The Vicious Circle of Documentation

Posted: Sun, 15 February 2015 | permalink | 2 Comments

Ever worked at a company (or on a codebase, or whatever) where it seemed like, no matter what the question was, the answer was written down somewhere you could easily find it? Most people haven’t, sadly, but they do exist, and I can assure you that it is an absolute pleasure.

On the other hand, practically everyone has experienced completely undocumented systems and processes, where knowledge is shared by word-of-mouth, or lost every time someone quits.

Why are there so many more undocumented systems than documented ones out there, and how can we cause more well-documented systems to exist? The answer isn’t “people are lazy”, and the solution is simple – though not easy.

Why Johnny Doesn’t Read

When someone needs to know something, they might go look for some documentation, or they might ask someone else or just guess wildly. The behaviour “look for documentation” is often reinforced negatively, by the result “documentation doesn’t exist”.

At the same time, the behaviours “ask someone” and “guess wildly” are positively reinforced, by the results “I get my question answered” and/or “at least I can get on with my work”. Over time, people optimise their behaviour by skipping the “look for documentation” step, and just go straight to asking other people (or guessing wildly).

Why Johnny Doesn’t Write

When someone writes documentation, they’re hoping that people will read it and not have to ask them questions in order to be productive and do the right thing. Hence, the behaviour “write documentation” is negatively reinforced by the results “I still get asked questions”, and “nobody does things the right way around here, dammit!”

Worse, though, is that there is very little positive reinforcement for the author: when someone does read the docs, and thus doesn’t ask a question, the author almost certainly doesn’t know they dodged a bullet. Similarly, when someone does things the right way, it’s unlikely that anyone will notice. It’s only the mistakes that catch the attention.

Given that the experience of writing documentation tends to skew towards the negative, it’s not surprising that eventually, the time spent writing documentation is reallocated to other, more utility-producing activities.

Death Spiral

The combination of these two situations is self-reinforcing. While a suitably motivated reader might start by strictly looking for documentation, or an author initially be enthused to always fully documenting their work, over time the “reflex” will be for readers to just go ask someone, because “there’s never any documentation!”, and for authors to not write documentation because “nobody bothers to read what I write anyway!”.

It is important to recognise that this iterative feedback loop is the “natural state” of the reader/author ecosystem, resulting in something akin to thermodynamic entropy. To avoid the system descending into chaos, energy needs to be constantly applied to keep the system in order.

The Solution

Effective methods for avoiding the vicious circle can be derived from the things that cause it. Change the forces that apply themselves to readers and authors, and they will behave differently.

On the reader’s side, the most effective way to encourage people to read documentation is for it to consistently exist. This means that those in control of a project or system mustn’t consider something “done” until the documentation is in a good state. Patches shouldn’t be landed, and releases shouldn’t be made, unless the documentation is altered to match the functional changes being made. Yes, this requires discipline, which is just a form of energy application to prevent entropic decay.

Writing documentation should be an explicit and well-understood part of somebody’s job description. Whoever is responsible for documentation needs to be given the time to do it properly. Writing well takes time and mental energy, and that time needs to be factored into the plans. Never forget that skimping on documentation, like short-changing QA or customer support, is a false economy that will cost more in the long term than it saves in the short term.

Even if the documentation exists, though, some people are going to tend towards asking people rather than consulting the documentation. This isn’t a moral failing on their part, but only happens when they believe that asking someone is more beneficial to them than going to the documentation. To change the behaviour, you need to change the belief.

You could change the belief by increasing the “cost” of asking. You could fire (or hellban) anyone who ever asks a question that is answered in the documentation. But you shouldn’t. You could yell “RTFM!” at everyone who asks a question. Thankfully that’s one acronym that’s falling out of favour.

Alternately, you can reduce the “cost” of getting the answer from the documentation. Possibly the largest single productivity boost for programmers, for example, has been the existence of Google. Whatever your problem, there’s a pretty good chance that a search or two will find a solution. For your private documentation, you probably don’t have the power of Google available, but decent full-text search systems are available. Use them.

