Re: Memory mapped IO in video drivers

On Tue, 12 Jul 2005 [email protected] wrote:

> I have an xFree86 video driver for a custom gfx board, and am running into a
> problem with the memory mapped IO.  The display portion of the driver works
> just fine, however we are also trying to support I2C/DDC.

> The quick explanation of the problem is that the IO address range to the
> video registers we need for I2C is getting mapped in and out multiple times
> during the X session.  Sometimes the registers are there, other times they
> aren't (i.e., we read 0xFF for all registers instead of valid data).

> The problem occurs with all of the mapped registers, not just the I2C ones,
> but the IO space always seems to be mapped in properly during accelerated
> drawing functions, it's just when we try to access I2C through our back door
> that it is rather unpredictable.

> The pointer to the address range itself is not changing, just the validity
> of the data.  Also, it doesn't appear that the BAR for the IO registers is
> being moved.

> The registers are being mapped in during ScreenInit with a call to
> xf86MapPciMem() with the VIDMEM_MMIO_32BIT flag.  Our code to unmap the
> registers with xf86UnMapVidMem() doesn't get called until the X session is
> closed.

> So, is there some part of the XFree86 server that would be mapping and
> unmapping the address space behind my back?  How can I indicate to the X
> server that this memory is non-relocatable and has to be permenant for the
> duration of the X session?

> Also, I have noticed that the memory mapped IO gets unmapped whenever I
> temporarily switch out of the X server to a virtual terminal in VGA mode,
> and then the IO gets mapped in again when I switch back to the X server.  I
> know our driver code for EnterVT() and LeaveVT() isn't doing this, so where
> does this happen in the server code?  Perhaps it could be a related issue?

> This is a multiple board configuration (four monitors, two per board), so
> perhaps its not something that has really been tested heavily before?

> BTW, if it matters, the video driver is being loaded as a module, not
> compiled into the kernel.

> Any help or tips here would be appreciated.  I have already been perusing
> the only XF86 video driver resource I could find at
> http://xfree86.activeventure.org/design/DESIGN.html
> <http://xfree86.activeventure.org/design/DESIGN.html> .

It sounds like your driver isn't sharing VGA properly.  For examples, look 
for calls to xf86SetOperatingState() in other drivers.  This function is 
documented in xc/programs/Xserver/hw/xfree86/RAC.Notes in the source.

Marc.

+----------------------------------+-----------------------------------+
|  Marc Aurele La France           |  work:   1-780-492-9310           |
|  Academic Information and        |  fax:    1-780-492-1729           |
|    Communications Technologies   |  email:  [email protected]          |
|  352 General Services Building   +-----------------------------------+
|  University of Alberta           |                                   |
|  Edmonton, Alberta               |     Standard disclaimers apply    |
|  T6G 2H1                         |                                   |
|  CANADA                          |                                   |
+----------------------------------+-----------------------------------+
XFree86 developer and VP.  ATI driver and X server internals.