Bug 4797

Reassigning to myself.

Created attachment 3573 [details] [review]
Almost fix AGP problems in MGA DRM.

Converts the remaining drm_agp_foo functions to be a drm_agp_foo and
drm_agp_foo_ioctl pair.  Modifies the MGA DRM to use the drm_agp_foo functions
instead of the drm_foo_agp functions.  The drm_foo_agp functions are no longer
exported by drm.ko.

The bsd-core side of things will also need these (or similar) changes applied.

When this patch was applied (with some conflicts) to a '-D "20050804 23:59
GMT"' DRM check out, everything worked correctly.  However, against current
CVS, everything seems to work fine in the kernel, but I get the following error
message in Xorg.0.log:

(EE) MGA(0): [drm] Failed to map DMA buffers list

Adding 'Option "ForcePciDma" "TRUE"' to xorg.conf with attachment #3573 [details] [review] causes
direct render to be enabled.  However, the glxgears window is all black.  There
are obviously still some problems.

Created attachment 3586 [details] [review]
Almost fix AGP problems in MGA DRM, take 2.

This patch includes everything of the previous patch, but it also makes sure
that dma->seg_count and dma->page_count are properly set in
drm_addbufs_{agp,sg,fb}.  drm_addbufs_pci was already correct.

Direct rendering is now enabled in the AGP case, but it's still not right.  The
Linux VM *kills* any direct rendering apps that run.

Created attachment 3587 [details]
/var/log/messages with attachment #3586 [details] [review]

Output of /var/log/messages of starting X and running glxgears when attachment
#3586 [details] [review] is applied.

(In reply to comment #0)
> Unbeknownst to me, the MGA DRM has been broken since August 4th (yes, over two
> months).

Vanilla kernels > 2.6.12 crashes here xorg (6.8.2) and Matrox G400 always: "EE
MGA(0): [dri] Idle timed out, resetting engine...". The crash occurs if glxgears
is executed. Exec of glxinfo doesn't crash xorg and shows direct rendering as
enabled.

Currently I have to switch back to kernel 2.6.11.12, this one works with the
same kernel config/xorg combination without any problem. Is there any patch on
top of 2.6.14?

Regards from Germany
                   Matthias

It looks like the commit of version 1.65 of linux-core/drm_bufs.c is responsible
for breaking PCI MGA.

(In reply to comment #6)
> (In reply to comment #0)
> > Unbeknownst to me, the MGA DRM has been broken since August 4th (yes, over two
> > months).
> 
> Vanilla kernels > 2.6.12 crashes here xorg (6.8.2) and Matrox G400 always: "EE
> MGA(0): [dri] Idle timed out, resetting engine...". The crash occurs if glxgears
> is executed. Exec of glxinfo doesn't crash xorg and shows direct rendering as
> enabled.
> 
> Currently I have to switch back to kernel 2.6.11.12, this one works with the
> same kernel config/xorg combination without any problem. Is there any patch on
> top of 2.6.14?

Can you try 2.6.14? I fixed a number of issues with MGA before 2.6.14 went out..
but not with MGA PCI.

Dave.

(In reply to comment #8)
> (In reply to comment #6)
> > (In reply to comment #0)
> > > Unbeknownst to me, the MGA DRM has been broken since August 4th (yes, over two
> > > months).
> > 
> > Vanilla kernels > 2.6.12 crashes here xorg (6.8.2) and Matrox G400 always: "EE
> > MGA(0): [dri] Idle timed out, resetting engine...". The crash occurs if glxgears
> > is executed. Exec of glxinfo doesn't crash xorg and shows direct rendering as
> > enabled.
> > 
> > Currently I have to switch back to kernel 2.6.11.12, this one works with the
> > same kernel config/xorg combination without any problem. Is there any patch on
> > top of 2.6.14?
> 
> Can you try 2.6.14? I fixed a number of issues with MGA before 2.6.14 went out..
> but not with MGA PCI.

Dave...

...yes, because I have seen these changes in the Changelog I've tried 2.6.14
last week: Without distinction :-( The G400 card resides in an AGP-Slot, using
IRQ 11, the kernel is compiled for SMP always, the server comes with two PIII
CPU's at 1,00 gHz per CPU. IO-APIC is used.

$ cat /proc/interrupts says on kernel 2.6.11.12 (as I said, no probs there with
DRI):
           CPU0       CPU1
  0:     773469     751432    IO-APIC-edge  timer
  1:       1411       1115    IO-APIC-edge  i8042
  2:          0          0          XT-PIC  cascade
  8:          1          0    IO-APIC-edge  rtc
  9:          0          0    IO-APIC-edge  acpi
 10:        732         84   IO-APIC-level  EMU10K1
 11:     149525          1   IO-APIC-level  eth0, mga@pci:0000:01:00.0
 12:      21780      18458    IO-APIC-edge  i8042
 14:     101359      96633   IO-APIC-level  aic7xxx, uhci_hcd, uhci_hcd, uhci_hcd
 15:          6          1   IO-APIC-level  tmscsim
NMI:          0          0
LOC:    1524674    1524700
ERR:          0
MIS:          0

Regards
      Matthias

(In reply to comment #4)

> Direct rendering is now enabled in the AGP case, but it's still not right.  The
> Linux VM *kills* any direct rendering apps that run.

I've tracked this down to mga_draw_triangle (in
src/mesa/drivers/dri/mga/mgatris.c).  After calling mgaAllocDmaLow, the pointer
returned is not accessable by the process.

After I discovered that, I verified that each of
mmesa->mgaScreen->bufs->list[].address was not accessable.  I thought the whole
point of calling drmMapBufs was to get accessable pointers.

This *only* happens if the buffers are in AGP memory.  If I put back the
virt_to_bus call in the kernel, PCI mode works perfectly.

So now I'm not sure whether the problem lies in the MGA DRM or in the
client-side driver (or the DDX?).  My guess is that DRM isn't doing something
that it should do.  Suggestions?

(In reply to comment #10)

> So now I'm not sure whether the problem lies in the MGA DRM or in the
> client-side driver (or the DDX?).  My guess is that DRM isn't doing something
> that it should do.  Suggestions?

The problem is definitely in the kernel somewhere.  If I add 'Option
"OldDmaInit" "TRUE"' to my xorg.conf everything works correctly.  There's
obviously a step being missed in the kernel that used to be done by the DDX.  Ugh...

Created attachment 3685 [details] [review]
Acquire AGP and allocate AGP memory, take 3

This patch does everything that patch #3586 does, but it also makes sure that
mga_do_agp_dma_bootstrap correctly sets agp_buffer_token.  This seems to fix
all the problems with the MGA DRM on AGP cards on Linux.

Are similar changes needed for BSD?

I have committed attachment #3685 [details] [review], and I am closing this bug.  I will submit
another bug for the PCI DMA issue.