Finally, authors would benefit from more positive reinforcement. If you find good documentation, let the author know! It requires a lot of effort (comparatively) to look up an author’s contact details and send them a nice e-mail. The “like” button is a more low-energy way of achieving a similar outcome – you click the button, and the author gets a warm, fuzzy feeling. If your internal documentation system doesn’t have some way to “close the loop” and let readers easily give authors a bit of kudos, fix it so it does.

Heck, even if authors just know that a page they wrote was loaded N times in the past week, that’s better than the current situation, in which deafening silence persists, punctuated by the occasional plaintive cry of “Hey, do you know how to…?”.

Do you have any other ideas for how to encourage readers to read, and for authors to write?


You stay classy, Uber

Posted: Sun, 23 November 2014 | permalink | 3 Comments

You may have heard that Uber has been under a bit of fire lately for its desires to hire private investigators to dig up “dirt” on journalists who are critical of Uber. From using users’ ride data for party entertainment, putting the assistance dogs of blind passengers in the trunk, adding a surcharge to reduce the number of dodgy drivers, or even booking rides with competitors and then cancelling, or using the ride to try and convince the driver to change teams, it’s pretty clear that Uber is a pretty good example of how companies are inherently sociopathic.

However, most of those examples are internal stupidities that happened to be made public. It’s a very rare company that doesn’t do all sorts of shady things, on the assumption that the world will never find out about them. Uber goes quite a bit further, though, and is so out-of-touch with the world that it blogs about analysing people’s sexual activity for amusement.

You’ll note that if you follow the above link, it sends you to the Wayback Machine, and not Uber’s own site. That’s because the original page has recently turned into a 404. Why? Probably because someone at Uber realised that bragging about how Uber employees can amuse themselves by perving on your one night stands might not be a great idea. That still leaves the question open of what sort of a corporate culture makes anyone ever think that inspecting user data for amusement would be a good thing, let alone publicising it? It’s horrific.

Thankfully, despite Uber’s fairly transparent attempt at whitewashing (“clearwashing”?), the good ol’ Wayback Machine helps us to remember what really went on. It would be amusing if Uber tried to pressure the Internet Archive to remove their copies of this blog post (don’t bother, Uber; I’ve got a “Save As” button and I’m not afraid to use it).

In any event, I’ve never used Uber (not that I’ve got one-night stands to analyse, anyway), and I’ll certainly not be patronising them in the future. If you’re not keen on companies amusing themselves with your private data, I suggest you might consider doing the same.


Multi-level prefix delegation is not a myth! I've seen it!

Posted: Thu, 20 November 2014 | permalink | No comments

Unless you’ve been living under a firewalled rock, you know that IPv6 is coming. There’s also a good chance that you’ve heard that IPv6 doesn’t have NAT. Or, if you pay close attention to the minutiae of IPv6 development, you’ve heard that IPv6 does have NAT, but you don’t have to (and shouldn’t) use it.

So let’s say we’ll skip NAT for IPv6. Fair enough. However, let’s say you have this use case:

  1. A bunch of containers that need Internet access…

  2. That are running in a VM…

  3. On your laptop…

  4. Behind your home router!

For IPv4, you’d just layer on the NAT, right? While SIP and IPsec might have kittens trying to work through three layers of NAT, for most things it’ll Just Work.

In the Grand Future of IPv6, without NAT, how the hell do you make that happen? The answer is “Prefix Delegation”, which allows routers to “delegate” management of a chunk of address space to downstream routers, and allow those downstream routers to, in turn, delegate pieces of that chunk to downstream routers.

In the case of our not-so-hypothetical containers-in-VM-on-laptop-at-home scenario, it would look like this:

  1. My “border router” (a DNS-323 running Debian) asks my ISP for a delegated prefix, using DHCPv6. The ISP delegates a /561. One /64 out of that is allocated to the network directly attached to the internal interface, and the rest goes into “the pool”, as /60 blocks (so I’ve got 15 of them to delegate, if required).

  2. My laptop gets an address on the LAN between itself and the DNS-323 via stateless auto-addressing (“SLAAC”). It also uses DHCPv6 to request one of the /60 blocks from the DNS-323. The laptop puts one /64 from that block as the address space for the “virtual LAN” (actually a Linux bridge) that connects the laptop to all my VMs, and puts the other 15 /64 blocks into a pool for delegation.

