRequest(req)
+	}
+}
--- a/protocol/protocol.go
+++ b/protocol/protocol.go
@ -0,0 +1,239 @@
+package protocol
+
+import (
+	"compress/flate"
+	"errors"
+	"io"
+	"sync"
+
+	"github.com/calmh/syncthing/buffers"
+)
+
+const (
+	messageTypeReserved = iota
+	messageTypeIndex
+	messageTypeRequest
+	messageTypeResponse
+	messageTypePing
+	messageTypePong
+)
+
+var ErrClosed = errors.New("Connection closed")
+
+type FileInfo struct {
+	Name     string
+	Flags    uint32
+	Modified int64
+	Blocks   []BlockInfo
+}
+
+type BlockInfo struct {
+	Length uint32
+	Hash   []byte
+}
+
+type Model interface {
+	// An index was received from the peer node
+	Index(nodeID string, files []FileInfo)
+	// A request was made by the peer node
+	Request(nodeID, name string, offset uint64, size uint32, hash []byte) ([]byte, error)
+	// The peer node closed the connection
+	Close(nodeID string)
+}
+
+type Connection struct {
+	receiver   Model
+	reader     io.Reader
+	mreader    *marshalReader
+	writer     io.Writer
+	mwriter    *marshalWriter
+	wLock      sync.RWMutex
+	closed     bool
+	closedLock sync.RWMutex
+	awaiting   map[int]chan interface{}
+	nextId     int
+	ID         string
+}
+
+func NewConnection(nodeID string, reader io.Reader, writer io.Writer, receiver Model) *Connection {
+	flrd := flate.NewReader(reader)
+	flwr, err := flate.NewWriter(writer, flate.BestSpeed)
+	if err != nil {
+		panic(err)
+	}
+
+	c := Connection{
+		receiver: receiver,
+		reader:   flrd,
+		mreader:  &marshalReader{flrd, 0, nil},
+		writer:   flwr,
+		mwriter:  &marshalWriter{flwr, 0, nil},
+		awaiting: make(map[int]chan interface{}),
+		ID:       nodeID,
+	}
+
+	go c.readerLoop()
+
+	return &c
+}
+
+// Index writes the list of file information to the connected peer node
+func (c *Connection) Index(idx []FileInfo) {
+	c.wLock.Lock()
+	defer c.wLock.Unlock()
+
+	c.mwriter.writeHeader(header{0, c.nextId, messageTypeIndex})
+	c.nextId = (c.nextId + 1) & 0xfff
+	c.mwriter.writeIndex(idx)
+	c.flush()
+}
+
+// Request returns the bytes for the specified block after fetching them from the connected peer.
+func (c *Connection) Request(name string, offset uint64, size uint32, hash []byte) ([]byte, error) {
+	c.wLock.Lock()
+	rc := make(chan interface{})
+	c.awaiting[c.nextId] = rc
+	c.mwriter.writeHeader(header{0, c.nextId, messageTypeRequest})
+	c.mwriter.writeRequest(request{name, offset, size, hash})
+	c.flush()
+	c.nextId = (c.nextId + 1) & 0xfff
+	c.wLock.Unlock()
+
+	// Reading something that might be nil from a possibly closed channel...
