torstai 15. kesäkuuta 2017

Of XML, GUIDs and installers

Every now and then I have to use Windows. I'd really like to use Emacs for editing so I need to install it. Unfortunately the official releases from the FSF are zip archives that you need to manually unpack, add shortcuts and all that stuff. This gets tedious and it would be nice if there were MSI installer packages to do all that for you. So I figured I'd create some.

Conceptually what the installer needs to do is the following:

  1. Create the c:\Program Files\GNU Emacs directory
  2. Unzip contents of the official release zip file in said directory
Seems simple, right? There are two main tools for creating installers. The first one is NSIS, but it only creates exe installers, not msi. It also has a scripting language designed by a shaman who has licked a few frogs too many. So let's ignore that.

The other tool is WiX. It creates nice installers but it is Enterprise. Very, Very Enterprise. And XML. But mostly Enterprise. For starters the installer needs to have a GUID (basically a 128 bit random number). It also needs an "upgrade GUID". And a "package GUID". The first two must be the same over all installer versions but the latter must not be.

Having conjured the necessary amount of GUIDs (but not too many), you are ready to tackle copying files around. As you probably guessed, each one of them needs its own GUID. But if you though that each one would require an XML element of their own, you are sadly mistaken. Every file needs two nested XML elements. The files also need a container. And a container-container.

Did I mention that the documentation consists almost entirely of XML fragments? So it is all but impossible to tell which tag should follow which and which ones should be nested?

The Emacs distribution has a lot of files, so obviously writing the XML by hand is out of the question. Unfortunately the WiX toolkit ships a helper utility called heat to generate the XML from a directory tree. Yes, I really did say unfortunately. While the script pretends to work in reality it does not. Following the documentation you might try doing something like this (simplified for purposes of exposition):

heat <directory with unpacked files> <output xml file> other_flags

This does create an installer which installs all the files. But it also does a little bit extra. If your unpack directory was called unpack, then the files will be installed to c:\Program Files\GNU Emacs\unpack. How might we get rid of that extra directory segment? The correct answer is that the system is optimized for in-source Visual Studio builds and trying to do anything else is doomed to fail. But let's try it anyway.

First one might look for a command line switch like -P1 for patch to discard the first path segment. It does not seem to exist.

Next one might try to be clever and cd inside the unpack dir and do this (again simplified):

heat . <output xml file> other_flags

The system reports no error but the output is identical to the previous attempt. The script helpfully notices that you are using a period for the directory name and will do a lookup in the parent directory to see what it would be called and substitutes it in. Because!

Since there are only a few directories in the top level one might try something along the lines of:

heat dir1 dir2 dir3 dir4 <output xml file> other args

Which again succeeds without error. However the output installer only has files from dir1. The tool parses all the input and then dutifully throws away all entries but the first without so much as a warning.

The way to make this work is to generate the XML files and then monkey patch all paths in them before passing them on to the installer generator. For That Is the Enterprise Way! Just be glad there are no files at the root directory.

Join us next time to learn how a randomly generated daddy GUID comes together with a randomly generated mommy GUID to produce a new entry in the start menu.

Is this available somewhere?

The script is here. There are no downloadable binaries because I don't have the resources to maintain those. It would be cool if someone (possibly even the FSF) would provide them.

5 kommenttia:

  1. I use chocolatey for installing software on Windows. Emacs is just "choco install emacs".

    Lately I've actually been using WSL + XMing instead. :D

    1. Package managers are nice and all but in many cases I don't want one (or have to choose which one). All I want is a single installer for a single appl.

  2. If you would prefer to create the .msi on linux you can use msitools.

    1. That's really cool. It still uses the WiX XML format dumpster fire, though.

  3. A few years ago I picked WiX-generated MSI files for distributing some software on Windows, and I still think it's the least-awful solution. I ended up writing my own Python script to do the work of heat, to avoid precisely the problems you document here. It was actually rather nice once it all worked. If you really want your eyebrows raised, look into the details of the MSI format! I was amazed to discover that they are not too dissimilar from Microsoft Word documents. The designers made the fascinating decision to reuse the fields that normally display things like the word count in Explorer to store properties used by the Windows Installer service; for instance, the Word Count field becomes a bitfield of flags...