Friday, December 2, 2011

A Qt hat, isn't it?

Fedora users may have noticed that F16 ships with Qt 4.8-RC. While I'm not sure what the rush was, it did cause an interesting problem when it came to the cygwin-qt cross-compiling package.

Building anything based on Qt requires more than just headers and libraries; Qt's object system requires code to be processed by moc, which generates additional C++ code for signal/slot handling and the like. There is also a number of other tools which "compile" various types of data into C++ code. I had been relying on the native tools in qt-devel to do this, since the generated code is not arch-specific.

The problem here is that moc-generated code is very internal-specific; code generated with one version of Qt will not compile with headers from another version. Since the version of cygwin-qt needs to match that of the Cygwin (Ports) qt4 package, and I have no plans of updating that to 4.8 until its proven stable in conjunction with KDE, this created a bit of an impasse. There was another issue with relying on qt-devel, in that qmake requires patches to build libraries or plugins correctly on Cygwin, which limited cygwin-qt cross-compiling to CMake-based packages.

The solution to all this was to change the cygwin-qt-qmake package, which had previously just included qmake mkspecs, to provide native-built tools from the same version of Qt as cygwin-qt and with the necessary patches. While qmake/moc/rcc/uic are always statically linked with the necessary Qt code, the linguist, qdbus, and qt3support build tools are linked against the Qt libraries; for now, they are also statically linked with their Qt prerequisites.

The updated cygwin-qt and cygwin-qt-qmake are now available for Fedora 14 and up on both arches, along with an updated cygwin-filesystem to support these changes. Even better, thanks to a patched qmake, I was also able to add cygwin-qscintilla and cygwin-qwt packages.

Thursday, December 1, 2011

Spicy Vinagre

One of the more eventful version bumps during the GNOME 3.2 update was with Vinagre, the Remote Desktop Viewer, which supports a number of protocols via plugins.  These plugins used to be managed by libpeas, a GNOME framework which makes it easier to develop and manage plugins.  However, in the case of Vinagre, it was a bit of an overkill, as the only plugins that were needed were those that shipped with Vinagre itself.

So, during the 3.2 development cycle, the devs decided to replace libpeas with their own VinagreStaticExtension class. In theory, a sensible move, but they made one major mistake: they never added any code to actually load the plugins!

Now that got your attention, didn't it?

Of course, if you try vinagre 3.2 on Linux, you may find that it works perfectly (but we'll get back to that).  But only because they used the most unportable, Linux/ELF-centric hack that I've ever seen (and trust me, I've seen my fair share of them).  It took me a while to figure it out at first, but here is what they did:
  1. the plugin init functions are marked __attribute__((constructor)), a GCC extension which causes the indicated function to be called automatically;
  2. the plugins rely on symbols in the main binary, of course, but furthermore...
  3. the executable also relied on symbols in the vnc plugin, and to top it off...
  4. the executable is linked against the plugins!! (If you don't think you misread that, then try again.)
In short, the only reason that this worked is that, by linking the plugins into the executable regardless of symbol dependency, the ELF runtime loader would load the plugin as a runtime dependency, at which time the plugin init function would be called due to being marked as a ctor.

To be fair, technique #1 is also used by LADSPA plugins, and #2 isn't uncommon and can be made to work even on PE platforms. However, #3, while theoretically possible, is impractical with plugins due to a lack of rpath in PE linkage, and relying on #4 wouldn't work at all because with PE only those DLLs whose symbols are required are hard-coded as runtime dependencies. And here's the kicker: even on ELF platforms, this doesn't work if linked with -Wl,--as-needed.

Fortunately, the fix is fairly easy: make the plugins static instead, link them into the executable as before, and actually call the ctors in main(). So not only is vinagre 3.2 working and available with the rest of GNOME 3.2 in Ports, but while I was at it, I added SPICE protocol support as well.