Docs and translation update

Refactor: move folder prepare to it's own function
Merge pull request #2430 from calmh/jsondecode
2015-11-05 15:47:06 +00:00 · 2015-11-05 08:01:47 +00:00 · 2015-11-04 20:44:56 +00:00 · 2015-11-04 20:33:10 +00:00 · 2015-11-04 20:15:36 +00:00 · 2015-11-04 20:15:11 +00:00
851 changed files with 144928 additions and 8879 deletions
@@ -0,0 +1,9 @@
+# Text files use LF line endings in this repository
+*   text=auto
+
+# Except the dependencies, which we leave alone
+Godeps/**   -text=auto
+
+# Diffs on these files are meaningless
+gui.files.go    -diff
+*.svg           -diff
@@ -1,3 +1,16 @@
 syncthing
+!gui/syncthing
+!Godeps/_workspace/src/github.com/syncthing
+syncthing.exe
 *.tar.gz
-build
+*.zip
+*.asc
+.jshintrc
+coverage.out
+files/pidx
+bin
+perfstats*.csv
+coverage.xml
+syncthing.sig
+RELEASE
+deb
@@ -0,0 +1 @@
+NICKS
@@ -0,0 +1,70 @@
+# This is the official list of Syncthing authors for copyright purposes.
+
+Aaron Bieber <qbit@deftly.net>
+Adam Piggott	<aD@simplypeachy.co.uk> <simplypeachy@users.noreply.github.com>
+Alexander Graf <register-github@alex-graf.de>
+Andrew Dunham <andrew@du.nham.ca>
+Antony Male <antony.male@gmail.com>
+Arthur Axel fREW Schmidt <frew@afoolishmanifesto.com> <frioux@gmail.com>
+Audrius Butkevicius <audrius.butkevicius@gmail.com>
+Bart De Vries <devriesb@gmail.com>
+Ben Curthoys <ben@bencurthoys.com>
+Ben Schulz <ueomkail@gmail.com> <uok@users.noreply.github.com>
+Ben Sidhom <bsidhom@gmail.com>
+Brandon Philips <brandon@ifup.org>
+Brendan Long <self@brendanlong.com>
+Brian R. Becker <brbecker@gmail.com>
+Caleb Callaway <enlightened.despot@gmail.com>
+Carsten Hagemann <moter8@gmail.com>
+Cathryne Linenweaver <cathryne.linenweaver@gmail.com> <Cathryne@users.noreply.github.com>
+Chris Howie <me@chrishowie.com>
+Chris Joel <chris@scriptolo.gy>
+Colin Kennedy <moshen.colin@gmail.com>
+Daniel Bergmann <dan.arne.bergmann@gmail.com> <brgmnn@users.noreply.github.com>
+Daniel Martí <mvdan@mvdan.cc>
+Denis A. <denisva@gmail.com>
+Dennis Wilson <dw@risu.io>
+Dominik Heidler <dominik@heidler.eu>
+Elias Jarlebring <jarlebring@gmail.com>
+Emil Hessman <emil@hessman.se>
+Erik Meitner <e.meitner@willystreet.coop>
+Federico Castagnini <federico.castagnini@gmail.com>
+Felix Ableitner <me@nutomic.com>
+Felix Unterpaintner <bigbear2nd@gmail.com>
+Francois-Xavier Gsell <fxgsell@gmail.com>
+Frank Isemann <frank@isemann.name>
+Gilli Sigurdsson <gilli@vx.is>
+Jacek Szafarkiewicz <szafar@linux.pl>
+Jakob Borg <jakob@nym.se>
+Jake Peterson <jake@acogdev.com>
+James Patterson <jamespatterson@operamail.com> <jpjp@users.noreply.github.com>
+Jaroslav Malec <dzardacz@gmail.com>
+Jens Diemer <github.com@jensdiemer.de> <git@jensdiemer.de>
+Jochen Voss <voss@seehuhn.de>
+Johan Vromans <jvromans@squirrel.nl>
+Karol Różycki <rozycki.karol@gmail.com>
+Ken'ichi Kamada <kamada@nanohz.org>
+Lode Hoste <zillode@zillode.be>
+Lord Landon Agahnim <lordlandon@gmail.com>
+Marc Laporte <marc@marclaporte.com> <marc@laporte.name>
+Marc Pujol <kilburn@la3.org>
+Marcin Dziadus <dziadus.marcin@gmail.com>
+Mateusz Naściszewski <matin1111@wp.pl>
+Matt Burke <mburke@amplify.com> <burkemw3@gmail.com>
+Michael Jephcote <rewt0r@gmx.com> <Rewt0r@users.noreply.github.com>
+Michael Tilli <pyfisch@gmail.com>
+Pascal Jungblut <github@pascalj.com> <mail@pascal-jungblut.com>
+Peter Hoeg <peter@speartail.com>
+Philippe Schommers <philippe@schommers.be>
+Phill Luby <phill.luby@newredo.com>
+Piotr Bejda <piotrb10@gmail.com>
+Ryan Sullivan <kayoticsully@gmail.com>
+Sergey Mishin <ralder@yandex.ru>
+Stefan Tatschner <stefan@sevenbyte.org> <rumpelsepp@sevenbyte.org>
+Tim Abell <tim@timwise.co.uk>
+Tobias Nygren <tnn@nygren.pp.se>
+Tomas Cerveny <kozec@kozec.com>
+Tully Robinson <tully@tojr.org>
+Veeti Paananen <veeti.paananen@rojekti.fi>
+Vil Brekin <vilbrekin@gmail.com>
+Yannic A. <eipiminusone+github@gmail.com> <eipiminus1@users.noreply.github.com>
@@ -0,0 +1,92 @@
+## Conduct
+
+* We are committed to providing a friendly, safe and welcoming
+  environment for all, regardless of gender, sexual orientation,
+  disability, ethnicity, religion, or similar personal characteristic.
+
+* On IRC, please avoid using overtly sexual nicknames or other nicknames
+  that might detract from a friendly, safe and welcoming environment for
+  all.
+
+* Please be kind and courteous. There's no need to be mean or rude.
+
+* Respect that people have differences of opinion and that every design
+  or implementation choice carries a trade-off and numerous costs. There
+  is seldom a right answer.
+
+* Please keep unstructured critique to a minimum. If you have solid
+  ideas you want to experiment with, make a fork and see how it works.
+
+* We will exclude you from interaction if you insult, demean or harass
+  anyone. That is not welcome behaviour. We interpret the term
+  "harassment" as including the definition in the <a
+  href="http://citizencodeofconduct.org/">Citizen Code of Conduct</a>;
+  if you have any lack of clarity about what might be included in that
+  concept, please read their definition. In particular, we don't
+  tolerate behavior that excludes people in socially marginalized
+  groups.
+
+* Private harassment is also unacceptable. No matter who you are, if you
+  feel you have been or are being harassed or made uncomfortable by a
+  community member, please contact one of the channel ops or any of the
+  Syncthing core team immediately. Whether you're a regular  contributor
+  or a newcomer, we care about making this community a safe place for
+  you and we've got your back.
+
+* Likewise any spamming, trolling, flaming, baiting or other
+  attention-stealing behaviour is not welcome.
+
+## Moderation
+
+These are the policies for upholding our community's standards of
+conduct in our communication channels, most notably in Syncthing-related
+IRC channels and on the web forum.
+
+1. Remarks that violate the Syncthing standards of conduct, including
+   hateful, hurtful, oppressive, or exclusionary remarks, are not
+   allowed. (Cursing is allowed, but never targeting another user, and
+   never in a hateful manner.)
+
+2. Remarks that moderators find inappropriate, whether listed in the
+   code of conduct or not, are also not allowed.
+
+3. Moderators will first respond to such remarks with a warning.
+
+4. If the warning is unheeded, the user will be "kicked," i.e., kicked
+   out of the communication channel to cool off.
+
+5. If the user comes back and continues to make trouble, they will be
+   banned, i.e., indefinitely excluded.
+
+6. Moderators may choose at their discretion to un-ban the user if it
+   was a first offense and they offer the offended party a genuine
+   apology.
+
+7. If a moderator bans someone and you think it was unjustified, please
+   take it up with that moderator, or with a different moderator, **in
+   private**. Complaints about bans in-channel are not allowed.
+
+8. Moderators are held to a higher standard than other community
+   members. If a moderator creates an inappropriate situation, they
+   should expect less leeway than others.
+
+In the Syncthing community we strive to go the extra step to look out
+for each other. Don't just aim to be technically unimpeachable, try to
+be your best self. In particular, avoid flirting with offensive or
+sensitive issues, particularly if they're off-topic; this all too
+often leads to unnecessary fights, hurt feelings, and damaged trust;
+worse, it can drive people away from the community entirely.
+
+And if someone takes issue with something you said or did, resist the
+urge to be defensive. Just stop doing what it was they complained about
+and apologize. Even if you feel you were misinterpreted or unfairly
+accused, chances are good there was something you could've communicated
+better — remember that it's your responsibility to make your fellow
+community members comfortable. Everyone wants to get along and we are
+all here first and foremost because we want to talk about cool
+technology. You will find that people will be eager to assume good
+intent and forgive as long as you earn their trust.
+
+*Adapted from the [Rust Code of Conduct](https://github.com/rust-lang/rust/wiki/Note-development-policy#conduct)*
+
+*Adapted from the [Node.js Policy on Trolling](http://blog.izs.me/post/30036893703/policy-on-trolling)*
@@ -0,0 +1,52 @@
+## Reporting Bugs
+
+Please file bugs in the [Github Issue
+Tracker](https://github.com/syncthing/syncthing/issues). Include at
+least the following:
+
+ - What happened
+
+ - What did you expect to happen instead of what *did* happen, if it's
+   not crazy obvious
+
+ - What operating system, operating system version and version of
+   Syncthing you are running
+
+ - The same for other connected devices, where relevant
+
+ - Screenshot if the issue concerns something visible in the GUI
+
+ - Console log entries, where possible and relevant
+
+If you're not sure whether something is relevant, erring on the side of
+too much information will never get you yelled at. :)
+
+## Contributing Translations
+
+All translations are done via
+[Transifex](https://www.transifex.com/projects/p/syncthing/). If you
+wish to contribute to a translation, just head over there and sign up.
+Before every release, the language resources are updated from the
+latest info on Transifex.
+
+## Contributing Code
+
+Every contribution is welcome. If you want to contribute but are unsure
+where to start, any open issues are fair game! See the [Contribution
+Guidelines](http://docs.syncthing.net/dev/contributing.html) for the full
+story on committing code.
+
+## Contributing Documentation
+
+Updates to the [documentation site](http://docs.syncthing.net/) can be
+made as pull requests on the [documentation
+repository](https://github.com/syncthing/docs).
+
+## Licensing
+
+All contributions are made under the same MPLv2 license as the rest of
+the project, except documentation, user interface text and translation
+strings which are licensed under the Creative Commons Attribution 4.0
+International License. You retain the copyright to code you have
+written.
+
@@ -0,0 +1,86 @@
+{
+	"ImportPath": "github.com/syncthing/syncthing",
+	"GoVersion": "go1.5.1",
+	"Packages": [
+		"./cmd/..."
+	],
+	"Deps": [
+		{
+			"ImportPath": "github.com/bkaradzic/go-lz4",
+			"Rev": "74ddf82598bc4745b965729e9c6a463bedd33049"
+		},
+		{
+			"ImportPath": "github.com/calmh/du",
+			"Rev": "3c0690cca16228b97741327b1b6781397afbdb24"
+		},
+		{
+			"ImportPath": "github.com/calmh/luhn",
+			"Rev": "0c8388ff95fa92d4094011e5a04fc99dea3d1632"
+		},
+		{
+			"ImportPath": "github.com/calmh/xdr",
+			"Rev": "47c0042d09a827b81ee62497f99e5e0c7f0bd31c"
+		},
+		{
+			"ImportPath": "github.com/golang/snappy",
+			"Rev": "723cc1e459b8eea2dea4583200fd60757d40097a"
+		},
+		{
+			"ImportPath": "github.com/juju/ratelimit",
+			"Rev": "772f5c38e468398c4511514f4f6aa9a4185bc0a0"
+		},
+		{
+			"ImportPath": "github.com/kardianos/osext",
+			"Rev": "6e7f843663477789fac7c02def0d0909e969b4e5"
+		},
+		{
+			"ImportPath": "github.com/syndtr/goleveldb/leveldb",
+			"Rev": "1a9d62f03ea92815b46fcaab357cfd4df264b1a0"
+		},
+		{
+			"ImportPath": "github.com/thejerf/suture",
+			"Comment": "v1.0.1",
+			"Rev": "99c1f2d613756768fc4299acd9dc621e11ed3fd7"
+		},
+		{
+			"ImportPath": "github.com/vitrun/qart/coding",
+			"Rev": "ccb109cf25f0cd24474da73b9fee4e7a3e8a8ce0"
+		},
+		{
+			"ImportPath": "github.com/vitrun/qart/gf256",
+			"Rev": "ccb109cf25f0cd24474da73b9fee4e7a3e8a8ce0"
+		},
+		{
+			"ImportPath": "github.com/vitrun/qart/qr",
+			"Rev": "ccb109cf25f0cd24474da73b9fee4e7a3e8a8ce0"
+		},
+		{
+			"ImportPath": "golang.org/x/crypto/bcrypt",
+			"Rev": "81bf7719a6b7ce9b665598222362b50122dfc13b"
+		},
+		{
+			"ImportPath": "golang.org/x/crypto/blowfish",
+			"Rev": "81bf7719a6b7ce9b665598222362b50122dfc13b"
+		},
+		{
+			"ImportPath": "golang.org/x/net/internal/iana",
+			"Rev": "4b709d93778b93d2f34943e3142c71578d83ad31"
+		},
+		{
+			"ImportPath": "golang.org/x/net/ipv6",
+			"Rev": "4b709d93778b93d2f34943e3142c71578d83ad31"
+		},
+		{
+			"ImportPath": "golang.org/x/net/proxy",
+			"Rev": "4b709d93778b93d2f34943e3142c71578d83ad31"
+		},
+		{
+			"ImportPath": "golang.org/x/text/transform",
+			"Rev": "723492b65e225eafcba054e76ba18bb9c5ac1ea2"
+		},
+		{
+			"ImportPath": "golang.org/x/text/unicode/norm",
+			"Rev": "723492b65e225eafcba054e76ba18bb9c5ac1ea2"
+		}
+	]
+}
@@ -0,0 +1,5 @@
+This directory tree is generated automatically by godep.
+
+Please do not edit.
+
+See https://github.com/tools/godep for more information.
@@ -0,0 +1,2 @@
+/pkg
+/bin
@@ -0,0 +1 @@
+/lz4-example/lz4-example
@@ -0,0 +1,9 @@
+language: go
+
+go:
+  - 1.1
+  - 1.2
+  - 1.3
+  - 1.4
+  - 1.5
+  - tip
@@ -0,0 +1,24 @@
+Copyright 2011-2012 Branimir Karadzic. All rights reserved.
+Copyright 2013 Damian Gryski. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+   1. Redistributions of source code must retain the above copyright notice, this
+      list of conditions and the following disclaimer.
+
+   2. Redistributions in binary form must reproduce the above copyright notice,
+      this list of conditions and the following disclaimer in the documentation
+      and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR
+IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
+SHALL COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+THE POSSIBILITY OF SUCH DAMAGE.
+
@@ -0,0 +1,71 @@
+go-lz4
+======
+
+go-lz4 is port of LZ4 lossless compression algorithm to Go. The original C code
+is located at:
+
+https://github.com/Cyan4973/lz4
+
+Status
+------
+[![Build Status](https://secure.travis-ci.org/bkaradzic/go-lz4.png)](http://travis-ci.org/bkaradzic/go-lz4)  
+[![GoDoc](https://godoc.org/github.com/bkaradzic/go-lz4?status.png)](https://godoc.org/github.com/bkaradzic/go-lz4)
+
+Usage
+-----
+
+    go get github.com/bkaradzic/go-lz4
+
+    import "github.com/bkaradzic/go-lz4"
+
+The package name is `lz4`
+
+Notes
+-----
+
+* go-lz4 saves a uint32 with the original uncompressed length at the beginning
+  of the encoded buffer.  They may get in the way of interoperability with
+  other implementations.
+
+Contributors
+------------
+
+Damian Gryski ([@dgryski](https://github.com/dgryski))  
+Dustin Sallings ([@dustin](https://github.com/dustin))
+
+Contact
+-------
+
+[@bkaradzic](https://twitter.com/bkaradzic)  
+http://www.stuckingeometry.com
+
+Project page  
+https://github.com/bkaradzic/go-lz4
+
+License
+-------
+
+Copyright 2011-2012 Branimir Karadzic. All rights reserved.  
+Copyright 2013 Damian Gryski. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification,
+are permitted provided that the following conditions are met:
+
+   1. Redistributions of source code must retain the above copyright notice, this
+      list of conditions and the following disclaimer.
+
+   2. Redistributions in binary form must reproduce the above copyright notice,
+      this list of conditions and the following disclaimer in the documentation
+      and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR
+IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
+SHALL COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+THE POSSIBILITY OF SUCH DAMAGE.
+
@@ -0,0 +1,23 @@
+// +build gofuzz
+
+package lz4
+
+import "encoding/binary"
+
+func Fuzz(data []byte) int {
+
+	if len(data) < 4 {
+		return 0
+	}
+
+	ln := binary.LittleEndian.Uint32(data)
+	if ln > (1 << 21) {
+		return 0
+	}
+
+	if _, err := Decode(nil, data); err != nil {
+		return 0
+	}
+
+	return 1
+}
@@ -0,0 +1,74 @@
+package main
+
+import (
+	"math/rand"
+
+	"github.com/bkaradzic/go-lz4"
+
+	// lz4's API matches snappy's, so we can easily see how it performs
+	// lz4 "code.google.com/p/snappy-go/snappy"
+)
+
+var input = `
+ADVENTURE I. A SCANDAL IN BOHEMIA
+
+I.
+
+To Sherlock Holmes she is always THE woman. I have seldom heard
+him mention her under any other name. In his eyes she eclipses
+and predominates the whole of her sex. It was not that he felt
+any emotion akin to love for Irene Adler. All emotions, and that
+one particularly, were abhorrent to his cold, precise but
+admirably balanced mind. He was, I take it, the most perfect
+reasoning and observing machine that the world has seen, but as a
+lover he would have placed himself in a false position. He never
+spoke of the softer passions, save with a gibe and a sneer. They
+were admirable things for the observer--excellent for drawing the
+veil from men's motives and actions. But for the trained reasoner
+to admit such intrusions into his own delicate and finely
+adjusted temperament was to introduce a distracting factor which
+might throw a doubt upon all his mental results. Grit in a
+sensitive instrument, or a crack in one of his own high-power
+lenses, would not be more disturbing than a strong emotion in a
+nature such as his. And yet there was but one woman to him, and
+that woman was the late Irene Adler, of dubious and questionable
+memory.
+
+I had seen little of Holmes lately. My marriage had drifted us
+away from each other. My own complete happiness, and the
+home-centred interests which rise up around the man who first
+finds himself master of his own establishment, were sufficient to
+absorb all my attention, while Holmes, who loathed every form of
+society with his whole Bohemian soul, remained in our lodgings in
+Baker Street, buried among his old books, and alternating from
+week to week between cocaine and ambition, the drowsiness of the
+drug, and the fierce energy of his own keen nature. He was still,
+as ever, deeply attracted by the study of crime, and occupied his
+immense faculties and extraordinary powers of observation in
+following out those clues, and clearing up those mysteries which
+had been abandoned as hopeless by the official police. From time
+to time I heard some vague account of his doings: of his summons
+to Odessa in the case of the Trepoff murder, of his clearing up
+of the singular tragedy of the Atkinson brothers at Trincomalee,
+and finally of the mission which he had accomplished so
+delicately and successfully for the reigning family of Holland.
+Beyond these signs of his activity, however, which I merely
+shared with all the readers of the daily press, I knew little of
+my former friend and companion.
+`
+
+func main() {
+
+	compressed, _ := lz4.Encode(nil, []byte(input))
+
+	modified := make([]byte, len(compressed))
+
+	for {
+		copy(modified, compressed)
+		for i := 0; i < 100; i++ {
+			modified[rand.Intn(len(compressed)-4)+4] = byte(rand.Intn(256))
+		}
+		lz4.Decode(nil, modified)
+	}
+
+}
@@ -0,0 +1,94 @@
+/*
+ * Copyright 2011 Branimir Karadzic. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification,
+ * are permitted provided that the following conditions are met:
+ *
+ *    1. Redistributions of source code must retain the above copyright notice, this
+ *       list of conditions and the following disclaimer.
+ *
+ *    2. Redistributions in binary form must reproduce the above copyright notice,
+ *       this list of conditions and the following disclaimer in the documentation
+ *       and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
+ * SHALL COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+ * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+ * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+ * THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+package main
+
+import (
+	"flag"
+	"fmt"
+	"io/ioutil"
+	"log"
+	"os"
+	"runtime/pprof"
+
+	lz4 "github.com/bkaradzic/go-lz4"
+)
+
+var (
+	decompress = flag.Bool("d", false, "decompress")
+)
+
+func main() {
+
+	var optCPUProfile = flag.String("cpuprofile", "", "profile")
+	flag.Parse()
+
+	if *optCPUProfile != "" {
+		f, err := os.Create(*optCPUProfile)
+		if err != nil {
+			log.Fatal(err)
+		}
+		pprof.StartCPUProfile(f)
+		defer pprof.StopCPUProfile()
+	}
+
+	args := flag.Args()
+
+	var data []byte
+
+	if len(args) < 2 {
+		fmt.Print("Usage: lz4 [-d] <input> <output>\n")
+		os.Exit(1)
+	}
+
+	input, err := os.OpenFile(args[0], os.O_RDONLY, 0644)
+	if err != nil {
+		fmt.Printf("Failed to open input file %s\n", args[0])
+		os.Exit(1)
+	}
+	defer input.Close()
+
+	if *decompress {
+		data, _ = ioutil.ReadAll(input)
+		data, err = lz4.Decode(nil, data)
+		if err != nil {
+			fmt.Println("Failed to decode:", err)
+			return
+		}
+	} else {
+		data, _ = ioutil.ReadAll(input)
+		data, err = lz4.Encode(nil, data)
+		if err != nil {
+			fmt.Println("Failed to encode:", err)
+			return
+		}
+	}
+
+	err = ioutil.WriteFile(args[1], data, 0644)
+	if err != nil {
+		fmt.Printf("Failed to open output file %s\n", args[1])
+		os.Exit(1)
+	}
+}
@@ -0,0 +1,63 @@
+package lz4
+
+import (
+	"bytes"
+	"io/ioutil"
+	"testing"
+)
+
+var testfile, _ = ioutil.ReadFile("testdata/pg1661.txt")
+
+func roundtrip(t *testing.T, input []byte) {
+
+	dst, err := Encode(nil, input)
+	if err != nil {
+		t.Errorf("got error during compression: %s", err)
+	}
+
+	output, err := Decode(nil, dst)
+
+	if err != nil {
+		t.Errorf("got error during decompress: %s", err)
+	}
+
+	if !bytes.Equal(output, input) {
+		t.Errorf("roundtrip failed")
+	}
+}
+
+func TestEmpty(t *testing.T) {
+	roundtrip(t, nil)
+}
+
+func TestLengths(t *testing.T) {
+
+	for i := 0; i < 1024; i++ {
+		roundtrip(t, testfile[:i])
+	}
+
+	for i := 1024; i < 4096; i += 23 {
+		roundtrip(t, testfile[:i])
+	}
+}
+
+func TestWords(t *testing.T) {
+	roundtrip(t, testfile)
+}
+
+func BenchmarkLZ4Encode(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		Encode(nil, testfile)
+	}
+}
+
+func BenchmarkLZ4Decode(b *testing.B) {
+
+	var compressed, _ = Encode(nil, testfile)
+
+	b.ResetTimer()
+
+	for i := 0; i < b.N; i++ {
+		Decode(nil, compressed)
+	}
+}
@@ -0,0 +1,199 @@
+/*
+ * Copyright 2011-2012 Branimir Karadzic. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification,
+ * are permitted provided that the following conditions are met:
+ *
+ *    1. Redistributions of source code must retain the above copyright notice, this
+ *       list of conditions and the following disclaimer.
+ *
+ *    2. Redistributions in binary form must reproduce the above copyright notice,
+ *       this list of conditions and the following disclaimer in the documentation
+ *       and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
+ * SHALL COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+ * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+ * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+ * THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+package lz4
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+var (
+	// ErrCorrupt indicates the input was corrupt
+	ErrCorrupt = errors.New("corrupt input")
+)
+
+const (
+	mlBits  = 4
+	mlMask  = (1 << mlBits) - 1
+	runBits = 8 - mlBits
+	runMask = (1 << runBits) - 1
+)
+
+type decoder struct {
+	src  []byte
+	dst  []byte
+	spos uint32
+	dpos uint32
+	ref  uint32
+}
+
+func (d *decoder) readByte() (uint8, error) {
+	if int(d.spos) == len(d.src) {
+		return 0, io.EOF
+	}
+	b := d.src[d.spos]
+	d.spos++
+	return b, nil
+}
+
+func (d *decoder) getLen() (uint32, error) {
+
+	length := uint32(0)
+	ln, err := d.readByte()
+	if err != nil {
+		return 0, ErrCorrupt
+	}
+	for ln == 255 {
+		length += 255
+		ln, err = d.readByte()
+		if err != nil {
+			return 0, ErrCorrupt
+		}
+	}
+	length += uint32(ln)
+
+	return length, nil
+}
+
+func (d *decoder) cp(length, decr uint32) {
+
+	if int(d.ref+length) < int(d.dpos) {
+		copy(d.dst[d.dpos:], d.dst[d.ref:d.ref+length])
+	} else {
+		for ii := uint32(0); ii < length; ii++ {
+			d.dst[d.dpos+ii] = d.dst[d.ref+ii]
+		}
+	}
+	d.dpos += length
+	d.ref += length - decr
+}
+
+func (d *decoder) finish(err error) error {
+	if err == io.EOF {
+		return nil
+	}
+
+	return err
+}
+
+// Decode returns the decoded form of src.  The returned slice may be a
+// subslice of dst if it was large enough to hold the entire decoded block.
+func Decode(dst, src []byte) ([]byte, error) {
+
+	if len(src) < 4 {
+		return nil, ErrCorrupt
+	}
+
+	uncompressedLen := binary.LittleEndian.Uint32(src)
+
+	if uncompressedLen == 0 {
+		return nil, nil
+	}
+
+	if uncompressedLen > MaxInputSize {
+		return nil, ErrTooLarge
+	}
+
+	if dst == nil || len(dst) < int(uncompressedLen) {
+		dst = make([]byte, uncompressedLen)
+	}
+
+	d := decoder{src: src, dst: dst[:uncompressedLen], spos: 4}
+
+	decr := []uint32{0, 3, 2, 3}
+
+	for {
+		code, err := d.readByte()
+		if err != nil {
+			return d.dst, d.finish(err)
+		}
+
+		length := uint32(code >> mlBits)
+		if length == runMask {
+			ln, err := d.getLen()
+			if err != nil {
+				return nil, ErrCorrupt
+			}
+			length += ln
+		}
+
+		if int(d.spos+length) > len(d.src) || int(d.dpos+length) > len(d.dst) {
+			return nil, ErrCorrupt
+		}
+
+		for ii := uint32(0); ii < length; ii++ {
+			d.dst[d.dpos+ii] = d.src[d.spos+ii]
+		}
+
+		d.spos += length
+		d.dpos += length
+
+		if int(d.spos) == len(d.src) {
+			return d.dst, nil
+		}
+
+		if int(d.spos+2) >= len(d.src) {
+			return nil, ErrCorrupt
+		}
+
+		back := uint32(d.src[d.spos]) | uint32(d.src[d.spos+1])<<8
+
+		if back > d.dpos {
+			return nil, ErrCorrupt
+		}
+
+		d.spos += 2
+		d.ref = d.dpos - back
+
+		length = uint32(code & mlMask)
+		if length == mlMask {
+			ln, err := d.getLen()
+			if err != nil {
+				return nil, ErrCorrupt
+			}
+			length += ln
+		}
+
+		literal := d.dpos - d.ref
+
+		if literal < 4 {
+			if int(d.dpos+4) > len(d.dst) {
+				return nil, ErrCorrupt
+			}
+
+			d.cp(4, decr[literal])
+		} else {
+			length += 4
+		}
+
+		if d.dpos+length > uncompressedLen {
+			return nil, ErrCorrupt
+		}
+
+		d.cp(length, 0)
+	}
+}
@@ -0,0 +1,190 @@
+/*
+ * Copyright 2011-2012 Branimir Karadzic. All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without modification,
+ * are permitted provided that the following conditions are met:
+ *
+ *    1. Redistributions of source code must retain the above copyright notice, this
+ *       list of conditions and the following disclaimer.
+ *
+ *    2. Redistributions in binary form must reproduce the above copyright notice,
+ *       this list of conditions and the following disclaimer in the documentation
+ *       and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY COPYRIGHT HOLDER ``AS IS'' AND ANY EXPRESS OR
+ * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
+ * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT
+ * SHALL COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+ * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+ * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
+ * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+ * THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+package lz4
+
+import (
+	"encoding/binary"
+	"errors"
+)
+
+const (
+	minMatch              = 4
+	hashLog               = 17
+	hashTableSize         = 1 << hashLog
+	hashShift             = (minMatch * 8) - hashLog
+	incompressible uint32 = 128
+	uninitHash            = 0x88888888
+
+	// MaxInputSize is the largest buffer than can be compressed in a single block
+	MaxInputSize = 0x7E000000
+)
+
+var (
+	// ErrTooLarge indicates the input buffer was too large
+	ErrTooLarge = errors.New("input too large")
+)
+
+type encoder struct {
+	src       []byte
+	dst       []byte
+	hashTable []uint32
+	pos       uint32
+	anchor    uint32
+	dpos      uint32
+}
+
+// CompressBound returns the maximum length of a lz4 block, given it's uncompressed length
+func CompressBound(isize int) int {
+	if isize > MaxInputSize {
+		return 0
+	}
+	return isize + ((isize) / 255) + 16 + 4
+}
+
+func (e *encoder) writeLiterals(length, mlLen, pos uint32) {
+
+	ln := length
+
+	var code byte
+	if ln > runMask-1 {
+		code = runMask
+	} else {
+		code = byte(ln)
+	}
+
+	if mlLen > mlMask-1 {
+		e.dst[e.dpos] = (code << mlBits) + byte(mlMask)
+	} else {
+		e.dst[e.dpos] = (code << mlBits) + byte(mlLen)
+	}
+	e.dpos++
+
+	if code == runMask {
+		ln -= runMask
+		for ; ln > 254; ln -= 255 {
+			e.dst[e.dpos] = 255
+			e.dpos++
+		}
+
+		e.dst[e.dpos] = byte(ln)
+		e.dpos++
+	}
+
+	for ii := uint32(0); ii < length; ii++ {
+		e.dst[e.dpos+ii] = e.src[pos+ii]
+	}
+
+	e.dpos += length
+}
+
+// Encode returns the encoded form of src.  The returned array may be a
+// sub-slice of dst if it was large enough to hold the entire output.
+func Encode(dst, src []byte) ([]byte, error) {
+
+	if len(src) >= MaxInputSize {
+		return nil, ErrTooLarge
+	}
+
+	if n := CompressBound(len(src)); len(dst) < n {
+		dst = make([]byte, n)
+	}
+
+	e := encoder{src: src, dst: dst, hashTable: make([]uint32, hashTableSize)}
+
+	binary.LittleEndian.PutUint32(dst, uint32(len(src)))
+	e.dpos = 4
+
+	var (
+		step  uint32 = 1
+		limit        = incompressible
+	)
+
+	for {
+		if int(e.pos)+12 >= len(e.src) {
+			e.writeLiterals(uint32(len(e.src))-e.anchor, 0, e.anchor)
+			return e.dst[:e.dpos], nil
+		}
+
+		sequence := uint32(e.src[e.pos+3])<<24 | uint32(e.src[e.pos+2])<<16 | uint32(e.src[e.pos+1])<<8 | uint32(e.src[e.pos+0])
+
+		hash := (sequence * 2654435761) >> hashShift
+		ref := e.hashTable[hash] + uninitHash
+		e.hashTable[hash] = e.pos - uninitHash
+
+		if ((e.pos-ref)>>16) != 0 || uint32(e.src[ref+3])<<24|uint32(e.src[ref+2])<<16|uint32(e.src[ref+1])<<8|uint32(e.src[ref+0]) != sequence {
+			if e.pos-e.anchor > limit {
+				limit <<= 1
+				step += 1 + (step >> 2)
+			}
+			e.pos += step
+			continue
+		}
+
+		if step > 1 {
+			e.hashTable[hash] = ref - uninitHash
+			e.pos -= step - 1
+			step = 1
+			continue
+		}
+		limit = incompressible
+
+		ln := e.pos - e.anchor
+		back := e.pos - ref
+
+		anchor := e.anchor
+
+		e.pos += minMatch
+		ref += minMatch
+		e.anchor = e.pos
+
+		for int(e.pos) < len(e.src)-5 && e.src[e.pos] == e.src[ref] {
+			e.pos++
+			ref++
+		}
+
+		mlLen := e.pos - e.anchor
+
+		e.writeLiterals(ln, mlLen, anchor)
+		e.dst[e.dpos] = uint8(back)
+		e.dst[e.dpos+1] = uint8(back >> 8)
+		e.dpos += 2
+
+		if mlLen > mlMask-1 {
+			mlLen -= mlMask
+			for mlLen > 254 {
+				mlLen -= 255
+
+				e.dst[e.dpos] = 255
+				e.dpos++
+			}
+
+			e.dst[e.dpos] = byte(mlLen)
+			e.dpos++
+		}
+
+		e.anchor = e.pos
+	}
+}
@@ -0,0 +1,24 @@
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or
+distribute this software, either in source code form or as a compiled
+binary, for any purpose, commercial or non-commercial, and by any
+means.
+
+In jurisdictions that recognize copyright laws, the author or authors
+of this software dedicate any and all copyright interest in the
+software to the public domain. We make this dedication for the benefit
+of the public at large and to the detriment of our heirs and
+successors. We intend this dedication to be an overt act of
+relinquishment in perpetuity of all present and future rights to this
+software under copyright law.
+
+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 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.
+
+For more information, please refer to <http://unlicense.org>
@@ -0,0 +1,14 @@
+du
+==
+
+Get total and available disk space on a given volume.
+
+Documentation
+-------------
+
+http://godoc.org/github.com/calmh/du
+
+License
+-------
+
+Public Domain
@@ -0,0 +1,21 @@
+package main
+
+import (
+	"fmt"
+	"log"
+	"os"
+
+	"github.com/calmh/du"
+)
+
+var KB = int64(1024)
+
+func main() {
+	usage, err := du.Get(os.Args[1])
+	if err != nil {
+		log.Fatal(err)
+	}
+	fmt.Println("Free:", usage.FreeBytes/(KB*KB), "MiB")
+	fmt.Println("Available:", usage.AvailBytes/(KB*KB), "MiB")
+	fmt.Println("Size:", usage.TotalBytes/(KB*KB), "MiB")
+}
@@ -0,0 +1,8 @@
+package du
+
+// Usage holds information about total and available storage on a volume.
+type Usage struct {
+	TotalBytes int64 // Size of volume
+	FreeBytes  int64 // Unused size
+	AvailBytes int64 // Available to a non-privileged user
+}
@@ -0,0 +1,24 @@
+// +build !windows,!netbsd,!openbsd,!solaris
+
+package du
+
+import (
+	"path/filepath"
+	"syscall"
+)
+
+// Get returns the Usage of a given path, or an error if usage data is
+// unavailable.
+func Get(path string) (Usage, error) {
+	var stat syscall.Statfs_t
+	err := syscall.Statfs(filepath.Clean(path), &stat)
+	if err != nil {
+		return Usage{}, err
+	}
+	u := Usage{
+		FreeBytes:  int64(stat.Bfree) * int64(stat.Bsize),
+		TotalBytes: int64(stat.Blocks) * int64(stat.Bsize),
+		AvailBytes: int64(stat.Bavail) * int64(stat.Bsize),
+	}
+	return u, nil
+}
@@ -0,0 +1,13 @@
+// +build netbsd openbsd solaris
+
+package du
+
+import "errors"
+
+var ErrUnsupported = errors.New("unsupported platform")
+
+// Get returns the Usage of a given path, or an error if usage data is
+// unavailable.
+func Get(path string) (Usage, error) {
+	return Usage{}, ErrUnsupported
+}
@@ -0,0 +1,27 @@
+package du
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+// Get returns the Usage of a given path, or an error if usage data is
+// unavailable.
+func Get(path string) (Usage, error) {
+	h := syscall.MustLoadDLL("kernel32.dll")
+	c := h.MustFindProc("GetDiskFreeSpaceExW")
+
+	var u Usage
+
+	ret, _, err := c.Call(
+		uintptr(unsafe.Pointer(syscall.StringToUTF16Ptr(path))),
+		uintptr(unsafe.Pointer(&u.FreeBytes)),
+		uintptr(unsafe.Pointer(&u.TotalBytes)),
+		uintptr(unsafe.Pointer(&u.AvailBytes)))
+
+	if ret == 0 {
+		return u, err
+	}
+
+	return u, nil
+}
@@ -0,0 +1,19 @@
+Copyright (C) 2014 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.
@@ -0,0 +1,70 @@
+// Copyright (C) 2014 Jakob Borg
+
+// Package luhn generates and validates Luhn mod N check digits.
+package luhn
+
+import (
+	"fmt"
+	"strings"
+)
+
+// An alphabet is a string of N characters, representing the digits of a given
+// base N.
+type Alphabet string
+
+var (
+	Base32 Alphabet = "ABCDEFGHIJKLMNOPQRSTUVWXYZ234567"
+)
+
+// Generate returns a check digit for the string s, which should be composed
+// of characters from the Alphabet a.
+func (a Alphabet) Generate(s string) (rune, error) {
+	if err := a.check(); err != nil {
+		return 0, err
+	}
+
+	factor := 1
+	sum := 0
+	n := len(a)
+
+	for i := range s {
+		codepoint := strings.IndexByte(string(a), s[i])
+		if codepoint == -1 {
+			return 0, fmt.Errorf("Digit %q not valid in alphabet %q", s[i], a)
+		}
+		addend := factor * codepoint
+		if factor == 2 {
+			factor = 1
+		} else {
+			factor = 2
+		}
+		addend = (addend / n) + (addend % n)
+		sum += addend
+	}
+	remainder := sum % n
+	checkCodepoint := (n - remainder) % n
+	return rune(a[checkCodepoint]), nil
+}
+
+// Validate returns true if the last character of the string s is correct, for
+// a string s composed of characters in the alphabet a.