  3. The VM that will be running the set of containers under test gets an address on the “all VMs virtual LAN” via SLAAC, and then requests a delegated /64 to use for the “all containers virtual LAN” (another bridge, this one running on the VM itself) that the containers will each connect to themselves.

Now, almost all of this Just Works. The current releases of ISC DHCP support prefix delegation just fine, and a bit of shell script plumbing between the client and server seals the deal – the client needs to rewrite the server’s config file to tell it the netblock from which it can delegate.

Except for one teensy, tiny problem – routing. When the DHCP server delegates a netblock to a particular machine, the routing table needs to get updated so that packets going to that netblock actually get sent to the machine the netblock was delegated to. Without that, traffic destined for the containers (or the VM) won’t actually make it to its destination, and a one-way Internet connection isn’t a whole lot of use.

I cannot understand why this problem hasn’t been tripped over before. It’s absolutely fundamental to the correct operation of the delegation system. Some people advocate running a dynamic routing protocol, but that’s a sledgehammer to crack a nut if ever I saw one.

Actually, I know this problem has been tripped over before, by OpenWrt. Their solution, however, was to use a PHP script to scan logfiles and add routes. Suffice it to say, that wasn’t an option I was keen on exploring.

Instead, I decided to patch ISC DHCP so that the server can run an external script to add the necessary routes, and perhaps modify firewall rules – and also to reverse the process when the delegation is released (or expired). If anyone else wants to play around with it, I’ve put it up on Github. I don’t make any promises that it’s the right way to do it, necessarily, but it works, and the script I’ve added in contrib/prefix-delegation-routing.rb shows how it can be used to good effect. By the way, if anyone knows how pull requests work over at ISC, drop me a line. From the look of their website, they don’t appear to accept (or at least encourage) external contributions.

So, that’s one small patch for DHCP, one giant leap for my home network.

  1. The standard recommendation is for ISPs to delegate each end-user customer a /48 (giving 65,536 /64 networks); my ISP is being a little conservative in “only” giving me 256 /64s. It works fine for my purposes, but if you’re an ISP getting set for deploying IPv6, make life easy on your customers and give them a /48.


A benefit of running an alternate init in Debian Jessie

Posted: Sun, 16 November 2014 | permalink | 1 Comment

If you’re someone who doesn’t like Debian’s policy of automatically starting on install (or its heinous cousin, the RUN or ENABLE variable in /etc/default/<service>), then running an init system other than systemd should work out nicely.


My entry in the "Least Used Software EVAH" competition

Posted: Wed, 15 October 2014 | permalink | 2 Comments

For some reason, I seem to end up writing software for very esoteric use-cases. Today, though, I think I’ve outdone myself: I sat down and wrote a Ruby library to get and set process resource limits – those things that nobody ever thinks about except when they run out of file descriptors.

I didn’t even have a direct need for it. Recently I was grovelling through the EventMachine codebase, looking at the filehandle limit code, and noticed that the pure-ruby implementation didn’t manipulate filehandle limits. I considered adding it, then realised that there wasn’t a library available to do it. Since I haven’t berked around with FFI for a while, I decided to write rlimit. Now to find the time to write that patch for EventMachine…

Since I doubt there are many people who have a burning need to manipulate rlimits in Ruby, this gem will no doubt sit quiet and undisturbed in the dark, dusty corners of rubygems.org. However, for the three people on earth who find this useful: you’re welcome.


Chromium tabs crashing and not rendering correctly?

Posted: Sat, 30 August 2014 | permalink | No comments

If you’ve noticed your chrome/chromium on Linux having problems since you upgraded to somewhere around version 35/36, you’re not alone. Thankfully, it’s relatively easy to workaround. It will hit people who keep their browser open for a long time, or who have lots of tabs (or if you’re like me, and do both).