+	// r0<~
+
+	var data []byte
+	i, ok := <-rc
+	if ok {
+		if d, ok := i.([]byte); ok {
+			data = d
+		}
+	}
+
+	var err error
+	i, ok = <-rc
+	if ok {
+		if e, ok := i.(error); ok {
+			err = e
+		}
+	}
+	return data, err
+}
+
+func (c *Connection) Ping() bool {
+	c.wLock.Lock()
+	rc := make(chan interface{})
+	c.awaiting[c.nextId] = rc
+	c.mwriter.writeHeader(header{0, c.nextId, messageTypePing})
+	c.flush()
+	c.nextId = (c.nextId + 1) & 0xfff
+	c.wLock.Unlock()
+
+	_, ok := <-rc
+	return ok
+}
+
+func (c *Connection) Stop() {
+}
+
+type flusher interface {
+	Flush() error
+}
+
+func (c *Connection) flush() {
+	if f, ok := c.writer.(flusher); ok {
+		f.Flush()
+	}
+}
+
+func (c *Connection) close() {
+	c.closedLock.Lock()
+	c.closed = true
+	c.closedLock.Unlock()
+	c.wLock.Lock()
+	for _, ch := range c.awaiting {
+		close(ch)
+	}
+	c.awaiting = nil
+	c.wLock.Unlock()
+	c.receiver.Close(c.ID)
+}
+
+func (c *Connection) isClosed() bool {
+	c.closedLock.RLock()
+	defer c.closedLock.RUnlock()
+	return c.closed
+}
+
+func (c *Connection) readerLoop() {
+	for !c.isClosed() {
+		hdr := c.mreader.readHeader()
+		if c.mreader.err != nil {
+			c.close()
+			break
+		}
+
+		switch hdr.msgType {
+		case messageTypeIndex:
+			files := c.mreader.readIndex()
+			if c.mreader.err != nil {
+				c.close()
+			} else {
+				c.receiver.Index(c.ID, files)
+			}
+
+		case messageTypeRequest:
+			c.processRequest(hdr.msgID)
+
+		case messageTypeResponse:
+			data := c.mreader.readResponse()
+
+			if c.mreader.err != nil {
+				c.close()
+			} else {
+				c.wLock.RLock()
+				rc, ok := c.awaiting[hdr.msgID]
+				c.wLock.RUnlock()
+
+				if ok {
+					rc <- data
+					rc <- c.mreader.err
+					delete(c.awaiting, hdr.msgID)
+					close(rc)
+				}
+			}
+
+		case messageTypePing:
+			c.wLock.Lock()
+			c.mwriter.writeUint32(encodeHeader(header{0, hdr.msgID, messageTypePong}))
+			c.flush()
+			c.wLock.Unlock()
+
+		case messageTypePong:
+			c.wLock.Lock()
+			if rc, ok := c.awaiting[hdr.msgID]; ok {
+				rc <- true
+				close(rc)
+				delete(c.awaiting, hdr.msgID)
+			}
+			c.wLock.Unlock()
+		}
+	}
+}
+
+func (c *Connection) processRequest(msgID int) {
+	req := c.mreader.readRequest()
+	if c.mreader.err != nil {
+		c.close()
+	} else {
+		go func() {
+			data, _ := c.receiver.Request(c.ID, req.name, req.offset, req.size, req.hash)
+			c.wLock.Lock()
+			c.mwriter.writeUint32(encodeHeader(header{0, msgID, messageTypeResponse}))
+			c.mwriter.writeResponse(data)
+			buffers.Put(data)
+			c.flush()
+			c.wLock.Unlock()
+		}()
+	}
+}
--- a/protocol/protocol_test.go
+++ b/protocol/protocol_test.go
@ -0,0 +1,37 @@
+package protocol
+
+import (
+	"testing"
+	"testing/quick"
+)
+
+func TestHeaderFunctions(t *testing.T) {
+	f := func(ver, id, typ int) bool {
+		ver = int(uint(ver) % 16)
+		id = int(uint(id) % 4096)
+		typ = int(uint(typ) % 256)
+		h0 := header{ver, id, typ}
+		h1 := decodeHeader(encodeHeader(h0))
+		return h0 == h1
+	}
+	if err := quick.Check(f, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestPad(t *testing.T) {
+	tests := [][]int{
+		{0, 0},
+		{1, 3},
+		{2, 2},
+		{3, 1},
+		{4, 0},
+		{32, 0},
+		{33, 3},
+	}
+	for _, tc := range tests {
+		if p := pad(tc[0]); p != tc[1] {
+			t.Errorf("Incorrect padding for %d bytes, %d != %d", tc[0], p, tc[1])
+		}
+	}
+}
--- a/syncthing.ini
+++ b/syncthing.ini
@ -0,0 +1,11 @@
+[repository]
+dir = /Users/jb/Synced
+
+# The nodes section lists the nodes that make up the cluster. The format is
+# <certificate id> = <space separated list of addresses>
+# The special address "dynamic" means that outbound connections will not be
+# attempted, but inbound connections are accepted.