+func (a Alphabet) Validate(s string) bool {
+	t := s[:len(s)-1]
+	c, err := a.Generate(t)
+	if err != nil {
+		return false
+	}
+	return rune(s[len(s)-1]) == c
+}
+
+// check returns an error if the given alphabet does not consist of unique characters
+func (a Alphabet) check() error {
+	cm := make(map[byte]bool, len(a))
+	for i := range a {
+		if cm[a[i]] {
+			return fmt.Errorf("Digit %q non-unique in alphabet %q", a[i], a)
+		}
+		cm[a[i]] = true
+	}
+	return nil
+}
@@ -0,0 +1,59 @@
+// Copyright (C) 2014 Jakob Borg
+
+package luhn_test
+
+import (
+	"testing"
+
+	"github.com/calmh/luhn"
+)
+
+func TestGenerate(t *testing.T) {
+	// Base 6 Luhn
+	a := luhn.Alphabet("abcdef")
+	c, err := a.Generate("abcdef")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if c != 'e' {
+		t.Errorf("Incorrect check digit %c != e", c)
+	}
+
+	// Base 10 Luhn
+	a = luhn.Alphabet("0123456789")
+	c, err = a.Generate("7992739871")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if c != '3' {
+		t.Errorf("Incorrect check digit %c != 3", c)
+	}
+}
+
+func TestInvalidString(t *testing.T) {
+	a := luhn.Alphabet("ABC")
+	_, err := a.Generate("7992739871")
+	t.Log(err)
+	if err == nil {
+		t.Error("Unexpected nil error")
+	}
+}
+
+func TestBadAlphabet(t *testing.T) {
+	a := luhn.Alphabet("01234566789")
+	_, err := a.Generate("7992739871")
+	t.Log(err)
+	if err == nil {
+		t.Error("Unexpected nil error")
+	}
+}
+
+func TestValidate(t *testing.T) {
+	a := luhn.Alphabet("abcdef")
+	if !a.Validate("abcdefe") {
+		t.Errorf("Incorrect validation response for abcdefe")
+	}
+	if a.Validate("abcdefd") {
+		t.Errorf("Incorrect validation response for abcdefd")
+	}
+}
@@ -0,0 +1 @@
+coverage.out
@@ -0,0 +1,19 @@
+language: go
+go:
+- tip
+
+install:
+- export PATH=$PATH:$HOME/gopath/bin
+- go get golang.org/x/tools/cover
+- go get github.com/mattn/goveralls
+
+script:
+- ./generate.sh
+- go test -coverprofile=coverage.out
+
+after_success:
+- goveralls -coverprofile=coverage.out -service=travis-ci -package=calmh/xdr -repotoken="$COVERALLS_TOKEN"
+
+env:
+  global:
+    secure: SmgnrGfp2zLrA44ChRMpjPeujubt9veZ8Fx/OseMWECmacyV5N/TuDhzIbwo6QwV4xB0sBacoPzvxQbJRVjNKsPiSu72UbcQmQ7flN4Tf7nW09tSh1iW8NgrpBCq/3UYLoBu2iPBEBKm93IK0aGNAKs6oEkB0fU27iTVBwiTXOY=
@@ -0,0 +1,19 @@
+Copyright (C) 2014 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.
@@ -0,0 +1,12 @@
+xdr
+===
+
+[![Build Status](https://img.shields.io/circleci/project/calmh/xdr.svg?style=flat-square)](https://circleci.com/gh/calmh/xdr)
+[![Coverage Status](https://img.shields.io/coveralls/calmh/xdr.svg?style=flat)](https://coveralls.io/r/calmh/xdr?branch=master)
+[![API Documentation](http://img.shields.io/badge/api-Godoc-blue.svg?style=flat)](http://godoc.org/github.com/calmh/xdr)
+[![MIT License](http://img.shields.io/badge/license-MIT-blue.svg?style=flat)](http://opensource.org/licenses/MIT)
+
+This is an XDR encoding/decoding library. It uses code generation and
+not reflection. It supports the IPDR bastardized XDR format when built
+with `-tags ipdr`.
+
@@ -0,0 +1,117 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+package xdr_test
+
+import (
+	"io"
+	"io/ioutil"
+	"testing"
+
+	"github.com/calmh/xdr"
+)
+
+type XDRBenchStruct struct {
+	I1  uint64
+	I2  uint32
+	I3  uint16
+	I4  uint8
+	Bs0 []byte // max:128
+	Bs1 []byte
+	S0  string // max:128
+	S1  string
+}
+
+var res []byte // no to be optimized away
+var s = XDRBenchStruct{
+	I1:  42,
+	I2:  43,
+	I3:  44,
+	I4:  45,
+	Bs0: []byte{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18},
+	Bs1: []byte{11, 12, 13, 14, 15, 16, 17, 18, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10},
+	S0:  "Hello World! String one.",
+	S1:  "Hello World! String two.",
+}
+var e []byte
+
+func init() {
+	e, _ = s.MarshalXDR()
+}
+
+func BenchmarkThisMarshal(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		res, _ = s.MarshalXDR()
+	}
+}
+
+func BenchmarkThisUnmarshal(b *testing.B) {
+	var t XDRBenchStruct
+	for i := 0; i < b.N; i++ {
+		err := t.UnmarshalXDR(e)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkThisEncode(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		_, err := s.EncodeXDR(ioutil.Discard)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkThisEncoder(b *testing.B) {
+	w := xdr.NewWriter(ioutil.Discard)
+	for i := 0; i < b.N; i++ {
+		_, err := s.EncodeXDRInto(w)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+type repeatReader struct {
+	data []byte
+}
+
+func (r *repeatReader) Read(bs []byte) (n int, err error) {
+	if len(bs) > len(r.data) {
+		err = io.EOF
+	}
+	n = copy(bs, r.data)
+	r.data = r.data[n:]
+	return n, err
+}
+
+func (r *repeatReader) Reset(bs []byte) {
+	r.data = bs
+}
+
+func BenchmarkThisDecode(b *testing.B) {
+	rr := &repeatReader{e}
+	var t XDRBenchStruct
+	for i := 0; i < b.N; i++ {
+		err := t.DecodeXDR(rr)
+		if err != nil {
+			b.Fatal(err)
+		}
+		rr.Reset(e)
+	}
+}
+
+func BenchmarkThisDecoder(b *testing.B) {
+	rr := &repeatReader{e}
+	r := xdr.NewReader(rr)
+	var t XDRBenchStruct
+	for i := 0; i < b.N; i++ {
+		err := t.DecodeXDRFrom(r)
+		if err != nil {
+			b.Fatal(err)
+		}
+		rr.Reset(e)
+	}
+}
@@ -0,0 +1,201 @@
+// ************************************************************
+// This file is automatically generated by genxdr. Do not edit.
+// ************************************************************
+
+package xdr_test
+
+import (
+	"bytes"
+	"io"
+
+	"github.com/calmh/xdr"
+)
+
+/*
+
+XDRBenchStruct Structure:
+
+ 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
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                                                               |
+                         I1 (64 bits)                          +
+|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                              I2                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|            0x0000             |              I3               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                        uint8 Structure                        \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Length of Bs0                         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                     Bs0 (variable length)                     \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Length of Bs1                         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                     Bs1 (variable length)                     \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Length of S0                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                     S0 (variable length)                      \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Length of S1                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                     S1 (variable length)                      \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+
+
+struct XDRBenchStruct {
+	unsigned hyper I1;
+	unsigned int I2;
+	unsigned int I3;
+	uint8 I4;
+	opaque Bs0<128>;
+	opaque Bs1<>;
+	string S0<128>;
+	string S1<>;
+}
+
+*/
+
+func (o XDRBenchStruct) EncodeXDR(w io.Writer) (int, error) {
+	var xw = xdr.NewWriter(w)
+	return o.EncodeXDRInto(xw)
+}
+
+func (o XDRBenchStruct) MarshalXDR() ([]byte, error) {
+	return o.AppendXDR(make([]byte, 0, 128))
+}
+
+func (o XDRBenchStruct) MustMarshalXDR() []byte {
+	bs, err := o.MarshalXDR()
+	if err != nil {
+		panic(err)
+	}
+	return bs
+}
+
+func (o XDRBenchStruct) AppendXDR(bs []byte) ([]byte, error) {
+	var aw = xdr.AppendWriter(bs)
+	var xw = xdr.NewWriter(&aw)
+	_, err := o.EncodeXDRInto(xw)
+	return []byte(aw), err
+}
+
+func (o XDRBenchStruct) EncodeXDRInto(xw *xdr.Writer) (int, error) {
+	xw.WriteUint64(o.I1)
+	xw.WriteUint32(o.I2)
+	xw.WriteUint16(o.I3)
+	xw.WriteUint8(o.I4)
+	if l := len(o.Bs0); l > 128 {
+		return xw.Tot(), xdr.ElementSizeExceeded("Bs0", l, 128)
+	}
+	xw.WriteBytes(o.Bs0)
+	xw.WriteBytes(o.Bs1)
+	if l := len(o.S0); l > 128 {
+		return xw.Tot(), xdr.ElementSizeExceeded("S0", l, 128)
+	}
+	xw.WriteString(o.S0)
+	xw.WriteString(o.S1)
+	return xw.Tot(), xw.Error()
+}
+
+func (o *XDRBenchStruct) DecodeXDR(r io.Reader) error {
+	xr := xdr.NewReader(r)
+	return o.DecodeXDRFrom(xr)
+}
+
+func (o *XDRBenchStruct) UnmarshalXDR(bs []byte) error {
+	var br = bytes.NewReader(bs)
+	var xr = xdr.NewReader(br)
+	return o.DecodeXDRFrom(xr)
+}
+
+func (o *XDRBenchStruct) DecodeXDRFrom(xr *xdr.Reader) error {
+	o.I1 = xr.ReadUint64()
+	o.I2 = xr.ReadUint32()
+	o.I3 = xr.ReadUint16()
+	o.I4 = xr.ReadUint8()
+	o.Bs0 = xr.ReadBytesMax(128)
+	o.Bs1 = xr.ReadBytes()
+	o.S0 = xr.ReadStringMax(128)
+	o.S1 = xr.ReadString()
+	return xr.Error()
+}
+
+/*
+
+repeatReader Structure:
+
+ 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
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                        Length of data                         |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                    data (variable length)                     \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+
+
+struct repeatReader {
+	opaque data<>;
+}
+
+*/
+
+func (o repeatReader) EncodeXDR(w io.Writer) (int, error) {
+	var xw = xdr.NewWriter(w)
+	return o.EncodeXDRInto(xw)
+}
+
+func (o repeatReader) MarshalXDR() ([]byte, error) {
+	return o.AppendXDR(make([]byte, 0, 128))
+}
+
+func (o repeatReader) MustMarshalXDR() []byte {
+	bs, err := o.MarshalXDR()
+	if err != nil {
+		panic(err)
+	}
+	return bs
+}
+
+func (o repeatReader) AppendXDR(bs []byte) ([]byte, error) {
+	var aw = xdr.AppendWriter(bs)
+	var xw = xdr.NewWriter(&aw)
+	_, err := o.EncodeXDRInto(xw)
+	return []byte(aw), err
+}
+
+func (o repeatReader) EncodeXDRInto(xw *xdr.Writer) (int, error) {
+	xw.WriteBytes(o.data)
+	return xw.Tot(), xw.Error()
+}
+
+func (o *repeatReader) DecodeXDR(r io.Reader) error {
+	xr := xdr.NewReader(r)
+	return o.DecodeXDRFrom(xr)
+}
+
+func (o *repeatReader) UnmarshalXDR(bs []byte) error {
+	var br = bytes.NewReader(bs)
+	var xr = xdr.NewReader(br)
+	return o.DecodeXDRFrom(xr)
+}
+
+func (o *repeatReader) DecodeXDRFrom(xr *xdr.Reader) error {
+	o.data = xr.ReadBytes()
+	return xr.Error()
+}
@@ -0,0 +1,482 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+package main
+
+import (
+	"bytes"
+	"flag"
+	"fmt"
+	"go/ast"
+	"go/format"
+	"go/parser"
+	"go/token"
+	"io"
+	"log"
+	"os"
+	"regexp"
+	"strconv"
+	"strings"
+	"text/template"
+)
+
+type fieldInfo struct {
+	Name      string
+	IsBasic   bool   // handled by one the native Read/WriteUint64 etc functions
+	IsSlice   bool   // field is a slice of FieldType
+	FieldType string // original type of field, i.e. "int"
+	Encoder   string // the encoder name, i.e. "Uint64" for Read/WriteUint64
+	Convert   string // what to convert to when encoding, i.e. "uint64"
+	Max       int    // max size for slices and strings
+	Submax    int    // max size for strings inside slices
+}
+
+type structInfo struct {
+	Name   string
+	Fields []fieldInfo
+}
+
+var headerTpl = template.Must(template.New("header").Parse(`// ************************************************************
+// This file is automatically generated by genxdr. Do not edit.
+// ************************************************************
+
+package {{.Package}}
+
+import (
+	"bytes"
+	"io"
+
+	"github.com/calmh/xdr"
+)
+`))
+
+var encodeTpl = template.Must(template.New("encoder").Parse(`
+func (o {{.TypeName}}) EncodeXDR(w io.Writer) (int, error) {
+	var xw = xdr.NewWriter(w)
+	return o.EncodeXDRInto(xw)
+}//+n
+
+func (o {{.TypeName}}) MarshalXDR() ([]byte, error) {
+	return o.AppendXDR(make([]byte, 0, 128))
+}//+n
+
+func (o {{.TypeName}}) MustMarshalXDR() []byte {
+	bs, err := o.MarshalXDR()
+	if err != nil {
+		panic(err)
+	}
+	return bs
+}//+n
+
+func (o {{.TypeName}}) AppendXDR(bs []byte) ([]byte, error) {
+	var aw = xdr.AppendWriter(bs)
+	var xw = xdr.NewWriter(&aw)
+	_, err := o.EncodeXDRInto(xw)
+	return []byte(aw), err
+}//+n
+
+func (o {{.TypeName}}) EncodeXDRInto(xw *xdr.Writer) (int, error) {
+	{{range $fieldInfo := .Fields}}
+		{{if not $fieldInfo.IsSlice}}
+			{{if ne $fieldInfo.Convert ""}}
+				xw.Write{{$fieldInfo.Encoder}}({{$fieldInfo.Convert}}(o.{{$fieldInfo.Name}}))
+			{{else if $fieldInfo.IsBasic}}
+				{{if ge $fieldInfo.Max 1}}
+					if l := len(o.{{$fieldInfo.Name}}); l > {{$fieldInfo.Max}} {
+						return xw.Tot(), xdr.ElementSizeExceeded("{{$fieldInfo.Name}}", l, {{$fieldInfo.Max}})
+					}
+				{{end}}
+				xw.Write{{$fieldInfo.Encoder}}(o.{{$fieldInfo.Name}})
+			{{else}}
+				_, err := o.{{$fieldInfo.Name}}.EncodeXDRInto(xw)
+				if err != nil {
+					return xw.Tot(), err
+				}
+			{{end}}
+		{{else}}
+			{{if ge $fieldInfo.Max 1}}
+				if l := len(o.{{$fieldInfo.Name}}); l > {{$fieldInfo.Max}} {
+					return xw.Tot(), xdr.ElementSizeExceeded("{{$fieldInfo.Name}}", l, {{$fieldInfo.Max}})
+				}
+			{{end}}
+			xw.WriteUint32(uint32(len(o.{{$fieldInfo.Name}})))
+			for i := range o.{{$fieldInfo.Name}} {
+			{{if ne $fieldInfo.Convert ""}}
+				xw.Write{{$fieldInfo.Encoder}}({{$fieldInfo.Convert}}(o.{{$fieldInfo.Name}}[i]))
+			{{else if $fieldInfo.IsBasic}}
+				xw.Write{{$fieldInfo.Encoder}}(o.{{$fieldInfo.Name}}[i])
+			{{else}}
+				_, err := o.{{$fieldInfo.Name}}[i].EncodeXDRInto(xw)
+				if err != nil {
+					return xw.Tot(), err
+				}
+			{{end}}
+			}
+		{{end}}
+	{{end}}
+	return xw.Tot(), xw.Error()
+}//+n
+
+func (o *{{.TypeName}}) DecodeXDR(r io.Reader) error {
+	xr := xdr.NewReader(r)
+	return o.DecodeXDRFrom(xr)
+}//+n
+
+func (o *{{.TypeName}}) UnmarshalXDR(bs []byte) error {
+	var br = bytes.NewReader(bs)
+	var xr = xdr.NewReader(br)
+	return o.DecodeXDRFrom(xr)
+}//+n
+
+func (o *{{.TypeName}}) DecodeXDRFrom(xr *xdr.Reader) error {
+	{{range $fieldInfo := .Fields}}
+		{{if not $fieldInfo.IsSlice}}
+			{{if ne $fieldInfo.Convert ""}}
+				o.{{$fieldInfo.Name}} = {{$fieldInfo.FieldType}}(xr.Read{{$fieldInfo.Encoder}}())
+			{{else if $fieldInfo.IsBasic}}
+				{{if ge $fieldInfo.Max 1}}
+					o.{{$fieldInfo.Name}} = xr.Read{{$fieldInfo.Encoder}}Max({{$fieldInfo.Max}})
+				{{else}}
+					o.{{$fieldInfo.Name}} = xr.Read{{$fieldInfo.Encoder}}()
+				{{end}}
+			{{else}}
+				(&o.{{$fieldInfo.Name}}).DecodeXDRFrom(xr)
+			{{end}}
+		{{else}}
+			_{{$fieldInfo.Name}}Size := int(xr.ReadUint32())
+			if _{{$fieldInfo.Name}}Size < 0 {
+				return xdr.ElementSizeExceeded("{{$fieldInfo.Name}}", _{{$fieldInfo.Name}}Size, {{$fieldInfo.Max}})
+			}
+			{{if ge $fieldInfo.Max 1}}
+				if _{{$fieldInfo.Name}}Size > {{$fieldInfo.Max}} {
+					return xdr.ElementSizeExceeded("{{$fieldInfo.Name}}", _{{$fieldInfo.Name}}Size, {{$fieldInfo.Max}})
+				}
+			{{end}}
+			o.{{$fieldInfo.Name}} = make([]{{$fieldInfo.FieldType}}, _{{$fieldInfo.Name}}Size)
+			for i := range o.{{$fieldInfo.Name}} {
+				{{if ne $fieldInfo.Convert ""}}
+					o.{{$fieldInfo.Name}}[i] = {{$fieldInfo.FieldType}}(xr.Read{{$fieldInfo.Encoder}}())
+				{{else if $fieldInfo.IsBasic}}
+					{{if ge $fieldInfo.Submax 1}}
+						o.{{$fieldInfo.Name}}[i] = xr.Read{{$fieldInfo.Encoder}}Max({{$fieldInfo.Submax}})
+					{{else}}
+						o.{{$fieldInfo.Name}}[i] = xr.Read{{$fieldInfo.Encoder}}()
+					{{end}}
+				{{else}}
+					(&o.{{$fieldInfo.Name}}[i]).DecodeXDRFrom(xr)
+				{{end}}
+			}
+		{{end}}
+	{{end}}
+	return xr.Error()
+}`))
+
+var maxRe = regexp.MustCompile(`(?:\Wmax:)(\d+)(?:\s*,\s*(\d+))?`)
+
+type typeSet struct {
+	Type    string
+	Encoder string
+}
+
+var xdrEncoders = map[string]typeSet{
+	"int8":   typeSet{"uint8", "Uint8"},
+	"uint8":  typeSet{"", "Uint8"},
+	"int16":  typeSet{"uint16", "Uint16"},
+	"uint16": typeSet{"", "Uint16"},
+	"int32":  typeSet{"uint32", "Uint32"},
+	"uint32": typeSet{"", "Uint32"},
+	"int64":  typeSet{"uint64", "Uint64"},
+	"uint64": typeSet{"", "Uint64"},
+	"int":    typeSet{"uint64", "Uint64"},
+	"string": typeSet{"", "String"},
+	"[]byte": typeSet{"", "Bytes"},
+	"bool":   typeSet{"", "Bool"},
+}
+
+func handleStruct(t *ast.StructType) []fieldInfo {
+	var fs []fieldInfo
+
+	for _, sf := range t.Fields.List {
+		if len(sf.Names) == 0 {
+			// We don't handle anonymous fields
+			continue
+		}
+
+		fn := sf.Names[0].Name
+		var max1, max2 int
+		if sf.Comment != nil {
+			c := sf.Comment.List[0].Text
+			m := maxRe.FindStringSubmatch(c)
+			if len(m) >= 2 {
+				max1, _ = strconv.Atoi(m[1])
+			}
+			if len(m) >= 3 {
+				max2, _ = strconv.Atoi(m[2])
+			}
+			if strings.Contains(c, "noencode") {
+				continue
+			}
+		}
+
+		var f fieldInfo
+		switch ft := sf.Type.(type) {
+		case *ast.Ident:
+			tn := ft.Name
+			if enc, ok := xdrEncoders[tn]; ok {
+				f = fieldInfo{
+					Name:      fn,
+					IsBasic:   true,
+					FieldType: tn,
+					Encoder:   enc.Encoder,
+					Convert:   enc.Type,
+					Max:       max1,
+					Submax:    max2,
+				}
+			} else {
+				f = fieldInfo{
+					Name:      fn,
+					IsBasic:   false,
+					FieldType: tn,
+					Max:       max1,
+					Submax:    max2,
+				}
+			}
+
+		case *ast.ArrayType:
+			if ft.Len != nil {
+				// We don't handle arrays
+				continue
+			}
+
+			tn := ft.Elt.(*ast.Ident).Name
+			if enc, ok := xdrEncoders["[]"+tn]; ok {
+				f = fieldInfo{
+					Name:      fn,
+					IsBasic:   true,
+					FieldType: tn,
+					Encoder:   enc.Encoder,
+					Convert:   enc.Type,
+					Max:       max1,
+					Submax:    max2,
+				}
+			} else if enc, ok := xdrEncoders[tn]; ok {
+				f = fieldInfo{
+					Name:      fn,
+					IsBasic:   true,
+					IsSlice:   true,
+					FieldType: tn,
+					Encoder:   enc.Encoder,
+					Convert:   enc.Type,
+					Max:       max1,
+					Submax:    max2,
+				}
+			} else {
+				f = fieldInfo{
+					Name:      fn,
+					IsSlice:   true,
+					FieldType: tn,
+					Max:       max1,
+					Submax:    max2,
+				}
+			}
+
+		case *ast.SelectorExpr:
+			f = fieldInfo{
+				Name:      fn,
+				FieldType: ft.Sel.Name,
+				Max:       max1,
+				Submax:    max2,
+			}
+		}
+
+		fs = append(fs, f)
+	}
+
+	return fs
+}
+
+func generateCode(output io.Writer, s structInfo) {
+	name := s.Name
+	fs := s.Fields
+
+	var buf bytes.Buffer
+	err := encodeTpl.Execute(&buf, map[string]interface{}{"TypeName": name, "Fields": fs})
+	if err != nil {
+		panic(err)
+	}
+
+	bs := regexp.MustCompile(`(\s*\n)+`).ReplaceAll(buf.Bytes(), []byte("\n"))
+	bs = bytes.Replace(bs, []byte("//+n"), []byte("\n"), -1)
+
+	bs, err = format.Source(bs)
+	if err != nil {
+		panic(err)
+	}
+	fmt.Fprintln(output, string(bs))
+}
+
+func uncamelize(s string) string {
+	return regexp.MustCompile("[a-z][A-Z]").ReplaceAllStringFunc(s, func(camel string) string {
+		return camel[:1] + " " + camel[1:]
+	})
+}
+
+func generateDiagram(output io.Writer, s structInfo) {
+	sn := s.Name
+	fs := s.Fields
+
+	fmt.Fprintln(output, sn+" Structure:")
+	fmt.Fprintln(output)
+	fmt.Fprintln(output, " 0                   1                   2                   3")
+	fmt.Fprintln(output, " 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")
+	line := "+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+"
+	fmt.Fprintln(output, line)
+
+	for _, f := range fs {
+		tn := f.FieldType
+		name := uncamelize(f.Name)
+
+		if f.IsSlice {
+			fmt.Fprintf(output, "| %s |\n", center("Number of "+name, 61))
+			fmt.Fprintln(output, line)
+		}
+		switch tn {
+		case "bool":
+			fmt.Fprintf(output, "| %s |V|\n", center(name+" (V=0 or 1)", 59))
+			fmt.Fprintln(output, line)
+		case "int16", "uint16":
+			fmt.Fprintf(output, "| %s | %s |\n", center("0x0000", 29), center(name, 29))
+			fmt.Fprintln(output, line)
+		case "int32", "uint32":
+			fmt.Fprintf(output, "| %s |\n", center(name, 61))
+			fmt.Fprintln(output, line)
+		case "int64", "uint64":
+			fmt.Fprintf(output, "| %-61s |\n", "")
+			fmt.Fprintf(output, "+ %s +\n", center(name+" (64 bits)", 61))
+			fmt.Fprintf(output, "| %-61s |\n", "")
+			fmt.Fprintln(output, line)
+		case "string", "byte": // XXX We assume slice of byte!
+			fmt.Fprintf(output, "| %s |\n", center("Length of "+name, 61))
+			fmt.Fprintln(output, line)
+			fmt.Fprintf(output, "/ %61s /\n", "")
+			fmt.Fprintf(output, "\\ %s \\\n", center(name+" (variable length)", 61))
+			fmt.Fprintf(output, "/ %61s /\n", "")
+			fmt.Fprintln(output, line)
+		default:
+			if f.IsSlice {
+				tn = "Zero or more " + tn + " Structures"
+				fmt.Fprintf(output, "/ %s /\n", center("", 61))
+				fmt.Fprintf(output, "\\ %s \\\n", center(tn, 61))
+				fmt.Fprintf(output, "/ %s /\n", center("", 61))
+			} else {
+				tn = tn + " Structure"
+				fmt.Fprintf(output, "/ %s /\n", center("", 61))
+				fmt.Fprintf(output, "\\ %s \\\n", center(tn, 61))
+				fmt.Fprintf(output, "/ %s /\n", center("", 61))
+			}
+			fmt.Fprintln(output, line)
+		}
+	}
+	fmt.Fprintln(output)
+	fmt.Fprintln(output)
+}
+
+func generateXdr(output io.Writer, s structInfo) {
+	sn := s.Name
+	fs := s.Fields
+
+	fmt.Fprintf(output, "struct %s {\n", sn)
+
+	for _, f := range fs {
+		tn := f.FieldType
+		fn := f.Name
+		suf := ""
+		l := ""
+		if f.Max > 0 {
+			l = strconv.Itoa(f.Max)
+		}
+		if f.IsSlice {
+			suf = "<" + l + ">"
+		}
+
+		switch tn {
+		case "int16", "int32":
+			fmt.Fprintf(output, "\tint %s%s;\n", fn, suf)
+		case "uint16", "uint32":
+			fmt.Fprintf(output, "\tunsigned int %s%s;\n", fn, suf)
+		case "int64":
+			fmt.Fprintf(output, "\thyper %s%s;\n", fn, suf)
+		case "uint64":
+			fmt.Fprintf(output, "\tunsigned hyper %s%s;\n", fn, suf)
+		case "string":
+			fmt.Fprintf(output, "\tstring %s<%s>;\n", fn, l)
+		case "byte":
+			fmt.Fprintf(output, "\topaque %s<%s>;\n", fn, l)
+		default:
+			fmt.Fprintf(output, "\t%s %s%s;\n", tn, fn, suf)
+		}
+	}
+	fmt.Fprintln(output, "}")
+	fmt.Fprintln(output)
+}
+
+func center(s string, w int) string {
+	w -= len(s)
+	l := w / 2
+	r := l
+	if l+r < w {
+		r++
+	}
+	return strings.Repeat(" ", l) + s + strings.Repeat(" ", r)
+}
+
+func inspector(structs *[]structInfo) func(ast.Node) bool {
+	return func(n ast.Node) bool {
+		switch n := n.(type) {
+		case *ast.TypeSpec:
+			switch t := n.Type.(type) {
+			case *ast.StructType:
+				name := n.Name.Name
+				fs := handleStruct(t)
+				*structs = append(*structs, structInfo{name, fs})
+			}
+			return false
+		default:
+			return true
+		}
+	}
+}
+
+func main() {
+	outputFile := flag.String("o", "", "Output file, blank for stdout")
+	flag.Parse()
+	fname := flag.Arg(0)
+
+	fset := token.NewFileSet()
+	f, err := parser.ParseFile(fset, fname, nil, parser.ParseComments)
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	var structs []structInfo
+	i := inspector(&structs)
+	ast.Inspect(f, i)
+
+	var output io.Writer = os.Stdout
+	if *outputFile != "" {
+		fd, err := os.Create(*outputFile)
+		if err != nil {
+			log.Fatal(err)
+		}
+		output = fd
+	}
+
+	headerTpl.Execute(output, map[string]string{"Package": f.Name.Name})
+	for _, s := range structs {
+		fmt.Fprintf(output, "\n/*\n\n")
+		generateDiagram(output, s)
+		generateXdr(output, s)
+		fmt.Fprintf(output, "*/\n")
+		generateCode(output, s)
+	}
+}
@@ -0,0 +1,16 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+package xdr
+
+import (
+	"log"
+	"os"
+)
+
+var (
+	debug = len(os.Getenv("XDRTRACE")) > 0
+	dl    = log.New(os.Stdout, "xdr: ", log.Lshortfile|log.Ltime|log.Lmicroseconds)
+)
+
+const maxDebugBytes = 32
@@ -0,0 +1,5 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+// Package xdr implements an XDR (RFC 4506) encoder/decoder.
+package xdr
@@ -0,0 +1,79 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+package xdr_test
+
+import (
+	"bytes"
+	"math/rand"
+	"reflect"
+	"testing"
+	"testing/quick"
+
+	"github.com/calmh/xdr"
+)
+
+// Contains all supported types
+type TestStruct struct {
+	I    int
+	I8   int8
+	UI8  uint8
+	I16  int16
+	UI16 uint16
+	I32  int32
+	UI32 uint32
+	I64  int64
+	UI64 uint64
+	BS   []byte // max:1024
+	S    string // max:1024
+	C    Opaque
+	SS   []string // max:1024
+}
+
+type Opaque [32]byte
+
+func (u *Opaque) EncodeXDRInto(w *xdr.Writer) (int, error) {
+	return w.WriteRaw(u[:])
+}
+
+func (u *Opaque) DecodeXDRFrom(r *xdr.Reader) (int, error) {
+	return r.ReadRaw(u[:])
+}
+
+func (Opaque) Generate(rand *rand.Rand, size int) reflect.Value {
+	var u Opaque
+	for i := range u[:] {
+		u[i] = byte(rand.Int())
+	}
+	return reflect.ValueOf(u)
+}
+
+func TestEncDec(t *testing.T) {
+	fn := func(t0 TestStruct) bool {
+		bs, err := t0.MarshalXDR()
+		if err != nil {
+			t.Fatal(err)
+		}
+		var t1 TestStruct
+		err = t1.UnmarshalXDR(bs)
+		if err != nil {
+			t.Fatal(err)
+		}
+
+		// Not comparing with DeepEqual since we'll unmarshal nil slices as empty
+		if t0.I != t1.I ||
+			t0.I16 != t1.I16 || t0.UI16 != t1.UI16 ||
+			t0.I32 != t1.I32 || t0.UI32 != t1.UI32 ||
+			t0.I64 != t1.I64 || t0.UI64 != t1.UI64 ||
+			bytes.Compare(t0.BS, t1.BS) != 0 ||
+			t0.S != t1.S || t0.C != t1.C {
+			t.Logf("%#v", t0)
+			t.Logf("%#v", t1)
+			return false
+		}
+		return true
+	}
+	if err := quick.Check(fn, nil); err != nil {
+		t.Error(err)
+	}
+}
@@ -0,0 +1,185 @@
+// ************************************************************
+// This file is automatically generated by genxdr. Do not edit.
+// ************************************************************
+
+package xdr_test
+
+import (
+	"bytes"
+	"io"
+
+	"github.com/calmh/xdr"
+)
+
+/*
+
+TestStruct Structure:
+
+ 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
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                         int Structure                         \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                        int8 Structure                         \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                        uint8 Structure                        \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|            0x0000             |              I16              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|            0x0000             |             UI16              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                              I32                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                             UI32                              |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                                                               |
+                         I64 (64 bits)                         +
+|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                                                               |
+                        UI64 (64 bits)                         +
+|                                                               |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Length of BS                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                     BS (variable length)                      \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                          Length of S                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                      S (variable length)                      \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                       Opaque Structure                        \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Number of SS                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+|                         Length of SS                          |
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+/                                                               /
+\                     SS (variable length)                      \
+/                                                               /
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+
+
+struct TestStruct {
+	int I;
+	int8 I8;
+	uint8 UI8;
+	int I16;
+	unsigned int UI16;
+	int I32;
+	unsigned int UI32;
+	hyper I64;
+	unsigned hyper UI64;
+	opaque BS<1024>;
+	string S<1024>;
+	Opaque C;
+	string SS<1024>;
+}
+
+*/
+
+func (o TestStruct) EncodeXDR(w io.Writer) (int, error) {
+	var xw = xdr.NewWriter(w)
+	return o.EncodeXDRInto(xw)
+}
+
+func (o TestStruct) MarshalXDR() ([]byte, error) {
+	return o.AppendXDR(make([]byte, 0, 128))
+}
+
+func (o TestStruct) MustMarshalXDR() []byte {
+	bs, err := o.MarshalXDR()
+	if err != nil {
+		panic(err)
+	}
+	return bs
+}
+
+func (o TestStruct) AppendXDR(bs []byte) ([]byte, error) {
+	var aw = xdr.AppendWriter(bs)
+	var xw = xdr.NewWriter(&aw)
+	_, err := o.EncodeXDRInto(xw)
+	return []byte(aw), err
+}
+
+func (o TestStruct) EncodeXDRInto(xw *xdr.Writer) (int, error) {
+	xw.WriteUint64(uint64(o.I))
+	xw.WriteUint8(uint8(o.I8))
+	xw.WriteUint8(o.UI8)
+	xw.WriteUint16(uint16(o.I16))
+	xw.WriteUint16(o.UI16)
+	xw.WriteUint32(uint32(o.I32))
+	xw.WriteUint32(o.UI32)
+	xw.WriteUint64(uint64(o.I64))
+	xw.WriteUint64(o.UI64)
+	if l := len(o.BS); l > 1024 {
+		return xw.Tot(), xdr.ElementSizeExceeded("BS", l, 1024)
+	}
+	xw.WriteBytes(o.BS)
+	if l := len(o.S); l > 1024 {
+		return xw.Tot(), xdr.ElementSizeExceeded("S", l, 1024)
+	}
+	xw.WriteString(o.S)
+	_, err := o.C.EncodeXDRInto(xw)
+	if err != nil {
+		return xw.Tot(), err
+	}
+	if l := len(o.SS); l > 1024 {
+		return xw.Tot(), xdr.ElementSizeExceeded("SS", l, 1024)
+	}
+	xw.WriteUint32(uint32(len(o.SS)))
+	for i := range o.SS {
+		xw.WriteString(o.SS[i])
+	}
+	return xw.Tot(), xw.Error()
+}
+
+func (o *TestStruct) DecodeXDR(r io.Reader) error {
+	xr := xdr.NewReader(r)
+	return o.DecodeXDRFrom(xr)
+}
+
+func (o *TestStruct) UnmarshalXDR(bs []byte) error {
+	var br = bytes.NewReader(bs)
+	var xr = xdr.NewReader(br)
+	return o.DecodeXDRFrom(xr)
+}
+
+func (o *TestStruct) DecodeXDRFrom(xr *xdr.Reader) error {
+	o.I = int(xr.ReadUint64())
+	o.I8 = int8(xr.ReadUint8())
+	o.UI8 = xr.ReadUint8()
+	o.I16 = int16(xr.ReadUint16())
+	o.UI16 = xr.ReadUint16()
+	o.I32 = int32(xr.ReadUint32())
+	o.UI32 = xr.ReadUint32()
+	o.I64 = int64(xr.ReadUint64())
+	o.UI64 = xr.ReadUint64()
+	o.BS = xr.ReadBytesMax(1024)
+	o.S = xr.ReadStringMax(1024)
+	(&o.C).DecodeXDRFrom(xr)
+	_SSSize := int(xr.ReadUint32())
+	if _SSSize < 0 {
+		return xdr.ElementSizeExceeded("SS", _SSSize, 1024)
+	}
+	if _SSSize > 1024 {
+		return xdr.ElementSizeExceeded("SS", _SSSize, 1024)
+	}
+	o.SS = make([]string, _SSSize)
+	for i := range o.SS {
+		o.SS[i] = xr.ReadString()
+	}
+	return xr.Error()
+}
@@ -0,0 +1,4 @@
+#!/bin/sh
+
+go run cmd/genxdr/main.go -- bench_test.go > bench_xdr_test.go
+go run cmd/genxdr/main.go -- encdec_test.go > encdec_xdr_test.go
@@ -0,0 +1,10 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+// +build ipdr
+
+package xdr
+
+func pad(l int) int {
+	return 0
+}
@@ -0,0 +1,14 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+// +build !ipdr
+
+package xdr
+
+func pad(l int) int {
+	d := l % 4
+	if d == 0 {
+		return 0
+	}
+	return 4 - d
+}
@@ -0,0 +1,171 @@
+// Copyright (C) 2014 Jakob Borg and Contributors (see the CONTRIBUTORS file).
+// All rights reserved. Use of this source code is governed by an MIT-style
+// license that can be found in the LICENSE file.
+
+package xdr
+
+import (
+	"fmt"
+	"io"
+	"reflect"
+	"unsafe"
+)
+
+type Reader struct {
+	r   io.Reader
+	err error
+	b   [8]byte
+}
+
+func NewReader(r io.Reader) *Reader {
+	return &Reader{
+		r: r,
+	}
+}
+
+func (r *Reader) ReadRaw(bs []byte) (int, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+
+	var n int
+	n, r.err = io.ReadFull(r.r, bs)
+	return n, r.err
+}
+
+func (r *Reader) ReadString() string {
+	return r.ReadStringMax(0)
+}
+
+func (r *Reader) ReadStringMax(max int) string {
+	buf := r.ReadBytesMaxInto(max, nil)
+	bh := (*reflect.SliceHeader)(unsafe.Pointer(&buf))
+	sh := reflect.StringHeader{
+		Data: bh.Data,
+		Len:  bh.Len,
+	}
+	return *((*string)(unsafe.Pointer(&sh)))
+}
+
+func (r *Reader) ReadBytes() []byte {
+	return r.ReadBytesInto(nil)
+}
+
+func (r *Reader) ReadBytesMax(max int) []byte {
+	return r.ReadBytesMaxInto(max, nil)
+}
+
+func (r *Reader) ReadBytesInto(dst []byte) []byte {
+	return r.ReadBytesMaxInto(0, dst)
+}
+
+func (r *Reader) ReadBytesMaxInto(max int, dst []byte) []byte {
+	if r.err != nil {
+		return nil
+	}
+
+	l := int(r.ReadUint32())
+	if r.err != nil {
+		return nil
+	}
+	if l < 0 || max > 0 && l > max {
+		// l may be negative on 32 bit builds
+		r.err = ElementSizeExceeded("bytes field", l, max)
+		return nil
+	}
+
+	if fullLen := l + pad(l); fullLen > len(dst) {
+		dst = make([]byte, fullLen)
+	} else {
+		dst = dst[:fullLen]
+	}
+
+	var n int
+	n, r.err = io.ReadFull(r.r, dst)
+	if r.err != nil {
+		if debug {
+			dl.Printf("rd bytes (%d): %v", len(dst), r.err)
+		}
+		return nil
+	}
+
+	if debug {
+		if n > maxDebugBytes {
+			dl.Printf("rd bytes (%d): %x...", len(dst), dst[:maxDebugBytes])
+		} else {
+			dl.Printf("rd bytes (%d): %x", len(dst), dst)
+		}
+	}
+	return dst[:l]
+}
+
+func (r *Reader) ReadBool() bool {
+	return r.ReadUint8() != 0
+}
+
+func (r *Reader) ReadUint32() uint32 {
+	if r.err != nil {
+		return 0
+	}
+
+	_, r.err = io.ReadFull(r.r, r.b[:4])
+	if r.err != nil {
+		if debug {
+			dl.Printf("rd uint32: %v", r.err)
+		}
+		return 0
+	}
+
+	v := uint32(r.b[3]) | uint32(r.b[2])<<8 | uint32(r.b[1])<<16 | uint32(r.b[0])<<24
+
+	if debug {
+		dl.Printf("rd uint32=%d (0x%08x)", v, v)
+	}
+	return v
+}
+
+func (r *Reader) ReadUint64() uint64 {
+	if r.err != nil {
+		return 0
+	}
+
+	_, r.err = io.ReadFull(r.r, r.b[:8])
+	if r.err != nil {
+		if debug {
+			dl.Printf("rd uint64: %v", r.err)
+		}
+		return 0
+	}
+
+	v := uint64(r.b[7]) | uint64(r.b[6])<<8 | uint64(r.b[5])<<16 | uint64(r.b[4])<<24 |
+		uint64(r.b[3])<<32 | uint64(r.b[2])<<40 | uint64(r.b[1])<<48 | uint64(r.b[0])<<56
+
+	if debug {
+		dl.Printf("rd uint64=%d (0x%016x)", v, v)
+	}
+	return v
+}
+
+type XDRError struct {
+	op  string
+	err error
+}
+
+func (e XDRError) Error() string {
+	return "xdr " + e.op + ": " + e.err.Error()
+}
+
+func (e XDRError) IsEOF() bool {
+	return e.err == io.EOF
+}
+
+func (r *Reader) Error() error {
+	if r.err == nil {
+		return nil
+	}
+	return XDRError{"read", r.err}
+}
+
+func ElementSizeExceeded(field string, size, limit int) error {
+	return fmt.Errorf("%s exceeds size limit; %d > %d", field, size, limit)
+}
@@ -0,0 +1,49 @@
+// Copyright (C) 2014 Jakob Borg and Contributors (see the CONTRIBUTORS file).