To tell if you’re suffering from this particular problem, crack open your ~/.xsession-errors file (or wherever your system logs stdout/stderr from programs running under X), and look for lines that look like this:

[22161:22185:0830/124533:ERROR:shared_memory_posix.cc(231)]
Creating shared memory in /dev/shm/.org.chromium.Chromium.gFTQSy
failed: Too many open files

And

[22161:22185:0830/124601:ERROR:host_shared_bitmap_manager.cc(122)]
Cannot create shared memory buffer

If you see those errors, congratulations! The rest of this blog post will be of use to you.

There’s probably a myriad of bugs open about this problem, but the one I found was #367037: Shared memory-related tab crash. It turns out there’s a file handle leak in the chromium codebase somewhere, relating to shared memory handling. There’s no fix available, but the workaround is quite simple: increase the number of files that processes are allowed to have open.

System-wide, you can do this by creating a file /etc/security/limits.d/local-nofile.conf, containing this line:

* - nofile 65535

You could also edit /etc/security/limits.conf to contain the same line, if you were so inclined. Note that this will only take effect next time you login, or perhaps even only when you restart X (or, at worst, your entire machine).

This doesn’t help you if you’ve got Chromium already open and you’d like to stop it from crashing Right Now (perhaps restarting your machine would be a terrible hardship, causing you to lose your hard-won uptime record), then you can use a magical tool called prlimit.

The prlimit syscall is available if you’re running a Linux 2.6.36 or later kernel, and running at least glibc 2.13. You’ll have a prlimit command line program if you’ve got util-linux 2.21 or later. If not, you can use the example source code in the prlimit(2) manpage, changing RLIMIT_CPU to RLIMIT_NOFILE, and then running it like this:

prlimit <PID> 65535 65535

The <PID> argument is taken from the first number in the log messages from .xsession-errors – in the example above, it’s 22161.

And now, you can go back to using your tabs as ersatz bookmarks, like I do.


First Step with Clojure: Terror

Posted: Thu, 24 July 2014 | permalink | 6 Comments

$ sudo apt-get install -y leiningen
[...]
$ lein new scratch
[...]
$ cd scratch
$ lein repl
Downloading: org/clojure/clojure/1.3.0/clojure-1.3.0.pom from repository central at http://repo1.maven.org/maven2
Transferring 5K from central
Downloading: org/sonatype/oss/oss-parent/5/oss-parent-5.pom from repository central at http://repo1.maven.org/maven2
Transferring 4K from central
Downloading: org/clojure/clojure/1.3.0/clojure-1.3.0.jar from repository central at http://repo1.maven.org/maven2
Transferring 3311K from central
[...]

Wait… what? lein downloads some random JARs from a website over HTTP1, with, as far as far I can tell, no verification that what I’m asking for is what I’m getting (has nobody ever heard of Man-in-the-Middle attacks in Maven land?). It downloads a .sha1 file to (presumably) do integrity checking, but that’s no safety net – if I can serve you a dodgy .jar, I can serve you an equally-dodgy .sha1 file, too (also, SHA256 is where all the cool kids are at these days). Finally, jarsigner tells me that there’s no signature on the .jar itself, either.

It gets better, though. The repo1.maven.org site is served by the fastly.net2 pseudo-CDN3, which adds another set of points in the chain which can be subverted to hijack and spoof traffic. More routers, more DNS zones, and more servers.

I’ve seen Debian take a kicking more than once because packages aren’t individually signed, or because packages aren’t served over HTTPS. But at least Debian’s packages can be verified by chaining to a signature made by a well-known, widely-distributed key, signed by two Debian Developers with very well-connected keys.

This repository, on the other hand… oy gevalt. There are OpenPGP (GPG) signatures available for each package (tack .asc onto the end of the .jar URL), but no attempt was made to download the signatures for the .jar I downloaded. Even if the signature was downloaded and checked, there’s no way for me (or anyone) to trust the signature – the signature was made by a key that’s signed by one other key, which itself has no signatures. If I were an attacker, it wouldn’t be hard for me to replace that key chain with one of my own devising.

