> Moving the MD parts into the kernel would also have the side benefit
 > of allowing us to have a single X server for all platforms (cpus,
 > graphics chips, etc).
 > 
 > The problems that were pointed out to me include:
 > 1) Doing a system call every time you want to draw a line or set a 
 >    pixel or whatever would be very very slow. 
 > 2) It would be very hard to come up with an API that is general enough to
 >    not require special casing in applications but at the same time would
 >    allow for the use of special features of some video cards.

I think the problems aren't so hard to solve if you gave up 
accelerations. The basic idea is 'damn frame buffer' which
you can mmap from user process.
I'm trying this. Please see sys/arch/hpcmips/dev/hpcfbio.h.
(it is not yet complete.)

Takemura