+// All rights reserved. Use of this source code is governed by an MIT-style
+// license that can be found in the LICENSE file.
+
+// +build ipdr
+
+package xdr
+
+import "io"
+
+func (r *Reader) ReadUint8() uint8 {
+	if r.err != nil {
+		return 0
+	}
+
+	_, r.err = io.ReadFull(r.r, r.b[:1])
+	if r.err != nil {
+		if debug {
+			dl.Printf("rd uint8: %v", r.err)
+		}
+		return 0
+	}
+
+	if debug {
+		dl.Printf("rd uint8=%d (0x%02x)", r.b[0], r.b[0])
+	}
+	return r.b[0]
+}
+
+func (r *Reader) ReadUint16() uint16 {
+	if r.err != nil {
+		return 0
+	}
+
+	_, r.err = io.ReadFull(r.r, r.b[:2])
+	if r.err != nil {
+		if debug {
+			dl.Printf("rd uint16: %v", r.err)
+		}
+		return 0
+	}
+
+	v := uint16(r.b[1]) | uint16(r.b[0])<<8
+
+	if debug {
+		dl.Printf("rd uint16=%d (0x%04x)", v, v)
+	}
+	return v
+}
@@ -0,0 +1,15 @@
+// Copyright (C) 2014 Jakob Borg and Contributors (see the CONTRIBUTORS file).
+// All rights reserved. Use of this source code is governed by an MIT-style
+// license that can be found in the LICENSE file.
+
+// +build !ipdr
+
+package xdr
+
+func (r *Reader) ReadUint8() uint8 {
+	return uint8(r.ReadUint32())
+}
+
+func (r *Reader) ReadUint16() uint16 {
+	return uint16(r.ReadUint32())
+}
@@ -0,0 +1,44 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+// +build refl
+
+package xdr_test
+
+import (
+	"bytes"
+	"testing"
+
+	refl "github.com/davecgh/go-xdr/xdr"
+)
+
+func TestCompareMarshals(t *testing.T) {
+	e0 := s.MarshalXDR()
+	e1, err := refl.Marshal(s)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if bytes.Compare(e0, e1) != 0 {
+		t.Fatalf("Encoding mismatch;\n\t%x (this)\n\t%x (refl)", e0, e1)
+	}
+}
+
+func BenchmarkReflMarshal(b *testing.B) {
+	var err error
+	for i := 0; i < b.N; i++ {
+		res, err = refl.Marshal(s)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkReflUnmarshal(b *testing.B) {
+	var t XDRBenchStruct
+	for i := 0; i < b.N; i++ {
+		_, err := refl.Unmarshal(e, &t)
+		if err != nil {
+			b.Fatal(err)
+		}
+	}
+}
@@ -0,0 +1,146 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+package xdr
+
+import (
+	"io"
+	"reflect"
+	"unsafe"
+)
+
+var padBytes = []byte{0, 0, 0}
+
+type Writer struct {
+	w   io.Writer
+	tot int
+	err error
+	b   [8]byte
+}
+
+type AppendWriter []byte
+
+func (w *AppendWriter) Write(bs []byte) (int, error) {
+	*w = append(*w, bs...)
+	return len(bs), nil
+}
+
+func NewWriter(w io.Writer) *Writer {
+	return &Writer{
+		w: w,
+	}
+}
+
+func (w *Writer) WriteRaw(bs []byte) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	var n int
+	n, w.err = w.w.Write(bs)
+	return n, w.err
+}
+
+func (w *Writer) WriteString(s string) (int, error) {
+	sh := *((*reflect.StringHeader)(unsafe.Pointer(&s)))
+	bh := reflect.SliceHeader{
+		Data: sh.Data,
+		Len:  sh.Len,
+		Cap:  sh.Len,
+	}
+	return w.WriteBytes(*(*[]byte)(unsafe.Pointer(&bh)))
+}
+
+func (w *Writer) WriteBytes(bs []byte) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	w.WriteUint32(uint32(len(bs)))
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	if debug {
+		if len(bs) > maxDebugBytes {
+			dl.Printf("wr bytes (%d): %x...", len(bs), bs[:maxDebugBytes])
+		} else {
+			dl.Printf("wr bytes (%d): %x", len(bs), bs)
+		}
+	}
+
+	var l, n int
+	n, w.err = w.w.Write(bs)
+	l += n
+
+	if p := pad(len(bs)); w.err == nil && p > 0 {
+		n, w.err = w.w.Write(padBytes[:p])
+		l += n
+	}
+
+	w.tot += l
+	return l, w.err
+}
+
+func (w *Writer) WriteBool(v bool) (int, error) {
+	if v {
+		return w.WriteUint8(1)
+	} else {
+		return w.WriteUint8(0)
+	}
+}
+
+func (w *Writer) WriteUint32(v uint32) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	if debug {
+		dl.Printf("wr uint32=%d", v)
+	}
+
+	w.b[0] = byte(v >> 24)
+	w.b[1] = byte(v >> 16)
+	w.b[2] = byte(v >> 8)
+	w.b[3] = byte(v)
+
+	var l int
+	l, w.err = w.w.Write(w.b[:4])
+	w.tot += l
+	return l, w.err
+}
+
+func (w *Writer) WriteUint64(v uint64) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	if debug {
+		dl.Printf("wr uint64=%d", v)
+	}
+
+	w.b[0] = byte(v >> 56)
+	w.b[1] = byte(v >> 48)
+	w.b[2] = byte(v >> 40)
+	w.b[3] = byte(v >> 32)
+	w.b[4] = byte(v >> 24)
+	w.b[5] = byte(v >> 16)
+	w.b[6] = byte(v >> 8)
+	w.b[7] = byte(v)
+
+	var l int
+	l, w.err = w.w.Write(w.b[:8])
+	w.tot += l
+	return l, w.err
+}
+
+func (w *Writer) Tot() int {
+	return w.tot
+}
+
+func (w *Writer) Error() error {
+	if w.err == nil {
+		return nil
+	}
+	return XDRError{"write", w.err}
+}
@@ -0,0 +1,41 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+// +build ipdr
+
+package xdr
+
+func (w *Writer) WriteUint8(v uint8) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	if debug {
+		dl.Printf("wr uint8=%d", v)
+	}
+
+	w.b[0] = byte(v)
+
+	var l int
+	l, w.err = w.w.Write(w.b[:1])
+	w.tot += l
+	return l, w.err
+}
+
+func (w *Writer) WriteUint16(v uint16) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
+	if debug {
+		dl.Printf("wr uint8=%d", v)
+	}
+
+	w.b[0] = byte(v >> 8)
+	w.b[1] = byte(v)
+
+	var l int
+	l, w.err = w.w.Write(w.b[:2])
+	w.tot += l
+	return l, w.err
+}
@@ -0,0 +1,14 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+// +build !ipdr
+
+package xdr
+
+func (w *Writer) WriteUint8(v uint8) (int, error) {
+	return w.WriteUint32(uint32(v))
+}
+
+func (w *Writer) WriteUint16(v uint16) (int, error) {
+	return w.WriteUint32(uint32(v))
+}
@@ -0,0 +1,93 @@
+// Copyright (C) 2014 Jakob Borg. All rights reserved. Use of this source code
+// is governed by an MIT-style license that can be found in the LICENSE file.
+
+package xdr
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+	"testing/quick"
+)
+
+func TestBytesNil(t *testing.T) {
+	fn := func(bs []byte) bool {
+		var b = new(bytes.Buffer)
+		var w = NewWriter(b)
+		var r = NewReader(b)
+		w.WriteBytes(bs)
+		w.WriteBytes(bs)
+		r.ReadBytes()
+		res := r.ReadBytes()
+		return bytes.Compare(bs, res) == 0
+	}
+	if err := quick.Check(fn, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestBytesGiven(t *testing.T) {
+	fn := func(bs []byte) bool {
+		var b = new(bytes.Buffer)
+		var w = NewWriter(b)
+		var r = NewReader(b)
+		w.WriteBytes(bs)
+		w.WriteBytes(bs)
+		res := make([]byte, 12)
+		res = r.ReadBytesInto(res)
+		res = r.ReadBytesInto(res)
+		return bytes.Compare(bs, res) == 0
+	}
+	if err := quick.Check(fn, nil); err != nil {
+		t.Error(err)
+	}
+}
+
+func TestReadBytesMaxInto(t *testing.T) {
+	var max = 64
+	for tot := 32; tot < 128; tot++ {
+		for diff := -32; diff <= 32; diff++ {
+			var b = new(bytes.Buffer)
+			var r = NewReader(b)
+			var w = NewWriter(b)
+
+			var toWrite = make([]byte, tot)
+			w.WriteBytes(toWrite)
+
+			var buf = make([]byte, tot+diff)
+			var bs = r.ReadBytesMaxInto(max, buf)
+
+			if tot <= max {
+				if read := len(bs); read != tot {
+					t.Errorf("Incorrect read bytes, wrote=%d, buf=%d, max=%d, read=%d", tot, tot+diff, max, read)
+				}
+			} else if !strings.Contains(r.err.Error(), "exceeds size") {
+				t.Errorf("Unexpected non-ErrElementSizeExceeded error for wrote=%d, max=%d: %v", tot, max, r.err)
+			}
+		}
+	}
+}
+
+func TestReadStringMax(t *testing.T) {
+	for tot := 42; tot < 72; tot++ {
+		for max := 0; max < 128; max++ {
+			var b = new(bytes.Buffer)
+			var r = NewReader(b)
+			var w = NewWriter(b)
+
+			var toWrite = make([]byte, tot)
+			w.WriteBytes(toWrite)
+
+			var str = r.ReadStringMax(max)
+			var read = len(str)
+
+			if max == 0 || tot <= max {
+				if read != tot {
+					t.Errorf("Incorrect read bytes, wrote=%d, max=%d, read=%d", tot, max, read)
+				}
+			} else if !strings.Contains(r.err.Error(), "exceeds size") {
+				t.Errorf("Unexpected non-ErrElementSizeExceeded error for wrote=%d, max=%d, read=%d: %v", tot, max, read, r.err)
+			}
+		}
+	}
+}
@@ -0,0 +1,14 @@
+# This is the official list of Snappy-Go authors for copyright purposes.
+# This file is distinct from the CONTRIBUTORS files.
+# See the latter for an explanation.
+
+# Names should be added to this file as
+#	Name or Organization <email address>
+# The email address is not required for organizations.
+
+# Please keep the list sorted.
+
+Damian Gryski <dgryski@gmail.com>
+Google Inc.
+Jan Mercl <0xjnml@gmail.com>
+Sebastien Binet <seb.binet@gmail.com>
@@ -0,0 +1,36 @@
+# This is the official list of people who can contribute
+# (and typically have contributed) code to the Snappy-Go repository.
+# The AUTHORS file lists the copyright holders; this file
+# lists people.  For example, Google employees are listed here
+# but not in AUTHORS, because Google holds the copyright.
+#
+# The submission process automatically checks to make sure
+# that people submitting code are listed in this file (by email address).
+#
+# Names should be added to this file only after verifying that
+# the individual or the individual's organization has agreed to
+# the appropriate Contributor License Agreement, found here:
+#
+#     http://code.google.com/legal/individual-cla-v1.0.html
+#     http://code.google.com/legal/corporate-cla-v1.0.html
+#
+# The agreement for individuals can be filled out on the web.
+#
+# When adding J Random Contributor's name to this file,
+# either J's name or J's organization's name should be
+# added to the AUTHORS file, depending on whether the
+# individual or corporate CLA was used.
+
+# Names should be added to this file like so:
+#     Name <email address>
+
+# Please keep the list sorted.
+
+Damian Gryski <dgryski@gmail.com>
+Jan Mercl <0xjnml@gmail.com>
+Kai Backman <kaib@golang.org>
+Marc-Antoine Ruel <maruel@chromium.org>
+Nigel Tao <nigeltao@golang.org>
+Rob Pike <r@golang.org>
+Russ Cox <rsc@golang.org>
+Sebastien Binet <seb.binet@gmail.com>
@@ -1,17 +1,18 @@
-Copyright (c) 2012 Jesse van den Kieboom. All rights reserved.
+Copyright (c) 2011 The Snappy-Go Authors. All rights reserved.
+
 Redistribution and use in source and binary forms, with or without
 modification, are permitted provided that the following conditions are
 met:

   * Redistributions of source code must retain the above copyright
-     notice, this list of conditions and the following disclaimer.
+notice, this list of conditions and the following disclaimer.
   * Redistributions in binary form must reproduce the above
-     copyright notice, this list of conditions and the following disclaimer
-     in the documentation and/or other materials provided with the
-     distribution.
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
   * Neither the name of Google Inc. nor the names of its
-     contributors may be used to endorse or promote products derived from
-     this software without specific prior written permission.
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.

 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
@@ -0,0 +1,7 @@
+The Snappy compression format in the Go programming language.
+
+To download and install from source:
+$ go get github.com/golang/snappy
+
+Unless otherwise noted, the Snappy-Go source files are distributed
+under the BSD-style license found in the LICENSE file.
@@ -0,0 +1,294 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"encoding/binary"
+	"errors"
+	"io"
+)
+
+var (
+	// ErrCorrupt reports that the input is invalid.
+	ErrCorrupt = errors.New("snappy: corrupt input")
+	// ErrTooLarge reports that the uncompressed length is too large.
+	ErrTooLarge = errors.New("snappy: decoded block is too large")
+	// ErrUnsupported reports that the input isn't supported.
+	ErrUnsupported = errors.New("snappy: unsupported input")
+)
+
+// DecodedLen returns the length of the decoded block.
+func DecodedLen(src []byte) (int, error) {
+	v, _, err := decodedLen(src)
+	return v, err
+}
+
+// decodedLen returns the length of the decoded block and the number of bytes
+// that the length header occupied.
+func decodedLen(src []byte) (blockLen, headerLen int, err error) {
+	v, n := binary.Uvarint(src)
+	if n <= 0 || v > 0xffffffff {
+		return 0, 0, ErrCorrupt
+	}
+
+	const wordSize = 32 << (^uint(0) >> 32 & 1)
+	if wordSize == 32 && v > 0x7fffffff {
+		return 0, 0, ErrTooLarge
+	}
+	return int(v), n, nil
+}
+
+// Decode returns the decoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire decoded block.
+// Otherwise, a newly allocated slice will be returned.
+// It is valid to pass a nil dst.
+func Decode(dst, src []byte) ([]byte, error) {
+	dLen, s, err := decodedLen(src)
+	if err != nil {
+		return nil, err
+	}
+	if len(dst) < dLen {
+		dst = make([]byte, dLen)
+	}
+
+	var d, offset, length int
+	for s < len(src) {
+		switch src[s] & 0x03 {
+		case tagLiteral:
+			x := uint(src[s] >> 2)
+			switch {
+			case x < 60:
+				s++
+			case x == 60:
+				s += 2
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-1])
+			case x == 61:
+				s += 3
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-2]) | uint(src[s-1])<<8
+			case x == 62:
+				s += 4
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-3]) | uint(src[s-2])<<8 | uint(src[s-1])<<16
+			case x == 63:
+				s += 5
+				if s > len(src) {
+					return nil, ErrCorrupt
+				}
+				x = uint(src[s-4]) | uint(src[s-3])<<8 | uint(src[s-2])<<16 | uint(src[s-1])<<24
+			}
+			length = int(x + 1)
+			if length <= 0 {
+				return nil, errors.New("snappy: unsupported literal length")
+			}
+			if length > len(dst)-d || length > len(src)-s {
+				return nil, ErrCorrupt
+			}
+			copy(dst[d:], src[s:s+length])
+			d += length
+			s += length
+			continue
+
+		case tagCopy1:
+			s += 2
+			if s > len(src) {
+				return nil, ErrCorrupt
+			}
+			length = 4 + int(src[s-2])>>2&0x7
+			offset = int(src[s-2])&0xe0<<3 | int(src[s-1])
+
+		case tagCopy2:
+			s += 3
+			if s > len(src) {
+				return nil, ErrCorrupt
+			}
+			length = 1 + int(src[s-3])>>2
+			offset = int(src[s-2]) | int(src[s-1])<<8
+
+		case tagCopy4:
+			return nil, errors.New("snappy: unsupported COPY_4 tag")
+		}
+
+		end := d + length
+		if offset > d || end > len(dst) {
+			return nil, ErrCorrupt
+		}
+		for ; d < end; d++ {
+			dst[d] = dst[d-offset]
+		}
+	}
+	if d != dLen {
+		return nil, ErrCorrupt
+	}
+	return dst[:d], nil
+}
+
+// NewReader returns a new Reader that decompresses from r, using the framing
+// format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func NewReader(r io.Reader) *Reader {
+	return &Reader{
+		r:       r,
+		decoded: make([]byte, maxUncompressedChunkLen),
+		buf:     make([]byte, MaxEncodedLen(maxUncompressedChunkLen)+checksumSize),
+	}
+}
+
+// Reader is an io.Reader than can read Snappy-compressed bytes.
+type Reader struct {
+	r       io.Reader
+	err     error
+	decoded []byte
+	buf     []byte
+	// decoded[i:j] contains decoded bytes that have not yet been passed on.
+	i, j       int
+	readHeader bool
+}
+
+// Reset discards any buffered data, resets all state, and switches the Snappy
+// reader to read from r. This permits reusing a Reader rather than allocating
+// a new one.
+func (r *Reader) Reset(reader io.Reader) {
+	r.r = reader
+	r.err = nil
+	r.i = 0
+	r.j = 0
+	r.readHeader = false
+}
+
+func (r *Reader) readFull(p []byte) (ok bool) {
+	if _, r.err = io.ReadFull(r.r, p); r.err != nil {
+		if r.err == io.ErrUnexpectedEOF {
+			r.err = ErrCorrupt
+		}
+		return false
+	}
+	return true
+}
+
+// Read satisfies the io.Reader interface.
+func (r *Reader) Read(p []byte) (int, error) {
+	if r.err != nil {
+		return 0, r.err
+	}
+	for {
+		if r.i < r.j {
+			n := copy(p, r.decoded[r.i:r.j])
+			r.i += n
+			return n, nil
+		}
+		if !r.readFull(r.buf[:4]) {
+			return 0, r.err
+		}
+		chunkType := r.buf[0]
+		if !r.readHeader {
+			if chunkType != chunkTypeStreamIdentifier {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			r.readHeader = true
+		}
+		chunkLen := int(r.buf[1]) | int(r.buf[2])<<8 | int(r.buf[3])<<16
+		if chunkLen > len(r.buf) {
+			r.err = ErrUnsupported
+			return 0, r.err
+		}
+
+		// The chunk types are specified at
+		// https://github.com/google/snappy/blob/master/framing_format.txt
+		switch chunkType {
+		case chunkTypeCompressedData:
+			// Section 4.2. Compressed data (chunk type 0x00).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			buf := r.buf[:chunkLen]
+			if !r.readFull(buf) {
+				return 0, r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			buf = buf[checksumSize:]
+
+			n, err := DecodedLen(buf)
+			if err != nil {
+				r.err = err
+				return 0, r.err
+			}
+			if n > len(r.decoded) {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			if _, err := Decode(r.decoded, buf); err != nil {
+				r.err = err
+				return 0, r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeUncompressedData:
+			// Section 4.3. Uncompressed data (chunk type 0x01).
+			if chunkLen < checksumSize {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			buf := r.buf[:checksumSize]
+			if !r.readFull(buf) {
+				return 0, r.err
+			}
+			checksum := uint32(buf[0]) | uint32(buf[1])<<8 | uint32(buf[2])<<16 | uint32(buf[3])<<24
+			// Read directly into r.decoded instead of via r.buf.
+			n := chunkLen - checksumSize
+			if !r.readFull(r.decoded[:n]) {
+				return 0, r.err
+			}
+			if crc(r.decoded[:n]) != checksum {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			r.i, r.j = 0, n
+			continue
+
+		case chunkTypeStreamIdentifier:
+			// Section 4.1. Stream identifier (chunk type 0xff).
+			if chunkLen != len(magicBody) {
+				r.err = ErrCorrupt
+				return 0, r.err
+			}
+			if !r.readFull(r.buf[:len(magicBody)]) {
+				return 0, r.err
+			}
+			for i := 0; i < len(magicBody); i++ {
+				if r.buf[i] != magicBody[i] {
+					r.err = ErrCorrupt
+					return 0, r.err
+				}
+			}
+			continue
+		}
+
+		if chunkType <= 0x7f {
+			// Section 4.5. Reserved unskippable chunks (chunk types 0x02-0x7f).
+			r.err = ErrUnsupported
+			return 0, r.err
+		}
+		// Section 4.4 Padding (chunk type 0xfe).
+		// Section 4.6. Reserved skippable chunks (chunk types 0x80-0xfd).
+		if !r.readFull(r.buf[:chunkLen]) {
+			return 0, r.err
+		}
+	}
+}
@@ -0,0 +1,254 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"encoding/binary"
+	"io"
+)
+
+// We limit how far copy back-references can go, the same as the C++ code.
+const maxOffset = 1 << 15
+
+// emitLiteral writes a literal chunk and returns the number of bytes written.
+func emitLiteral(dst, lit []byte) int {
+	i, n := 0, uint(len(lit)-1)
+	switch {
+	case n < 60:
+		dst[0] = uint8(n)<<2 | tagLiteral
+		i = 1
+	case n < 1<<8:
+		dst[0] = 60<<2 | tagLiteral
+		dst[1] = uint8(n)
+		i = 2
+	case n < 1<<16:
+		dst[0] = 61<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		i = 3
+	case n < 1<<24:
+		dst[0] = 62<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		dst[3] = uint8(n >> 16)
+		i = 4
+	case int64(n) < 1<<32:
+		dst[0] = 63<<2 | tagLiteral
+		dst[1] = uint8(n)
+		dst[2] = uint8(n >> 8)
+		dst[3] = uint8(n >> 16)
+		dst[4] = uint8(n >> 24)
+		i = 5
+	default:
+		panic("snappy: source buffer is too long")
+	}
+	if copy(dst[i:], lit) != len(lit) {
+		panic("snappy: destination buffer is too short")
+	}
+	return i + len(lit)
+}
+
+// emitCopy writes a copy chunk and returns the number of bytes written.
+func emitCopy(dst []byte, offset, length int) int {
+	i := 0
+	for length > 0 {
+		x := length - 4
+		if 0 <= x && x < 1<<3 && offset < 1<<11 {
+			dst[i+0] = uint8(offset>>8)&0x07<<5 | uint8(x)<<2 | tagCopy1
+			dst[i+1] = uint8(offset)
+			i += 2
+			break
+		}
+
+		x = length
+		if x > 1<<6 {
+			x = 1 << 6
+		}
+		dst[i+0] = uint8(x-1)<<2 | tagCopy2
+		dst[i+1] = uint8(offset)
+		dst[i+2] = uint8(offset >> 8)
+		i += 3
+		length -= x
+	}
+	return i
+}
+
+// Encode returns the encoded form of src. The returned slice may be a sub-
+// slice of dst if dst was large enough to hold the entire encoded block.
+// Otherwise, a newly allocated slice will be returned.
+// It is valid to pass a nil dst.
+func Encode(dst, src []byte) []byte {
+	if n := MaxEncodedLen(len(src)); len(dst) < n {
+		dst = make([]byte, n)
+	}
+
+	// The block starts with the varint-encoded length of the decompressed bytes.
+	d := binary.PutUvarint(dst, uint64(len(src)))
+
+	// Return early if src is short.
+	if len(src) <= 4 {
+		if len(src) != 0 {
+			d += emitLiteral(dst[d:], src)
+		}
+		return dst[:d]
+	}
+
+	// Initialize the hash table. Its size ranges from 1<<8 to 1<<14 inclusive.
+	const maxTableSize = 1 << 14
+	shift, tableSize := uint(32-8), 1<<8
+	for tableSize < maxTableSize && tableSize < len(src) {
+		shift--
+		tableSize *= 2
+	}
+	var table [maxTableSize]int
+
+	// Iterate over the source bytes.
+	var (
+		s   int // The iterator position.
+		t   int // The last position with the same hash as s.
+		lit int // The start position of any pending literal bytes.
+	)
+	for s+3 < len(src) {
+		// Update the hash table.
+		b0, b1, b2, b3 := src[s], src[s+1], src[s+2], src[s+3]
+		h := uint32(b0) | uint32(b1)<<8 | uint32(b2)<<16 | uint32(b3)<<24
+		p := &table[(h*0x1e35a7bd)>>shift]
+		// We need to to store values in [-1, inf) in table. To save
+		// some initialization time, (re)use the table's zero value
+		// and shift the values against this zero: add 1 on writes,
+		// subtract 1 on reads.
+		t, *p = *p-1, s+1
+		// If t is invalid or src[s:s+4] differs from src[t:t+4], accumulate a literal byte.
+		if t < 0 || s-t >= maxOffset || b0 != src[t] || b1 != src[t+1] || b2 != src[t+2] || b3 != src[t+3] {
+			s++
+			continue
+		}
+		// Otherwise, we have a match. First, emit any pending literal bytes.
+		if lit != s {
+			d += emitLiteral(dst[d:], src[lit:s])
+		}
+		// Extend the match to be as long as possible.
+		s0 := s
+		s, t = s+4, t+4
+		for s < len(src) && src[s] == src[t] {
+			s++
+			t++
+		}
+		// Emit the copied bytes.
+		d += emitCopy(dst[d:], s-t, s-s0)
+		lit = s
+	}
+
+	// Emit any final pending literal bytes and return.
+	if lit != len(src) {
+		d += emitLiteral(dst[d:], src[lit:])
+	}
+	return dst[:d]
+}
+
+// MaxEncodedLen returns the maximum length of a snappy block, given its
+// uncompressed length.
+func MaxEncodedLen(srcLen int) int {
+	// Compressed data can be defined as:
+	//    compressed := item* literal*
+	//    item       := literal* copy
+	//
+	// The trailing literal sequence has a space blowup of at most 62/60
+	// since a literal of length 60 needs one tag byte + one extra byte
+	// for length information.
+	//
+	// Item blowup is trickier to measure. Suppose the "copy" op copies
+	// 4 bytes of data. Because of a special check in the encoding code,
+	// we produce a 4-byte copy only if the offset is < 65536. Therefore
+	// the copy op takes 3 bytes to encode, and this type of item leads
+	// to at most the 62/60 blowup for representing literals.
+	//
+	// Suppose the "copy" op copies 5 bytes of data. If the offset is big
+	// enough, it will take 5 bytes to encode the copy op. Therefore the
+	// worst case here is a one-byte literal followed by a five-byte copy.
+	// That is, 6 bytes of input turn into 7 bytes of "compressed" data.
+	//
+	// This last factor dominates the blowup, so the final estimate is:
+	return 32 + srcLen + srcLen/6
+}
+
+// NewWriter returns a new Writer that compresses to w, using the framing
+// format described at
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func NewWriter(w io.Writer) *Writer {
+	return &Writer{
+		w:   w,
+		enc: make([]byte, MaxEncodedLen(maxUncompressedChunkLen)),
+	}
+}
+
+// Writer is an io.Writer than can write Snappy-compressed bytes.
+type Writer struct {
+	w           io.Writer
+	err         error
+	enc         []byte
+	buf         [checksumSize + chunkHeaderSize]byte
+	wroteHeader bool
+}
+
+// Reset discards the writer's state and switches the Snappy writer to write to
+// w. This permits reusing a Writer rather than allocating a new one.
+func (w *Writer) Reset(writer io.Writer) {
+	w.w = writer
+	w.err = nil
+	w.wroteHeader = false
+}
+
+// Write satisfies the io.Writer interface.
+func (w *Writer) Write(p []byte) (n int, errRet error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+	if !w.wroteHeader {
+		copy(w.enc, magicChunk)
+		if _, err := w.w.Write(w.enc[:len(magicChunk)]); err != nil {
+			w.err = err
+			return n, err
+		}
+		w.wroteHeader = true
+	}
+	for len(p) > 0 {
+		var uncompressed []byte
+		if len(p) > maxUncompressedChunkLen {
+			uncompressed, p = p[:maxUncompressedChunkLen], p[maxUncompressedChunkLen:]
+		} else {
+			uncompressed, p = p, nil
+		}
+		checksum := crc(uncompressed)
+
+		// Compress the buffer, discarding the result if the improvement
+		// isn't at least 12.5%.
+		chunkType := uint8(chunkTypeCompressedData)
+		chunkBody := Encode(w.enc, uncompressed)
+		if len(chunkBody) >= len(uncompressed)-len(uncompressed)/8 {
+			chunkType, chunkBody = chunkTypeUncompressedData, uncompressed
+		}
+
+		chunkLen := 4 + len(chunkBody)
+		w.buf[0] = chunkType
+		w.buf[1] = uint8(chunkLen >> 0)
+		w.buf[2] = uint8(chunkLen >> 8)
+		w.buf[3] = uint8(chunkLen >> 16)
+		w.buf[4] = uint8(checksum >> 0)
+		w.buf[5] = uint8(checksum >> 8)
+		w.buf[6] = uint8(checksum >> 16)
+		w.buf[7] = uint8(checksum >> 24)
+		if _, err := w.w.Write(w.buf[:]); err != nil {
+			w.err = err
+			return n, err
+		}
+		if _, err := w.w.Write(chunkBody); err != nil {
+			w.err = err
+			return n, err
+		}
+		n += len(uncompressed)
+	}
+	return n, nil
+}
@@ -0,0 +1,68 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package snappy implements the snappy block-based compression format.
+// It aims for very high speeds and reasonable compression.
+//
+// The C++ snappy implementation is at https://github.com/google/snappy
+package snappy
+
+import (
+	"hash/crc32"
+)
+
+/*
+Each encoded block begins with the varint-encoded length of the decoded data,
+followed by a sequence of chunks. Chunks begin and end on byte boundaries. The
+first byte of each chunk is broken into its 2 least and 6 most significant bits
+called l and m: l ranges in [0, 4) and m ranges in [0, 64). l is the chunk tag.
+Zero means a literal tag. All other values mean a copy tag.
+
+For literal tags:
+  - If m < 60, the next 1 + m bytes are literal bytes.
+  - Otherwise, let n be the little-endian unsigned integer denoted by the next
+    m - 59 bytes. The next 1 + n bytes after that are literal bytes.
+
+For copy tags, length bytes are copied from offset bytes ago, in the style of
+Lempel-Ziv compression algorithms. In particular:
+  - For l == 1, the offset ranges in [0, 1<<11) and the length in [4, 12).
+    The length is 4 + the low 3 bits of m. The high 3 bits of m form bits 8-10
+    of the offset. The next byte is bits 0-7 of the offset.
+  - For l == 2, the offset ranges in [0, 1<<16) and the length in [1, 65).
+    The length is 1 + m. The offset is the little-endian unsigned integer
+    denoted by the next 2 bytes.
+  - For l == 3, this tag is a legacy format that is no longer supported.
+*/
+const (
+	tagLiteral = 0x00
+	tagCopy1   = 0x01
+	tagCopy2   = 0x02
+	tagCopy4   = 0x03
+)
+
+const (
+	checksumSize    = 4
+	chunkHeaderSize = 4
+	magicChunk      = "\xff\x06\x00\x00" + magicBody
+	magicBody       = "sNaPpY"
+	// https://github.com/google/snappy/blob/master/framing_format.txt says
+	// that "the uncompressed data in a chunk must be no longer than 65536 bytes".
+	maxUncompressedChunkLen = 65536
+)
+
+const (
+	chunkTypeCompressedData   = 0x00
+	chunkTypeUncompressedData = 0x01
+	chunkTypePadding          = 0xfe
+	chunkTypeStreamIdentifier = 0xff
+)
+
+var crcTable = crc32.MakeTable(crc32.Castagnoli)
+
+// crc implements the checksum specified in section 3 of
+// https://github.com/google/snappy/blob/master/framing_format.txt
+func crc(b []byte) uint32 {
+	c := crc32.Update(0, crcTable, b)
+	return uint32(c>>15|c<<17) + 0xa282ead8
+}
@@ -0,0 +1,377 @@
+// Copyright 2011 The Snappy-Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package snappy
+
+import (
+	"bytes"
+	"flag"
+	"fmt"
+	"io"
+	"io/ioutil"
+	"math/rand"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+var (
+	download = flag.Bool("download", false, "If true, download any missing files before running benchmarks")
+	testdata = flag.String("testdata", "testdata", "Directory containing the test data")
+)
+
+func roundtrip(b, ebuf, dbuf []byte) error {
+	d, err := Decode(dbuf, Encode(ebuf, b))
+	if err != nil {
+		return fmt.Errorf("decoding error: %v", err)
+	}
+	if !bytes.Equal(b, d) {
+		return fmt.Errorf("roundtrip mismatch:\n\twant %v\n\tgot  %v", b, d)
+	}
+	return nil
+}
+
+func TestEmpty(t *testing.T) {
+	if err := roundtrip(nil, nil, nil); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestSmallCopy(t *testing.T) {
+	for _, ebuf := range [][]byte{nil, make([]byte, 20), make([]byte, 64)} {
+		for _, dbuf := range [][]byte{nil, make([]byte, 20), make([]byte, 64)} {
+			for i := 0; i < 32; i++ {
+				s := "aaaa" + strings.Repeat("b", i) + "aaaabbbb"
+				if err := roundtrip([]byte(s), ebuf, dbuf); err != nil {
+					t.Errorf("len(ebuf)=%d, len(dbuf)=%d, i=%d: %v", len(ebuf), len(dbuf), i, err)
+				}
+			}
+		}
+	}
+}
+
+func TestSmallRand(t *testing.T) {
+	rng := rand.New(rand.NewSource(27354294))
+	for n := 1; n < 20000; n += 23 {
+		b := make([]byte, n)
+		for i := range b {
+			b[i] = uint8(rng.Uint32())
+		}
+		if err := roundtrip(b, nil, nil); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+func TestSmallRegular(t *testing.T) {
+	for n := 1; n < 20000; n += 23 {
+		b := make([]byte, n)
+		for i := range b {
+			b[i] = uint8(i%10 + 'a')
+		}
+		if err := roundtrip(b, nil, nil); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+func TestInvalidVarint(t *testing.T) {
+	data := []byte("\xff\xff\xff\xff\xff\xff\xff\xff\xff\xff\x00")
+	if _, err := DecodedLen(data); err != ErrCorrupt {
+		t.Errorf("DecodedLen: got %v, want ErrCorrupt", err)
+	}
+	if _, err := Decode(nil, data); err != ErrCorrupt {
+		t.Errorf("Decode: got %v, want ErrCorrupt", err)
+	}
+
+	// The encoded varint overflows 32 bits
+	data = []byte("\xff\xff\xff\xff\xff\x00")
+
+	if _, err := DecodedLen(data); err != ErrCorrupt {
+		t.Errorf("DecodedLen: got %v, want ErrCorrupt", err)
+	}
+	if _, err := Decode(nil, data); err != ErrCorrupt {
+		t.Errorf("Decode: got %v, want ErrCorrupt", err)
+	}
+}
+
+func cmp(a, b []byte) error {
+	if len(a) != len(b) {
+		return fmt.Errorf("got %d bytes, want %d", len(a), len(b))
+	}
+	for i := range a {
+		if a[i] != b[i] {
+			return fmt.Errorf("byte #%d: got 0x%02x, want 0x%02x", i, a[i], b[i])
+		}
+	}
+	return nil
+}
+
+func TestFramingFormat(t *testing.T) {
+	// src is comprised of alternating 1e5-sized sequences of random
+	// (incompressible) bytes and repeated (compressible) bytes. 1e5 was chosen
+	// because it is larger than maxUncompressedChunkLen (64k).
+	src := make([]byte, 1e6)
+	rng := rand.New(rand.NewSource(1))
+	for i := 0; i < 10; i++ {
+		if i%2 == 0 {
+			for j := 0; j < 1e5; j++ {
+				src[1e5*i+j] = uint8(rng.Intn(256))
+			}
+		} else {
+			for j := 0; j < 1e5; j++ {
+				src[1e5*i+j] = uint8(i)
+			}
+		}
+	}
+
+	buf := new(bytes.Buffer)
+	if _, err := NewWriter(buf).Write(src); err != nil {
+		t.Fatalf("Write: encoding: %v", err)
+	}
+	dst, err := ioutil.ReadAll(NewReader(buf))
+	if err != nil {
+		t.Fatalf("ReadAll: decoding: %v", err)
+	}
+	if err := cmp(dst, src); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestReaderReset(t *testing.T) {
+	gold := bytes.Repeat([]byte("All that is gold does not glitter,\n"), 10000)
+	buf := new(bytes.Buffer)
+	if _, err := NewWriter(buf).Write(gold); err != nil {
+		t.Fatalf("Write: %v", err)
+	}
+	encoded, invalid, partial := buf.String(), "invalid", "partial"
+	r := NewReader(nil)
+	for i, s := range []string{encoded, invalid, partial, encoded, partial, invalid, encoded, encoded} {
+		if s == partial {
+			r.Reset(strings.NewReader(encoded))
+			if _, err := r.Read(make([]byte, 101)); err != nil {
+				t.Errorf("#%d: %v", i, err)
+				continue
+			}
+			continue
+		}
+		r.Reset(strings.NewReader(s))
+		got, err := ioutil.ReadAll(r)
+		switch s {
+		case encoded:
+			if err != nil {
+				t.Errorf("#%d: %v", i, err)
+				continue
+			}
+			if err := cmp(got, gold); err != nil {
+				t.Errorf("#%d: %v", i, err)
+				continue
+			}
+		case invalid:
+			if err == nil {
+				t.Errorf("#%d: got nil error, want non-nil", i)
+				continue
+			}
+		}
+	}
+}
+
+func TestWriterReset(t *testing.T) {
+	gold := bytes.Repeat([]byte("Not all those who wander are lost;\n"), 10000)
+	var gots, wants [][]byte
+	const n = 20
+	w, failed := NewWriter(nil), false
+	for i := 0; i <= n; i++ {
+		buf := new(bytes.Buffer)
+		w.Reset(buf)
+		want := gold[:len(gold)*i/n]
+		if _, err := w.Write(want); err != nil {
+			t.Errorf("#%d: Write: %v", i, err)
+			failed = true
+			continue
+		}
+		got, err := ioutil.ReadAll(NewReader(buf))
+		if err != nil {
+			t.Errorf("#%d: ReadAll: %v", i, err)
+			failed = true
+			continue
+		}
+		gots = append(gots, got)
+		wants = append(wants, want)
+	}
+	if failed {
+		return
+	}
+	for i := range gots {
+		if err := cmp(gots[i], wants[i]); err != nil {
+			t.Errorf("#%d: %v", i, err)
+		}
+	}
+}
+
+func benchDecode(b *testing.B, src []byte) {
+	encoded := Encode(nil, src)
+	// Bandwidth is in amount of uncompressed data.