+
+[nodes]
+ITZXTZ7A32DWV3NLNR5W4M3CHVBW56NA = 172.16.32.1:22000 192.23.34.56:22000
+CUGAE43Y5N64CRJU26YFH6MTWPSBLSUL = dynamic
--- a/testdata/bar
+++ b/testdata/bar
@ -0,0 +1 @@
+foobarbaz
--- a/testdata/baz/quux
+++ b/testdata/baz/quux
@ -0,0 +1 @@
+baazquux
--- a/testdata/foo
+++ b/testdata/foo
@ -0,0 +1 @@
+foobar
--- a/tls.go
+++ b/tls.go
@ -0,0 +1,68 @@
+package main
+
+import (
+	"crypto/rand"
+	"crypto/rsa"
+	"crypto/sha1"
+	"crypto/tls"
+	"crypto/x509"
+	"crypto/x509/pkix"
+	"encoding/base32"
+	"encoding/pem"
+	"math/big"
+	"os"
+	"path"
+	"time"
+)
+
+const (
+	tlsRSABits = 2048
+	tlsName    = "syncthing"
+)
+
+func loadCert(dir string) (tls.Certificate, error) {
+	return tls.LoadX509KeyPair(path.Join(dir, "cert.pem"), path.Join(dir, "key.pem"))
+}
+
+func certId(bs []byte) string {
+	hf := sha1.New()
+	hf.Write(bs)
+	id := hf.Sum(nil)
+	return base32.StdEncoding.EncodeToString(id)
+}
+
+func newCertificate(dir string) {
+	priv, err := rsa.GenerateKey(rand.Reader, tlsRSABits)
+	fatalErr(err)
+
+	notBefore := time.Now()
+	notAfter := time.Date(2049, 12, 31, 23, 59, 59, 0, time.UTC)
+
+	template := x509.Certificate{
+		SerialNumber: new(big.Int).SetInt64(0),
+		Subject: pkix.Name{
+			CommonName: tlsName,
+		},
+		NotBefore: notBefore,
+		NotAfter:  notAfter,
+
+		KeyUsage:              x509.KeyUsageKeyEncipherment | x509.KeyUsageDigitalSignature,
+		ExtKeyUsage:           []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
+		BasicConstraintsValid: true,
+	}
+
+	derBytes, err := x509.CreateCertificate(rand.Reader, &template, &template, &priv.PublicKey, priv)
+	fatalErr(err)
+
+	certOut, err := os.Create(path.Join(dir, "cert.pem"))
+	fatalErr(err)
+	pem.Encode(certOut, &pem.Block{Type: "CERTIFICATE", Bytes: derBytes})
+	certOut.Close()
+	okln("wrote cert.pem")
+
+	keyOut, err := os.OpenFile(path.Join(dir, "key.pem"), os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0600)
+	fatalErr(err)
+	pem.Encode(keyOut, &pem.Block{Type: "RSA PRIVATE KEY", Bytes: x509.MarshalPKCS1PrivateKey(priv)})
+	keyOut.Close()
+	okln("wrote key.pem")
+}
--- a/util.go
+++ b/util.go
@ -0,0 +1,7 @@
+package main
+
+import "time"
+
+func timing(name string, t0 time.Time) {
+	debugf("%s: %.02f ms", name, time.Since(t0).Seconds()*1000)
+}
--- a/walk.go
+++ b/walk.go
@ -0,0 +1,117 @@
+package main
+
+import (
+	"fmt"
+	"os"
+	"path"
+	"path/filepath"
+	"strings"
+)
+
+const BlockSize = 128 * 1024
+
+type File struct {
+	Name     string
+	Flags    uint32
+	Modified int64
+	Blocks   BlockList
+}
+
+func (f File) Dump() {
+	fmt.Printf("%s\n", f.Name)
+	for _, b := range f.Blocks {
+		fmt.Printf("  %dB @ %d: %x\n", b.Length, b.Offset, b.Hash)
+	}
+	fmt.Println()
+}
+
+func isTempName(name string) bool {
+	return strings.HasPrefix(path.Base(name), ".syncthing.")
