I don't know if this is specifically a problem with the AGP driver or
with X, but basically the AGP driver detects an aperture base address
different from that of the prefetchable memory area of the video card
and the mtrr's are not in agreement.

I'm using this with a Dual P3-500mhz and a Matrox 400 under XFree86
4.0.1 with the Matrox supplied driver. Both X and /proc/pci report the
same frame buffer addresses:

Bus  1, device   0, function  0:
  VGA compatible controller: Matrox Graphics, Inc. MGA G400 AGP (rev 5).
    IRQ 16.
    Master Capable.  Latency=64.  Min Gnt=16.Max Lat=32.
    Prefetchable 32 bit memory at 0xf6000000 [0xf7ffffff].
    Non-prefetchable 32 bit memory at 0xfeafc000 [0xfeafffff].
    Non-prefetchable 32 bit memory at 0xfe000000 [0xfe7fffff].

But agpgart sets up:

agpgart: Maximum main memory to use for agp memory: 440M
agpgart: Detected Intel 440BX chipset
agpgart: AGP aperture is 32M @ 0xfa000000

Needless to say, the two disagree and direct rendering is disabled.
Attached is my .config for 2.4.0-test11-pre7. I've been wading through
the code for AGP and DRM support, but nothing jumps out at me.

Attached is my .config. The hardware is:

Dual P3-500 mhz
Intel 440bx AGP set
512mb of RAM
Matrox G400 dual head card
XFree86 4.0.1
Matrox driver 1.00.003 (beta)
(http://www.matrox.com/mga/support/drivers/latest/home.htm)

This is with X configured to run multi-head, not Xinerama and the kernel
built with egcs 1.1.2.

Also, I noticed that in arch/i386/kernel/mtrr.c around line 1888 that
the mtrr_add call for Intel chips falls through to the next case. When
that happens, X cannot register the mtrr for the main display on the
Matrox card, though it can register the secondary display. I slipped a
break in there to bypass the fall through and the mtrr registered (with
performance improvement). Is the fall through correct? I haven't had any
failures or lockups since putting in the bypass, but I suspect that it
is incorrect and related to the misdetection of the base address of the
video card.

John