+	b.SetBytes(int64(len(src)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		Decode(src, encoded)
+	}
+}
+
+func benchEncode(b *testing.B, src []byte) {
+	// Bandwidth is in amount of uncompressed data.
+	b.SetBytes(int64(len(src)))
+	dst := make([]byte, MaxEncodedLen(len(src)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		Encode(dst, src)
+	}
+}
+
+func readFile(b testing.TB, filename string) []byte {
+	src, err := ioutil.ReadFile(filename)
+	if err != nil {
+		b.Skipf("skipping benchmark: %v", err)
+	}
+	if len(src) == 0 {
+		b.Fatalf("%s has zero length", filename)
+	}
+	return src
+}
+
+// expand returns a slice of length n containing repeated copies of src.
+func expand(src []byte, n int) []byte {
+	dst := make([]byte, n)
+	for x := dst; len(x) > 0; {
+		i := copy(x, src)
+		x = x[i:]
+	}
+	return dst
+}
+
+func benchWords(b *testing.B, n int, decode bool) {
+	// Note: the file is OS-language dependent so the resulting values are not
+	// directly comparable for non-US-English OS installations.
+	data := expand(readFile(b, "/usr/share/dict/words"), n)
+	if decode {
+		benchDecode(b, data)
+	} else {
+		benchEncode(b, data)
+	}
+}
+
+func BenchmarkWordsDecode1e3(b *testing.B) { benchWords(b, 1e3, true) }
+func BenchmarkWordsDecode1e4(b *testing.B) { benchWords(b, 1e4, true) }
+func BenchmarkWordsDecode1e5(b *testing.B) { benchWords(b, 1e5, true) }
+func BenchmarkWordsDecode1e6(b *testing.B) { benchWords(b, 1e6, true) }
+func BenchmarkWordsEncode1e3(b *testing.B) { benchWords(b, 1e3, false) }
+func BenchmarkWordsEncode1e4(b *testing.B) { benchWords(b, 1e4, false) }
+func BenchmarkWordsEncode1e5(b *testing.B) { benchWords(b, 1e5, false) }
+func BenchmarkWordsEncode1e6(b *testing.B) { benchWords(b, 1e6, false) }
+
+// testFiles' values are copied directly from
+// https://raw.githubusercontent.com/google/snappy/master/snappy_unittest.cc
+// The label field is unused in snappy-go.
+var testFiles = []struct {
+	label    string
+	filename string
+}{
+	{"html", "html"},
+	{"urls", "urls.10K"},
+	{"jpg", "fireworks.jpeg"},
+	{"jpg_200", "fireworks.jpeg"},
+	{"pdf", "paper-100k.pdf"},
+	{"html4", "html_x_4"},
+	{"txt1", "alice29.txt"},
+	{"txt2", "asyoulik.txt"},
+	{"txt3", "lcet10.txt"},
+	{"txt4", "plrabn12.txt"},
+	{"pb", "geo.protodata"},
+	{"gaviota", "kppkn.gtb"},
+}
+
+// The test data files are present at this canonical URL.
+const baseURL = "https://raw.githubusercontent.com/google/snappy/master/testdata/"
+
+func downloadTestdata(b *testing.B, basename string) (errRet error) {
+	filename := filepath.Join(*testdata, basename)
+	if stat, err := os.Stat(filename); err == nil && stat.Size() != 0 {
+		return nil
+	}
+
+	if !*download {
+		b.Skipf("test data not found; skipping benchmark without the -download flag")
+	}
+	// Download the official snappy C++ implementation reference test data
+	// files for benchmarking.
+	if err := os.Mkdir(*testdata, 0777); err != nil && !os.IsExist(err) {
+		return fmt.Errorf("failed to create testdata: %s", err)
+	}
+
+	f, err := os.Create(filename)
+	if err != nil {
+		return fmt.Errorf("failed to create %s: %s", filename, err)
+	}
+	defer f.Close()
+	defer func() {
+		if errRet != nil {
+			os.Remove(filename)
+		}
+	}()
+	url := baseURL + basename
+	resp, err := http.Get(url)
+	if err != nil {
+		return fmt.Errorf("failed to download %s: %s", url, err)
+	}
+	defer resp.Body.Close()
+	if s := resp.StatusCode; s != http.StatusOK {
+		return fmt.Errorf("downloading %s: HTTP status code %d (%s)", url, s, http.StatusText(s))
+	}
+	_, err = io.Copy(f, resp.Body)
+	if err != nil {
+		return fmt.Errorf("failed to download %s to %s: %s", url, filename, err)
+	}
+	return nil
+}
+
+func benchFile(b *testing.B, n int, decode bool) {
+	if err := downloadTestdata(b, testFiles[n].filename); err != nil {
+		b.Fatalf("failed to download testdata: %s", err)
+	}
+	data := readFile(b, filepath.Join(*testdata, testFiles[n].filename))
+	if decode {
+		benchDecode(b, data)
+	} else {
+		benchEncode(b, data)
+	}
+}
+
+// Naming convention is kept similar to what snappy's C++ implementation uses.
+func Benchmark_UFlat0(b *testing.B)  { benchFile(b, 0, true) }
+func Benchmark_UFlat1(b *testing.B)  { benchFile(b, 1, true) }
+func Benchmark_UFlat2(b *testing.B)  { benchFile(b, 2, true) }
+func Benchmark_UFlat3(b *testing.B)  { benchFile(b, 3, true) }
+func Benchmark_UFlat4(b *testing.B)  { benchFile(b, 4, true) }
+func Benchmark_UFlat5(b *testing.B)  { benchFile(b, 5, true) }
+func Benchmark_UFlat6(b *testing.B)  { benchFile(b, 6, true) }
+func Benchmark_UFlat7(b *testing.B)  { benchFile(b, 7, true) }
+func Benchmark_UFlat8(b *testing.B)  { benchFile(b, 8, true) }
+func Benchmark_UFlat9(b *testing.B)  { benchFile(b, 9, true) }
+func Benchmark_UFlat10(b *testing.B) { benchFile(b, 10, true) }
+func Benchmark_UFlat11(b *testing.B) { benchFile(b, 11, true) }
+func Benchmark_ZFlat0(b *testing.B)  { benchFile(b, 0, false) }
+func Benchmark_ZFlat1(b *testing.B)  { benchFile(b, 1, false) }
+func Benchmark_ZFlat2(b *testing.B)  { benchFile(b, 2, false) }
+func Benchmark_ZFlat3(b *testing.B)  { benchFile(b, 3, false) }
+func Benchmark_ZFlat4(b *testing.B)  { benchFile(b, 4, false) }
+func Benchmark_ZFlat5(b *testing.B)  { benchFile(b, 5, false) }
+func Benchmark_ZFlat6(b *testing.B)  { benchFile(b, 6, false) }
+func Benchmark_ZFlat7(b *testing.B)  { benchFile(b, 7, false) }
+func Benchmark_ZFlat8(b *testing.B)  { benchFile(b, 8, false) }
+func Benchmark_ZFlat9(b *testing.B)  { benchFile(b, 9, false) }
+func Benchmark_ZFlat10(b *testing.B) { benchFile(b, 10, false) }
+func Benchmark_ZFlat11(b *testing.B) { benchFile(b, 11, false) }
@@ -0,0 +1,191 @@
+All files in this repository are licensed as follows. If you contribute
+to this repository, it is assumed that you license your contribution
+under the same license unless you state otherwise.
+
+All files Copyright (C) 2015 Canonical Ltd. unless otherwise specified in the file.
+
+This software is licensed under the LGPLv3, included below.
+
+As a special exception to the GNU Lesser General Public License version 3
+("LGPL3"), the copyright holders of this Library give you permission to
+convey to a third party a Combined Work that links statically or dynamically
+to this Library without providing any Minimal Corresponding Source or
+Minimal Application Code as set out in 4d or providing the installation
+information set out in section 4e, provided that you comply with the other
+provisions of LGPL3 and provided that you meet, for the Application the
+terms and conditions of the license(s) which apply to the Application.
+
+Except as stated in this special exception, the provisions of LGPL3 will
+continue to comply in full to this Library. If you modify this Library, you
+may apply this exception to your version of this Library, but you are not
+obliged to do so. If you do not wish to do so, delete this exception
+statement from your version. This exception does not (and cannot) modify any
+license terms which apply to the Application, with which you must still
+comply.
+
+
+                   GNU LESSER GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+
+  This version of the GNU Lesser General Public License incorporates
+the terms and conditions of version 3 of the GNU General Public
+License, supplemented by the additional permissions listed below.
+
+  0. Additional Definitions.
+
+  As used herein, "this License" refers to version 3 of the GNU Lesser
+General Public License, and the "GNU GPL" refers to version 3 of the GNU
+General Public License.
+
+  "The Library" refers to a covered work governed by this License,
+other than an Application or a Combined Work as defined below.
+
+  An "Application" is any work that makes use of an interface provided
+by the Library, but which is not otherwise based on the Library.
+Defining a subclass of a class defined by the Library is deemed a mode
+of using an interface provided by the Library.
+
+  A "Combined Work" is a work produced by combining or linking an
+Application with the Library.  The particular version of the Library
+with which the Combined Work was made is also called the "Linked
+Version".
+
+  The "Minimal Corresponding Source" for a Combined Work means the
+Corresponding Source for the Combined Work, excluding any source code
+for portions of the Combined Work that, considered in isolation, are
+based on the Application, and not on the Linked Version.
+
+  The "Corresponding Application Code" for a Combined Work means the
+object code and/or source code for the Application, including any data
+and utility programs needed for reproducing the Combined Work from the
+Application, but excluding the System Libraries of the Combined Work.
+
+  1. Exception to Section 3 of the GNU GPL.
+
+  You may convey a covered work under sections 3 and 4 of this License
+without being bound by section 3 of the GNU GPL.
+
+  2. Conveying Modified Versions.
+
+  If you modify a copy of the Library, and, in your modifications, a
+facility refers to a function or data to be supplied by an Application
+that uses the facility (other than as an argument passed when the
+facility is invoked), then you may convey a copy of the modified
+version:
+
+   a) under this License, provided that you make a good faith effort to
+   ensure that, in the event an Application does not supply the
+   function or data, the facility still operates, and performs
+   whatever part of its purpose remains meaningful, or
+
+   b) under the GNU GPL, with none of the additional permissions of
+   this License applicable to that copy.
+
+  3. Object Code Incorporating Material from Library Header Files.
+
+  The object code form of an Application may incorporate material from
+a header file that is part of the Library.  You may convey such object
+code under terms of your choice, provided that, if the incorporated
+material is not limited to numerical parameters, data structure
+layouts and accessors, or small macros, inline functions and templates
+(ten or fewer lines in length), you do both of the following:
+
+   a) Give prominent notice with each copy of the object code that the
+   Library is used in it and that the Library and its use are
+   covered by this License.
+
+   b) Accompany the object code with a copy of the GNU GPL and this license
+   document.
+
+  4. Combined Works.
+
+  You may convey a Combined Work under terms of your choice that,
+taken together, effectively do not restrict modification of the
+portions of the Library contained in the Combined Work and reverse
+engineering for debugging such modifications, if you also do each of
+the following:
+
+   a) Give prominent notice with each copy of the Combined Work that
+   the Library is used in it and that the Library and its use are
+   covered by this License.
+
+   b) Accompany the Combined Work with a copy of the GNU GPL and this license
+   document.
+
+   c) For a Combined Work that displays copyright notices during
+   execution, include the copyright notice for the Library among
+   these notices, as well as a reference directing the user to the
+   copies of the GNU GPL and this license document.
+
+   d) Do one of the following:
+
+       0) Convey the Minimal Corresponding Source under the terms of this
+       License, and the Corresponding Application Code in a form
+       suitable for, and under terms that permit, the user to
+       recombine or relink the Application with a modified version of
+       the Linked Version to produce a modified Combined Work, in the
+       manner specified by section 6 of the GNU GPL for conveying
+       Corresponding Source.
+
+       1) Use a suitable shared library mechanism for linking with the
+       Library.  A suitable mechanism is one that (a) uses at run time
+       a copy of the Library already present on the user's computer
+       system, and (b) will operate properly with a modified version
+       of the Library that is interface-compatible with the Linked
+       Version.
+
+   e) Provide Installation Information, but only if you would otherwise
+   be required to provide such information under section 6 of the
+   GNU GPL, and only to the extent that such information is
+   necessary to install and execute a modified version of the
+   Combined Work produced by recombining or relinking the
+   Application with a modified version of the Linked Version. (If
+   you use option 4d0, the Installation Information must accompany
+   the Minimal Corresponding Source and Corresponding Application
+   Code. If you use option 4d1, you must provide the Installation
+   Information in the manner specified by section 6 of the GNU GPL
+   for conveying Corresponding Source.)
+
+  5. Combined Libraries.
+
+  You may place library facilities that are a work based on the
+Library side by side in a single library together with other library
+facilities that are not Applications and are not covered by this
+License, and convey such a combined library under terms of your
+choice, if you do both of the following:
+
+   a) Accompany the combined library with a copy of the same work based
+   on the Library, uncombined with any other library facilities,
+   conveyed under the terms of this License.
+
+   b) Give prominent notice with the combined library that part of it
+   is a work based on the Library, and explaining where to find the
+   accompanying uncombined form of the same work.
+
+  6. Revised Versions of the GNU Lesser General Public License.
+
+  The Free Software Foundation may publish revised and/or new versions
+of the GNU Lesser General Public License from time to time. Such new
+versions will be similar in spirit to the present version, but may
+differ in detail to address new problems or concerns.
+
+  Each version is given a distinguishing version number. If the
+Library as you received it specifies that a certain numbered version
+of the GNU Lesser General Public License "or any later version"
+applies to it, you have the option of following the terms and
+conditions either of that published version or of any later version
+published by the Free Software Foundation. If the Library as you
+received it does not specify a version number of the GNU Lesser
+General Public License, you may choose any version of the GNU Lesser
+General Public License ever published by the Free Software Foundation.
+
+  If the Library as you received it specifies that a proxy can decide
+whether future versions of the GNU Lesser General Public License shall
+apply, that proxy's public statement of acceptance of any version is
+permanent authorization for you to choose that version for the
+Library.
@@ -0,0 +1,117 @@
+# ratelimit
+--
+    import "github.com/juju/ratelimit"
+
+The ratelimit package provides an efficient token bucket implementation. See
+http://en.wikipedia.org/wiki/Token_bucket.
+
+## Usage
+
+#### func  Reader
+
+```go
+func Reader(r io.Reader, bucket *Bucket) io.Reader
+```
+Reader returns a reader that is rate limited by the given token bucket. Each
+token in the bucket represents one byte.
+
+#### func  Writer
+
+```go
+func Writer(w io.Writer, bucket *Bucket) io.Writer
+```
+Writer returns a writer that is rate limited by the given token bucket. Each
+token in the bucket represents one byte.
+
+#### type Bucket
+
+```go
+type Bucket struct {
+}
+```
+
+Bucket represents a token bucket that fills at a predetermined rate. Methods on
+Bucket may be called concurrently.
+
+#### func  NewBucket
+
+```go
+func NewBucket(fillInterval time.Duration, capacity int64) *Bucket
+```
+NewBucket returns a new token bucket that fills at the rate of one token every
+fillInterval, up to the given maximum capacity. Both arguments must be positive.
+The bucket is initially full.
+
+#### func  NewBucketWithQuantum
+
+```go
+func NewBucketWithQuantum(fillInterval time.Duration, capacity, quantum int64) *Bucket
+```
+NewBucketWithQuantum is similar to NewBucket, but allows the specification of
+the quantum size - quantum tokens are added every fillInterval.
+
+#### func  NewBucketWithRate
+
+```go
+func NewBucketWithRate(rate float64, capacity int64) *Bucket
+```
+NewBucketWithRate returns a token bucket that fills the bucket at the rate of
+rate tokens per second up to the given maximum capacity. Because of limited
+clock resolution, at high rates, the actual rate may be up to 1% different from
+the specified rate.
+
+#### func (*Bucket) Rate
+
+```go
+func (tb *Bucket) Rate() float64
+```
+Rate returns the fill rate of the bucket, in tokens per second.
+
+#### func (*Bucket) Take
+
+```go
+func (tb *Bucket) Take(count int64) time.Duration
+```
+Take takes count tokens from the bucket without blocking. It returns the time
+that the caller should wait until the tokens are actually available.
+
+Note that if the request is irrevocable - there is no way to return tokens to
+the bucket once this method commits us to taking them.
+
+#### func (*Bucket) TakeAvailable
+
+```go
+func (tb *Bucket) TakeAvailable(count int64) int64
+```
+TakeAvailable takes up to count immediately available tokens from the bucket. It
+returns the number of tokens removed, or zero if there are no available tokens.
+It does not block.
+
+#### func (*Bucket) TakeMaxDuration
+
+```go
+func (tb *Bucket) TakeMaxDuration(count int64, maxWait time.Duration) (time.Duration, bool)
+```
+TakeMaxDuration is like Take, except that it will only take tokens from the
+bucket if the wait time for the tokens is no greater than maxWait.
+
+If it would take longer than maxWait for the tokens to become available, it does
+nothing and reports false, otherwise it returns the time that the caller should
+wait until the tokens are actually available, and reports true.
+
+#### func (*Bucket) Wait
+
+```go
+func (tb *Bucket) Wait(count int64)
+```
+Wait takes count tokens from the bucket, waiting until they are available.
+
+#### func (*Bucket) WaitMaxDuration
+
+```go
+func (tb *Bucket) WaitMaxDuration(count int64, maxWait time.Duration) bool
+```
+WaitMaxDuration is like Wait except that it will only take tokens from the
+bucket if it needs to wait for no greater than maxWait. It reports whether any
+tokens have been removed from the bucket If no tokens have been removed, it
+returns immediately.
@@ -0,0 +1,221 @@
+// Copyright 2014 Canonical Ltd.
+// Licensed under the LGPLv3 with static-linking exception.
+// See LICENCE file for details.
+
+// The ratelimit package provides an efficient token bucket implementation
+// that can be used to limit the rate of arbitrary things.
+// See http://en.wikipedia.org/wiki/Token_bucket.
+package ratelimit
+
+import (
+	"strconv"
+	"sync"
+	"time"
+	"math"
+)
+
+// Bucket represents a token bucket that fills at a predetermined rate.
+// Methods on Bucket may be called concurrently.
+type Bucket struct {
+	startTime    time.Time
+	capacity     int64
+	quantum      int64
+	fillInterval time.Duration
+
+	// The mutex guards the fields following it.
+	mu sync.Mutex
+
+	// avail holds the number of available tokens
+	// in the bucket, as of availTick ticks from startTime.
+	// It will be negative when there are consumers
+	// waiting for tokens.
+	avail     int64
+	availTick int64
+}
+
+// NewBucket returns a new token bucket that fills at the
+// rate of one token every fillInterval, up to the given
+// maximum capacity. Both arguments must be
+// positive. The bucket is initially full.
+func NewBucket(fillInterval time.Duration, capacity int64) *Bucket {
+	return NewBucketWithQuantum(fillInterval, capacity, 1)
+}
+
+// rateMargin specifes the allowed variance of actual
+// rate from specified rate. 1% seems reasonable.
+const rateMargin = 0.01
+
+// NewBucketWithRate returns a token bucket that fills the bucket
+// at the rate of rate tokens per second up to the given
+// maximum capacity. Because of limited clock resolution,
+// at high rates, the actual rate may be up to 1% different from the
+// specified rate.
+func NewBucketWithRate(rate float64, capacity int64) *Bucket {
+	for quantum := int64(1); quantum < 1<<50; quantum = nextQuantum(quantum) {
+		fillInterval := time.Duration(1e9 * float64(quantum) / rate)
+		if fillInterval <= 0 {
+			continue
+		}
+		tb := NewBucketWithQuantum(fillInterval, capacity, quantum)
+		if diff := math.Abs(tb.Rate() - rate); diff/rate <= rateMargin {
+			return tb
+		}
+	}
+	panic("cannot find suitable quantum for " + strconv.FormatFloat(rate, 'g', -1, 64))
+}
+
+// nextQuantum returns the next quantum to try after q.
+// We grow the quantum exponentially, but slowly, so we
+// get a good fit in the lower numbers.
+func nextQuantum(q int64) int64 {
+	q1 := q * 11 / 10
+	if q1 == q {
+		q1++
+	}
+	return q1
+}
+
+// NewBucketWithQuantum is similar to NewBucket, but allows
+// the specification of the quantum size - quantum tokens
+// are added every fillInterval.
+func NewBucketWithQuantum(fillInterval time.Duration, capacity, quantum int64) *Bucket {
+	if fillInterval <= 0 {
+		panic("token bucket fill interval is not > 0")
+	}
+	if capacity <= 0 {
+		panic("token bucket capacity is not > 0")
+	}
+	if quantum <= 0 {
+		panic("token bucket quantum is not > 0")
+	}
+	return &Bucket{
+		startTime:    time.Now(),
+		capacity:     capacity,
+		quantum:      quantum,
+		avail:        capacity,
+		fillInterval: fillInterval,
+	}
+}
+
+// Wait takes count tokens from the bucket, waiting until they are
+// available.
+func (tb *Bucket) Wait(count int64) {
+	if d := tb.Take(count); d > 0 {
+		time.Sleep(d)
+	}
+}
+
+// WaitMaxDuration is like Wait except that it will
+// only take tokens from the bucket if it needs to wait
+// for no greater than maxWait. It reports whether
+// any tokens have been removed from the bucket
+// If no tokens have been removed, it returns immediately.
+func (tb *Bucket) WaitMaxDuration(count int64, maxWait time.Duration) bool {
+	d, ok := tb.TakeMaxDuration(count, maxWait)
+	if d > 0 {
+		time.Sleep(d)
+	}
+	return ok
+}
+
+const infinityDuration time.Duration = 0x7fffffffffffffff
+
+// Take takes count tokens from the bucket without blocking. It returns
+// the time that the caller should wait until the tokens are actually
+// available.
+//
+// Note that if the request is irrevocable - there is no way to return
+// tokens to the bucket once this method commits us to taking them.
+func (tb *Bucket) Take(count int64) time.Duration {
+	d, _ := tb.take(time.Now(), count, infinityDuration)
+	return d
+}
+
+// TakeMaxDuration is like Take, except that
+// it will only take tokens from the bucket if the wait
+// time for the tokens is no greater than maxWait.
+//
+// If it would take longer than maxWait for the tokens
+// to become available, it does nothing and reports false,
+// otherwise it returns the time that the caller should
+// wait until the tokens are actually available, and reports
+// true.
+func (tb *Bucket) TakeMaxDuration(count int64, maxWait time.Duration) (time.Duration, bool) {
+	return tb.take(time.Now(), count, maxWait)
+}
+
+// TakeAvailable takes up to count immediately available tokens from the
+// bucket. It returns the number of tokens removed, or zero if there are
+// no available tokens. It does not block.
+func (tb *Bucket) TakeAvailable(count int64) int64 {
+	return tb.takeAvailable(time.Now(), count)
+}
+
+// takeAvailable is the internal version of TakeAvailable - it takes the
+// current time as an argument to enable easy testing.
+func (tb *Bucket) takeAvailable(now time.Time, count int64) int64 {
+	if count <= 0 {
+		return 0
+	}
+	tb.mu.Lock()
+	defer tb.mu.Unlock()
+
+	tb.adjust(now)
+	if tb.avail <= 0 {
+		return 0
+	}
+	if count > tb.avail {
+		count = tb.avail
+	}
+	tb.avail -= count
+	return count
+}
+
+// Rate returns the fill rate of the bucket, in tokens per second.
+func (tb *Bucket) Rate() float64 {
+	return 1e9 * float64(tb.quantum) / float64(tb.fillInterval)
+}
+
+// take is the internal version of Take - it takes the current time as
+// an argument to enable easy testing.
+func (tb *Bucket) take(now time.Time, count int64, maxWait time.Duration) (time.Duration, bool) {
+	if count <= 0 {
+		return 0, true
+	}
+	tb.mu.Lock()
+	defer tb.mu.Unlock()
+
+	currentTick := tb.adjust(now)
+	avail := tb.avail - count
+	if avail >= 0 {
+		tb.avail = avail
+		return 0, true
+	}
+	// Round up the missing tokens to the nearest multiple
+	// of quantum - the tokens won't be available until
+	// that tick.
+	endTick := currentTick + (-avail+tb.quantum-1)/tb.quantum
+	endTime := tb.startTime.Add(time.Duration(endTick) * tb.fillInterval)
+	waitTime := endTime.Sub(now)
+	if waitTime > maxWait {
+		return 0, false
+	}
+	tb.avail = avail
+	return waitTime, true
+}
+
+// adjust adjusts the current bucket capacity based on the current time.
+// It returns the current tick.
+func (tb *Bucket) adjust(now time.Time) (currentTick int64) {
+	currentTick = int64(now.Sub(tb.startTime) / tb.fillInterval)
+
+	if tb.avail >= tb.capacity {
+		return
+	}
+	tb.avail += (currentTick - tb.availTick) * tb.quantum
+	if tb.avail > tb.capacity {
+		tb.avail = tb.capacity
+	}
+	tb.availTick = currentTick
+	return
+}
@@ -0,0 +1,328 @@
+// Copyright 2014 Canonical Ltd.
+// Licensed under the LGPLv3 with static-linking exception.
+// See LICENCE file for details.
+
+package ratelimit
+
+import (
+	gc "launchpad.net/gocheck"
+
+	"testing"
+	"time"
+)
+
+func TestPackage(t *testing.T) {
+	gc.TestingT(t)
+}
+
+type rateLimitSuite struct{}
+
+var _ = gc.Suite(rateLimitSuite{})
+
+type takeReq struct {
+	time       time.Duration
+	count      int64
+	expectWait time.Duration
+}
+
+var takeTests = []struct {
+	about        string
+	fillInterval time.Duration
+	capacity     int64
+	reqs         []takeReq
+}{{
+	about:        "serial requests",
+	fillInterval: 250 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeReq{{
+		time:       0,
+		count:      0,
+		expectWait: 0,
+	}, {
+		time:       0,
+		count:      10,
+		expectWait: 0,
+	}, {
+		time:       0,
+		count:      1,
+		expectWait: 250 * time.Millisecond,
+	}, {
+		time:       250 * time.Millisecond,
+		count:      1,
+		expectWait: 250 * time.Millisecond,
+	}},
+}, {
+	about:        "concurrent requests",
+	fillInterval: 250 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeReq{{
+		time:       0,
+		count:      10,
+		expectWait: 0,
+	}, {
+		time:       0,
+		count:      2,
+		expectWait: 500 * time.Millisecond,
+	}, {
+		time:       0,
+		count:      2,
+		expectWait: 1000 * time.Millisecond,
+	}, {
+		time:       0,
+		count:      1,
+		expectWait: 1250 * time.Millisecond,
+	}},
+}, {
+	about:        "more than capacity",
+	fillInterval: 1 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeReq{{
+		time:       0,
+		count:      10,
+		expectWait: 0,
+	}, {
+		time:       20 * time.Millisecond,
+		count:      15,
+		expectWait: 5 * time.Millisecond,
+	}},
+}, {
+	about:        "sub-quantum time",
+	fillInterval: 10 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeReq{{
+		time:       0,
+		count:      10,
+		expectWait: 0,
+	}, {
+		time:       7 * time.Millisecond,
+		count:      1,
+		expectWait: 3 * time.Millisecond,
+	}, {
+		time:       8 * time.Millisecond,
+		count:      1,
+		expectWait: 12 * time.Millisecond,
+	}},
+}, {
+	about:        "within capacity",
+	fillInterval: 10 * time.Millisecond,
+	capacity:     5,
+	reqs: []takeReq{{
+		time:       0,
+		count:      5,
+		expectWait: 0,
+	}, {
+		time:       60 * time.Millisecond,
+		count:      5,
+		expectWait: 0,
+	}, {
+		time:       60 * time.Millisecond,
+		count:      1,
+		expectWait: 10 * time.Millisecond,
+	}, {
+		time:       80 * time.Millisecond,
+		count:      2,
+		expectWait: 10 * time.Millisecond,
+	}},
+}}
+
+func (rateLimitSuite) TestTake(c *gc.C) {
+	for i, test := range takeTests {
+		tb := NewBucket(test.fillInterval, test.capacity)
+		for j, req := range test.reqs {
+			d, ok := tb.take(tb.startTime.Add(req.time), req.count, infinityDuration)
+			c.Assert(ok, gc.Equals, true)
+			if d != req.expectWait {
+				c.Fatalf("test %d.%d, %s, got %v want %v", i, j, test.about, d, req.expectWait)
+			}
+		}
+	}
+}
+
+func (rateLimitSuite) TestTakeMaxDuration(c *gc.C) {
+	for i, test := range takeTests {
+		tb := NewBucket(test.fillInterval, test.capacity)
+		for j, req := range test.reqs {
+			if req.expectWait > 0 {
+				d, ok := tb.take(tb.startTime.Add(req.time), req.count, req.expectWait-1)
+				c.Assert(ok, gc.Equals, false)
+				c.Assert(d, gc.Equals, time.Duration(0))
+			}
+			d, ok := tb.take(tb.startTime.Add(req.time), req.count, req.expectWait)
+			c.Assert(ok, gc.Equals, true)
+			if d != req.expectWait {
+				c.Fatalf("test %d.%d, %s, got %v want %v", i, j, test.about, d, req.expectWait)
+			}
+		}
+	}
+}
+
+type takeAvailableReq struct {
+	time   time.Duration
+	count  int64
+	expect int64
+}
+
+var takeAvailableTests = []struct {
+	about        string
+	fillInterval time.Duration
+	capacity     int64
+	reqs         []takeAvailableReq
+}{{
+	about:        "serial requests",
+	fillInterval: 250 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeAvailableReq{{
+		time:   0,
+		count:  0,
+		expect: 0,
+	}, {
+		time:   0,
+		count:  10,
+		expect: 10,
+	}, {
+		time:   0,
+		count:  1,
+		expect: 0,
+	}, {
+		time:   250 * time.Millisecond,
+		count:  1,
+		expect: 1,
+	}},
+}, {
+	about:        "concurrent requests",
+	fillInterval: 250 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeAvailableReq{{
+		time:   0,
+		count:  5,
+		expect: 5,
+	}, {
+		time:   0,
+		count:  2,
+		expect: 2,
+	}, {
+		time:   0,
+		count:  5,
+		expect: 3,
+	}, {
+		time:   0,
+		count:  1,
+		expect: 0,
+	}},
+}, {
+	about:        "more than capacity",
+	fillInterval: 1 * time.Millisecond,
+	capacity:     10,
+	reqs: []takeAvailableReq{{
+		time:   0,
+		count:  10,
+		expect: 10,
+	}, {
+		time:   20 * time.Millisecond,
+		count:  15,
+		expect: 10,
+	}},
+}, {
+	about:        "within capacity",
+	fillInterval: 10 * time.Millisecond,
+	capacity:     5,
+	reqs: []takeAvailableReq{{
+		time:   0,
+		count:  5,
+		expect: 5,
+	}, {
+		time:   60 * time.Millisecond,
+		count:  5,
+		expect: 5,
+	}, {
+		time:   70 * time.Millisecond,
+		count:  1,
+		expect: 1,
+	}},
+}}
+
+func (rateLimitSuite) TestTakeAvailable(c *gc.C) {
+	for i, test := range takeAvailableTests {
+		tb := NewBucket(test.fillInterval, test.capacity)
+		for j, req := range test.reqs {
+			d := tb.takeAvailable(tb.startTime.Add(req.time), req.count)
+			if d != req.expect {
+				c.Fatalf("test %d.%d, %s, got %v want %v", i, j, test.about, d, req.expect)
+			}
+		}
+	}
+}
+
+func (rateLimitSuite) TestPanics(c *gc.C) {
+	c.Assert(func() { NewBucket(0, 1) }, gc.PanicMatches, "token bucket fill interval is not > 0")
+	c.Assert(func() { NewBucket(-2, 1) }, gc.PanicMatches, "token bucket fill interval is not > 0")
+	c.Assert(func() { NewBucket(1, 0) }, gc.PanicMatches, "token bucket capacity is not > 0")
+	c.Assert(func() { NewBucket(1, -2) }, gc.PanicMatches, "token bucket capacity is not > 0")
+}
+
+func isCloseTo(x, y, tolerance float64) bool {
+	return abs(x-y)/y < tolerance
+}
+
+func (rateLimitSuite) TestRate(c *gc.C) {
+	tb := NewBucket(1, 1)
+	if !isCloseTo(tb.Rate(), 1e9, 0.00001) {
+		c.Fatalf("got %v want 1e9", tb.Rate())
+	}
+	tb = NewBucket(2*time.Second, 1)
+	if !isCloseTo(tb.Rate(), 0.5, 0.00001) {
+		c.Fatalf("got %v want 0.5", tb.Rate())
+	}
+	tb = NewBucketWithQuantum(100*time.Millisecond, 1, 5)
+	if !isCloseTo(tb.Rate(), 50, 0.00001) {
+		c.Fatalf("got %v want 50", tb.Rate())
+	}
+}
+
+func checkRate(c *gc.C, rate float64) {
+	tb := NewBucketWithRate(rate, 1<<62)
+	if !isCloseTo(tb.Rate(), rate, rateMargin) {
+		c.Fatalf("got %g want %v", tb.Rate(), rate)
+	}
+	d, ok := tb.take(tb.startTime, 1<<62, infinityDuration)
+	c.Assert(ok, gc.Equals, true)
+	c.Assert(d, gc.Equals, time.Duration(0))
+
+	// Check that the actual rate is as expected by
+	// asking for a not-quite multiple of the bucket's
+	// quantum and checking that the wait time
+	// correct.
+	d, ok = tb.take(tb.startTime, tb.quantum*2-tb.quantum/2, infinityDuration)
+	c.Assert(ok, gc.Equals, true)
+	expectTime := 1e9 * float64(tb.quantum) * 2 / rate
+	if !isCloseTo(float64(d), expectTime, rateMargin) {
+		c.Fatalf("rate %g: got %g want %v", rate, float64(d), expectTime)
+	}
+}
+
+func (rateLimitSuite) TestNewWithRate(c *gc.C) {
+	for rate := float64(1); rate < 1e6; rate += 7 {
+		checkRate(c, rate)
+	}
+	for _, rate := range []float64{
+		1024 * 1024 * 1024,
+		1e-5,
+		0.9e-5,
+		0.5,
+		0.9,
+		0.9e8,
+		3e12,
+		4e18,
+	} {
+		checkRate(c, rate)
+		checkRate(c, rate/3)
+		checkRate(c, rate*1.3)
+	}
+}
+
+func BenchmarkWait(b *testing.B) {
+	tb := NewBucket(1, 16*1024)
+	for i := b.N - 1; i >= 0; i-- {
+		tb.Wait(1)
+	}
+}
@@ -0,0 +1,51 @@
+// Copyright 2014 Canonical Ltd.
+// Licensed under the LGPLv3 with static-linking exception.
+// See LICENCE file for details.
+
+package ratelimit
+
+import "io"
+
+type reader struct {
+	r      io.Reader
+	bucket *Bucket
+}
+
+// Reader returns a reader that is rate limited by
+// the given token bucket. Each token in the bucket
+// represents one byte.
+func Reader(r io.Reader, bucket *Bucket) io.Reader {
+	return &reader{
+		r:      r,
+		bucket: bucket,
+	}
+}
+
+func (r *reader) Read(buf []byte) (int, error) {
+	n, err := r.r.Read(buf)
+	if n <= 0 {
+		return n, err
+	}
+	r.bucket.Wait(int64(n))
+	return n, err
+}
+
+type writer struct {
+	w      io.Writer
+	bucket *Bucket
+}
+
+// Writer returns a reader that is rate limited by
+// the given token bucket. Each token in the bucket
+// represents one byte.
+func Writer(w io.Writer, bucket *Bucket) io.Writer {
+	return &writer{
+		w:      w,
+		bucket: bucket,
+	}
+}
+
+func (w *writer) Write(buf []byte) (int, error) {
+	w.bucket.Wait(int64(len(buf)))
+	return w.w.Write(buf)
+}
@@ -0,0 +1,27 @@
+Copyright (c) 2012 The Go Authors. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+   * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+   * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+   * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
@@ -0,0 +1,16 @@
+### Extensions to the "os" package.
+
+## Find the current Executable and ExecutableFolder.
+
+There is sometimes utility in finding the current executable file
+that is running. This can be used for upgrading the current executable
+or finding resources located relative to the executable file. Both
+working directory and the os.Args[0] value are arbitrary and cannot
+be relied on; os.Args[0] can be "faked".
+
+Multi-platform and supports:
+ * Linux
+ * OS X
+ * Windows
+ * Plan 9
+ * BSDs.
@@ -0,0 +1,27 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Extensions to the standard "os" package.
+package osext
+
+import "path/filepath"
+
+// Executable returns an absolute path that can be used to
+// re-invoke the current program.
+// It may not be valid after the current program exits.
+func Executable() (string, error) {
+	p, err := executable()
+	return filepath.Clean(p), err
+}
+
+// Returns same path as Executable, returns just the folder
+// path. Excludes the executable name and any trailing slash.