Even ignoring everyone living behind a government- or company-run intercepting proxy, and everyone using public wifi, it’s pretty well common knowledge by now (thanks to Edward Snowden) that playing silly-buggers with Internet traffic isn’t hard to do, and there’s no shortage of evidence that it is, in fact, done on a routine basis by all manner of people. Serving up executable code to a large number of people, in that threat environment, with no way for them to have any reasonable assurance that code is trustworthy, is very disappointing.

Please, for the good of the Internet, improve your act, Maven. Putting HTTPS on your distribution would be a bare minimum. There are attacks on SSL, sure, but they’re a lot harder to pull off than sitting on public wifi hijacking TCP connections. Far better would be to start mandating signatures, requiring signature checks to pass, and having all signatures chain to a well-known, widely-trusted, and properly secured trust root. Signing all keys that are allowed to upload to maven.org with a “maven.org distribution root” key (itself kept in hardware and only used offline), and then verifying that all signatures chain to that key, wouldn’t be insanely difficult, and would greatly improve the security of the software supply chain. Sure, it wouldn’t be perfect, but don’t make the perfect the enemy of the good. Cost-effective improvements are possible here.

Yes, security is hard. But you don’t get to ignore it just because of that, when you’re creating an attractive nuisance for anyone who wants to own up a whole passel of machines by slipping some dodgy code into a widely-used package.


  1. To add insult to injury, it appears to ignore my http_proxy environment variable, and the repo1.maven.org server returns plain-text error responses with Content-Type: text/xml. But at this point, that’s just icing on the shit cake.

  2. At one point in the past, my then-employer (a hosting provider) blocked Fastly’s caching servers from their network because they took down a customer site with a massive number of requests to a single resource, and the incoming request traffic was indistinguishable from a botnet-sourced DDoS attack. The requests were coming from IP space registered to a number of different ISPs, with no distinguishing rDNS (184-106-82-243.static.cloud-ips.com doesn’t help me to distinguish between “I’m a professionally-run distributed proxy” and “I’m a pwned box here to hammer your site into the ground”).

  3. Pretty much all of the new breed of so-called CDNs aren’t actually pro-actively distributing content, they’re just proxies. That isn’t a bad thing, per se, but I rather dislike the far-too-common practice of installing varnish (and perhaps mod_pagespeed, if they’re providing “advanced” capabilities) on a couple of AWS instances, and hanging out your shingle as a CDN. I prefer a bit of truth in my advertising.


Per-repo update hooks with gitolite

Posted: Wed, 23 July 2014 | permalink | 3 Comments

Gitolite is a popular way to manage collections of git repositories entirely from the command line – it’s configured using configuration stored in a git repo, which is nicely self-referential. Providing per-branch access control and a wide range of addons, it’s quite a valuable system.

In recent versions (3.6), it added support for configuring per-repository git hooks from within the gitolite-admin repo itself – something which previously required directly jiggering around with the repo metadata on the filesystem. It allows you to “chain” multiple hooks together, too, which is a nice touch. You can, for example, define hooks for “validate style guidelines”, “submit patch to code review” and “push to the CI server”. Then for each repo you can pick which of those hooks to execute. It’s neat.

There’s one glaring problem, though – you can only use these chained, per-repo hooks on the pre-receive, post-receive, and post-update hooks. The update hook is special, and gitolite wants to make sure you never, ever forget it. You can hook into the update processing chain by using something called a “virtual ref”; they’re stored in a separate configuration directory, use a different syntax in the config file, and if you’re trying to learn what they do, you’ll spend a fair bit of time on them. The documentation describes VREFs as “a mechanism to add additional constraints to a push”. The association between that and the update hook is one you get to make for yourself.

The interesting thing is that there’s no need for this gratuitous difference in configuration methods between the different hooks. I wrote a very small and simple patch that makes the update hook configurable in exactly the same way as the other server-side hooks, with no loss of existing functionality.

The reason I’m posting it here is that I tried to submit it to the primary gitolite developer, and was told “I’m not touching the update hook […] I’m not discussing this […] take it or leave it”. So instead, I’m publicising this patch for anyone who wants to locally patch their gitolite installation to have a consistent per-repo hook UI. Share and enjoy!