+}
+
+func tempName(name string, modified int64) string {
+	tdir := path.Dir(name)
+	tname := fmt.Sprintf(".syncthing.%s.%d", path.Base(name), modified)
+	return path.Join(tdir, tname)
+}
+
+func genWalker(base string, res *[]File, model *Model) filepath.WalkFunc {
+	return func(p string, info os.FileInfo, err error) error {
+		if err != nil {
+			return err
+		}
+
+		if isTempName(p) {
+			return nil
+		}
+
+		if info.Mode()&os.ModeType == 0 {
+			rn, err := filepath.Rel(base, p)
+			if err != nil {
+				return err
+			}
+
+			fi, err := os.Stat(p)
+			if err != nil {
+				return err
+			}
+			modified := fi.ModTime().Unix()
+
+			hf, ok := model.LocalFile(rn)
+			if ok && hf.Modified == modified {
+				// No change
+				*res = append(*res, hf)
+			} else {
+				if traceFile {
+					debugf("FILE: Hash %q", p)
+				}
+				fd, err := os.Open(p)
+				if err != nil {
+					return err
+				}
+				defer fd.Close()
+
+				blocks, err := Blocks(fd, BlockSize)
+				if err != nil {
+					return err
+				}
+				f := File{
+					Name:     rn,
+					Flags:    uint32(info.Mode()),
+					Modified: modified,
+					Blocks:   blocks,
+				}
+				*res = append(*res, f)
+			}
+		}
+
+		return nil
+	}
+}
+
+func Walk(dir string, model *Model) []File {
+	var files []File
+	fn := genWalker(dir, &files, model)
+	err := filepath.Walk(dir, fn)
+	if err != nil {
+		warnln(err)
+	}
+	return files
+}
+
+func cleanTempFile(path string, info os.FileInfo, err error) error {
+	if err != nil {
+		return err
+	}
+	if info.Mode()&os.ModeType == 0 && isTempName(path) {
+		if traceFile {
+			debugf("FILE: Remove %q", path)
+		}
+		os.Remove(path)
+	}
+	return nil
+}
+
+func CleanTempFiles(dir string) {
+	filepath.Walk(dir, cleanTempFile)
+}
--- a/walk_test.go
+++ b/walk_test.go
@ -0,0 +1,42 @@
+package main
+
+import (
+	"fmt"
+	"testing"
+	"time"
+)
+
+var testdata = []struct {
+	name string
+	size int
+	hash string
+}{
+	{"bar", 10, "2f72cc11a6fcd0271ecef8c61056ee1eb1243be3805bf9a9df98f92f7636b05c"},
+	{"baz/quux", 9, "c154d94e94ba7298a6adb0523afe34d1b6a581d6b893a763d45ddc5e209dcb83"},
+	{"foo", 7, "aec070645fe53ee3b3763059376134f058cc337247c978add178b6ccdfb0019f"},
+}
+
+func TestWalk(t *testing.T) {
+	m := new(Model)
+	files := Walk("testdata", m)
+
+	if l1, l2 := len(files), len(testdata); l1 != l2 {
+		t.Fatalf("Incorrect number of walked files %d != %d", l1, l2)
+	}
+
+	for i := range testdata {
+		if n1, n2 := testdata[i].name, files[i].Name; n1 != n2 {
+			t.Errorf("Incorrect file name %q != %q for case #%d", n1, n2, i)
+		}
+
+		if h1, h2 := fmt.Sprintf("%x", files[i].Blocks[0].Hash), testdata[i].hash; h1 != h2 {
+			t.Errorf("Incorrect hash %q != %q for case #%d", h1, h2, i)
+		}
+
+		t0 := time.Date(2010, 1, 1, 0, 0, 0, 0, time.UTC).Unix()
+		t1 := time.Date(2020, 1, 1, 0, 0, 0, 0, time.UTC).Unix()
+		if mt := files[i].Modified; mt < t0 || mt > t1 {
+			t.Errorf("Unrealistic modtime %d for test %d", mt, i)
+		}
+	}
+}