+func ExecutableFolder() (string, error) {
+	p, err := Executable()
+	if err != nil {
+		return "", err
+	}
+
+	return filepath.Dir(p), nil
+}
@@ -0,0 +1,20 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package osext
+
+import (
+	"os"
+	"strconv"
+	"syscall"
+)
+
+func executable() (string, error) {
+	f, err := os.Open("/proc/" + strconv.Itoa(os.Getpid()) + "/text")
+	if err != nil {
+		return "", err
+	}
+	defer f.Close()
+	return syscall.Fd2path(int(f.Fd()))
+}
@@ -0,0 +1,36 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build linux netbsd openbsd solaris dragonfly
+
+package osext
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"runtime"
+	"strings"
+)
+
+func executable() (string, error) {
+	switch runtime.GOOS {
+	case "linux":
+		const deletedTag = " (deleted)"
+		execpath, err := os.Readlink("/proc/self/exe")
+		if err != nil {
+			return execpath, err
+		}
+		execpath = strings.TrimSuffix(execpath, deletedTag)
+		execpath = strings.TrimPrefix(execpath, deletedTag)
+		return execpath, nil
+	case "netbsd":
+		return os.Readlink("/proc/curproc/exe")
+	case "openbsd", "dragonfly":
+		return os.Readlink("/proc/curproc/file")
+	case "solaris":
+		return os.Readlink(fmt.Sprintf("/proc/%d/path/a.out", os.Getpid()))
+	}
+	return "", errors.New("ExecPath not implemented for " + runtime.GOOS)
+}
@@ -0,0 +1,79 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build darwin freebsd
+
+package osext
+
+import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"syscall"
+	"unsafe"
+)
+
+var initCwd, initCwdErr = os.Getwd()
+
+func executable() (string, error) {
+	var mib [4]int32
+	switch runtime.GOOS {
+	case "freebsd":
+		mib = [4]int32{1 /* CTL_KERN */, 14 /* KERN_PROC */, 12 /* KERN_PROC_PATHNAME */, -1}
+	case "darwin":
+		mib = [4]int32{1 /* CTL_KERN */, 38 /* KERN_PROCARGS */, int32(os.Getpid()), -1}
+	}
+
+	n := uintptr(0)
+	// Get length.
+	_, _, errNum := syscall.Syscall6(syscall.SYS___SYSCTL, uintptr(unsafe.Pointer(&mib[0])), 4, 0, uintptr(unsafe.Pointer(&n)), 0, 0)
+	if errNum != 0 {
+		return "", errNum
+	}
+	if n == 0 { // This shouldn't happen.
+		return "", nil
+	}
+	buf := make([]byte, n)
+	_, _, errNum = syscall.Syscall6(syscall.SYS___SYSCTL, uintptr(unsafe.Pointer(&mib[0])), 4, uintptr(unsafe.Pointer(&buf[0])), uintptr(unsafe.Pointer(&n)), 0, 0)
+	if errNum != 0 {
+		return "", errNum
+	}
+	if n == 0 { // This shouldn't happen.
+		return "", nil
+	}
+	for i, v := range buf {
+		if v == 0 {
+			buf = buf[:i]
+			break
+		}
+	}
+	var err error
+	execPath := string(buf)
+	// execPath will not be empty due to above checks.
+	// Try to get the absolute path if the execPath is not rooted.
+	if execPath[0] != '/' {
+		execPath, err = getAbs(execPath)
+		if err != nil {
+			return execPath, err
+		}
+	}
+	// For darwin KERN_PROCARGS may return the path to a symlink rather than the
+	// actual executable.
+	if runtime.GOOS == "darwin" {
+		if execPath, err = filepath.EvalSymlinks(execPath); err != nil {
+			return execPath, err
+		}
+	}
+	return execPath, nil
+}
+
+func getAbs(execPath string) (string, error) {
+	if initCwdErr != nil {
+		return execPath, initCwdErr
+	}
+	// The execPath may begin with a "../" or a "./" so clean it first.
+	// Join the two paths, trailing and starting slashes undetermined, so use
+	// the generic Join function.
+	return filepath.Join(initCwd, filepath.Clean(execPath)), nil
+}
@@ -0,0 +1,203 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build darwin linux freebsd netbsd windows
+
+package osext
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"testing"
+)
+
+const (
+	executableEnvVar = "OSTEST_OUTPUT_EXECUTABLE"
+
+	executableEnvValueMatch  = "match"
+	executableEnvValueDelete = "delete"
+)
+
+func TestPrintExecutable(t *testing.T) {
+	ef, err := Executable()
+	if err != nil {
+		t.Fatalf("Executable failed: %v", err)
+	}
+	t.Log("Executable:", ef)
+}
+func TestPrintExecutableFolder(t *testing.T) {
+	ef, err := ExecutableFolder()
+	if err != nil {
+		t.Fatalf("ExecutableFolder failed: %v", err)
+	}
+	t.Log("Executable Folder:", ef)
+}
+func TestExecutableFolder(t *testing.T) {
+	ef, err := ExecutableFolder()
+	if err != nil {
+		t.Fatalf("ExecutableFolder failed: %v", err)
+	}
+	if ef[len(ef)-1] == filepath.Separator {
+		t.Fatal("ExecutableFolder ends with a trailing slash.")
+	}
+}
+func TestExecutableMatch(t *testing.T) {
+	ep, err := Executable()
+	if err != nil {
+		t.Fatalf("Executable failed: %v", err)
+	}
+
+	// fullpath to be of the form "dir/prog".
+	dir := filepath.Dir(filepath.Dir(ep))
+	fullpath, err := filepath.Rel(dir, ep)
+	if err != nil {
+		t.Fatalf("filepath.Rel: %v", err)
+	}
+	// Make child start with a relative program path.
+	// Alter argv[0] for child to verify getting real path without argv[0].
+	cmd := &exec.Cmd{
+		Dir:  dir,
+		Path: fullpath,
+		Env:  []string{fmt.Sprintf("%s=%s", executableEnvVar, executableEnvValueMatch)},
+	}
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("exec(self) failed: %v", err)
+	}
+	outs := string(out)
+	if !filepath.IsAbs(outs) {
+		t.Fatalf("Child returned %q, want an absolute path", out)
+	}
+	if !sameFile(outs, ep) {
+		t.Fatalf("Child returned %q, not the same file as %q", out, ep)
+	}
+}
+
+func TestExecutableDelete(t *testing.T) {
+	if runtime.GOOS != "linux" {
+		t.Skip()
+	}
+	fpath, err := Executable()
+	if err != nil {
+		t.Fatalf("Executable failed: %v", err)
+	}
+
+	r, w := io.Pipe()
+	stderrBuff := &bytes.Buffer{}
+	stdoutBuff := &bytes.Buffer{}
+	cmd := &exec.Cmd{
+		Path:   fpath,
+		Env:    []string{fmt.Sprintf("%s=%s", executableEnvVar, executableEnvValueDelete)},
+		Stdin:  r,
+		Stderr: stderrBuff,
+		Stdout: stdoutBuff,
+	}
+	err = cmd.Start()
+	if err != nil {
+		t.Fatalf("exec(self) start failed: %v", err)
+	}
+
+	tempPath := fpath + "_copy"
+	_ = os.Remove(tempPath)
+
+	err = copyFile(tempPath, fpath)
+	if err != nil {
+		t.Fatalf("copy file failed: %v", err)
+	}
+	err = os.Remove(fpath)
+	if err != nil {
+		t.Fatalf("remove running test file failed: %v", err)
+	}
+	err = os.Rename(tempPath, fpath)
+	if err != nil {
+		t.Fatalf("rename copy to previous name failed: %v", err)
+	}
+
+	w.Write([]byte{0})
+	w.Close()
+
+	err = cmd.Wait()
+	if err != nil {
+		t.Fatalf("exec wait failed: %v", err)
+	}
+
+	childPath := stderrBuff.String()
+	if !filepath.IsAbs(childPath) {
+		t.Fatalf("Child returned %q, want an absolute path", childPath)
+	}
+	if !sameFile(childPath, fpath) {
+		t.Fatalf("Child returned %q, not the same file as %q", childPath, fpath)
+	}
+}
+
+func sameFile(fn1, fn2 string) bool {
+	fi1, err := os.Stat(fn1)
+	if err != nil {
+		return false
+	}
+	fi2, err := os.Stat(fn2)
+	if err != nil {
+		return false
+	}
+	return os.SameFile(fi1, fi2)
+}
+func copyFile(dest, src string) error {
+	df, err := os.Create(dest)
+	if err != nil {
+		return err
+	}
+	defer df.Close()
+
+	sf, err := os.Open(src)
+	if err != nil {
+		return err
+	}
+	defer sf.Close()
+
+	_, err = io.Copy(df, sf)
+	return err
+}
+
+func TestMain(m *testing.M) {
+	env := os.Getenv(executableEnvVar)
+	switch env {
+	case "":
+		os.Exit(m.Run())
+	case executableEnvValueMatch:
+		// First chdir to another path.
+		dir := "/"
+		if runtime.GOOS == "windows" {
+			dir = filepath.VolumeName(".")
+		}
+		os.Chdir(dir)
+		if ep, err := Executable(); err != nil {
+			fmt.Fprint(os.Stderr, "ERROR: ", err)
+		} else {
+			fmt.Fprint(os.Stderr, ep)
+		}
+	case executableEnvValueDelete:
+		bb := make([]byte, 1)
+		var err error
+		n, err := os.Stdin.Read(bb)
+		if err != nil {
+			fmt.Fprint(os.Stderr, "ERROR: ", err)
+			os.Exit(2)
+		}
+		if n != 1 {
+			fmt.Fprint(os.Stderr, "ERROR: n != 1, n == ", n)
+			os.Exit(2)
+		}
+		if ep, err := Executable(); err != nil {
+			fmt.Fprint(os.Stderr, "ERROR: ", err)
+		} else {
+			fmt.Fprint(os.Stderr, ep)
+		}
+	}
+	os.Exit(0)
+}
@@ -0,0 +1,34 @@
+// Copyright 2012 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package osext
+
+import (
+	"syscall"
+	"unicode/utf16"
+	"unsafe"
+)
+
+var (
+	kernel                = syscall.MustLoadDLL("kernel32.dll")
+	getModuleFileNameProc = kernel.MustFindProc("GetModuleFileNameW")
+)
+
+// GetModuleFileName() with hModule = NULL
+func executable() (exePath string, err error) {
+	return getModuleFileName()
+}
+
+func getModuleFileName() (string, error) {
+	var n uint32
+	b := make([]uint16, syscall.MAX_PATH)
+	size := uint32(len(b))
+
+	r0, _, e1 := getModuleFileNameProc.Call(0, uintptr(unsafe.Pointer(&b[0])), uintptr(size))
+	n = uint32(r0)
+	if n == 0 {
+		return "", e1
+	}
+	return string(utf16.Decode(b[0:n])), nil
+}
@@ -0,0 +1,252 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"encoding/binary"
+	"fmt"
+
+	"github.com/syndtr/goleveldb/leveldb/errors"
+	"github.com/syndtr/goleveldb/leveldb/memdb"
+)
+
+type ErrBatchCorrupted struct {
+	Reason string
+}
+
+func (e *ErrBatchCorrupted) Error() string {
+	return fmt.Sprintf("leveldb: batch corrupted: %s", e.Reason)
+}
+
+func newErrBatchCorrupted(reason string) error {
+	return errors.NewErrCorrupted(nil, &ErrBatchCorrupted{reason})
+}
+
+const (
+	batchHdrLen  = 8 + 4
+	batchGrowRec = 3000
+)
+
+type BatchReplay interface {
+	Put(key, value []byte)
+	Delete(key []byte)
+}
+
+// Batch is a write batch.
+type Batch struct {
+	data       []byte
+	rLen, bLen int
+	seq        uint64
+	sync       bool
+}
+
+func (b *Batch) grow(n int) {
+	off := len(b.data)
+	if off == 0 {
+		off = batchHdrLen
+		if b.data != nil {
+			b.data = b.data[:off]
+		}
+	}
+	if cap(b.data)-off < n {
+		if b.data == nil {
+			b.data = make([]byte, off, off+n)
+		} else {
+			odata := b.data
+			div := 1
+			if b.rLen > batchGrowRec {
+				div = b.rLen / batchGrowRec
+			}
+			b.data = make([]byte, off, off+n+(off-batchHdrLen)/div)
+			copy(b.data, odata)
+		}
+	}
+}
+
+func (b *Batch) appendRec(kt kType, key, value []byte) {
+	n := 1 + binary.MaxVarintLen32 + len(key)
+	if kt == ktVal {
+		n += binary.MaxVarintLen32 + len(value)
+	}
+	b.grow(n)
+	off := len(b.data)
+	data := b.data[:off+n]
+	data[off] = byte(kt)
+	off += 1
+	off += binary.PutUvarint(data[off:], uint64(len(key)))
+	copy(data[off:], key)
+	off += len(key)
+	if kt == ktVal {
+		off += binary.PutUvarint(data[off:], uint64(len(value)))
+		copy(data[off:], value)
+		off += len(value)
+	}
+	b.data = data[:off]
+	b.rLen++
+	//  Include 8-byte ikey header
+	b.bLen += len(key) + len(value) + 8
+}
+
+// Put appends 'put operation' of the given key/value pair to the batch.
+// It is safe to modify the contents of the argument after Put returns.
+func (b *Batch) Put(key, value []byte) {
+	b.appendRec(ktVal, key, value)
+}
+
+// Delete appends 'delete operation' of the given key to the batch.
+// It is safe to modify the contents of the argument after Delete returns.
+func (b *Batch) Delete(key []byte) {
+	b.appendRec(ktDel, key, nil)
+}
+
+// Dump dumps batch contents. The returned slice can be loaded into the
+// batch using Load method.
+// The returned slice is not its own copy, so the contents should not be
+// modified.
+func (b *Batch) Dump() []byte {
+	return b.encode()
+}
+
+// Load loads given slice into the batch. Previous contents of the batch
+// will be discarded.
+// The given slice will not be copied and will be used as batch buffer, so
+// it is not safe to modify the contents of the slice.
+func (b *Batch) Load(data []byte) error {
+	return b.decode(0, data)
+}
+
+// Replay replays batch contents.
+func (b *Batch) Replay(r BatchReplay) error {
+	return b.decodeRec(func(i int, kt kType, key, value []byte) {
+		switch kt {
+		case ktVal:
+			r.Put(key, value)
+		case ktDel:
+			r.Delete(key)
+		}
+	})
+}
+
+// Len returns number of records in the batch.
+func (b *Batch) Len() int {
+	return b.rLen
+}
+
+// Reset resets the batch.
+func (b *Batch) Reset() {
+	b.data = b.data[:0]
+	b.seq = 0
+	b.rLen = 0
+	b.bLen = 0
+	b.sync = false
+}
+
+func (b *Batch) init(sync bool) {
+	b.sync = sync
+}
+
+func (b *Batch) append(p *Batch) {
+	if p.rLen > 0 {
+		b.grow(len(p.data) - batchHdrLen)
+		b.data = append(b.data, p.data[batchHdrLen:]...)
+		b.rLen += p.rLen
+	}
+	if p.sync {
+		b.sync = true
+	}
+}
+
+// size returns sums of key/value pair length plus 8-bytes ikey.
+func (b *Batch) size() int {
+	return b.bLen
+}
+
+func (b *Batch) encode() []byte {
+	b.grow(0)
+	binary.LittleEndian.PutUint64(b.data, b.seq)
+	binary.LittleEndian.PutUint32(b.data[8:], uint32(b.rLen))
+
+	return b.data
+}
+
+func (b *Batch) decode(prevSeq uint64, data []byte) error {
+	if len(data) < batchHdrLen {
+		return newErrBatchCorrupted("too short")
+	}
+
+	b.seq = binary.LittleEndian.Uint64(data)
+	if b.seq < prevSeq {
+		return newErrBatchCorrupted("invalid sequence number")
+	}
+	b.rLen = int(binary.LittleEndian.Uint32(data[8:]))
+	if b.rLen < 0 {
+		return newErrBatchCorrupted("invalid records length")
+	}
+	// No need to be precise at this point, it won't be used anyway
+	b.bLen = len(data) - batchHdrLen
+	b.data = data
+
+	return nil
+}
+
+func (b *Batch) decodeRec(f func(i int, kt kType, key, value []byte)) (err error) {
+	off := batchHdrLen
+	for i := 0; i < b.rLen; i++ {
+		if off >= len(b.data) {
+			return newErrBatchCorrupted("invalid records length")
+		}
+
+		kt := kType(b.data[off])
+		if kt > ktVal {
+			return newErrBatchCorrupted("bad record: invalid type")
+		}
+		off += 1
+
+		x, n := binary.Uvarint(b.data[off:])
+		off += n
+		if n <= 0 || off+int(x) > len(b.data) {
+			return newErrBatchCorrupted("bad record: invalid key length")
+		}
+		key := b.data[off : off+int(x)]
+		off += int(x)
+		var value []byte
+		if kt == ktVal {
+			x, n := binary.Uvarint(b.data[off:])
+			off += n
+			if n <= 0 || off+int(x) > len(b.data) {
+				return newErrBatchCorrupted("bad record: invalid value length")
+			}
+			value = b.data[off : off+int(x)]
+			off += int(x)
+		}
+
+		f(i, kt, key, value)
+	}
+
+	return nil
+}
+
+func (b *Batch) memReplay(to *memdb.DB) error {
+	return b.decodeRec(func(i int, kt kType, key, value []byte) {
+		ikey := newIkey(key, b.seq+uint64(i), kt)
+		to.Put(ikey, value)
+	})
+}
+
+func (b *Batch) memDecodeAndReplay(prevSeq uint64, data []byte, to *memdb.DB) error {
+	if err := b.decode(prevSeq, data); err != nil {
+		return err
+	}
+	return b.memReplay(to)
+}
+
+func (b *Batch) revertMemReplay(to *memdb.DB) error {
+	return b.decodeRec(func(i int, kt kType, key, value []byte) {
+		ikey := newIkey(key, b.seq+uint64(i), kt)
+		to.Delete(ikey)
+	})
+}
@@ -0,0 +1,120 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/syndtr/goleveldb/leveldb/comparer"
+	"github.com/syndtr/goleveldb/leveldb/memdb"
+)
+
+type tbRec struct {
+	kt         kType
+	key, value []byte
+}
+
+type testBatch struct {
+	rec []*tbRec
+}
+
+func (p *testBatch) Put(key, value []byte) {
+	p.rec = append(p.rec, &tbRec{ktVal, key, value})
+}
+
+func (p *testBatch) Delete(key []byte) {
+	p.rec = append(p.rec, &tbRec{ktDel, key, nil})
+}
+
+func compareBatch(t *testing.T, b1, b2 *Batch) {
+	if b1.seq != b2.seq {
+		t.Errorf("invalid seq number want %d, got %d", b1.seq, b2.seq)
+	}
+	if b1.Len() != b2.Len() {
+		t.Fatalf("invalid record length want %d, got %d", b1.Len(), b2.Len())
+	}
+	p1, p2 := new(testBatch), new(testBatch)
+	err := b1.Replay(p1)
+	if err != nil {
+		t.Fatal("error when replaying batch 1: ", err)
+	}
+	err = b2.Replay(p2)
+	if err != nil {
+		t.Fatal("error when replaying batch 2: ", err)
+	}
+	for i := range p1.rec {
+		r1, r2 := p1.rec[i], p2.rec[i]
+		if r1.kt != r2.kt {
+			t.Errorf("invalid type on record '%d' want %d, got %d", i, r1.kt, r2.kt)
+		}
+		if !bytes.Equal(r1.key, r2.key) {
+			t.Errorf("invalid key on record '%d' want %s, got %s", i, string(r1.key), string(r2.key))
+		}
+		if r1.kt == ktVal {
+			if !bytes.Equal(r1.value, r2.value) {
+				t.Errorf("invalid value on record '%d' want %s, got %s", i, string(r1.value), string(r2.value))
+			}
+		}
+	}
+}
+
+func TestBatch_EncodeDecode(t *testing.T) {
+	b1 := new(Batch)
+	b1.seq = 10009
+	b1.Put([]byte("key1"), []byte("value1"))
+	b1.Put([]byte("key2"), []byte("value2"))
+	b1.Delete([]byte("key1"))
+	b1.Put([]byte("k"), []byte(""))
+	b1.Put([]byte("zzzzzzzzzzz"), []byte("zzzzzzzzzzzzzzzzzzzzzzzz"))
+	b1.Delete([]byte("key10000"))
+	b1.Delete([]byte("k"))
+	buf := b1.encode()
+	b2 := new(Batch)
+	err := b2.decode(0, buf)
+	if err != nil {
+		t.Error("error when decoding batch: ", err)
+	}
+	compareBatch(t, b1, b2)
+}
+
+func TestBatch_Append(t *testing.T) {
+	b1 := new(Batch)
+	b1.seq = 10009
+	b1.Put([]byte("key1"), []byte("value1"))
+	b1.Put([]byte("key2"), []byte("value2"))
+	b1.Delete([]byte("key1"))
+	b1.Put([]byte("foo"), []byte("foovalue"))
+	b1.Put([]byte("bar"), []byte("barvalue"))
+	b2a := new(Batch)
+	b2a.seq = 10009
+	b2a.Put([]byte("key1"), []byte("value1"))
+	b2a.Put([]byte("key2"), []byte("value2"))
+	b2a.Delete([]byte("key1"))
+	b2b := new(Batch)
+	b2b.Put([]byte("foo"), []byte("foovalue"))
+	b2b.Put([]byte("bar"), []byte("barvalue"))
+	b2a.append(b2b)
+	compareBatch(t, b1, b2a)
+}
+
+func TestBatch_Size(t *testing.T) {
+	b := new(Batch)
+	for i := 0; i < 2; i++ {
+		b.Put([]byte("key1"), []byte("value1"))
+		b.Put([]byte("key2"), []byte("value2"))
+		b.Delete([]byte("key1"))
+		b.Put([]byte("foo"), []byte("foovalue"))
+		b.Put([]byte("bar"), []byte("barvalue"))
+		mem := memdb.New(&iComparer{comparer.DefaultComparer}, 0)
+		b.memReplay(mem)
+		if b.size() != mem.Size() {
+			t.Errorf("invalid batch size calculation, want=%d got=%d", mem.Size(), b.size())
+		}
+		b.Reset()
+	}
+}
@@ -0,0 +1,58 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build !go1.2
+
+package leveldb
+
+import (
+	"sync/atomic"
+	"testing"
+)
+
+func BenchmarkDBReadConcurrent(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+	defer p.close()
+
+	b.ResetTimer()
+	b.SetBytes(116)
+
+	b.RunParallel(func(pb *testing.PB) {
+		iter := p.newIter()
+		defer iter.Release()
+		for pb.Next() && iter.Next() {
+		}
+	})
+}
+
+func BenchmarkDBReadConcurrent2(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+	defer p.close()
+
+	b.ResetTimer()
+	b.SetBytes(116)
+
+	var dir uint32
+	b.RunParallel(func(pb *testing.PB) {
+		iter := p.newIter()
+		defer iter.Release()
+		if atomic.AddUint32(&dir, 1)%2 == 0 {
+			for pb.Next() && iter.Next() {
+			}
+		} else {
+			if pb.Next() && iter.Last() {
+				for pb.Next() && iter.Prev() {
+				}
+			}
+		}
+	})
+}
@@ -0,0 +1,464 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"bytes"
+	"fmt"
+	"math/rand"
+	"os"
+	"path/filepath"
+	"runtime"
+	"testing"
+
+	"github.com/syndtr/goleveldb/leveldb/iterator"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/storage"
+)
+
+func randomString(r *rand.Rand, n int) []byte {
+	b := new(bytes.Buffer)
+	for i := 0; i < n; i++ {
+		b.WriteByte(' ' + byte(r.Intn(95)))
+	}
+	return b.Bytes()
+}
+
+func compressibleStr(r *rand.Rand, frac float32, n int) []byte {
+	nn := int(float32(n) * frac)
+	rb := randomString(r, nn)
+	b := make([]byte, 0, n+nn)
+	for len(b) < n {
+		b = append(b, rb...)
+	}
+	return b[:n]
+}
+
+type valueGen struct {
+	src []byte
+	pos int
+}
+
+func newValueGen(frac float32) *valueGen {
+	v := new(valueGen)
+	r := rand.New(rand.NewSource(301))
+	v.src = make([]byte, 0, 1048576+100)
+	for len(v.src) < 1048576 {
+		v.src = append(v.src, compressibleStr(r, frac, 100)...)
+	}
+	return v
+}
+
+func (v *valueGen) get(n int) []byte {
+	if v.pos+n > len(v.src) {
+		v.pos = 0
+	}
+	v.pos += n
+	return v.src[v.pos-n : v.pos]
+}
+
+var benchDB = filepath.Join(os.TempDir(), fmt.Sprintf("goleveldbbench-%d", os.Getuid()))
+
+type dbBench struct {
+	b    *testing.B
+	stor storage.Storage
+	db   *DB
+
+	o  *opt.Options
+	ro *opt.ReadOptions
+	wo *opt.WriteOptions
+
+	keys, values [][]byte
+}
+
+func openDBBench(b *testing.B, noCompress bool) *dbBench {
+	_, err := os.Stat(benchDB)
+	if err == nil {
+		err = os.RemoveAll(benchDB)
+		if err != nil {
+			b.Fatal("cannot remove old db: ", err)
+		}
+	}
+
+	p := &dbBench{
+		b:  b,
+		o:  &opt.Options{},
+		ro: &opt.ReadOptions{},
+		wo: &opt.WriteOptions{},
+	}
+	p.stor, err = storage.OpenFile(benchDB)
+	if err != nil {
+		b.Fatal("cannot open stor: ", err)
+	}
+	if noCompress {
+		p.o.Compression = opt.NoCompression
+	}
+
+	p.db, err = Open(p.stor, p.o)
+	if err != nil {
+		b.Fatal("cannot open db: ", err)
+	}
+
+	runtime.GOMAXPROCS(runtime.NumCPU())
+	return p
+}
+
+func (p *dbBench) reopen() {
+	p.db.Close()
+	var err error
+	p.db, err = Open(p.stor, p.o)
+	if err != nil {
+		p.b.Fatal("Reopen: got error: ", err)
+	}
+}
+
+func (p *dbBench) populate(n int) {
+	p.keys, p.values = make([][]byte, n), make([][]byte, n)
+	v := newValueGen(0.5)
+	for i := range p.keys {
+		p.keys[i], p.values[i] = []byte(fmt.Sprintf("%016d", i)), v.get(100)
+	}
+}
+
+func (p *dbBench) randomize() {
+	m := len(p.keys)
+	times := m * 2
+	r1, r2 := rand.New(rand.NewSource(0xdeadbeef)), rand.New(rand.NewSource(0xbeefface))
+	for n := 0; n < times; n++ {
+		i, j := r1.Int()%m, r2.Int()%m
+		if i == j {
+			continue
+		}
+		p.keys[i], p.keys[j] = p.keys[j], p.keys[i]
+		p.values[i], p.values[j] = p.values[j], p.values[i]
+	}
+}
+
+func (p *dbBench) writes(perBatch int) {
+	b := p.b
+	db := p.db
+
+	n := len(p.keys)
+	m := n / perBatch
+	if n%perBatch > 0 {
+		m++
+	}
+	batches := make([]Batch, m)
+	j := 0
+	for i := range batches {
+		first := true
+		for ; j < n && ((j+1)%perBatch != 0 || first); j++ {
+			first = false
+			batches[i].Put(p.keys[j], p.values[j])
+		}
+	}
+	runtime.GC()
+
+	b.ResetTimer()
+	b.StartTimer()
+	for i := range batches {
+		err := db.Write(&(batches[i]), p.wo)
+		if err != nil {
+			b.Fatal("write failed: ", err)
+		}
+	}
+	b.StopTimer()
+	b.SetBytes(116)
+}
+
+func (p *dbBench) gc() {
+	p.keys, p.values = nil, nil
+	runtime.GC()
+}
+
+func (p *dbBench) puts() {
+	b := p.b
+	db := p.db
+
+	b.ResetTimer()
+	b.StartTimer()
+	for i := range p.keys {
+		err := db.Put(p.keys[i], p.values[i], p.wo)
+		if err != nil {
+			b.Fatal("put failed: ", err)
+		}
+	}
+	b.StopTimer()
+	b.SetBytes(116)
+}
+
+func (p *dbBench) fill() {
+	b := p.b
+	db := p.db
+
+	perBatch := 10000
+	batch := new(Batch)
+	for i, n := 0, len(p.keys); i < n; {
+		first := true
+		for ; i < n && ((i+1)%perBatch != 0 || first); i++ {
+			first = false
+			batch.Put(p.keys[i], p.values[i])
+		}
+		err := db.Write(batch, p.wo)
+		if err != nil {
+			b.Fatal("write failed: ", err)
+		}
+		batch.Reset()
+	}
+}
+
+func (p *dbBench) gets() {
+	b := p.b
+	db := p.db
+
+	b.ResetTimer()
+	for i := range p.keys {
+		_, err := db.Get(p.keys[i], p.ro)
+		if err != nil {
+			b.Error("got error: ", err)
+		}
+	}
+	b.StopTimer()
+}
+
+func (p *dbBench) seeks() {
+	b := p.b
+
+	iter := p.newIter()
+	defer iter.Release()
+	b.ResetTimer()
+	for i := range p.keys {
+		if !iter.Seek(p.keys[i]) {
+			b.Error("value not found for: ", string(p.keys[i]))
+		}
+	}
+	b.StopTimer()
+}
+
+func (p *dbBench) newIter() iterator.Iterator {
+	iter := p.db.NewIterator(nil, p.ro)
+	err := iter.Error()
+	if err != nil {
+		p.b.Fatal("cannot create iterator: ", err)
+	}
+	return iter
+}
+
+func (p *dbBench) close() {
+	if bp, err := p.db.GetProperty("leveldb.blockpool"); err == nil {
+		p.b.Log("Block pool stats: ", bp)
+	}
+	p.db.Close()
+	p.stor.Close()
+	os.RemoveAll(benchDB)
+	p.db = nil
+	p.keys = nil
+	p.values = nil
+	runtime.GC()
+	runtime.GOMAXPROCS(1)
+}
+
+func BenchmarkDBWrite(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.writes(1)
+	p.close()
+}
+
+func BenchmarkDBWriteBatch(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.writes(1000)
+	p.close()
+}
+
+func BenchmarkDBWriteUncompressed(b *testing.B) {
+	p := openDBBench(b, true)
+	p.populate(b.N)
+	p.writes(1)
+	p.close()
+}
+
+func BenchmarkDBWriteBatchUncompressed(b *testing.B) {
+	p := openDBBench(b, true)
+	p.populate(b.N)
+	p.writes(1000)
+	p.close()
+}
+
+func BenchmarkDBWriteRandom(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.randomize()
+	p.writes(1)
+	p.close()
+}
+
+func BenchmarkDBWriteRandomSync(b *testing.B) {
+	p := openDBBench(b, false)
+	p.wo.Sync = true
+	p.populate(b.N)
+	p.writes(1)
+	p.close()
+}
+
+func BenchmarkDBOverwrite(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.writes(1)
+	p.writes(1)
+	p.close()
+}
+
+func BenchmarkDBOverwriteRandom(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.writes(1)
+	p.randomize()
+	p.writes(1)
+	p.close()
+}
+
+func BenchmarkDBPut(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.puts()
+	p.close()
+}
+
+func BenchmarkDBRead(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+
+	iter := p.newIter()
+	b.ResetTimer()
+	for iter.Next() {
+	}
+	iter.Release()
+	b.StopTimer()
+	b.SetBytes(116)
+	p.close()
+}
+
+func BenchmarkDBReadGC(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+
+	iter := p.newIter()
+	b.ResetTimer()
+	for iter.Next() {
+	}
+	iter.Release()
+	b.StopTimer()
+	b.SetBytes(116)
+	p.close()
+}
+
+func BenchmarkDBReadUncompressed(b *testing.B) {
+	p := openDBBench(b, true)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+
+	iter := p.newIter()
+	b.ResetTimer()
+	for iter.Next() {
+	}
+	iter.Release()
+	b.StopTimer()
+	b.SetBytes(116)
+	p.close()
+}
+
+func BenchmarkDBReadTable(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.reopen()
+	p.gc()
+
+	iter := p.newIter()
+	b.ResetTimer()
+	for iter.Next() {
+	}
+	iter.Release()
+	b.StopTimer()
+	b.SetBytes(116)
+	p.close()
+}
+
+func BenchmarkDBReadReverse(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gc()
+
+	iter := p.newIter()
+	b.ResetTimer()
+	iter.Last()
+	for iter.Prev() {
+	}
+	iter.Release()
+	b.StopTimer()
+	b.SetBytes(116)
+	p.close()
+}
+
+func BenchmarkDBReadReverseTable(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.reopen()
+	p.gc()
+
+	iter := p.newIter()
+	b.ResetTimer()
+	iter.Last()
+	for iter.Prev() {
+	}
+	iter.Release()
+	b.StopTimer()
+	b.SetBytes(116)
+	p.close()
+}
+
+func BenchmarkDBSeek(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.seeks()
+	p.close()
+}
+
+func BenchmarkDBSeekRandom(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.randomize()
+	p.seeks()
+	p.close()
+}
+
+func BenchmarkDBGet(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.gets()
+	p.close()
+}
+
+func BenchmarkDBGetRandom(b *testing.B) {
+	p := openDBBench(b, false)
+	p.populate(b.N)
+	p.fill()
+	p.randomize()
+	p.gets()
+	p.close()
+}
@@ -0,0 +1,30 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// +build !go1.2
+
+package cache
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func BenchmarkLRUCache(b *testing.B) {
+	c := NewCache(NewLRU(10000))
+
+	b.SetParallelism(10)
+	b.RunParallel(func(pb *testing.PB) {
+		r := rand.New(rand.NewSource(time.Now().UnixNano()))
+
+		for pb.Next() {
+			key := uint64(r.Intn(1000000))
+			c.Get(0, key, func() (int, Value) {
+				return 1, key
+			}).Release()
+		}
+	})
+}
@@ -0,0 +1,676 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// Package cache provides interface and implementation of a cache algorithms.
+package cache
+
+import (
+	"sync"
+	"sync/atomic"
+	"unsafe"
+
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+// Cacher provides interface to implements a caching functionality.
+// An implementation must be goroutine-safe.
+type Cacher interface {
+	// Capacity returns cache capacity.
+	Capacity() int
+
+	// SetCapacity sets cache capacity.
+	SetCapacity(capacity int)
+
+	// Promote promotes the 'cache node'.
+	Promote(n *Node)
+
+	// Ban evicts the 'cache node' and prevent subsequent 'promote'.
+	Ban(n *Node)
+
+	// Evict evicts the 'cache node'.
+	Evict(n *Node)
+
+	// EvictNS evicts 'cache node' with the given namespace.
+	EvictNS(ns uint64)
+
+	// EvictAll evicts all 'cache node'.
+	EvictAll()
+
+	// Close closes the 'cache tree'
+	Close() error
+}
+
+// Value is a 'cacheable object'. It may implements util.Releaser, if
+// so the the Release method will be called once object is released.
+type Value interface{}
+
+type CacheGetter struct {
+	Cache *Cache
+	NS    uint64
+}
+
+func (g *CacheGetter) Get(key uint64, setFunc func() (size int, value Value)) *Handle {
+	return g.Cache.Get(g.NS, key, setFunc)
+}
+
+// The hash tables implementation is based on:
+// "Dynamic-Sized Nonblocking Hash Tables", by Yujie Liu, Kunlong Zhang, and Michael Spear. ACM Symposium on Principles of Distributed Computing, Jul 2014.
+
+const (
+	mInitialSize           = 1 << 4
+	mOverflowThreshold     = 1 << 5
+	mOverflowGrowThreshold = 1 << 7
+)
+
+type mBucket struct {
+	mu     sync.Mutex
+	node   []*Node
+	frozen bool
+}
+
+func (b *mBucket) freeze() []*Node {
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	if !b.frozen {
+		b.frozen = true
+	}
+	return b.node
+}
+
+func (b *mBucket) get(r *Cache, h *mNode, hash uint32, ns, key uint64, noset bool) (done, added bool, n *Node) {
+	b.mu.Lock()
+
+	if b.frozen {
+		b.mu.Unlock()
+		return
+	}
+
+	// Scan the node.
+	for _, n := range b.node {
+		if n.hash == hash && n.ns == ns && n.key == key {
+			atomic.AddInt32(&n.ref, 1)
+			b.mu.Unlock()
+			return true, false, n
+		}
+	}
+
+	// Get only.
+	if noset {
+		b.mu.Unlock()
+		return true, false, nil
+	}
+
+	// Create node.
+	n = &Node{
+		r:    r,
+		hash: hash,
+		ns:   ns,
+		key:  key,
+		ref:  1,
+	}
+	// Add node to bucket.
+	b.node = append(b.node, n)
+	bLen := len(b.node)
+	b.mu.Unlock()
+
+	// Update counter.
+	grow := atomic.AddInt32(&r.nodes, 1) >= h.growThreshold
+	if bLen > mOverflowThreshold {
+		grow = grow || atomic.AddInt32(&h.overflow, 1) >= mOverflowGrowThreshold
+	}
+
+	// Grow.
+	if grow && atomic.CompareAndSwapInt32(&h.resizeInProgess, 0, 1) {
+		nhLen := len(h.buckets) << 1
+		nh := &mNode{
+			buckets:         make([]unsafe.Pointer, nhLen),
+			mask:            uint32(nhLen) - 1,
+			pred:            unsafe.Pointer(h),
+			growThreshold:   int32(nhLen * mOverflowThreshold),
+			shrinkThreshold: int32(nhLen >> 1),
+		}
+		ok := atomic.CompareAndSwapPointer(&r.mHead, unsafe.Pointer(h), unsafe.Pointer(nh))
+		if !ok {
+			panic("BUG: failed swapping head")
+		}
+		go nh.initBuckets()
+	}
+
+	return true, true, n
+}
+
+func (b *mBucket) delete(r *Cache, h *mNode, hash uint32, ns, key uint64) (done, deleted bool) {
+	b.mu.Lock()
+
+	if b.frozen {
+		b.mu.Unlock()
+		return
+	}
+
+	// Scan the node.
+	var (
+		n    *Node
+		bLen int
+	)
+	for i := range b.node {
+		n = b.node[i]
+		if n.ns == ns && n.key == key {
+			if atomic.LoadInt32(&n.ref) == 0 {
+				deleted = true
+
+				// Call releaser.
+				if n.value != nil {
+					if r, ok := n.value.(util.Releaser); ok {
+						r.Release()
+					}
+					n.value = nil
+				}
+
+				// Remove node from bucket.
+				b.node = append(b.node[:i], b.node[i+1:]...)
+				bLen = len(b.node)
+			}
+			break
+		}
+	}
+	b.mu.Unlock()
+
+	if deleted {
+		// Call OnDel.
+		for _, f := range n.onDel {
+			f()
+		}
+
+		// Update counter.
+		atomic.AddInt32(&r.size, int32(n.size)*-1)
+		shrink := atomic.AddInt32(&r.nodes, -1) < h.shrinkThreshold
+		if bLen >= mOverflowThreshold {
+			atomic.AddInt32(&h.overflow, -1)
+		}
+
+		// Shrink.
+		if shrink && len(h.buckets) > mInitialSize && atomic.CompareAndSwapInt32(&h.resizeInProgess, 0, 1) {
+			nhLen := len(h.buckets) >> 1
+			nh := &mNode{
+				buckets:         make([]unsafe.Pointer, nhLen),
+				mask:            uint32(nhLen) - 1,
+				pred:            unsafe.Pointer(h),
+				growThreshold:   int32(nhLen * mOverflowThreshold),
+				shrinkThreshold: int32(nhLen >> 1),
+			}
+			ok := atomic.CompareAndSwapPointer(&r.mHead, unsafe.Pointer(h), unsafe.Pointer(nh))
+			if !ok {
+				panic("BUG: failed swapping head")
+			}
+			go nh.initBuckets()
+		}
+	}
+
+	return true, deleted
+}
+
+type mNode struct {
+	buckets         []unsafe.Pointer // []*mBucket
+	mask            uint32
+	pred            unsafe.Pointer // *mNode
+	resizeInProgess int32
+
+	overflow        int32
+	growThreshold   int32
+	shrinkThreshold int32
+}
+
+func (n *mNode) initBucket(i uint32) *mBucket {
+	if b := (*mBucket)(atomic.LoadPointer(&n.buckets[i])); b != nil {
+		return b
+	}
+
+	p := (*mNode)(atomic.LoadPointer(&n.pred))
+	if p != nil {
+		var node []*Node
+		if n.mask > p.mask {
+			// Grow.
