[Dri-devel] Re: Some DRM related changes for running multiple i810/830M X servers

[David D. <dawes@XFree86.Org>]

I meant to add that it'd be good if something along the lines of Michel
D=E4nzer's work to allow DRI to be reinitialised at VT switch was added i=
n
after 4.3.  I'm curious why it didn't made it into the DRI trunk already.

David

On Fri, Feb 14, 2003 at 04:17:16PM -0500, David Dawes wrote:
>Egbert and I have been looking into the issues that are preventing a sec=
ond
>X server to be started for i810/830M platforms when DRI is enabled.  Sev=
eral
>issues were found:
>
>  1. The i810 support doesn't unbind/release the agpgart module when VT
>     switching away, and this prevents a second X server from acquiring
>     it.  Since the i810 driver requires agpgart access even for 2D
>     operation, this is prevents a second X server from running.  A fix
>     for this is to add the unbind/release calls at LeaveVT, and
>     acquire/rebind at EnterVT.  The attached patch does this.  It also
>     makes a small change to the unbind code in the DRM kernel modules
>     to handle this.
>
>  2. The 830M support does unbind/release the agpgart module, but code
>     in DRM(release) called when closing a /dev/dri/* device blocks
>     until it can obtain the lock.  Since the first server holds the
>     lock while switched away, the second one can never get it, so it
>     ends up blocking in close().  The second server opens/closes the
>     devices to find out whether DRI can be enabled.  DRI can't be
>     enabled on the second server (which isn't a problem), but this
>     blocking behaviour is preventing it from starting up at all.  I've
>     found that this can be solved by keeping track of whether the opene=
r
>     has ever tried to acquire the lock, and not try to acquire it at
>     close/release when it had never previously been acquired.  The patc=
h
>     below implements this.
>
>  3. The drm module AGP support code keeps track of a "handle" for
>     allocated AGP regions.  It uses the memory address returned from
>     the allocation for this purpose.  This would normally be unique,
>     but for the i810 driver case where "DCACHE" memory is "allocated".
>     In the DCACHE case, the allocated memory is freed immediately (sinc=
e
>     DCACHE doesn't come from the system memory pool), and the next
>     allocation often ends up getting the same memory address, and thus
>     the same "handle".  This showed up as a problem when the unbind/reb=
ind
>     code was added.
>
>     The user-level agpgart interfaces use a "key" value to uniquely
>     identify allocated AGP regions.  I don't know why the DRM module
>     doesn't do the same, since the key is guaranteed to be unique.
>     The patch below changes the handle to be the "key" value.
>
>I'm wary of making changes like this so close to the 4.3 release, but
>I would like to see this problem fixed in 4.3.  Does anyone see any
>potential problems with the attached patch?
>
>David
>--=20
>David Dawes
>Release Engineer/Architect                      The XFree86 Project
>www.XFree86.org/~dawes