Re: How to get the physical address of a DRM_SHM map

[Felix <fx...@gm...>]

Am Fr, den 15.10.2004 schrieb Keith Whitwell um 15:50:
> Felix K=FChling wrote:
> > Am Fr, den 15.10.2004 schrieb Vladimir Dergachev um 5:27:
> >=20
> >>On Fri, 15 Oct 2004, Felix [ISO-8859-1] K=FChling wrote:
> >>
> >>
> >>>Am Fr, den 15.10.2004 schrieb Jon Smirl um 3:58:
> >>>
> >>>>On Fri, 15 Oct 2004 03:55:37 +0200, Felix K=FChling <fx...@gm...> =
wrote:
> >>>>
> >>>>>I'm trying to get Shadow status working in the savage driver. I'd li=
ke
> >>>>>to put the status page into locked shared memory and tell the Savage
> >>>>>chips about its physical address. I used a drmAddMap call copied fro=
m
> >>>>>the MGA driver that looked like it did what I wanted. But I had to f=
ind
> >>>>>out that the handle returned by drmAddMap is a kernel virtual addres=
s,
> >>>>>not a physical address. Later I saw that the status page seems to be
> >>>>>unused in the MGA driver. The code to get the physical address from =
the
> >>>>>kernel virtual address is #if-ed out, probably for portability reaso=
ns.
> >>>>
> >>>>Does the code in drm_pci.c apply? It's only used by the via driver.
> >>>>I'm not sure what it does.
> >>>
> >>>I guess if I was only going to use that memory in kernel drm_pci_alloc
> >>>would be fine. But I want to map the status page at least in the
> >>>Xserver, maybe also in clients.
> >>
> >>Take a look at km or v4l drivers - maybe there is something you can use=
.
> >>
> >>With 2.4 I used rvmalloc - I don't recall whether there is a standard=20
> >>mechanism for this in 2.6.
> >=20
> >=20
> > The point is that I'm not just looking for a way to get a physical
> > address in Linux. In that case I'd just use virt_to_bus on the map
> > handle. Since I only need a single page I also don't need to worry abou=
t
> > allocating consistent memory. However, I want to use the mechanisms
> > available in the DRM abstraction layer so that the savage driver stays
> > (at least theoretically) portable to BSD and so that I can use
> > drmMap/drmMapBufs to map the status page in the Xserver and/or in DRI
> > clients.
>=20
> I'm not sure that such a mechanism exists.  If it doesn't, I think the on=
us=20
> falls on the porter to provide one, if a port is ever attempted.

I see two alternatives for providing what I need.

     1. A new map type _DRM_PCI (or _DRM_CONSISTENT if you prefer that
        in contrast to _DRM_SCATTER_GATHER) that allocates pages with
        drm_alloc_pages and returns the physical address as handle.
     2. A new map flag for the _DRM_SHM map type, say _DRM_CONSISTENT
        that makes it allocate with drm_alloc_pages and return the
        physical address as handle.

I think the first one sounds cleaner. If we can agree on this I'd
prepare a patch for the linux-core implementation and the corresponding
headers in Xorg for review.(*) When BSD gets ported to the
core/personality model it would need to get the new PCI map type too in
order to support any drivers that use it.

>=20
> Keith

Regards,
  Felix

(*) In fact I'm starting right now.

--=20
| Felix K=FChling <fx...@gm...>                     http://fxk.de.vu |
| PGP Fingerprint: 6A3C 9566 5B30 DDED 73C3  B152 151C 5CC1 D888 E595 |