+			pb := (*mBucket)(atomic.LoadPointer(&p.buckets[i&p.mask]))
+			if pb == nil {
+				pb = p.initBucket(i & p.mask)
+			}
+			m := pb.freeze()
+			// Split nodes.
+			for _, x := range m {
+				if x.hash&n.mask == i {
+					node = append(node, x)
+				}
+			}
+		} else {
+			// Shrink.
+			pb0 := (*mBucket)(atomic.LoadPointer(&p.buckets[i]))
+			if pb0 == nil {
+				pb0 = p.initBucket(i)
+			}
+			pb1 := (*mBucket)(atomic.LoadPointer(&p.buckets[i+uint32(len(n.buckets))]))
+			if pb1 == nil {
+				pb1 = p.initBucket(i + uint32(len(n.buckets)))
+			}
+			m0 := pb0.freeze()
+			m1 := pb1.freeze()
+			// Merge nodes.
+			node = make([]*Node, 0, len(m0)+len(m1))
+			node = append(node, m0...)
+			node = append(node, m1...)
+		}
+		b := &mBucket{node: node}
+		if atomic.CompareAndSwapPointer(&n.buckets[i], nil, unsafe.Pointer(b)) {
+			if len(node) > mOverflowThreshold {
+				atomic.AddInt32(&n.overflow, int32(len(node)-mOverflowThreshold))
+			}
+			return b
+		}
+	}
+
+	return (*mBucket)(atomic.LoadPointer(&n.buckets[i]))
+}
+
+func (n *mNode) initBuckets() {
+	for i := range n.buckets {
+		n.initBucket(uint32(i))
+	}
+	atomic.StorePointer(&n.pred, nil)
+}
+
+// Cache is a 'cache map'.
+type Cache struct {
+	mu     sync.RWMutex
+	mHead  unsafe.Pointer // *mNode
+	nodes  int32
+	size   int32
+	cacher Cacher
+	closed bool
+}
+
+// NewCache creates a new 'cache map'. The cacher is optional and
+// may be nil.
+func NewCache(cacher Cacher) *Cache {
+	h := &mNode{
+		buckets:         make([]unsafe.Pointer, mInitialSize),
+		mask:            mInitialSize - 1,
+		growThreshold:   int32(mInitialSize * mOverflowThreshold),
+		shrinkThreshold: 0,
+	}
+	for i := range h.buckets {
+		h.buckets[i] = unsafe.Pointer(&mBucket{})
+	}
+	r := &Cache{
+		mHead:  unsafe.Pointer(h),
+		cacher: cacher,
+	}
+	return r
+}
+
+func (r *Cache) getBucket(hash uint32) (*mNode, *mBucket) {
+	h := (*mNode)(atomic.LoadPointer(&r.mHead))
+	i := hash & h.mask
+	b := (*mBucket)(atomic.LoadPointer(&h.buckets[i]))
+	if b == nil {
+		b = h.initBucket(i)
+	}
+	return h, b
+}
+
+func (r *Cache) delete(n *Node) bool {
+	for {
+		h, b := r.getBucket(n.hash)
+		done, deleted := b.delete(r, h, n.hash, n.ns, n.key)
+		if done {
+			return deleted
+		}
+	}
+	return false
+}
+
+// Nodes returns number of 'cache node' in the map.
+func (r *Cache) Nodes() int {
+	return int(atomic.LoadInt32(&r.nodes))
+}
+
+// Size returns sums of 'cache node' size in the map.
+func (r *Cache) Size() int {
+	return int(atomic.LoadInt32(&r.size))
+}
+
+// Capacity returns cache capacity.
+func (r *Cache) Capacity() int {
+	if r.cacher == nil {
+		return 0
+	}
+	return r.cacher.Capacity()
+}
+
+// SetCapacity sets cache capacity.
+func (r *Cache) SetCapacity(capacity int) {
+	if r.cacher != nil {
+		r.cacher.SetCapacity(capacity)
+	}
+}
+
+// Get gets 'cache node' with the given namespace and key.
+// If cache node is not found and setFunc is not nil, Get will atomically creates
+// the 'cache node' by calling setFunc. Otherwise Get will returns nil.
+//
+// The returned 'cache handle' should be released after use by calling Release
+// method.
+func (r *Cache) Get(ns, key uint64, setFunc func() (size int, value Value)) *Handle {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return nil
+	}
+
+	hash := murmur32(ns, key, 0xf00)
+	for {
+		h, b := r.getBucket(hash)
+		done, _, n := b.get(r, h, hash, ns, key, setFunc == nil)
+		if done {
+			if n != nil {
+				n.mu.Lock()
+				if n.value == nil {
+					if setFunc == nil {
+						n.mu.Unlock()
+						n.unref()
+						return nil
+					}
+
+					n.size, n.value = setFunc()
+					if n.value == nil {
+						n.size = 0
+						n.mu.Unlock()
+						n.unref()
+						return nil
+					}
+					atomic.AddInt32(&r.size, int32(n.size))
+				}
+				n.mu.Unlock()
+				if r.cacher != nil {
+					r.cacher.Promote(n)
+				}
+				return &Handle{unsafe.Pointer(n)}
+			}
+
+			break
+		}
+	}
+	return nil
+}
+
+// Delete removes and ban 'cache node' with the given namespace and key.
+// A banned 'cache node' will never inserted into the 'cache tree'. Ban
+// only attributed to the particular 'cache node', so when a 'cache node'
+// is recreated it will not be banned.
+//
+// If onDel is not nil, then it will be executed if such 'cache node'
+// doesn't exist or once the 'cache node' is released.
+//
+// Delete return true is such 'cache node' exist.
+func (r *Cache) Delete(ns, key uint64, onDel func()) bool {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return false
+	}
+
+	hash := murmur32(ns, key, 0xf00)
+	for {
+		h, b := r.getBucket(hash)
+		done, _, n := b.get(r, h, hash, ns, key, true)
+		if done {
+			if n != nil {
+				if onDel != nil {
+					n.mu.Lock()
+					n.onDel = append(n.onDel, onDel)
+					n.mu.Unlock()
+				}
+				if r.cacher != nil {
+					r.cacher.Ban(n)
+				}
+				n.unref()
+				return true
+			}
+
+			break
+		}
+	}
+
+	if onDel != nil {
+		onDel()
+	}
+
+	return false
+}
+
+// Evict evicts 'cache node' with the given namespace and key. This will
+// simply call Cacher.Evict.
+//
+// Evict return true is such 'cache node' exist.
+func (r *Cache) Evict(ns, key uint64) bool {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return false
+	}
+
+	hash := murmur32(ns, key, 0xf00)
+	for {
+		h, b := r.getBucket(hash)
+		done, _, n := b.get(r, h, hash, ns, key, true)
+		if done {
+			if n != nil {
+				if r.cacher != nil {
+					r.cacher.Evict(n)
+				}
+				n.unref()
+				return true
+			}
+
+			break
+		}
+	}
+
+	return false
+}
+
+// EvictNS evicts 'cache node' with the given namespace. This will
+// simply call Cacher.EvictNS.
+func (r *Cache) EvictNS(ns uint64) {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return
+	}
+
+	if r.cacher != nil {
+		r.cacher.EvictNS(ns)
+	}
+}
+
+// EvictAll evicts all 'cache node'. This will simply call Cacher.EvictAll.
+func (r *Cache) EvictAll() {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	if r.closed {
+		return
+	}
+
+	if r.cacher != nil {
+		r.cacher.EvictAll()
+	}
+}
+
+// Close closes the 'cache map' and releases all 'cache node'.
+func (r *Cache) Close() error {
+	r.mu.Lock()
+	if !r.closed {
+		r.closed = true
+
+		if r.cacher != nil {
+			if err := r.cacher.Close(); err != nil {
+				return err
+			}
+		}
+
+		h := (*mNode)(r.mHead)
+		h.initBuckets()
+
+		for i := range h.buckets {
+			b := (*mBucket)(h.buckets[i])
+			for _, n := range b.node {
+				// Call releaser.
+				if n.value != nil {
+					if r, ok := n.value.(util.Releaser); ok {
+						r.Release()
+					}
+					n.value = nil
+				}
+
+				// Call OnDel.
+				for _, f := range n.onDel {
+					f()
+				}
+			}
+		}
+	}
+	r.mu.Unlock()
+	return nil
+}
+
+// Node is a 'cache node'.
+type Node struct {
+	r *Cache
+
+	hash    uint32
+	ns, key uint64
+
+	mu    sync.Mutex
+	size  int
+	value Value
+
+	ref   int32
+	onDel []func()
+
+	CacheData unsafe.Pointer
+}
+
+// NS returns this 'cache node' namespace.
+func (n *Node) NS() uint64 {
+	return n.ns
+}
+
+// Key returns this 'cache node' key.
+func (n *Node) Key() uint64 {
+	return n.key
+}
+
+// Size returns this 'cache node' size.
+func (n *Node) Size() int {
+	return n.size
+}
+
+// Value returns this 'cache node' value.
+func (n *Node) Value() Value {
+	return n.value
+}
+
+// Ref returns this 'cache node' ref counter.
+func (n *Node) Ref() int32 {
+	return atomic.LoadInt32(&n.ref)
+}
+
+// GetHandle returns an handle for this 'cache node'.
+func (n *Node) GetHandle() *Handle {
+	if atomic.AddInt32(&n.ref, 1) <= 1 {
+		panic("BUG: Node.GetHandle on zero ref")
+	}
+	return &Handle{unsafe.Pointer(n)}
+}
+
+func (n *Node) unref() {
+	if atomic.AddInt32(&n.ref, -1) == 0 {
+		n.r.delete(n)
+	}
+}
+
+func (n *Node) unrefLocked() {
+	if atomic.AddInt32(&n.ref, -1) == 0 {
+		n.r.mu.RLock()
+		if !n.r.closed {
+			n.r.delete(n)
+		}
+		n.r.mu.RUnlock()
+	}
+}
+
+type Handle struct {
+	n unsafe.Pointer // *Node
+}
+
+func (h *Handle) Value() Value {
+	n := (*Node)(atomic.LoadPointer(&h.n))
+	if n != nil {
+		return n.value
+	}
+	return nil
+}
+
+func (h *Handle) Release() {
+	nPtr := atomic.LoadPointer(&h.n)
+	if nPtr != nil && atomic.CompareAndSwapPointer(&h.n, nPtr, nil) {
+		n := (*Node)(nPtr)
+		n.unrefLocked()
+	}
+}
+
+func murmur32(ns, key uint64, seed uint32) uint32 {
+	const (
+		m = uint32(0x5bd1e995)
+		r = 24
+	)
+
+	k1 := uint32(ns >> 32)
+	k2 := uint32(ns)
+	k3 := uint32(key >> 32)
+	k4 := uint32(key)
+
+	k1 *= m
+	k1 ^= k1 >> r
+	k1 *= m
+
+	k2 *= m
+	k2 ^= k2 >> r
+	k2 *= m
+
+	k3 *= m
+	k3 ^= k3 >> r
+	k3 *= m
+
+	k4 *= m
+	k4 ^= k4 >> r
+	k4 *= m
+
+	h := seed
+
+	h *= m
+	h ^= k1
+	h *= m
+	h ^= k2
+	h *= m
+	h ^= k3
+	h *= m
+	h ^= k4
+
+	h ^= h >> 13
+	h *= m
+	h ^= h >> 15
+
+	return h
+}
@@ -0,0 +1,554 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package cache
+
+import (
+	"math/rand"
+	"runtime"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+	"unsafe"
+)
+
+type int32o int32
+
+func (o *int32o) acquire() {
+	if atomic.AddInt32((*int32)(o), 1) != 1 {
+		panic("BUG: invalid ref")
+	}
+}
+
+func (o *int32o) Release() {
+	if atomic.AddInt32((*int32)(o), -1) != 0 {
+		panic("BUG: invalid ref")
+	}
+}
+
+type releaserFunc struct {
+	fn    func()
+	value Value
+}
+
+func (r releaserFunc) Release() {
+	if r.fn != nil {
+		r.fn()
+	}
+}
+
+func set(c *Cache, ns, key uint64, value Value, charge int, relf func()) *Handle {
+	return c.Get(ns, key, func() (int, Value) {
+		if relf != nil {
+			return charge, releaserFunc{relf, value}
+		} else {
+			return charge, value
+		}
+	})
+}
+
+func TestCacheMap(t *testing.T) {
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	nsx := []struct {
+		nobjects, nhandles, concurrent, repeat int
+	}{
+		{10000, 400, 50, 3},
+		{100000, 1000, 100, 10},
+	}
+
+	var (
+		objects [][]int32o
+		handles [][]unsafe.Pointer
+	)
+
+	for _, x := range nsx {
+		objects = append(objects, make([]int32o, x.nobjects))
+		handles = append(handles, make([]unsafe.Pointer, x.nhandles))
+	}
+
+	c := NewCache(nil)
+
+	wg := new(sync.WaitGroup)
+	var done int32
+
+	for ns, x := range nsx {
+		for i := 0; i < x.concurrent; i++ {
+			wg.Add(1)
+			go func(ns, i, repeat int, objects []int32o, handles []unsafe.Pointer) {
+				defer wg.Done()
+				r := rand.New(rand.NewSource(time.Now().UnixNano()))
+
+				for j := len(objects) * repeat; j >= 0; j-- {
+					key := uint64(r.Intn(len(objects)))
+					h := c.Get(uint64(ns), key, func() (int, Value) {
+						o := &objects[key]
+						o.acquire()
+						return 1, o
+					})
+					if v := h.Value().(*int32o); v != &objects[key] {
+						t.Fatalf("#%d invalid value: want=%p got=%p", ns, &objects[key], v)
+					}
+					if objects[key] != 1 {
+						t.Fatalf("#%d invalid object %d: %d", ns, key, objects[key])
+					}
+					if !atomic.CompareAndSwapPointer(&handles[r.Intn(len(handles))], nil, unsafe.Pointer(h)) {
+						h.Release()
+					}
+				}
+			}(ns, i, x.repeat, objects[ns], handles[ns])
+		}
+
+		go func(handles []unsafe.Pointer) {
+			r := rand.New(rand.NewSource(time.Now().UnixNano()))
+
+			for atomic.LoadInt32(&done) == 0 {
+				i := r.Intn(len(handles))
+				h := (*Handle)(atomic.LoadPointer(&handles[i]))
+				if h != nil && atomic.CompareAndSwapPointer(&handles[i], unsafe.Pointer(h), nil) {
+					h.Release()
+				}
+				time.Sleep(time.Millisecond)
+			}
+		}(handles[ns])
+	}
+
+	go func() {
+		handles := make([]*Handle, 100000)
+		for atomic.LoadInt32(&done) == 0 {
+			for i := range handles {
+				handles[i] = c.Get(999999999, uint64(i), func() (int, Value) {
+					return 1, 1
+				})
+			}
+			for _, h := range handles {
+				h.Release()
+			}
+		}
+	}()
+
+	wg.Wait()
+
+	atomic.StoreInt32(&done, 1)
+
+	for _, handles0 := range handles {
+		for i := range handles0 {
+			h := (*Handle)(atomic.LoadPointer(&handles0[i]))
+			if h != nil && atomic.CompareAndSwapPointer(&handles0[i], unsafe.Pointer(h), nil) {
+				h.Release()
+			}
+		}
+	}
+
+	for ns, objects0 := range objects {
+		for i, o := range objects0 {
+			if o != 0 {
+				t.Fatalf("invalid object #%d.%d: ref=%d", ns, i, o)
+			}
+		}
+	}
+}
+
+func TestCacheMap_NodesAndSize(t *testing.T) {
+	c := NewCache(nil)
+	if c.Nodes() != 0 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 0, c.Nodes())
+	}
+	if c.Size() != 0 {
+		t.Errorf("invalid size counter: want=%d got=%d", 0, c.Size())
+	}
+	set(c, 0, 1, 1, 1, nil)
+	set(c, 0, 2, 2, 2, nil)
+	set(c, 1, 1, 3, 3, nil)
+	set(c, 2, 1, 4, 1, nil)
+	if c.Nodes() != 4 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 4, c.Nodes())
+	}
+	if c.Size() != 7 {
+		t.Errorf("invalid size counter: want=%d got=%d", 4, c.Size())
+	}
+}
+
+func TestLRUCache_Capacity(t *testing.T) {
+	c := NewCache(NewLRU(10))
+	if c.Capacity() != 10 {
+		t.Errorf("invalid capacity: want=%d got=%d", 10, c.Capacity())
+	}
+	set(c, 0, 1, 1, 1, nil).Release()
+	set(c, 0, 2, 2, 2, nil).Release()
+	set(c, 1, 1, 3, 3, nil).Release()
+	set(c, 2, 1, 4, 1, nil).Release()
+	set(c, 2, 2, 5, 1, nil).Release()
+	set(c, 2, 3, 6, 1, nil).Release()
+	set(c, 2, 4, 7, 1, nil).Release()
+	set(c, 2, 5, 8, 1, nil).Release()
+	if c.Nodes() != 7 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 7, c.Nodes())
+	}
+	if c.Size() != 10 {
+		t.Errorf("invalid size counter: want=%d got=%d", 10, c.Size())
+	}
+	c.SetCapacity(9)
+	if c.Capacity() != 9 {
+		t.Errorf("invalid capacity: want=%d got=%d", 9, c.Capacity())
+	}
+	if c.Nodes() != 6 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 6, c.Nodes())
+	}
+	if c.Size() != 8 {
+		t.Errorf("invalid size counter: want=%d got=%d", 8, c.Size())
+	}
+}
+
+func TestCacheMap_NilValue(t *testing.T) {
+	c := NewCache(NewLRU(10))
+	h := c.Get(0, 0, func() (size int, value Value) {
+		return 1, nil
+	})
+	if h != nil {
+		t.Error("cache handle is non-nil")
+	}
+	if c.Nodes() != 0 {
+		t.Errorf("invalid nodes counter: want=%d got=%d", 0, c.Nodes())
+	}
+	if c.Size() != 0 {
+		t.Errorf("invalid size counter: want=%d got=%d", 0, c.Size())
+	}
+}
+
+func TestLRUCache_GetLatency(t *testing.T) {
+	runtime.GOMAXPROCS(runtime.NumCPU())
+
+	const (
+		concurrentSet = 30
+		concurrentGet = 3
+		duration      = 3 * time.Second
+		delay         = 3 * time.Millisecond
+		maxkey        = 100000
+	)
+
+	var (
+		set, getHit, getAll        int32
+		getMaxLatency, getDuration int64
+	)
+
+	c := NewCache(NewLRU(5000))
+	wg := &sync.WaitGroup{}
+	until := time.Now().Add(duration)
+	for i := 0; i < concurrentSet; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			r := rand.New(rand.NewSource(time.Now().UnixNano()))
+			for time.Now().Before(until) {
+				c.Get(0, uint64(r.Intn(maxkey)), func() (int, Value) {
+					time.Sleep(delay)
+					atomic.AddInt32(&set, 1)
+					return 1, 1
+				}).Release()
+			}
+		}(i)
+	}
+	for i := 0; i < concurrentGet; i++ {
+		wg.Add(1)
+		go func(i int) {
+			defer wg.Done()
+			r := rand.New(rand.NewSource(time.Now().UnixNano()))
+			for {
+				mark := time.Now()
+				if mark.Before(until) {
+					h := c.Get(0, uint64(r.Intn(maxkey)), nil)
+					latency := int64(time.Now().Sub(mark))
+					m := atomic.LoadInt64(&getMaxLatency)
+					if latency > m {
+						atomic.CompareAndSwapInt64(&getMaxLatency, m, latency)
+					}
+					atomic.AddInt64(&getDuration, latency)
+					if h != nil {
+						atomic.AddInt32(&getHit, 1)
+						h.Release()
+					}
+					atomic.AddInt32(&getAll, 1)
+				} else {
+					break
+				}
+			}
+		}(i)
+	}
+
+	wg.Wait()
+	getAvglatency := time.Duration(getDuration) / time.Duration(getAll)
+	t.Logf("set=%d getHit=%d getAll=%d getMaxLatency=%v getAvgLatency=%v",
+		set, getHit, getAll, time.Duration(getMaxLatency), getAvglatency)
+
+	if getAvglatency > delay/3 {
+		t.Errorf("get avg latency > %v: got=%v", delay/3, getAvglatency)
+	}
+}
+
+func TestLRUCache_HitMiss(t *testing.T) {
+	cases := []struct {
+		key   uint64
+		value string
+	}{
+		{1, "vvvvvvvvv"},
+		{100, "v1"},
+		{0, "v2"},
+		{12346, "v3"},
+		{777, "v4"},
+		{999, "v5"},
+		{7654, "v6"},
+		{2, "v7"},
+		{3, "v8"},
+		{9, "v9"},
+	}
+
+	setfin := 0
+	c := NewCache(NewLRU(1000))
+	for i, x := range cases {
+		set(c, 0, x.key, x.value, len(x.value), func() {
+			setfin++
+		}).Release()
+		for j, y := range cases {
+			h := c.Get(0, y.key, nil)
+			if j <= i {
+				// should hit
+				if h == nil {
+					t.Errorf("case '%d' iteration '%d' is miss", i, j)
+				} else {
+					if x := h.Value().(releaserFunc).value.(string); x != y.value {
+						t.Errorf("case '%d' iteration '%d' has invalid value got '%s', want '%s'", i, j, x, y.value)
+					}
+				}
+			} else {
+				// should miss
+				if h != nil {
+					t.Errorf("case '%d' iteration '%d' is hit , value '%s'", i, j, h.Value().(releaserFunc).value.(string))
+				}
+			}
+			if h != nil {
+				h.Release()
+			}
+		}
+	}
+
+	for i, x := range cases {
+		finalizerOk := false
+		c.Delete(0, x.key, func() {
+			finalizerOk = true
+		})
+
+		if !finalizerOk {
+			t.Errorf("case %d delete finalizer not executed", i)
+		}
+
+		for j, y := range cases {
+			h := c.Get(0, y.key, nil)
+			if j > i {
+				// should hit
+				if h == nil {
+					t.Errorf("case '%d' iteration '%d' is miss", i, j)
+				} else {
+					if x := h.Value().(releaserFunc).value.(string); x != y.value {
+						t.Errorf("case '%d' iteration '%d' has invalid value got '%s', want '%s'", i, j, x, y.value)
+					}
+				}
+			} else {
+				// should miss
+				if h != nil {
+					t.Errorf("case '%d' iteration '%d' is hit, value '%s'", i, j, h.Value().(releaserFunc).value.(string))
+				}
+			}
+			if h != nil {
+				h.Release()
+			}
+		}
+	}
+
+	if setfin != len(cases) {
+		t.Errorf("some set finalizer may not be executed, want=%d got=%d", len(cases), setfin)
+	}
+}
+
+func TestLRUCache_Eviction(t *testing.T) {
+	c := NewCache(NewLRU(12))
+	o1 := set(c, 0, 1, 1, 1, nil)
+	set(c, 0, 2, 2, 1, nil).Release()
+	set(c, 0, 3, 3, 1, nil).Release()
+	set(c, 0, 4, 4, 1, nil).Release()
+	set(c, 0, 5, 5, 1, nil).Release()
+	if h := c.Get(0, 2, nil); h != nil { // 1,3,4,5,2
+		h.Release()
+	}
+	set(c, 0, 9, 9, 10, nil).Release() // 5,2,9
+
+	for _, key := range []uint64{9, 2, 5, 1} {
+		h := c.Get(0, key, nil)
+		if h == nil {
+			t.Errorf("miss for key '%d'", key)
+		} else {
+			if x := h.Value().(int); x != int(key) {
+				t.Errorf("invalid value for key '%d' want '%d', got '%d'", key, key, x)
+			}
+			h.Release()
+		}
+	}
+	o1.Release()
+	for _, key := range []uint64{1, 2, 5} {
+		h := c.Get(0, key, nil)
+		if h == nil {
+			t.Errorf("miss for key '%d'", key)
+		} else {
+			if x := h.Value().(int); x != int(key) {
+				t.Errorf("invalid value for key '%d' want '%d', got '%d'", key, key, x)
+			}
+			h.Release()
+		}
+	}
+	for _, key := range []uint64{3, 4, 9} {
+		h := c.Get(0, key, nil)
+		if h != nil {
+			t.Errorf("hit for key '%d'", key)
+			if x := h.Value().(int); x != int(key) {
+				t.Errorf("invalid value for key '%d' want '%d', got '%d'", key, key, x)
+			}
+			h.Release()
+		}
+	}
+}
+
+func TestLRUCache_Evict(t *testing.T) {
+	c := NewCache(NewLRU(6))
+	set(c, 0, 1, 1, 1, nil).Release()
+	set(c, 0, 2, 2, 1, nil).Release()
+	set(c, 1, 1, 4, 1, nil).Release()
+	set(c, 1, 2, 5, 1, nil).Release()
+	set(c, 2, 1, 6, 1, nil).Release()
+	set(c, 2, 2, 7, 1, nil).Release()
+
+	for ns := 0; ns < 3; ns++ {
+		for key := 1; key < 3; key++ {
+			if h := c.Get(uint64(ns), uint64(key), nil); h != nil {
+				h.Release()
+			} else {
+				t.Errorf("Cache.Get on #%d.%d return nil", ns, key)
+			}
+		}
+	}
+
+	if ok := c.Evict(0, 1); !ok {
+		t.Error("first Cache.Evict on #0.1 return false")
+	}
+	if ok := c.Evict(0, 1); ok {
+		t.Error("second Cache.Evict on #0.1 return true")
+	}
+	if h := c.Get(0, 1, nil); h != nil {
+		t.Errorf("Cache.Get on #0.1 return non-nil: %v", h.Value())
+	}
+
+	c.EvictNS(1)
+	if h := c.Get(1, 1, nil); h != nil {
+		t.Errorf("Cache.Get on #1.1 return non-nil: %v", h.Value())
+	}
+	if h := c.Get(1, 2, nil); h != nil {
+		t.Errorf("Cache.Get on #1.2 return non-nil: %v", h.Value())
+	}
+
+	c.EvictAll()
+	for ns := 0; ns < 3; ns++ {
+		for key := 1; key < 3; key++ {
+			if h := c.Get(uint64(ns), uint64(key), nil); h != nil {
+				t.Errorf("Cache.Get on #%d.%d return non-nil: %v", ns, key, h.Value())
+			}
+		}
+	}
+}
+
+func TestLRUCache_Delete(t *testing.T) {
+	delFuncCalled := 0
+	delFunc := func() {
+		delFuncCalled++
+	}
+
+	c := NewCache(NewLRU(2))
+	set(c, 0, 1, 1, 1, nil).Release()
+	set(c, 0, 2, 2, 1, nil).Release()
+
+	if ok := c.Delete(0, 1, delFunc); !ok {
+		t.Error("Cache.Delete on #1 return false")
+	}
+	if h := c.Get(0, 1, nil); h != nil {
+		t.Errorf("Cache.Get on #1 return non-nil: %v", h.Value())
+	}
+	if ok := c.Delete(0, 1, delFunc); ok {
+		t.Error("Cache.Delete on #1 return true")
+	}
+
+	h2 := c.Get(0, 2, nil)
+	if h2 == nil {
+		t.Error("Cache.Get on #2 return nil")
+	}
+	if ok := c.Delete(0, 2, delFunc); !ok {
+		t.Error("(1) Cache.Delete on #2 return false")
+	}
+	if ok := c.Delete(0, 2, delFunc); !ok {
+		t.Error("(2) Cache.Delete on #2 return false")
+	}
+
+	set(c, 0, 3, 3, 1, nil).Release()
+	set(c, 0, 4, 4, 1, nil).Release()
+	c.Get(0, 2, nil).Release()
+
+	for key := 2; key <= 4; key++ {
+		if h := c.Get(0, uint64(key), nil); h != nil {
+			h.Release()
+		} else {
+			t.Errorf("Cache.Get on #%d return nil", key)
+		}
+	}
+
+	h2.Release()
+	if h := c.Get(0, 2, nil); h != nil {
+		t.Errorf("Cache.Get on #2 return non-nil: %v", h.Value())
+	}
+
+	if delFuncCalled != 4 {
+		t.Errorf("delFunc isn't called 4 times: got=%d", delFuncCalled)
+	}
+}
+
+func TestLRUCache_Close(t *testing.T) {
+	relFuncCalled := 0
+	relFunc := func() {
+		relFuncCalled++
+	}
+	delFuncCalled := 0
+	delFunc := func() {
+		delFuncCalled++
+	}
+
+	c := NewCache(NewLRU(2))
+	set(c, 0, 1, 1, 1, relFunc).Release()
+	set(c, 0, 2, 2, 1, relFunc).Release()
+
+	h3 := set(c, 0, 3, 3, 1, relFunc)
+	if h3 == nil {
+		t.Error("Cache.Get on #3 return nil")
+	}
+	if ok := c.Delete(0, 3, delFunc); !ok {
+		t.Error("Cache.Delete on #3 return false")
+	}
+
+	c.Close()
+
+	if relFuncCalled != 3 {
+		t.Errorf("relFunc isn't called 3 times: got=%d", relFuncCalled)
+	}
+	if delFuncCalled != 1 {
+		t.Errorf("delFunc isn't called 1 times: got=%d", delFuncCalled)
+	}
+}
@@ -0,0 +1,195 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package cache
+
+import (
+	"sync"
+	"unsafe"
+)
+
+type lruNode struct {
+	n   *Node
+	h   *Handle
+	ban bool
+
+	next, prev *lruNode
+}
+
+func (n *lruNode) insert(at *lruNode) {
+	x := at.next
+	at.next = n
+	n.prev = at
+	n.next = x
+	x.prev = n
+}
+
+func (n *lruNode) remove() {
+	if n.prev != nil {
+		n.prev.next = n.next
+		n.next.prev = n.prev
+		n.prev = nil
+		n.next = nil
+	} else {
+		panic("BUG: removing removed node")
+	}
+}
+
+type lru struct {
+	mu       sync.Mutex
+	capacity int
+	used     int
+	recent   lruNode
+}
+
+func (r *lru) reset() {
+	r.recent.next = &r.recent
+	r.recent.prev = &r.recent
+	r.used = 0
+}
+
+func (r *lru) Capacity() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.capacity
+}
+
+func (r *lru) SetCapacity(capacity int) {
+	var evicted []*lruNode
+
+	r.mu.Lock()
+	r.capacity = capacity
+	for r.used > r.capacity {
+		rn := r.recent.prev
+		if rn == nil {
+			panic("BUG: invalid LRU used or capacity counter")
+		}
+		rn.remove()
+		rn.n.CacheData = nil
+		r.used -= rn.n.Size()
+		evicted = append(evicted, rn)
+	}
+	r.mu.Unlock()
+
+	for _, rn := range evicted {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) Promote(n *Node) {
+	var evicted []*lruNode
+
+	r.mu.Lock()
+	if n.CacheData == nil {
+		if n.Size() <= r.capacity {
+			rn := &lruNode{n: n, h: n.GetHandle()}
+			rn.insert(&r.recent)
+			n.CacheData = unsafe.Pointer(rn)
+			r.used += n.Size()
+
+			for r.used > r.capacity {
+				rn := r.recent.prev
+				if rn == nil {
+					panic("BUG: invalid LRU used or capacity counter")
+				}
+				rn.remove()
+				rn.n.CacheData = nil
+				r.used -= rn.n.Size()
+				evicted = append(evicted, rn)
+			}
+		}
+	} else {
+		rn := (*lruNode)(n.CacheData)
+		if !rn.ban {
+			rn.remove()
+			rn.insert(&r.recent)
+		}
+	}
+	r.mu.Unlock()
+
+	for _, rn := range evicted {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) Ban(n *Node) {
+	r.mu.Lock()
+	if n.CacheData == nil {
+		n.CacheData = unsafe.Pointer(&lruNode{n: n, ban: true})
+	} else {
+		rn := (*lruNode)(n.CacheData)
+		if !rn.ban {
+			rn.remove()
+			rn.ban = true
+			r.used -= rn.n.Size()
+			r.mu.Unlock()
+
+			rn.h.Release()
+			rn.h = nil
+			return
+		}
+	}
+	r.mu.Unlock()
+}
+
+func (r *lru) Evict(n *Node) {
+	r.mu.Lock()
+	rn := (*lruNode)(n.CacheData)
+	if rn == nil || rn.ban {
+		r.mu.Unlock()
+		return
+	}
+	n.CacheData = nil
+	r.mu.Unlock()
+
+	rn.h.Release()
+}
+
+func (r *lru) EvictNS(ns uint64) {
+	var evicted []*lruNode
+
+	r.mu.Lock()
+	for e := r.recent.prev; e != &r.recent; {
+		rn := e
+		e = e.prev
+		if rn.n.NS() == ns {
+			rn.remove()
+			rn.n.CacheData = nil
+			r.used -= rn.n.Size()
+			evicted = append(evicted, rn)
+		}
+	}
+	r.mu.Unlock()
+
+	for _, rn := range evicted {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) EvictAll() {
+	r.mu.Lock()
+	back := r.recent.prev
+	for rn := back; rn != &r.recent; rn = rn.prev {
+		rn.n.CacheData = nil
+	}
+	r.reset()
+	r.mu.Unlock()
+
+	for rn := back; rn != &r.recent; rn = rn.prev {
+		rn.h.Release()
+	}
+}
+
+func (r *lru) Close() error {
+	return nil
+}
+
+// NewLRU create a new LRU-cache.
+func NewLRU(capacity int) Cacher {
+	r := &lru{capacity: capacity}
+	r.reset()
+	return r
+}
@@ -0,0 +1,75 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import "github.com/syndtr/goleveldb/leveldb/comparer"
+
+type iComparer struct {
+	ucmp comparer.Comparer
+}
+
+func (icmp *iComparer) uName() string {
+	return icmp.ucmp.Name()
+}
+
+func (icmp *iComparer) uCompare(a, b []byte) int {
+	return icmp.ucmp.Compare(a, b)
+}
+
+func (icmp *iComparer) uSeparator(dst, a, b []byte) []byte {
+	return icmp.ucmp.Separator(dst, a, b)
+}
+
+func (icmp *iComparer) uSuccessor(dst, b []byte) []byte {
+	return icmp.ucmp.Successor(dst, b)
+}
+
+func (icmp *iComparer) Name() string {
+	return icmp.uName()
+}
+
+func (icmp *iComparer) Compare(a, b []byte) int {
+	x := icmp.ucmp.Compare(iKey(a).ukey(), iKey(b).ukey())
+	if x == 0 {
+		if m, n := iKey(a).num(), iKey(b).num(); m > n {
+			x = -1
+		} else if m < n {
+			x = 1
+		}
+	}
+	return x
+}
+
+func (icmp *iComparer) Separator(dst, a, b []byte) []byte {
+	ua, ub := iKey(a).ukey(), iKey(b).ukey()
+	dst = icmp.ucmp.Separator(dst, ua, ub)
+	if dst == nil {
+		return nil
+	}
+	if len(dst) < len(ua) && icmp.uCompare(ua, dst) < 0 {
+		dst = append(dst, kMaxNumBytes...)
+	} else {
+		// Did not close possibilities that n maybe longer than len(ub).
+		dst = append(dst, a[len(a)-8:]...)
+	}
+	return dst
+}
+
+func (icmp *iComparer) Successor(dst, b []byte) []byte {
+	ub := iKey(b).ukey()
+	dst = icmp.ucmp.Successor(dst, ub)
+	if dst == nil {
+		return nil
+	}
+	if len(dst) < len(ub) && icmp.uCompare(ub, dst) < 0 {
+		dst = append(dst, kMaxNumBytes...)
+	} else {
+		// Did not close possibilities that n maybe longer than len(ub).
+		dst = append(dst, b[len(b)-8:]...)
+	}
+	return dst
+}
@@ -0,0 +1,51 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package comparer
+
+import "bytes"
+
+type bytesComparer struct{}
+
+func (bytesComparer) Compare(a, b []byte) int {
+	return bytes.Compare(a, b)
+}
+
+func (bytesComparer) Name() string {
+	return "leveldb.BytewiseComparator"
+}
+
+func (bytesComparer) Separator(dst, a, b []byte) []byte {
+	i, n := 0, len(a)
+	if n > len(b) {
+		n = len(b)
+	}
+	for ; i < n && a[i] == b[i]; i++ {
+	}
+	if i >= n {
+		// Do not shorten if one string is a prefix of the other
+	} else if c := a[i]; c < 0xff && c+1 < b[i] {
+		dst = append(dst, a[:i+1]...)
+		dst[i]++
+		return dst
+	}
+	return nil
+}
+
+func (bytesComparer) Successor(dst, b []byte) []byte {
+	for i, c := range b {
+		if c != 0xff {
+			dst = append(dst, b[:i+1]...)
+			dst[i]++
+			return dst
+		}
+	}
+	return nil
+}
+
+// DefaultComparer are default implementation of the Comparer interface.
+// It uses the natural ordering, consistent with bytes.Compare.
+var DefaultComparer = bytesComparer{}
@@ -0,0 +1,57 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// Package comparer provides interface and implementation for ordering
+// sets of data.
+package comparer
+
+// BasicComparer is the interface that wraps the basic Compare method.
+type BasicComparer interface {
+	// Compare returns -1, 0, or +1 depending on whether a is 'less than',
+	// 'equal to' or 'greater than' b. The two arguments can only be 'equal'
+	// if their contents are exactly equal. Furthermore, the empty slice
+	// must be 'less than' any non-empty slice.
+	Compare(a, b []byte) int
+}
+
+// Comparer defines a total ordering over the space of []byte keys: a 'less
+// than' relationship.
+type Comparer interface {
+	BasicComparer
+
+	// Name returns name of the comparer.
+	//
+	// The Level-DB on-disk format stores the comparer name, and opening a
+	// database with a different comparer from the one it was created with
+	// will result in an error.
+	//
+	// An implementation to a new name whenever the comparer implementation
+	// changes in a way that will cause the relative ordering of any two keys
+	// to change.
+	//
+	// Names starting with "leveldb." are reserved and should not be used
+	// by any users of this package.
+	Name() string
+
+	// Bellow are advanced functions used used to reduce the space requirements
+	// for internal data structures such as index blocks.
+
+	// Separator appends a sequence of bytes x to dst such that a <= x && x < b,
+	// where 'less than' is consistent with Compare. An implementation should
+	// return nil if x equal to a.
+	//
+	// Either contents of a or b should not by any means modified. Doing so
+	// may cause corruption on the internal state.
+	Separator(dst, a, b []byte) []byte
+
+	// Successor appends a sequence of bytes x to dst such that x >= b, where
+	// 'less than' is consistent with Compare. An implementation should return
+	// nil if x equal to b.
+	//
+	// Contents of b should not by any means modified. Doing so may cause
+	// corruption on the internal state.
