Sooooo...let's see. From what I am hearing, the best way to improve the situation with Linux for Linux enthusiasts like myself is to change the market. Converting as many individuals who are likely to be satisfied with Linux is a good start. That way, there will be less Windows/Mac market share which would reduce the benefit to firms like ATI to keep certain parts of their drivers closed.
And seduce them from Windows 7 how? You'll be telling them that the extra cash they shelled out for a discrete card will only get them 70% of the capabilities of a Windows driver, unless of course, they can convince 100 of their friends to switch. Then, only then, will ATI respond with a driver that meets or exceeds the Windows driver.
if i were to redo KMS today there is few major things i would change.
If and when there will be a comprehensive radeon KMS documentation, this stuff would probably deserve to be included. As in, post-implementation thoughts, how could it be better if someone ever wants to go back at it.
I think the best way to improve things would be to find out how to attract more individual developers and companies to help in development of Mesa and Gallium, or fund such development (and/or doing any of these yourself).
Perhaps some cool tutorial on how to make a game work by implementing a new OpenGL extension for Mesa, or adding a feature to r600g or nv50g would help in that (obviously, it would be retrospective analysis on something already added).
Another interesting avenue could be attempting to add OpenCL support to server applications.
If significant gains can be demonstrated, good GPU drivers would become a must in the Linux server market, which is very significant.
I'm not sure however whether this can be done for a significant share of server applications.
And finally, getting third parties (or even employees of the GPU company itself, but that's harder) with NDA access to specifications and/or proprietary code to leak them will also obviously help (especially for very unfriendly companies like nVidia and Imagination).
Longer term, a move to Larrabee-style dumb GPUs with stable ISAs (and "graphics-on-compute" programming models) will likely make driver writing much easier and allow sharing much more between GPUs, hopefully obsoleting ad-hoc proprietary driver stacks.
Oh, and it would be nice if AMD could somehow speed up the release of their Evergreen Mesa code (and ideally r700->evergreen 3D register differences too), so people who bought recent cards can actually contribute...