Gone Horribly Wrong

While it’s true that in previous occasions I have switched to Debian testing awaiting the worst to happen to my computer and at the end it’s all been for the better minus the unavoidable annoyance of having to rebuild some of the software I run without the blessings of a package manager, this time the switch from Debian GNU/Linux 6.0 “squeeze” to “wheezy” (current Testing) surpassed my expectations.

I knew that something would go wrong but I expected it to be a temporary issue of the kind “oops, packager screwed up this build, uploading a fix to Sid shortly”. However, for all I know it’s a more complicated issue that just hit me in the face, and it’s called X.org server 1.10.

Stable uses X.org server 1.7.7, which is well tested and stable and worked well with my hideous mix of newer libraries from upstream (Mesa and libdrm from git master, xf86-video-intel 2.15.0) and in fact provided top-notch performance despite its age.

Some time around the end of 2010 while I was still using Bluecore, I switched to X.org 1.9.x from Experimental, of all things, and didn’t cause any problems with my ATI R6xx-based configuration. I was an idiot and switched to Debian testing now based on my previous experience with X.org 1.9.x instead of considering that the second version number is not attached to bug fixes but features. (A Wesnoth developer is supposed to know better, I know. I’m an idiot and I deserve this.)

Now the jump from Debian Stable to Testing was particularly rough because it’s left me swamped with a lot of bugs that are difficult to describe and leave me in the uneasy position where I’m unable to discern exactly what I should report and to whom.

Following the switch, I restarted X without rebooting and logged into KDE SC 4.4, noticing some annoying issues:

  • Kwin’s compositing became glitchy, with menu shadows being partially or completely missing at random.
  • Huge black rectangles or random garbage lines started appearing in place of popup menus or combo box popups for an instant; while the menus/popups are correctly drawn at the end (bar their shadows) the effect is pretty nauseating.
  • Cairo/Gtk2 applications’ performance went down the sink with compositing enabled.
  • Frogatto and other GL clients no longer run at near max framerate when compositing is enabled. (Uncannily, this also affects Wesnoth, a pure-software SDL client which I previously had to run under a compositing WM to make it playable at all on Bluecore!)

That was with X.org server 1.10.1. In rage, I grabbed version 1.10.2 from Sid, and KDE SC 4.6. Neither solved it.

I’d like to blame the X.org server itself since using Mesa, libdrm and xf86-video-intel from Debian did not help either. Neither did switching from Linux 2.6.38.8 to 2.6.39.1. All signs point to an issue that’s completely out of my control as a user. To make sure this was the case, I asked Ivanovic in the Wesnoth devs IRC channel about it. For the record, he uses Gentoo and the open-source Radeon stack on an ATI R6xx family GPU, while here I use the Intel stack on a GM45 IGP.

Me: are you using X.org server 1.10.x and KDE SC 4.6 there? do you see some garbage beneath the areas where a window will be drawn when compositing is enabled and very awful gtk2 drawing performance?

Me: (garbage that goes away once the window and its shadow are actually drawn)

Ivanovic: no idea about gtk2 drawing performance

Ivanovic: but yeah to the rest

The drawing glitches are only slightly more bearable when using QtCurve in place of Oxygen and Oxygen-gtk. Using Fluxbox helps with the Cairo/Gtk2 performance loss (which impacts GIMP hard) but not with the menu pre-paint garbage glitch, which seems to not be tied to any window manager, Gtk2 engine or client in particular.

So yeah, I got KDE SC 4.6 as I wanted. Is it an improvement over 4.4? Perhaps — whatever Kwin optimizations have been done in the mean time are pretty much irrelevant right now. Plasma is very dependent upon a graphics server that does shit right, and this is not one of those right now. The new (4.5?) notifications are nice, though.

Of course I really regret the idea of switching to Testing at this point, but I somehow can’t bring myself to nuke it and restore Stable from the pre-upgrade backup snapshot that I still have in my external hard disk. Perhaps I am masochist after all, or I have hopes that this will be solved in the next version of whatever component regressed so hard. Just… don’t try this at home!

Hardware, Software