+	Successor(dst, b []byte) []byte
+}
@@ -0,0 +1,500 @@
+// Copyright (c) 2013, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"bytes"
+	"fmt"
+	"github.com/syndtr/goleveldb/leveldb/filter"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/storage"
+	"io"
+	"math/rand"
+	"testing"
+)
+
+const ctValSize = 1000
+
+type dbCorruptHarness struct {
+	dbHarness
+}
+
+func newDbCorruptHarnessWopt(t *testing.T, o *opt.Options) *dbCorruptHarness {
+	h := new(dbCorruptHarness)
+	h.init(t, o)
+	return h
+}
+
+func newDbCorruptHarness(t *testing.T) *dbCorruptHarness {
+	return newDbCorruptHarnessWopt(t, &opt.Options{
+		BlockCacheCapacity: 100,
+		Strict:             opt.StrictJournalChecksum,
+	})
+}
+
+func (h *dbCorruptHarness) recover() {
+	p := &h.dbHarness
+	t := p.t
+
+	var err error
+	p.db, err = Recover(h.stor, h.o)
+	if err != nil {
+		t.Fatal("Repair: got error: ", err)
+	}
+}
+
+func (h *dbCorruptHarness) build(n int) {
+	p := &h.dbHarness
+	t := p.t
+	db := p.db
+
+	batch := new(Batch)
+	for i := 0; i < n; i++ {
+		batch.Reset()
+		batch.Put(tkey(i), tval(i, ctValSize))
+		err := db.Write(batch, p.wo)
+		if err != nil {
+			t.Fatal("write error: ", err)
+		}
+	}
+}
+
+func (h *dbCorruptHarness) buildShuffled(n int, rnd *rand.Rand) {
+	p := &h.dbHarness
+	t := p.t
+	db := p.db
+
+	batch := new(Batch)
+	for i := range rnd.Perm(n) {
+		batch.Reset()
+		batch.Put(tkey(i), tval(i, ctValSize))
+		err := db.Write(batch, p.wo)
+		if err != nil {
+			t.Fatal("write error: ", err)
+		}
+	}
+}
+
+func (h *dbCorruptHarness) deleteRand(n, max int, rnd *rand.Rand) {
+	p := &h.dbHarness
+	t := p.t
+	db := p.db
+
+	batch := new(Batch)
+	for i := 0; i < n; i++ {
+		batch.Reset()
+		batch.Delete(tkey(rnd.Intn(max)))
+		err := db.Write(batch, p.wo)
+		if err != nil {
+			t.Fatal("write error: ", err)
+		}
+	}
+}
+
+func (h *dbCorruptHarness) corrupt(ft storage.FileType, fi, offset, n int) {
+	p := &h.dbHarness
+	t := p.t
+
+	ff, _ := p.stor.GetFiles(ft)
+	sff := files(ff)
+	sff.sort()
+	if fi < 0 {
+		fi = len(sff) - 1
+	}
+	if fi >= len(sff) {
+		t.Fatalf("no such file with type %q with index %d", ft, fi)
+	}
+
+	file := sff[fi]
+
+	r, err := file.Open()
+	if err != nil {
+		t.Fatal("cannot open file: ", err)
+	}
+	x, err := r.Seek(0, 2)
+	if err != nil {
+		t.Fatal("cannot query file size: ", err)
+	}
+	m := int(x)
+	if _, err := r.Seek(0, 0); err != nil {
+		t.Fatal(err)
+	}
+
+	if offset < 0 {
+		if -offset > m {
+			offset = 0
+		} else {
+			offset = m + offset
+		}
+	}
+	if offset > m {
+		offset = m
+	}
+	if offset+n > m {
+		n = m - offset
+	}
+
+	buf := make([]byte, m)
+	_, err = io.ReadFull(r, buf)
+	if err != nil {
+		t.Fatal("cannot read file: ", err)
+	}
+	r.Close()
+
+	for i := 0; i < n; i++ {
+		buf[offset+i] ^= 0x80
+	}
+
+	err = file.Remove()
+	if err != nil {
+		t.Fatal("cannot remove old file: ", err)
+	}
+	w, err := file.Create()
+	if err != nil {
+		t.Fatal("cannot create new file: ", err)
+	}
+	_, err = w.Write(buf)
+	if err != nil {
+		t.Fatal("cannot write new file: ", err)
+	}
+	w.Close()
+}
+
+func (h *dbCorruptHarness) removeAll(ft storage.FileType) {
+	ff, err := h.stor.GetFiles(ft)
+	if err != nil {
+		h.t.Fatal("get files: ", err)
+	}
+	for _, f := range ff {
+		if err := f.Remove(); err != nil {
+			h.t.Error("remove file: ", err)
+		}
+	}
+}
+
+func (h *dbCorruptHarness) removeOne(ft storage.FileType) {
+	ff, err := h.stor.GetFiles(ft)
+	if err != nil {
+		h.t.Fatal("get files: ", err)
+	}
+	f := ff[rand.Intn(len(ff))]
+	h.t.Logf("removing file @%d", f.Num())
+	if err := f.Remove(); err != nil {
+		h.t.Error("remove file: ", err)
+	}
+}
+
+func (h *dbCorruptHarness) check(min, max int) {
+	p := &h.dbHarness
+	t := p.t
+	db := p.db
+
+	var n, badk, badv, missed, good int
+	iter := db.NewIterator(nil, p.ro)
+	for iter.Next() {
+		k := 0
+		fmt.Sscanf(string(iter.Key()), "%d", &k)
+		if k < n {
+			badk++
+			continue
+		}
+		missed += k - n
+		n = k + 1
+		if !bytes.Equal(iter.Value(), tval(k, ctValSize)) {
+			badv++
+		} else {
+			good++
+		}
+	}
+	err := iter.Error()
+	iter.Release()
+	t.Logf("want=%d..%d got=%d badkeys=%d badvalues=%d missed=%d, err=%v",
+		min, max, good, badk, badv, missed, err)
+	if good < min || good > max {
+		t.Errorf("good entries number not in range")
+	}
+}
+
+func TestCorruptDB_Journal(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.build(100)
+	h.check(100, 100)
+	h.closeDB()
+	h.corrupt(storage.TypeJournal, -1, 19, 1)
+	h.corrupt(storage.TypeJournal, -1, 32*1024+1000, 1)
+
+	h.openDB()
+	h.check(36, 36)
+
+	h.close()
+}
+
+func TestCorruptDB_Table(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.build(100)
+	h.compactMem()
+	h.compactRangeAt(0, "", "")
+	h.compactRangeAt(1, "", "")
+	h.closeDB()
+	h.corrupt(storage.TypeTable, -1, 100, 1)
+
+	h.openDB()
+	h.check(99, 99)
+
+	h.close()
+}
+
+func TestCorruptDB_TableIndex(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.build(10000)
+	h.compactMem()
+	h.closeDB()
+	h.corrupt(storage.TypeTable, -1, -2000, 500)
+
+	h.openDB()
+	h.check(5000, 9999)
+
+	h.close()
+}
+
+func TestCorruptDB_MissingManifest(t *testing.T) {
+	rnd := rand.New(rand.NewSource(0x0badda7a))
+	h := newDbCorruptHarnessWopt(t, &opt.Options{
+		BlockCacheCapacity: 100,
+		Strict:             opt.StrictJournalChecksum,
+		WriteBuffer:        1000 * 60,
+	})
+
+	h.build(1000)
+	h.compactMem()
+	h.buildShuffled(1000, rnd)
+	h.compactMem()
+	h.deleteRand(500, 1000, rnd)
+	h.compactMem()
+	h.buildShuffled(1000, rnd)
+	h.compactMem()
+	h.deleteRand(500, 1000, rnd)
+	h.compactMem()
+	h.buildShuffled(1000, rnd)
+	h.compactMem()
+	h.closeDB()
+
+	h.stor.SetIgnoreOpenErr(storage.TypeManifest)
+	h.removeAll(storage.TypeManifest)
+	h.openAssert(false)
+	h.stor.SetIgnoreOpenErr(0)
+
+	h.recover()
+	h.check(1000, 1000)
+	h.build(1000)
+	h.compactMem()
+	h.compactRange("", "")
+	h.closeDB()
+
+	h.recover()
+	h.check(1000, 1000)
+
+	h.close()
+}
+
+func TestCorruptDB_SequenceNumberRecovery(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.put("foo", "v1")
+	h.put("foo", "v2")
+	h.put("foo", "v3")
+	h.put("foo", "v4")
+	h.put("foo", "v5")
+	h.closeDB()
+
+	h.recover()
+	h.getVal("foo", "v5")
+	h.put("foo", "v6")
+	h.getVal("foo", "v6")
+
+	h.reopenDB()
+	h.getVal("foo", "v6")
+
+	h.close()
+}
+
+func TestCorruptDB_SequenceNumberRecoveryTable(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.put("foo", "v1")
+	h.put("foo", "v2")
+	h.put("foo", "v3")
+	h.compactMem()
+	h.put("foo", "v4")
+	h.put("foo", "v5")
+	h.compactMem()
+	h.closeDB()
+
+	h.recover()
+	h.getVal("foo", "v5")
+	h.put("foo", "v6")
+	h.getVal("foo", "v6")
+
+	h.reopenDB()
+	h.getVal("foo", "v6")
+
+	h.close()
+}
+
+func TestCorruptDB_CorruptedManifest(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.put("foo", "hello")
+	h.compactMem()
+	h.compactRange("", "")
+	h.closeDB()
+	h.corrupt(storage.TypeManifest, -1, 0, 1000)
+	h.openAssert(false)
+
+	h.recover()
+	h.getVal("foo", "hello")
+
+	h.close()
+}
+
+func TestCorruptDB_CompactionInputError(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.build(10)
+	h.compactMem()
+	h.closeDB()
+	h.corrupt(storage.TypeTable, -1, 100, 1)
+
+	h.openDB()
+	h.check(9, 9)
+
+	h.build(10000)
+	h.check(10000, 10000)
+
+	h.close()
+}
+
+func TestCorruptDB_UnrelatedKeys(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.build(10)
+	h.compactMem()
+	h.closeDB()
+	h.corrupt(storage.TypeTable, -1, 100, 1)
+
+	h.openDB()
+	h.put(string(tkey(1000)), string(tval(1000, ctValSize)))
+	h.getVal(string(tkey(1000)), string(tval(1000, ctValSize)))
+	h.compactMem()
+	h.getVal(string(tkey(1000)), string(tval(1000, ctValSize)))
+
+	h.close()
+}
+
+func TestCorruptDB_Level0NewerFileHasOlderSeqnum(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.put("a", "v1")
+	h.put("b", "v1")
+	h.compactMem()
+	h.put("a", "v2")
+	h.put("b", "v2")
+	h.compactMem()
+	h.put("a", "v3")
+	h.put("b", "v3")
+	h.compactMem()
+	h.put("c", "v0")
+	h.put("d", "v0")
+	h.compactMem()
+	h.compactRangeAt(1, "", "")
+	h.closeDB()
+
+	h.recover()
+	h.getVal("a", "v3")
+	h.getVal("b", "v3")
+	h.getVal("c", "v0")
+	h.getVal("d", "v0")
+
+	h.close()
+}
+
+func TestCorruptDB_RecoverInvalidSeq_Issue53(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.put("a", "v1")
+	h.put("b", "v1")
+	h.compactMem()
+	h.put("a", "v2")
+	h.put("b", "v2")
+	h.compactMem()
+	h.put("a", "v3")
+	h.put("b", "v3")
+	h.compactMem()
+	h.put("c", "v0")
+	h.put("d", "v0")
+	h.compactMem()
+	h.compactRangeAt(0, "", "")
+	h.closeDB()
+
+	h.recover()
+	h.getVal("a", "v3")
+	h.getVal("b", "v3")
+	h.getVal("c", "v0")
+	h.getVal("d", "v0")
+
+	h.close()
+}
+
+func TestCorruptDB_MissingTableFiles(t *testing.T) {
+	h := newDbCorruptHarness(t)
+
+	h.put("a", "v1")
+	h.put("b", "v1")
+	h.compactMem()
+	h.put("c", "v2")
+	h.put("d", "v2")
+	h.compactMem()
+	h.put("e", "v3")
+	h.put("f", "v3")
+	h.closeDB()
+
+	h.removeOne(storage.TypeTable)
+	h.openAssert(false)
+
+	h.close()
+}
+
+func TestCorruptDB_RecoverTable(t *testing.T) {
+	h := newDbCorruptHarnessWopt(t, &opt.Options{
+		WriteBuffer:         112 * opt.KiB,
+		CompactionTableSize: 90 * opt.KiB,
+		Filter:              filter.NewBloomFilter(10),
+	})
+
+	h.build(1000)
+	h.compactMem()
+	h.compactRangeAt(0, "", "")
+	h.compactRangeAt(1, "", "")
+	seq := h.db.seq
+	h.closeDB()
+	h.corrupt(storage.TypeTable, 0, 1000, 1)
+	h.corrupt(storage.TypeTable, 3, 10000, 1)
+	// Corrupted filter shouldn't affect recovery.
+	h.corrupt(storage.TypeTable, 3, 113888, 10)
+	h.corrupt(storage.TypeTable, -1, 20000, 1)
+
+	h.recover()
+	if h.db.seq != seq {
+		t.Errorf("invalid seq, want=%d got=%d", seq, h.db.seq)
+	}
+	h.check(985, 985)
+
+	h.close()
+}
@@ -0,0 +1,791 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"sync"
+	"time"
+
+	"github.com/syndtr/goleveldb/leveldb/errors"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+)
+
+var (
+	errCompactionTransactExiting = errors.New("leveldb: compaction transact exiting")
+)
+
+type cStats struct {
+	sync.Mutex
+	duration time.Duration
+	read     uint64
+	write    uint64
+}
+
+func (p *cStats) add(n *cStatsStaging) {
+	p.Lock()
+	p.duration += n.duration
+	p.read += n.read
+	p.write += n.write
+	p.Unlock()
+}
+
+func (p *cStats) get() (duration time.Duration, read, write uint64) {
+	p.Lock()
+	defer p.Unlock()
+	return p.duration, p.read, p.write
+}
+
+type cStatsStaging struct {
+	start    time.Time
+	duration time.Duration
+	on       bool
+	read     uint64
+	write    uint64
+}
+
+func (p *cStatsStaging) startTimer() {
+	if !p.on {
+		p.start = time.Now()
+		p.on = true
+	}
+}
+
+func (p *cStatsStaging) stopTimer() {
+	if p.on {
+		p.duration += time.Since(p.start)
+		p.on = false
+	}
+}
+
+func (db *DB) compactionError() {
+	var err error
+noerr:
+	// No error.
+	for {
+		select {
+		case err = <-db.compErrSetC:
+			switch {
+			case err == nil:
+			case err == ErrReadOnly, errors.IsCorrupted(err):
+				goto hasperr
+			default:
+				goto haserr
+			}
+		case _, _ = <-db.closeC:
+			return
+		}
+	}
+haserr:
+	// Transient error.
+	for {
+		select {
+		case db.compErrC <- err:
+		case err = <-db.compErrSetC:
+			switch {
+			case err == nil:
+				goto noerr
+			case err == ErrReadOnly, errors.IsCorrupted(err):
+				goto hasperr
+			default:
+			}
+		case _, _ = <-db.closeC:
+			return
+		}
+	}
+hasperr:
+	// Persistent error.
+	for {
+		select {
+		case db.compErrC <- err:
+		case db.compPerErrC <- err:
+		case db.writeLockC <- struct{}{}:
+			// Hold write lock, so that write won't pass-through.
+			db.compWriteLocking = true
+		case _, _ = <-db.closeC:
+			if db.compWriteLocking {
+				// We should release the lock or Close will hang.
+				<-db.writeLockC
+			}
+			return
+		}
+	}
+}
+
+type compactionTransactCounter int
+
+func (cnt *compactionTransactCounter) incr() {
+	*cnt++
+}
+
+type compactionTransactInterface interface {
+	run(cnt *compactionTransactCounter) error
+	revert() error
+}
+
+func (db *DB) compactionTransact(name string, t compactionTransactInterface) {
+	defer func() {
+		if x := recover(); x != nil {
+			if x == errCompactionTransactExiting {
+				if err := t.revert(); err != nil {
+					db.logf("%s revert error %q", name, err)
+				}
+			}
+			panic(x)
+		}
+	}()
+
+	const (
+		backoffMin = 1 * time.Second
+		backoffMax = 8 * time.Second
+		backoffMul = 2 * time.Second
+	)
+	var (
+		backoff  = backoffMin
+		backoffT = time.NewTimer(backoff)
+		lastCnt  = compactionTransactCounter(0)
+
+		disableBackoff = db.s.o.GetDisableCompactionBackoff()
+	)
+	for n := 0; ; n++ {
+		// Check wether the DB is closed.
+		if db.isClosed() {
+			db.logf("%s exiting", name)
+			db.compactionExitTransact()
+		} else if n > 0 {
+			db.logf("%s retrying N·%d", name, n)
+		}
+
+		// Execute.
+		cnt := compactionTransactCounter(0)
+		err := t.run(&cnt)
+		if err != nil {
+			db.logf("%s error I·%d %q", name, cnt, err)
+		}
+
+		// Set compaction error status.
+		select {
+		case db.compErrSetC <- err:
+		case perr := <-db.compPerErrC:
+			if err != nil {
+				db.logf("%s exiting (persistent error %q)", name, perr)
+				db.compactionExitTransact()
+			}
+		case _, _ = <-db.closeC:
+			db.logf("%s exiting", name)
+			db.compactionExitTransact()
+		}
+		if err == nil {
+			return
+		}
+		if errors.IsCorrupted(err) {
+			db.logf("%s exiting (corruption detected)", name)
+			db.compactionExitTransact()
+		}
+
+		if !disableBackoff {
+			// Reset backoff duration if counter is advancing.
+			if cnt > lastCnt {
+				backoff = backoffMin
+				lastCnt = cnt
+			}
+
+			// Backoff.
+			backoffT.Reset(backoff)
+			if backoff < backoffMax {
+				backoff *= backoffMul
+				if backoff > backoffMax {
+					backoff = backoffMax
+				}
+			}
+			select {
+			case <-backoffT.C:
+			case _, _ = <-db.closeC:
+				db.logf("%s exiting", name)
+				db.compactionExitTransact()
+			}
+		}
+	}
+}
+
+type compactionTransactFunc struct {
+	runFunc    func(cnt *compactionTransactCounter) error
+	revertFunc func() error
+}
+
+func (t *compactionTransactFunc) run(cnt *compactionTransactCounter) error {
+	return t.runFunc(cnt)
+}
+
+func (t *compactionTransactFunc) revert() error {
+	if t.revertFunc != nil {
+		return t.revertFunc()
+	}
+	return nil
+}
+
+func (db *DB) compactionTransactFunc(name string, run func(cnt *compactionTransactCounter) error, revert func() error) {
+	db.compactionTransact(name, &compactionTransactFunc{run, revert})
+}
+
+func (db *DB) compactionExitTransact() {
+	panic(errCompactionTransactExiting)
+}
+
+func (db *DB) memCompaction() {
+	mdb := db.getFrozenMem()
+	if mdb == nil {
+		return
+	}
+	defer mdb.decref()
+
+	db.logf("memdb@flush N·%d S·%s", mdb.Len(), shortenb(mdb.Size()))
+
+	// Don't compact empty memdb.
+	if mdb.Len() == 0 {
+		db.logf("memdb@flush skipping")
+		// drop frozen memdb
+		db.dropFrozenMem()
+		return
+	}
+
+	// Pause table compaction.
+	resumeC := make(chan struct{})
+	select {
+	case db.tcompPauseC <- (chan<- struct{})(resumeC):
+	case <-db.compPerErrC:
+		close(resumeC)
+		resumeC = nil
+	case _, _ = <-db.closeC:
+		return
+	}
+
+	var (
+		rec        = &sessionRecord{}
+		stats      = &cStatsStaging{}
+		flushLevel int
+	)
+
+	db.compactionTransactFunc("memdb@flush", func(cnt *compactionTransactCounter) (err error) {
+		stats.startTimer()
+		flushLevel, err = db.s.flushMemdb(rec, mdb.DB, -1)
+		stats.stopTimer()
+		return
+	}, func() error {
+		for _, r := range rec.addedTables {
+			db.logf("memdb@flush revert @%d", r.num)
+			f := db.s.getTableFile(r.num)
+			if err := f.Remove(); err != nil {
+				return err
+			}
+		}
+		return nil
+	})
+
+	db.compactionTransactFunc("memdb@commit", func(cnt *compactionTransactCounter) (err error) {
+		stats.startTimer()
+		rec.setJournalNum(db.journalFile.Num())
+		rec.setSeqNum(db.frozenSeq)
+		err = db.s.commit(rec)
+		stats.stopTimer()
+		return
+	}, nil)
+
+	db.logf("memdb@flush committed F·%d T·%v", len(rec.addedTables), stats.duration)
+
+	for _, r := range rec.addedTables {
+		stats.write += r.size
+	}
+	db.compStats[flushLevel].add(stats)
+
+	// Drop frozen memdb.
+	db.dropFrozenMem()
+
+	// Resume table compaction.
+	if resumeC != nil {
+		select {
+		case <-resumeC:
+			close(resumeC)
+		case _, _ = <-db.closeC:
+			return
+		}
+	}
+
+	// Trigger table compaction.
+	db.compSendTrigger(db.tcompCmdC)
+}
+
+type tableCompactionBuilder struct {
+	db           *DB
+	s            *session
+	c            *compaction
+	rec          *sessionRecord
+	stat0, stat1 *cStatsStaging
+
+	snapHasLastUkey bool
+	snapLastUkey    []byte
+	snapLastSeq     uint64
+	snapIter        int
+	snapKerrCnt     int
+	snapDropCnt     int
+
+	kerrCnt int
+	dropCnt int
+
+	minSeq    uint64
+	strict    bool
+	tableSize int
+
+	tw *tWriter
+}
+
+func (b *tableCompactionBuilder) appendKV(key, value []byte) error {
+	// Create new table if not already.
+	if b.tw == nil {
+		// Check for pause event.
+		if b.db != nil {
+			select {
+			case ch := <-b.db.tcompPauseC:
+				b.db.pauseCompaction(ch)
+			case _, _ = <-b.db.closeC:
+				b.db.compactionExitTransact()
+			default:
+			}
+		}
+
+		// Create new table.
+		var err error
+		b.tw, err = b.s.tops.create()
+		if err != nil {
+			return err
+		}
+	}
+
+	// Write key/value into table.
+	return b.tw.append(key, value)
+}
+
+func (b *tableCompactionBuilder) needFlush() bool {
+	return b.tw.tw.BytesLen() >= b.tableSize
+}
+
+func (b *tableCompactionBuilder) flush() error {
+	t, err := b.tw.finish()
+	if err != nil {
+		return err
+	}
+	b.rec.addTableFile(b.c.level+1, t)
+	b.stat1.write += t.size
+	b.s.logf("table@build created L%d@%d N·%d S·%s %q:%q", b.c.level+1, t.file.Num(), b.tw.tw.EntriesLen(), shortenb(int(t.size)), t.imin, t.imax)
+	b.tw = nil
+	return nil
+}
+
+func (b *tableCompactionBuilder) cleanup() {
+	if b.tw != nil {
+		b.tw.drop()
+		b.tw = nil
+	}
+}
+
+func (b *tableCompactionBuilder) run(cnt *compactionTransactCounter) error {
+	snapResumed := b.snapIter > 0
+	hasLastUkey := b.snapHasLastUkey // The key might has zero length, so this is necessary.
+	lastUkey := append([]byte{}, b.snapLastUkey...)
+	lastSeq := b.snapLastSeq
+	b.kerrCnt = b.snapKerrCnt
+	b.dropCnt = b.snapDropCnt
+	// Restore compaction state.
+	b.c.restore()
+
+	defer b.cleanup()
+
+	b.stat1.startTimer()
+	defer b.stat1.stopTimer()
+
+	iter := b.c.newIterator()
+	defer iter.Release()
+	for i := 0; iter.Next(); i++ {
+		// Incr transact counter.
+		cnt.incr()
+
+		// Skip until last state.
+		if i < b.snapIter {
+			continue
+		}
+
+		resumed := false
+		if snapResumed {
+			resumed = true
+			snapResumed = false
+		}
+
+		ikey := iter.Key()
+		ukey, seq, kt, kerr := parseIkey(ikey)
+
+		if kerr == nil {
+			shouldStop := !resumed && b.c.shouldStopBefore(ikey)
+
+			if !hasLastUkey || b.s.icmp.uCompare(lastUkey, ukey) != 0 {
+				// First occurrence of this user key.
+
+				// Only rotate tables if ukey doesn't hop across.
+				if b.tw != nil && (shouldStop || b.needFlush()) {
+					if err := b.flush(); err != nil {
+						return err
+					}
+
+					// Creates snapshot of the state.
+					b.c.save()
+					b.snapHasLastUkey = hasLastUkey
+					b.snapLastUkey = append(b.snapLastUkey[:0], lastUkey...)
+					b.snapLastSeq = lastSeq
+					b.snapIter = i
+					b.snapKerrCnt = b.kerrCnt
+					b.snapDropCnt = b.dropCnt
+				}
+
+				hasLastUkey = true
+				lastUkey = append(lastUkey[:0], ukey...)
+				lastSeq = kMaxSeq
+			}
+
+			switch {
+			case lastSeq <= b.minSeq:
+				// Dropped because newer entry for same user key exist
+				fallthrough // (A)
+			case kt == ktDel && seq <= b.minSeq && b.c.baseLevelForKey(lastUkey):
+				// For this user key:
+				// (1) there is no data in higher levels
+				// (2) data in lower levels will have larger seq numbers
+				// (3) data in layers that are being compacted here and have
+				//     smaller seq numbers will be dropped in the next
+				//     few iterations of this loop (by rule (A) above).
+				// Therefore this deletion marker is obsolete and can be dropped.
+				lastSeq = seq
+				b.dropCnt++
+				continue
+			default:
+				lastSeq = seq
+			}
+		} else {
+			if b.strict {
+				return kerr
+			}
+
+			// Don't drop corrupted keys.
+			hasLastUkey = false
+			lastUkey = lastUkey[:0]
+			lastSeq = kMaxSeq
+			b.kerrCnt++
+		}
+
+		if err := b.appendKV(ikey, iter.Value()); err != nil {
+			return err
+		}
+	}
+
+	if err := iter.Error(); err != nil {
+		return err
+	}
+
+	// Finish last table.
+	if b.tw != nil && !b.tw.empty() {
+		return b.flush()
+	}
+	return nil
+}
+
+func (b *tableCompactionBuilder) revert() error {
+	for _, at := range b.rec.addedTables {
+		b.s.logf("table@build revert @%d", at.num)
+		f := b.s.getTableFile(at.num)
+		if err := f.Remove(); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func (db *DB) tableCompaction(c *compaction, noTrivial bool) {
+	defer c.release()
+
+	rec := &sessionRecord{}
+	rec.addCompPtr(c.level, c.imax)
+
+	if !noTrivial && c.trivial() {
+		t := c.tables[0][0]
+		db.logf("table@move L%d@%d -> L%d", c.level, t.file.Num(), c.level+1)
+		rec.delTable(c.level, t.file.Num())
+		rec.addTableFile(c.level+1, t)
+		db.compactionTransactFunc("table@move", func(cnt *compactionTransactCounter) (err error) {
+			return db.s.commit(rec)
+		}, nil)
+		return
+	}
+
+	var stats [2]cStatsStaging
+	for i, tables := range c.tables {
+		for _, t := range tables {
+			stats[i].read += t.size
+			// Insert deleted tables into record
+			rec.delTable(c.level+i, t.file.Num())
+		}
+	}
+	sourceSize := int(stats[0].read + stats[1].read)
+	minSeq := db.minSeq()
+	db.logf("table@compaction L%d·%d -> L%d·%d S·%s Q·%d", c.level, len(c.tables[0]), c.level+1, len(c.tables[1]), shortenb(sourceSize), minSeq)
+
+	b := &tableCompactionBuilder{
+		db:        db,
+		s:         db.s,
+		c:         c,
+		rec:       rec,
+		stat1:     &stats[1],
+		minSeq:    minSeq,
+		strict:    db.s.o.GetStrict(opt.StrictCompaction),
+		tableSize: db.s.o.GetCompactionTableSize(c.level + 1),
+	}
+	db.compactionTransact("table@build", b)
+
+	// Commit changes
+	db.compactionTransactFunc("table@commit", func(cnt *compactionTransactCounter) (err error) {
+		stats[1].startTimer()
+		defer stats[1].stopTimer()
+		return db.s.commit(rec)
+	}, nil)
+
+	resultSize := int(stats[1].write)
+	db.logf("table@compaction committed F%s S%s Ke·%d D·%d T·%v", sint(len(rec.addedTables)-len(rec.deletedTables)), sshortenb(resultSize-sourceSize), b.kerrCnt, b.dropCnt, stats[1].duration)
+
+	// Save compaction stats
+	for i := range stats {
+		db.compStats[c.level+1].add(&stats[i])
+	}
+}
+
+func (db *DB) tableRangeCompaction(level int, umin, umax []byte) {
+	db.logf("table@compaction range L%d %q:%q", level, umin, umax)
+
+	if level >= 0 {
+		if c := db.s.getCompactionRange(level, umin, umax); c != nil {
+			db.tableCompaction(c, true)
+		}
+	} else {
+		v := db.s.version()
+		m := 1
+		for i, t := range v.tables[1:] {
+			if t.overlaps(db.s.icmp, umin, umax, false) {
+				m = i + 1
+			}
+		}
+		v.release()
+
+		for level := 0; level < m; level++ {
+			if c := db.s.getCompactionRange(level, umin, umax); c != nil {
+				db.tableCompaction(c, true)
+			}
+		}
+	}
+}
+
+func (db *DB) tableAutoCompaction() {
+	if c := db.s.pickCompaction(); c != nil {
+		db.tableCompaction(c, false)
+	}
+}
+
+func (db *DB) tableNeedCompaction() bool {
+	v := db.s.version()
+	defer v.release()
+	return v.needCompaction()
+}
+
+func (db *DB) pauseCompaction(ch chan<- struct{}) {
+	select {
+	case ch <- struct{}{}:
+	case _, _ = <-db.closeC:
+		db.compactionExitTransact()
+	}
+}
+
+type cCmd interface {
+	ack(err error)
+}
+
+type cIdle struct {
+	ackC chan<- error
+}
+
+func (r cIdle) ack(err error) {
+	if r.ackC != nil {
+		defer func() {
+			recover()
+		}()
+		r.ackC <- err
+	}
+}
+
+type cRange struct {
+	level    int
+	min, max []byte
+	ackC     chan<- error
+}
+
+func (r cRange) ack(err error) {
+	if r.ackC != nil {
+		defer func() {
+			recover()
+		}()
+		r.ackC <- err
+	}
+}
+
+// This will trigger auto compation and/or wait for all compaction to be done.
+func (db *DB) compSendIdle(compC chan<- cCmd) (err error) {
+	ch := make(chan error)
+	defer close(ch)
+	// Send cmd.
+	select {
+	case compC <- cIdle{ch}:
+	case err = <-db.compErrC:
+		return
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+	// Wait cmd.
+	select {
+	case err = <-ch:
+	case err = <-db.compErrC:
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+	return err
+}
+
+// This will trigger auto compaction but will not wait for it.
+func (db *DB) compSendTrigger(compC chan<- cCmd) {
+	select {
+	case compC <- cIdle{}:
+	default:
+	}
+}
+
+// Send range compaction request.
+func (db *DB) compSendRange(compC chan<- cCmd, level int, min, max []byte) (err error) {
+	ch := make(chan error)
+	defer close(ch)
+	// Send cmd.
+	select {
+	case compC <- cRange{level, min, max, ch}:
+	case err := <-db.compErrC:
+		return err
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+	// Wait cmd.
+	select {
+	case err = <-ch:
+	case err = <-db.compErrC:
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+	return err
+}
+
+func (db *DB) mCompaction() {
+	var x cCmd
+
+	defer func() {
+		if x := recover(); x != nil {
+			if x != errCompactionTransactExiting {
+				panic(x)
+			}
+		}
+		if x != nil {
+			x.ack(ErrClosed)
+		}
+		db.closeW.Done()
+	}()
+
+	for {
+		select {
+		case x = <-db.mcompCmdC:
+			switch x.(type) {
+			case cIdle:
+				db.memCompaction()
+				x.ack(nil)
+				x = nil
+			default:
+				panic("leveldb: unknown command")
+			}
+		case _, _ = <-db.closeC:
+			return
+		}
+	}
+}
+
+func (db *DB) tCompaction() {
+	var x cCmd
+	var ackQ []cCmd
+
+	defer func() {
+		if x := recover(); x != nil {
+			if x != errCompactionTransactExiting {
+				panic(x)
+			}
+		}
+		for i := range ackQ {
+			ackQ[i].ack(ErrClosed)
+			ackQ[i] = nil
+		}
+		if x != nil {
+			x.ack(ErrClosed)
+		}
+		db.closeW.Done()
+	}()
+
+	for {
+		if db.tableNeedCompaction() {
+			select {
+			case x = <-db.tcompCmdC:
+			case ch := <-db.tcompPauseC:
+				db.pauseCompaction(ch)
+				continue
+			case _, _ = <-db.closeC:
+				return
+			default:
+			}
+		} else {
+			for i := range ackQ {
+				ackQ[i].ack(nil)
+				ackQ[i] = nil
+			}
+			ackQ = ackQ[:0]
+			select {
+			case x = <-db.tcompCmdC:
+			case ch := <-db.tcompPauseC:
+				db.pauseCompaction(ch)
+				continue
+			case _, _ = <-db.closeC:
+				return
+			}
+		}
+		if x != nil {
+			switch cmd := x.(type) {
+			case cIdle:
+				ackQ = append(ackQ, x)
+			case cRange:
+				db.tableRangeCompaction(cmd.level, cmd.min, cmd.max)
+				x.ack(nil)
+			default:
+				panic("leveldb: unknown command")
+			}
+			x = nil
+		}
+		db.tableAutoCompaction()
+	}
+}
@@ -0,0 +1,350 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"errors"
+	"math/rand"
+	"runtime"
+	"sync"
+	"sync/atomic"
+
+	"github.com/syndtr/goleveldb/leveldb/iterator"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+var (
+	errInvalidIkey = errors.New("leveldb: Iterator: invalid internal key")
+)
+
+type memdbReleaser struct {
+	once sync.Once
+	m    *memDB
+}
+
+func (mr *memdbReleaser) Release() {
+	mr.once.Do(func() {
+		mr.m.decref()
+	})
+}
+
+func (db *DB) newRawIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
+	em, fm := db.getMems()
+	v := db.s.version()
+
+	ti := v.getIterators(slice, ro)
+	n := len(ti) + 2
+	i := make([]iterator.Iterator, 0, n)
+	emi := em.NewIterator(slice)
+	emi.SetReleaser(&memdbReleaser{m: em})
+	i = append(i, emi)
+	if fm != nil {
+		fmi := fm.NewIterator(slice)
+		fmi.SetReleaser(&memdbReleaser{m: fm})
+		i = append(i, fmi)
+	}
+	i = append(i, ti...)
+	strict := opt.GetStrict(db.s.o.Options, ro, opt.StrictReader)
+	mi := iterator.NewMergedIterator(i, db.s.icmp, strict)
+	mi.SetReleaser(&versionReleaser{v: v})
+	return mi
+}
+
+func (db *DB) newIterator(seq uint64, slice *util.Range, ro *opt.ReadOptions) *dbIter {
+	var islice *util.Range
+	if slice != nil {
+		islice = &util.Range{}
+		if slice.Start != nil {
+			islice.Start = newIkey(slice.Start, kMaxSeq, ktSeek)
+		}
+		if slice.Limit != nil {
+			islice.Limit = newIkey(slice.Limit, kMaxSeq, ktSeek)
+		}
+	}
+	rawIter := db.newRawIterator(islice, ro)
+	iter := &dbIter{
+		db:     db,
+		icmp:   db.s.icmp,
+		iter:   rawIter,
+		seq:    seq,
+		strict: opt.GetStrict(db.s.o.Options, ro, opt.StrictReader),
+		key:    make([]byte, 0),
+		value:  make([]byte, 0),
+	}
+	atomic.AddInt32(&db.aliveIters, 1)
+	runtime.SetFinalizer(iter, (*dbIter).Release)
+	return iter
+}
+
+func (db *DB) iterSamplingRate() int {
+	return rand.Intn(2 * db.s.o.GetIteratorSamplingRate())
+}
+
+type dir int
+
+const (
+	dirReleased dir = iota - 1
+	dirSOI
+	dirEOI
+	dirBackward
+	dirForward
+)
+
+// dbIter represent an interator states over a database session.
+type dbIter struct {
+	db     *DB
+	icmp   *iComparer
+	iter   iterator.Iterator
+	seq    uint64
+	strict bool
+
+	smaplingGap int
+	dir         dir
+	key         []byte
+	value       []byte
+	err         error
+	releaser    util.Releaser
+}
+
+func (i *dbIter) sampleSeek() {
+	ikey := i.iter.Key()
+	i.smaplingGap -= len(ikey) + len(i.iter.Value())
+	for i.smaplingGap < 0 {
+		i.smaplingGap += i.db.iterSamplingRate()
+		i.db.sampleSeek(ikey)
+	}
+}
+
+func (i *dbIter) setErr(err error) {
+	i.err = err
+	i.key = nil
+	i.value = nil
+}
+
+func (i *dbIter) iterErr() {
+	if err := i.iter.Error(); err != nil {
+		i.setErr(err)
+	}
+}
+
+func (i *dbIter) Valid() bool {
+	return i.err == nil && i.dir > dirEOI
+}
+
+func (i *dbIter) First() bool {
+	if i.err != nil {
+		return false
+	} else if i.dir == dirReleased {
+		i.err = ErrIterReleased
+		return false
+	}
+
+	if i.iter.First() {
+		i.dir = dirSOI
+		return i.next()
+	}
+	i.dir = dirEOI
+	i.iterErr()
+	return false
+}
+
+func (i *dbIter) Last() bool {
+	if i.err != nil {
+		return false
+	} else if i.dir == dirReleased {
+		i.err = ErrIterReleased
+		return false
+	}
+
+	if i.iter.Last() {
+		return i.prev()
+	}
+	i.dir = dirSOI
+	i.iterErr()
+	return false
+}
+
+func (i *dbIter) Seek(key []byte) bool {
+	if i.err != nil {
+		return false
+	} else if i.dir == dirReleased {
+		i.err = ErrIterReleased
+		return false
+	}
+
+	ikey := newIkey(key, i.seq, ktSeek)
+	if i.iter.Seek(ikey) {
+		i.dir = dirSOI
+		return i.next()
+	}
+	i.dir = dirEOI
+	i.iterErr()
+	return false
+}
+
+func (i *dbIter) next() bool {
+	for {
+		if ukey, seq, kt, kerr := parseIkey(i.iter.Key()); kerr == nil {
+			i.sampleSeek()
+			if seq <= i.seq {
+				switch kt {
+				case ktDel:
+					// Skip deleted key.
+					i.key = append(i.key[:0], ukey...)
+					i.dir = dirForward
+				case ktVal:
+					if i.dir == dirSOI || i.icmp.uCompare(ukey, i.key) > 0 {
+						i.key = append(i.key[:0], ukey...)
+						i.value = append(i.value[:0], i.iter.Value()...)
+						i.dir = dirForward
+						return true
+					}
+				}
+			}
+		} else if i.strict {
+			i.setErr(kerr)
+			break
+		}
+		if !i.iter.Next() {
+			i.dir = dirEOI
+			i.iterErr()
+			break
+		}
+	}
+	return false
+}
+
+func (i *dbIter) Next() bool {
+	if i.dir == dirEOI || i.err != nil {
+		return false
+	} else if i.dir == dirReleased {
+		i.err = ErrIterReleased
+		return false
+	}
+
+	if !i.iter.Next() || (i.dir == dirBackward && !i.iter.Next()) {
+		i.dir = dirEOI
+		i.iterErr()
+		return false
+	}
+	return i.next()
+}
+
+func (i *dbIter) prev() bool {
+	i.dir = dirBackward
+	del := true
+	if i.iter.Valid() {
+		for {
+			if ukey, seq, kt, kerr := parseIkey(i.iter.Key()); kerr == nil {
+				i.sampleSeek()
+				if seq <= i.seq {
+					if !del && i.icmp.uCompare(ukey, i.key) < 0 {
+						return true
+					}
+					del = (kt == ktDel)
+					if !del {
+						i.key = append(i.key[:0], ukey...)