Doing Password Complexity Wrong

Posted: Tue, 8 July 2014 | permalink | 2 Comments

I just made an account on yet another web service. On the suggestion of my password manager, I attempted to use the password “W:9[$X*F”. It was rejected because “Password must contain at least one non-alphabet character, one lowercase letter, one uppercase letter”. OK, how about “Passw0rd”? Yep, that’s fine.

Anyone want to guess which of those two passwords is going to fall victim to a brute-force attack first? Go on, don’t be shy, take a wild shot in the dark!


Witness the security of this fully DNSSEC-enabled zone!

Posted: Sun, 6 July 2014 | permalink | No comments

After dealing with the client side of the DNSSEC puzzle last week, I thought it behooved me to also go about getting DNSSEC going on the domains I run DNS for. Like the resolver configuration, the server side work is straightforward enough once you know how, but boy howdy are there some landmines to be aware of.

One thing that made my job a little less ordinary is that I use and love tinydns. It’s an amazingly small and simple authoritative DNS server, strong in the Unix tradition of “do one thing and do it well”. Unfortunately, DNSSEC is anything but “small and simple” and so tinydns doesn’t support DNSSEC out of the box. However, Peter Conrad has produced a patch for tinydns to do DNSSEC, and that does the trick very nicely.

A brief aside about tinydns and DNSSEC, if I may… Poor key security is probably the single biggest compromise vector for crypto. So you want to keep your keys secure. A great way to keep keys secure is to not put them on machines that run public-facing network services (like DNS servers). So, you want to keep your keys away from your public DNS servers. A really great way of doing that would be to have all of your DNS records somewhere out of the way, and when they change regenerate the zone file, re-sign it, and push it out to all your DNS servers. That happens to be exactly how tinydns works. I happen to think that tinydns fits very nicely into a DNSSEC-enabled world. Anyway, back to the story.

Once I’d patched the tinydns source and built updated packages, it was time to start DNSSEC-enabling zones. This breaks down into a few simple steps:

  1. Generate a key for each zone. This will produce a private key (which, as the name suggests, you should keep to yourself), a public key in a DNSKEY DNS record, and a DS DNS record. More on those in a minute.

    One thing to be wary of, if you’re like me and don’t want or need separate “Key Signing” and “Zone Signing” keys. You must generate a “Key Signing” key – this is a key with a “flags” value of 257. Doing this wrong will result in all sorts of odd-ball problems. I wanted to just sign zones, so I generated a “Zone Signing” key, which has a “flags” value of 256. Big mistake.

    Also, the DS record is a hash of everything in the DNSKEY record, so don’t just think you can change the 256 to a 257 and everything will still work. It won’t.

  2. Add the key records to the zone data. For tinydns, this is just a matter of copying the zone records from the generated key into the zone file itself, and adding an extra pseudo record (it’s all covered in the tinydnssec howto).

  3. Publish the zone data. Reload your BIND config, run tinydns-sign and tinydns-data then rsync, or do whatever it is PowerDNS people do (kick the database until replication starts working again?).

  4. Test everything. I found the Verisign Labs DNSSEC Debugger to be very helpful. You want ticks everywhere except for where it’s looking for DS records for your zone in the higher-level zone. If there are any other freak-outs, you’ll want to fix those – because broken DNSSEC will take your domain off the Internet in no time.

  5. Tell the world about your DNSSEC keys. This is simply a matter of giving your DS record to your domain registrar, for them to add it to the zone data for your domain’s parent. Wherever you’d normally go to edit the nameservers or contact details for your domain, you probably want to do to the same place and look for something about “DS” or “Domain Signer” records. Copy and paste the details from the DS record in your zone into there, submit, and wait a minute or two for the records to get published.

  6. Test again. Before you pat yourself on the back, make sure you’ve got a full board of green ticks in the DNSSEC Debugger. if anything’s wrong, you want to rollback immediately, because broken DNSSEC means that anyone using a DNSSEC-enabled resolver just lost the ability to see your domain.

That’s it! There’s a lot of complicated crypto going on behind the scenes, and DNSSEC seems to revel in the number of acronyms and concepts that it introduces, but the actual execution of DNSSEC-enabling your domains is quite straightforward.