+						i.value = append(i.value[:0], i.iter.Value()...)
+					}
+				}
+			} else if i.strict {
+				i.setErr(kerr)
+				return false
+			}
+			if !i.iter.Prev() {
+				break
+			}
+		}
+	}
+	if del {
+		i.dir = dirSOI
+		i.iterErr()
+		return false
+	}
+	return true
+}
+
+func (i *dbIter) Prev() bool {
+	if i.dir == dirSOI || i.err != nil {
+		return false
+	} else if i.dir == dirReleased {
+		i.err = ErrIterReleased
+		return false
+	}
+
+	switch i.dir {
+	case dirEOI:
+		return i.Last()
+	case dirForward:
+		for i.iter.Prev() {
+			if ukey, _, _, kerr := parseIkey(i.iter.Key()); kerr == nil {
+				i.sampleSeek()
+				if i.icmp.uCompare(ukey, i.key) < 0 {
+					goto cont
+				}
+			} else if i.strict {
+				i.setErr(kerr)
+				return false
+			}
+		}
+		i.dir = dirSOI
+		i.iterErr()
+		return false
+	}
+
+cont:
+	return i.prev()
+}
+
+func (i *dbIter) Key() []byte {
+	if i.err != nil || i.dir <= dirEOI {
+		return nil
+	}
+	return i.key
+}
+
+func (i *dbIter) Value() []byte {
+	if i.err != nil || i.dir <= dirEOI {
+		return nil
+	}
+	return i.value
+}
+
+func (i *dbIter) Release() {
+	if i.dir != dirReleased {
+		// Clear the finalizer.
+		runtime.SetFinalizer(i, nil)
+
+		if i.releaser != nil {
+			i.releaser.Release()
+			i.releaser = nil
+		}
+
+		i.dir = dirReleased
+		i.key = nil
+		i.value = nil
+		i.iter.Release()
+		i.iter = nil
+		atomic.AddInt32(&i.db.aliveIters, -1)
+		i.db = nil
+	}
+}
+
+func (i *dbIter) SetReleaser(releaser util.Releaser) {
+	if i.dir == dirReleased {
+		panic(util.ErrReleased)
+	}
+	if i.releaser != nil && releaser != nil {
+		panic(util.ErrHasReleaser)
+	}
+	i.releaser = releaser
+}
+
+func (i *dbIter) Error() error {
+	return i.err
+}
@@ -0,0 +1,183 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"container/list"
+	"fmt"
+	"runtime"
+	"sync"
+	"sync/atomic"
+
+	"github.com/syndtr/goleveldb/leveldb/iterator"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+type snapshotElement struct {
+	seq uint64
+	ref int
+	e   *list.Element
+}
+
+// Acquires a snapshot, based on latest sequence.
+func (db *DB) acquireSnapshot() *snapshotElement {
+	db.snapsMu.Lock()
+	defer db.snapsMu.Unlock()
+
+	seq := db.getSeq()
+
+	if e := db.snapsList.Back(); e != nil {
+		se := e.Value.(*snapshotElement)
+		if se.seq == seq {
+			se.ref++
+			return se
+		} else if seq < se.seq {
+			panic("leveldb: sequence number is not increasing")
+		}
+	}
+	se := &snapshotElement{seq: seq, ref: 1}
+	se.e = db.snapsList.PushBack(se)
+	return se
+}
+
+// Releases given snapshot element.
+func (db *DB) releaseSnapshot(se *snapshotElement) {
+	db.snapsMu.Lock()
+	defer db.snapsMu.Unlock()
+
+	se.ref--
+	if se.ref == 0 {
+		db.snapsList.Remove(se.e)
+		se.e = nil
+	} else if se.ref < 0 {
+		panic("leveldb: Snapshot: negative element reference")
+	}
+}
+
+// Gets minimum sequence that not being snapshoted.
+func (db *DB) minSeq() uint64 {
+	db.snapsMu.Lock()
+	defer db.snapsMu.Unlock()
+
+	if e := db.snapsList.Front(); e != nil {
+		return e.Value.(*snapshotElement).seq
+	}
+
+	return db.getSeq()
+}
+
+// Snapshot is a DB snapshot.
+type Snapshot struct {
+	db       *DB
+	elem     *snapshotElement
+	mu       sync.RWMutex
+	released bool
+}
+
+// Creates new snapshot object.
+func (db *DB) newSnapshot() *Snapshot {
+	snap := &Snapshot{
+		db:   db,
+		elem: db.acquireSnapshot(),
+	}
+	atomic.AddInt32(&db.aliveSnaps, 1)
+	runtime.SetFinalizer(snap, (*Snapshot).Release)
+	return snap
+}
+
+func (snap *Snapshot) String() string {
+	return fmt.Sprintf("leveldb.Snapshot{%d}", snap.elem.seq)
+}
+
+// Get gets the value for the given key. It returns ErrNotFound if
+// the DB does not contains the key.
+//
+// The caller should not modify the contents of the returned slice, but
+// it is safe to modify the contents of the argument after Get returns.
+func (snap *Snapshot) Get(key []byte, ro *opt.ReadOptions) (value []byte, err error) {
+	err = snap.db.ok()
+	if err != nil {
+		return
+	}
+	snap.mu.RLock()
+	defer snap.mu.RUnlock()
+	if snap.released {
+		err = ErrSnapshotReleased
+		return
+	}
+	return snap.db.get(key, snap.elem.seq, ro)
+}
+
+// Has returns true if the DB does contains the given key.
+//
+// It is safe to modify the contents of the argument after Get returns.
+func (snap *Snapshot) Has(key []byte, ro *opt.ReadOptions) (ret bool, err error) {
+	err = snap.db.ok()
+	if err != nil {
+		return
+	}
+	snap.mu.RLock()
+	defer snap.mu.RUnlock()
+	if snap.released {
+		err = ErrSnapshotReleased
+		return
+	}
+	return snap.db.has(key, snap.elem.seq, ro)
+}
+
+// NewIterator returns an iterator for the snapshot of the uderlying DB.
+// The returned iterator is not goroutine-safe, but it is safe to use
+// multiple iterators concurrently, with each in a dedicated goroutine.
+// It is also safe to use an iterator concurrently with modifying its
+// underlying DB. The resultant key/value pairs are guaranteed to be
+// consistent.
+//
+// Slice allows slicing the iterator to only contains keys in the given
+// range. A nil Range.Start is treated as a key before all keys in the
+// DB. And a nil Range.Limit is treated as a key after all keys in
+// the DB.
+//
+// The iterator must be released after use, by calling Release method.
+// Releasing the snapshot doesn't mean releasing the iterator too, the
+// iterator would be still valid until released.
+//
+// Also read Iterator documentation of the leveldb/iterator package.
+func (snap *Snapshot) NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator {
+	if err := snap.db.ok(); err != nil {
+		return iterator.NewEmptyIterator(err)
+	}
+	snap.mu.Lock()
+	defer snap.mu.Unlock()
+	if snap.released {
+		return iterator.NewEmptyIterator(ErrSnapshotReleased)
+	}
+	// Since iterator already hold version ref, it doesn't need to
+	// hold snapshot ref.
+	return snap.db.newIterator(snap.elem.seq, slice, ro)
+}
+
+// Release releases the snapshot. This will not release any returned
+// iterators, the iterators would still be valid until released or the
+// underlying DB is closed.
+//
+// Other methods should not be called after the snapshot has been released.
+func (snap *Snapshot) Release() {
+	snap.mu.Lock()
+	defer snap.mu.Unlock()
+
+	if !snap.released {
+		// Clear the finalizer.
+		runtime.SetFinalizer(snap, nil)
+
+		snap.released = true
+		snap.db.releaseSnapshot(snap.elem)
+		atomic.AddInt32(&snap.db.aliveSnaps, -1)
+		snap.db = nil
+		snap.elem = nil
+	}
+}
@@ -0,0 +1,211 @@
+// Copyright (c) 2013, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"sync/atomic"
+	"time"
+
+	"github.com/syndtr/goleveldb/leveldb/journal"
+	"github.com/syndtr/goleveldb/leveldb/memdb"
+)
+
+type memDB struct {
+	db *DB
+	*memdb.DB
+	ref int32
+}
+
+func (m *memDB) incref() {
+	atomic.AddInt32(&m.ref, 1)
+}
+
+func (m *memDB) decref() {
+	if ref := atomic.AddInt32(&m.ref, -1); ref == 0 {
+		// Only put back memdb with std capacity.
+		if m.Capacity() == m.db.s.o.GetWriteBuffer() {
+			m.Reset()
+			m.db.mpoolPut(m.DB)
+		}
+		m.db = nil
+		m.DB = nil
+	} else if ref < 0 {
+		panic("negative memdb ref")
+	}
+}
+
+// Get latest sequence number.
+func (db *DB) getSeq() uint64 {
+	return atomic.LoadUint64(&db.seq)
+}
+
+// Atomically adds delta to seq.
+func (db *DB) addSeq(delta uint64) {
+	atomic.AddUint64(&db.seq, delta)
+}
+
+func (db *DB) sampleSeek(ikey iKey) {
+	v := db.s.version()
+	if v.sampleSeek(ikey) {
+		// Trigger table compaction.
+		db.compSendTrigger(db.tcompCmdC)
+	}
+	v.release()
+}
+
+func (db *DB) mpoolPut(mem *memdb.DB) {
+	defer func() {
+		recover()
+	}()
+	select {
+	case db.memPool <- mem:
+	default:
+	}
+}
+
+func (db *DB) mpoolGet() *memdb.DB {
+	select {
+	case mem := <-db.memPool:
+		return mem
+	default:
+		return nil
+	}
+}
+
+func (db *DB) mpoolDrain() {
+	ticker := time.NewTicker(30 * time.Second)
+	for {
+		select {
+		case <-ticker.C:
+			select {
+			case <-db.memPool:
+			default:
+			}
+		case _, _ = <-db.closeC:
+			close(db.memPool)
+			return
+		}
+	}
+}
+
+// Create new memdb and froze the old one; need external synchronization.
+// newMem only called synchronously by the writer.
+func (db *DB) newMem(n int) (mem *memDB, err error) {
+	num := db.s.allocFileNum()
+	file := db.s.getJournalFile(num)
+	w, err := file.Create()
+	if err != nil {
+		db.s.reuseFileNum(num)
+		return
+	}
+
+	db.memMu.Lock()
+	defer db.memMu.Unlock()
+
+	if db.frozenMem != nil {
+		panic("still has frozen mem")
+	}
+
+	if db.journal == nil {
+		db.journal = journal.NewWriter(w)
+	} else {
+		db.journal.Reset(w)
+		db.journalWriter.Close()
+		db.frozenJournalFile = db.journalFile
+	}
+	db.journalWriter = w
+	db.journalFile = file
+	db.frozenMem = db.mem
+	mdb := db.mpoolGet()
+	if mdb == nil || mdb.Capacity() < n {
+		mdb = memdb.New(db.s.icmp, maxInt(db.s.o.GetWriteBuffer(), n))
+	}
+	mem = &memDB{
+		db:  db,
+		DB:  mdb,
+		ref: 2,
+	}
+	db.mem = mem
+	// The seq only incremented by the writer. And whoever called newMem
+	// should hold write lock, so no need additional synchronization here.
+	db.frozenSeq = db.seq
+	return
+}
+
+// Get all memdbs.
+func (db *DB) getMems() (e, f *memDB) {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	if db.mem == nil {
+		panic("nil effective mem")
+	}
+	db.mem.incref()
+	if db.frozenMem != nil {
+		db.frozenMem.incref()
+	}
+	return db.mem, db.frozenMem
+}
+
+// Get frozen memdb.
+func (db *DB) getEffectiveMem() *memDB {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	if db.mem == nil {
+		panic("nil effective mem")
+	}
+	db.mem.incref()
+	return db.mem
+}
+
+// Check whether we has frozen memdb.
+func (db *DB) hasFrozenMem() bool {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	return db.frozenMem != nil
+}
+
+// Get frozen memdb.
+func (db *DB) getFrozenMem() *memDB {
+	db.memMu.RLock()
+	defer db.memMu.RUnlock()
+	if db.frozenMem != nil {
+		db.frozenMem.incref()
+	}
+	return db.frozenMem
+}
+
+// Drop frozen memdb; assume that frozen memdb isn't nil.
+func (db *DB) dropFrozenMem() {
+	db.memMu.Lock()
+	if err := db.frozenJournalFile.Remove(); err != nil {
+		db.logf("journal@remove removing @%d %q", db.frozenJournalFile.Num(), err)
+	} else {
+		db.logf("journal@remove removed @%d", db.frozenJournalFile.Num())
+	}
+	db.frozenJournalFile = nil
+	db.frozenMem.decref()
+	db.frozenMem = nil
+	db.memMu.Unlock()
+}
+
+// Set closed flag; return true if not already closed.
+func (db *DB) setClosed() bool {
+	return atomic.CompareAndSwapUint32(&db.closed, 0, 1)
+}
+
+// Check whether DB was closed.
+func (db *DB) isClosed() bool {
+	return atomic.LoadUint32(&db.closed) != 0
+}
+
+// Check read ok status.
+func (db *DB) ok() error {
+	if db.isClosed() {
+		return ErrClosed
+	}
+	return nil
+}
@@ -0,0 +1,100 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"github.com/syndtr/goleveldb/leveldb/errors"
+	"github.com/syndtr/goleveldb/leveldb/iterator"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/storage"
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+// Reader is the interface that wraps basic Get and NewIterator methods.
+// This interface implemented by both DB and Snapshot.
+type Reader interface {
+	Get(key []byte, ro *opt.ReadOptions) (value []byte, err error)
+	NewIterator(slice *util.Range, ro *opt.ReadOptions) iterator.Iterator
+}
+
+type Sizes []uint64
+
+// Sum returns sum of the sizes.
+func (p Sizes) Sum() (n uint64) {
+	for _, s := range p {
+		n += s
+	}
+	return n
+}
+
+// Logging.
+func (db *DB) log(v ...interface{})                 { db.s.log(v...) }
+func (db *DB) logf(format string, v ...interface{}) { db.s.logf(format, v...) }
+
+// Check and clean files.
+func (db *DB) checkAndCleanFiles() error {
+	v := db.s.version()
+	defer v.release()
+
+	tablesMap := make(map[uint64]bool)
+	for _, tables := range v.tables {
+		for _, t := range tables {
+			tablesMap[t.file.Num()] = false
+		}
+	}
+
+	files, err := db.s.getFiles(storage.TypeAll)
+	if err != nil {
+		return err
+	}
+
+	var nTables int
+	var rem []storage.File
+	for _, f := range files {
+		keep := true
+		switch f.Type() {
+		case storage.TypeManifest:
+			keep = f.Num() >= db.s.manifestFile.Num()
+		case storage.TypeJournal:
+			if db.frozenJournalFile != nil {
+				keep = f.Num() >= db.frozenJournalFile.Num()
+			} else {
+				keep = f.Num() >= db.journalFile.Num()
+			}
+		case storage.TypeTable:
+			_, keep = tablesMap[f.Num()]
+			if keep {
+				tablesMap[f.Num()] = true
+				nTables++
+			}
+		}
+
+		if !keep {
+			rem = append(rem, f)
+		}
+	}
+
+	if nTables != len(tablesMap) {
+		var missing []*storage.FileInfo
+		for num, present := range tablesMap {
+			if !present {
+				missing = append(missing, &storage.FileInfo{Type: storage.TypeTable, Num: num})
+				db.logf("db@janitor table missing @%d", num)
+			}
+		}
+		return errors.NewErrCorrupted(nil, &errors.ErrMissingFiles{Files: missing})
+	}
+
+	db.logf("db@janitor F·%d G·%d", len(files), len(rem))
+	for _, f := range rem {
+		db.logf("db@janitor removing %s-%d", f.Type(), f.Num())
+		if err := f.Remove(); err != nil {
+			return err
+		}
+	}
+	return nil
+}
@@ -0,0 +1,338 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"time"
+
+	"github.com/syndtr/goleveldb/leveldb/memdb"
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+func (db *DB) writeJournal(b *Batch) error {
+	w, err := db.journal.Next()
+	if err != nil {
+		return err
+	}
+	if _, err := w.Write(b.encode()); err != nil {
+		return err
+	}
+	if err := db.journal.Flush(); err != nil {
+		return err
+	}
+	if b.sync {
+		return db.journalWriter.Sync()
+	}
+	return nil
+}
+
+func (db *DB) jWriter() {
+	defer db.closeW.Done()
+	for {
+		select {
+		case b := <-db.journalC:
+			if b != nil {
+				db.journalAckC <- db.writeJournal(b)
+			}
+		case _, _ = <-db.closeC:
+			return
+		}
+	}
+}
+
+func (db *DB) rotateMem(n int) (mem *memDB, err error) {
+	// Wait for pending memdb compaction.
+	err = db.compSendIdle(db.mcompCmdC)
+	if err != nil {
+		return
+	}
+
+	// Create new memdb and journal.
+	mem, err = db.newMem(n)
+	if err != nil {
+		return
+	}
+
+	// Schedule memdb compaction.
+	db.compSendTrigger(db.mcompCmdC)
+	return
+}
+
+func (db *DB) flush(n int) (mdb *memDB, mdbFree int, err error) {
+	delayed := false
+	flush := func() (retry bool) {
+		v := db.s.version()
+		defer v.release()
+		mdb = db.getEffectiveMem()
+		defer func() {
+			if retry {
+				mdb.decref()
+				mdb = nil
+			}
+		}()
+		mdbFree = mdb.Free()
+		switch {
+		case v.tLen(0) >= db.s.o.GetWriteL0SlowdownTrigger() && !delayed:
+			delayed = true
+			time.Sleep(time.Millisecond)
+		case mdbFree >= n:
+			return false
+		case v.tLen(0) >= db.s.o.GetWriteL0PauseTrigger():
+			delayed = true
+			err = db.compSendIdle(db.tcompCmdC)
+			if err != nil {
+				return false
+			}
+		default:
+			// Allow memdb to grow if it has no entry.
+			if mdb.Len() == 0 {
+				mdbFree = n
+			} else {
+				mdb.decref()
+				mdb, err = db.rotateMem(n)
+				if err == nil {
+					mdbFree = mdb.Free()
+				} else {
+					mdbFree = 0
+				}
+			}
+			return false
+		}
+		return true
+	}
+	start := time.Now()
+	for flush() {
+	}
+	if delayed {
+		db.writeDelay += time.Since(start)
+		db.writeDelayN++
+	} else if db.writeDelayN > 0 {
+		db.logf("db@write was delayed N·%d T·%v", db.writeDelayN, db.writeDelay)
+		db.writeDelay = 0
+		db.writeDelayN = 0
+	}
+	return
+}
+
+// Write apply the given batch to the DB. The batch will be applied
+// sequentially.
+//
+// It is safe to modify the contents of the arguments after Write returns.
+func (db *DB) Write(b *Batch, wo *opt.WriteOptions) (err error) {
+	err = db.ok()
+	if err != nil || b == nil || b.Len() == 0 {
+		return
+	}
+
+	b.init(wo.GetSync() && !db.s.o.GetNoSync())
+
+	// The write happen synchronously.
+	select {
+	case db.writeC <- b:
+		if <-db.writeMergedC {
+			return <-db.writeAckC
+		}
+	case db.writeLockC <- struct{}{}:
+	case err = <-db.compPerErrC:
+		return
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+
+	merged := 0
+	danglingMerge := false
+	defer func() {
+		if danglingMerge {
+			db.writeMergedC <- false
+		} else {
+			<-db.writeLockC
+		}
+		for i := 0; i < merged; i++ {
+			db.writeAckC <- err
+		}
+	}()
+
+	mdb, mdbFree, err := db.flush(b.size())
+	if err != nil {
+		return
+	}
+	defer mdb.decref()
+
+	// Calculate maximum size of the batch.
+	m := 1 << 20
+	if x := b.size(); x <= 128<<10 {
+		m = x + (128 << 10)
+	}
+	m = minInt(m, mdbFree)
+
+	// Merge with other batch.
+drain:
+	for b.size() < m && !b.sync {
+		select {
+		case nb := <-db.writeC:
+			if b.size()+nb.size() <= m {
+				b.append(nb)
+				db.writeMergedC <- true
+				merged++
+			} else {
+				danglingMerge = true
+				break drain
+			}
+		default:
+			break drain
+		}
+	}
+
+	// Set batch first seq number relative from last seq.
+	b.seq = db.seq + 1
+
+	// Write journal concurrently if it is large enough.
+	if b.size() >= (128 << 10) {
+		// Push the write batch to the journal writer
+		select {
+		case db.journalC <- b:
+			// Write into memdb
+			if berr := b.memReplay(mdb.DB); berr != nil {
+				panic(berr)
+			}
+		case err = <-db.compPerErrC:
+			return
+		case _, _ = <-db.closeC:
+			err = ErrClosed
+			return
+		}
+		// Wait for journal writer
+		select {
+		case err = <-db.journalAckC:
+			if err != nil {
+				// Revert memdb if error detected
+				if berr := b.revertMemReplay(mdb.DB); berr != nil {
+					panic(berr)
+				}
+				return
+			}
+		case _, _ = <-db.closeC:
+			err = ErrClosed
+			return
+		}
+	} else {
+		err = db.writeJournal(b)
+		if err != nil {
+			return
+		}
+		if berr := b.memReplay(mdb.DB); berr != nil {
+			panic(berr)
+		}
+	}
+
+	// Set last seq number.
+	db.addSeq(uint64(b.Len()))
+
+	if b.size() >= mdbFree {
+		db.rotateMem(0)
+	}
+	return
+}
+
+// Put sets the value for the given key. It overwrites any previous value
+// for that key; a DB is not a multi-map.
+//
+// It is safe to modify the contents of the arguments after Put returns.
+func (db *DB) Put(key, value []byte, wo *opt.WriteOptions) error {
+	b := new(Batch)
+	b.Put(key, value)
+	return db.Write(b, wo)
+}
+
+// Delete deletes the value for the given key.
+//
+// It is safe to modify the contents of the arguments after Delete returns.
+func (db *DB) Delete(key []byte, wo *opt.WriteOptions) error {
+	b := new(Batch)
+	b.Delete(key)
+	return db.Write(b, wo)
+}
+
+func isMemOverlaps(icmp *iComparer, mem *memdb.DB, min, max []byte) bool {
+	iter := mem.NewIterator(nil)
+	defer iter.Release()
+	return (max == nil || (iter.First() && icmp.uCompare(max, iKey(iter.Key()).ukey()) >= 0)) &&
+		(min == nil || (iter.Last() && icmp.uCompare(min, iKey(iter.Key()).ukey()) <= 0))
+}
+
+// CompactRange compacts the underlying DB for the given key range.
+// In particular, deleted and overwritten versions are discarded,
+// and the data is rearranged to reduce the cost of operations
+// needed to access the data. This operation should typically only
+// be invoked by users who understand the underlying implementation.
+//
+// A nil Range.Start is treated as a key before all keys in the DB.
+// And a nil Range.Limit is treated as a key after all keys in the DB.
+// Therefore if both is nil then it will compact entire DB.
+func (db *DB) CompactRange(r util.Range) error {
+	if err := db.ok(); err != nil {
+		return err
+	}
+
+	// Lock writer.
+	select {
+	case db.writeLockC <- struct{}{}:
+	case err := <-db.compPerErrC:
+		return err
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+
+	// Check for overlaps in memdb.
+	mdb := db.getEffectiveMem()
+	defer mdb.decref()
+	if isMemOverlaps(db.s.icmp, mdb.DB, r.Start, r.Limit) {
+		// Memdb compaction.
+		if _, err := db.rotateMem(0); err != nil {
+			<-db.writeLockC
+			return err
+		}
+		<-db.writeLockC
+		if err := db.compSendIdle(db.mcompCmdC); err != nil {
+			return err
+		}
+	} else {
+		<-db.writeLockC
+	}
+
+	// Table compaction.
+	return db.compSendRange(db.tcompCmdC, -1, r.Start, r.Limit)
+}
+
+// SetReadOnly makes DB read-only. It will stay read-only until reopened.
+func (db *DB) SetReadOnly() error {
+	if err := db.ok(); err != nil {
+		return err
+	}
+
+	// Lock writer.
+	select {
+	case db.writeLockC <- struct{}{}:
+		db.compWriteLocking = true
+	case err := <-db.compPerErrC:
+		return err
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+
+	// Set compaction read-only.
+	select {
+	case db.compErrSetC <- ErrReadOnly:
+	case perr := <-db.compPerErrC:
+		return perr
+	case _, _ = <-db.closeC:
+		return ErrClosed
+	}
+
+	return nil
+}
@@ -0,0 +1,90 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// Package leveldb provides implementation of LevelDB key/value database.
+//
+// Create or open a database:
+//
+//	db, err := leveldb.OpenFile("path/to/db", nil)
+//	...
+//	defer db.Close()
+//	...
+//
+// Read or modify the database content:
+//
+//	// Remember that the contents of the returned slice should not be modified.
+//	data, err := db.Get([]byte("key"), nil)
+//	...
+//	err = db.Put([]byte("key"), []byte("value"), nil)
+//	...
+//	err = db.Delete([]byte("key"), nil)
+//	...
+//
+// Iterate over database content:
+//
+//	iter := db.NewIterator(nil, nil)
+//	for iter.Next() {
+//		// Remember that the contents of the returned slice should not be modified, and
+//		// only valid until the next call to Next.
+//		key := iter.Key()
+//		value := iter.Value()
+//		...
+//	}
+//	iter.Release()
+//	err = iter.Error()
+//	...
+//
+// Iterate over subset of database content with a particular prefix:
+//	iter := db.NewIterator(util.BytesPrefix([]byte("foo-")), nil)
+//	for iter.Next() {
+//		// Use key/value.
+//		...
+//	}
+//	iter.Release()
+//	err = iter.Error()
+//	...
+//
+// Seek-then-Iterate:
+//
+// 	iter := db.NewIterator(nil, nil)
+// 	for ok := iter.Seek(key); ok; ok = iter.Next() {
+// 		// Use key/value.
+// 		...
+// 	}
+// 	iter.Release()
+// 	err = iter.Error()
+// 	...
+//
+// Iterate over subset of database content:
+//
+// 	iter := db.NewIterator(&util.Range{Start: []byte("foo"), Limit: []byte("xoo")}, nil)
+// 	for iter.Next() {
+// 		// Use key/value.
+// 		...
+// 	}
+// 	iter.Release()
+// 	err = iter.Error()
+// 	...
+//
+// Batch writes:
+//
+//	batch := new(leveldb.Batch)
+//	batch.Put([]byte("foo"), []byte("value"))
+//	batch.Put([]byte("bar"), []byte("another value"))
+//	batch.Delete([]byte("baz"))
+//	err = db.Write(batch, nil)
+//	...
+//
+// Use bloom filter:
+//
+//	o := &opt.Options{
+//		Filter: filter.NewBloomFilter(10),
+//	}
+//	db, err := leveldb.OpenFile("path/to/db", o)
+//	...
+//	defer db.Close()
+//	...
+package leveldb
@@ -0,0 +1,19 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"github.com/syndtr/goleveldb/leveldb/errors"
+)
+
+var (
+	ErrNotFound         = errors.ErrNotFound
+	ErrReadOnly         = errors.New("leveldb: read-only mode")
+	ErrSnapshotReleased = errors.New("leveldb: snapshot released")
+	ErrIterReleased     = errors.New("leveldb: iterator released")
+	ErrClosed           = errors.New("leveldb: closed")
+)
@@ -0,0 +1,78 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// Package errors provides common error types used throughout leveldb.
+package errors
+
+import (
+	"errors"
+	"fmt"
+
+	"github.com/syndtr/goleveldb/leveldb/storage"
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+var (
+	ErrNotFound    = New("leveldb: not found")
+	ErrReleased    = util.ErrReleased
+	ErrHasReleaser = util.ErrHasReleaser
+)
+
+// New returns an error that formats as the given text.
+func New(text string) error {
+	return errors.New(text)
+}
+
+// ErrCorrupted is the type that wraps errors that indicate corruption in
+// the database.
+type ErrCorrupted struct {
+	File *storage.FileInfo
+	Err  error
+}
+
+func (e *ErrCorrupted) Error() string {
+	if e.File != nil {
+		return fmt.Sprintf("%v [file=%v]", e.Err, e.File)
+	} else {
+		return e.Err.Error()
+	}
+}
+
+// NewErrCorrupted creates new ErrCorrupted error.
+func NewErrCorrupted(f storage.File, err error) error {
+	return &ErrCorrupted{storage.NewFileInfo(f), err}
+}
+
+// IsCorrupted returns a boolean indicating whether the error is indicating
+// a corruption.
+func IsCorrupted(err error) bool {
+	switch err.(type) {
+	case *ErrCorrupted:
+		return true
+	case *storage.ErrCorrupted:
+		return true
+	}
+	return false
+}
+
+// ErrMissingFiles is the type that indicating a corruption due to missing
+// files. ErrMissingFiles always wrapped with ErrCorrupted.
+type ErrMissingFiles struct {
+	Files []*storage.FileInfo
+}
+
+func (e *ErrMissingFiles) Error() string { return "file missing" }
+
+// SetFile sets 'file info' of the given error with the given file.
+// Currently only ErrCorrupted is supported, otherwise will do nothing.
+func SetFile(err error, f storage.File) error {
+	switch x := err.(type) {
+	case *ErrCorrupted:
+		x.File = storage.NewFileInfo(f)
+		return x
+	}
+	return err
+}
@@ -0,0 +1,58 @@
+// Copyright (c) 2014, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	. "github.com/onsi/ginkgo"
+	. "github.com/onsi/gomega"
+
+	"github.com/syndtr/goleveldb/leveldb/opt"
+	"github.com/syndtr/goleveldb/leveldb/testutil"
+)
+
+var _ = testutil.Defer(func() {
+	Describe("Leveldb external", func() {
+		o := &opt.Options{
+			DisableBlockCache:      true,
+			BlockRestartInterval:   5,
+			BlockSize:              80,
+			Compression:            opt.NoCompression,
+			OpenFilesCacheCapacity: -1,
+			Strict:                 opt.StrictAll,
+			WriteBuffer:            1000,
+			CompactionTableSize:    2000,
+		}
+
+		Describe("write test", func() {
+			It("should do write correctly", func(done Done) {
+				db := newTestingDB(o, nil, nil)
+				t := testutil.DBTesting{
+					DB:      db,
+					Deleted: testutil.KeyValue_Generate(nil, 500, 1, 50, 5, 5).Clone(),
+				}
+				testutil.DoDBTesting(&t)
+				db.TestClose()
+				done <- true
+			}, 20.0)
+		})
+
+		Describe("read test", func() {
+			testutil.AllKeyValueTesting(nil, nil, func(kv testutil.KeyValue) testutil.DB {
+				// Building the DB.
+				db := newTestingDB(o, nil, nil)
+				kv.IterateShuffled(nil, func(i int, key, value []byte) {
+					err := db.TestPut(key, value)
+					Expect(err).NotTo(HaveOccurred())
+				})
+
+				return db
+			}, func(db testutil.DB) {
+				db.(*testingDB).TestClose()
+			})
+		})
+	})
+})
@@ -0,0 +1,31 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package leveldb
+
+import (
+	"github.com/syndtr/goleveldb/leveldb/filter"
+)
+
+type iFilter struct {
+	filter.Filter
+}
+
+func (f iFilter) Contains(filter, key []byte) bool {
+	return f.Filter.Contains(filter, iKey(key).ukey())
+}
+
+func (f iFilter) NewGenerator() filter.FilterGenerator {
+	return iFilterGenerator{f.Filter.NewGenerator()}
+}
+
+type iFilterGenerator struct {
+	filter.FilterGenerator
+}
+
+func (g iFilterGenerator) Add(key []byte) {
+	g.FilterGenerator.Add(iKey(key).ukey())
+}
@@ -0,0 +1,116 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package filter
+
+import (
+	"github.com/syndtr/goleveldb/leveldb/util"
+)
+
+func bloomHash(key []byte) uint32 {
+	return util.Hash(key, 0xbc9f1d34)
+}
+
+type bloomFilter int
+
+// The bloom filter serializes its parameters and is backward compatible
+// with respect to them. Therefor, its parameters are not added to its
+// name.
+func (bloomFilter) Name() string {
+	return "leveldb.BuiltinBloomFilter"
+}
+
+func (f bloomFilter) Contains(filter, key []byte) bool {
+	nBytes := len(filter) - 1
+	if nBytes < 1 {
+		return false
+	}
+	nBits := uint32(nBytes * 8)
+
+	// Use the encoded k so that we can read filters generated by
+	// bloom filters created using different parameters.
+	k := filter[nBytes]
+	if k > 30 {
+		// Reserved for potentially new encodings for short bloom filters.
+		// Consider it a match.
+		return true
+	}
+
+	kh := bloomHash(key)
+	delta := (kh >> 17) | (kh << 15) // Rotate right 17 bits
+	for j := uint8(0); j < k; j++ {
+		bitpos := kh % nBits
+		if (uint32(filter[bitpos/8]) & (1 << (bitpos % 8))) == 0 {
+			return false
+		}
+		kh += delta
+	}
+	return true
+}
+
+func (f bloomFilter) NewGenerator() FilterGenerator {
+	// Round down to reduce probing cost a little bit.
+	k := uint8(f * 69 / 100) // 0.69 =~ ln(2)
+	if k < 1 {
+		k = 1
+	} else if k > 30 {
+		k = 30
+	}
+	return &bloomFilterGenerator{
+		n: int(f),
+		k: k,
+	}
+}
+
+type bloomFilterGenerator struct {
+	n int
+	k uint8
+
+	keyHashes []uint32
+}
+
+func (g *bloomFilterGenerator) Add(key []byte) {
+	// Use double-hashing to generate a sequence of hash values.
+	// See analysis in [Kirsch,Mitzenmacher 2006].
+	g.keyHashes = append(g.keyHashes, bloomHash(key))
+}
+
+func (g *bloomFilterGenerator) Generate(b Buffer) {
+	// Compute bloom filter size (in both bits and bytes)
+	nBits := uint32(len(g.keyHashes) * g.n)
+	// For small n, we can see a very high false positive rate.  Fix it
+	// by enforcing a minimum bloom filter length.
+	if nBits < 64 {
+		nBits = 64
+	}
+	nBytes := (nBits + 7) / 8
+	nBits = nBytes * 8
+
+	dest := b.Alloc(int(nBytes) + 1)
+	dest[nBytes] = g.k
+	for _, kh := range g.keyHashes {
+		delta := (kh >> 17) | (kh << 15) // Rotate right 17 bits
+		for j := uint8(0); j < g.k; j++ {
+			bitpos := kh % nBits
+			dest[bitpos/8] |= (1 << (bitpos % 8))
+			kh += delta
+		}
+	}
+
+	g.keyHashes = g.keyHashes[:0]
+}
+
+// NewBloomFilter creates a new initialized bloom filter for given
+// bitsPerKey.
+//
+// Since bitsPerKey is persisted individually for each bloom filter
+// serialization, bloom filters are backwards compatible with respect to
+// changing bitsPerKey. This means that no big performance penalty will
+// be experienced when changing the parameter. See documentation for
+// opt.Options.Filter for more information.
+func NewBloomFilter(bitsPerKey int) Filter {
+	return bloomFilter(bitsPerKey)
+}
@@ -0,0 +1,142 @@
+// Copyright (c) 2012, Suryandaru Triandana <syndtr@gmail.com>
+// All rights reserved.
+//
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+package filter
+
+import (
+	"encoding/binary"
+	"github.com/syndtr/goleveldb/leveldb/util"
+	"testing"
+)
+
+type harness struct {
+	t *testing.T
+
+	bloom     Filter
+	generator FilterGenerator
+	filter    []byte
+}
+
+func newHarness(t *testing.T) *harness {
+	bloom := NewBloomFilter(10)
+	return &harness{
+		t:         t,
+		bloom:     bloom,
+		generator: bloom.NewGenerator(),
+	}
+}
+
+func (h *harness) add(key []byte) {
+	h.generator.Add(key)
+}
+
+func (h *harness) addNum(key uint32) {
+	var b [4]byte
+	binary.LittleEndian.PutUint32(b[:], key)
+	h.add(b[:])
+}
+
+func (h *harness) build() {
+	b := &util.Buffer{}
+	h.generator.Generate(b)
+	h.filter = b.Bytes()
+}
+
+func (h *harness) reset() {
+	h.filter = nil
+}
+
+func (h *harness) filterLen() int {
+	return len(h.filter)
+}
+
+func (h *harness) assert(key []byte, want, silent bool) bool {
+	got := h.bloom.Contains(h.filter, key)
+	if !silent && got != want {
+		h.t.Errorf("assert on '%v' failed got '%v', want '%v'", key, got, want)
+	}
+	return got
+}
+
+func (h *harness) assertNum(key uint32, want, silent bool) bool {
+	var b [4]byte
+	binary.LittleEndian.PutUint32(b[:], key)
+	return h.assert(b[:], want, silent)
+}
+
+func TestBloomFilter_Empty(t *testing.T) {
+	h := newHarness(t)
+	h.build()
+	h.assert([]byte("hello"), false, false)
+	h.assert([]byte("world"), false, false)
+}
+
+func TestBloomFilter_Small(t *testing.T) {
+	h := newHarness(t)
+	h.add([]byte("hello"))
+	h.add([]byte("world"))
+	h.build()
+	h.assert([]byte("hello"), true, false)
+	h.assert([]byte("world"), true, false)
+	h.assert([]byte("x"), false, false)
+	h.assert([]byte("foo"), false, false)
+}
+
+func nextN(n int) int {
+	switch {
+	case n < 10:
+		n += 1
+	case n < 100:
+		n += 10
+	case n < 1000:
+		n += 100
+	default:
+		n += 1000
+	}
+	return n
+}
+
+func TestBloomFilter_VaryingLengths(t *testing.T) {
+	h := newHarness(t)
+	var mediocre, good int
+	for n := 1; n < 10000; n = nextN(n) {
+		h.reset()
+		for i := 0; i < n; i++ {
+			h.addNum(uint32(i))
+		}
+		h.build()
+
+		got := h.filterLen()
+		want := (n * 10 / 8) + 40
+		if got > want {
+			t.Errorf("filter len test failed, '%d' > '%d'", got, want)
+		}
+
+		for i := 0; i < n; i++ {
+			h.assertNum(uint32(i), true, false)
+		}
+
+		var rate float32
+		for i := 0; i < 10000; i++ {
+			if h.assertNum(uint32(i+1000000000), true, true) {
+				rate++
+			}
+		}
+		rate /= 10000
+		if rate > 0.02 {
+			t.Errorf("false positive rate is more than 2%%, got %v, at len %d", rate, n)
+		}
+		if rate > 0.0125 {
+			mediocre++
+		} else {
+			good++
+		}
+	}
+	t.Logf("false positive rate: %d good, %d mediocre", good, mediocre)
+	if mediocre > good/5 {
+		t.Error("mediocre false positive rate is more than expected")
+	}
+}
--- a/Show More
+++ b/Show More