Date: Wed, 14 May 2008 17:49:33 -0400 From: Coleman Kane <cokane@FreeBSD.org> To: Eric Anholt <eric@anholt.net> Cc: arch@FreeBSD.org Subject: Re: Feature requests for open-source graphics Message-ID: <1210801773.2466.12.camel@localhost> In-Reply-To: <1210352688.2152.78.camel@localhost> References: <1210352688.2152.78.camel@localhost>
next in thread | previous in thread | raw e-mail | index | archive | help
[-- Attachment #1 --] On Fri, 2008-05-09 at 10:04 -0700, Eric Anholt wrote: > (Re-sent after I dragged a discussion of NVIDIA on FreeBSD off-topic on > developers@) > > On Wed, 2008-05-07 at 10:37 -0700, Marcel Moolenaar wrote: > > On May 7, 2008, at 8:29 AM, John Baldwin wrote: > > > > > Quite, and this has been covered several times already. In fact, > > > they engage > > > in several hacks to support Linux. Other OS's such as Solaris and > > > OS X have > > > cleaner interfaces for many of the things they wish to do. > > > > I think this is where I normally say that we need kernel drivers > > for graphics hardware. I'm not going to do that anymore; I've been > > stating the obvious too often already... > > It's OK, we're finally listening. By the end of the year, major Linux > distributions should be shipping "DRM modesetting" -- the DRM device > takes over your graphics card and manages memory, execution, and > suspend/resume. Your kernel console and X Server then sit on top of > that, submitting video mode setting and command execution requests to > the DRM. The chips I would expect to be supported by then are all > Intel, pre-r500 ATI at least, and a subset of nvidia through nouveau. > > What FreeBSD needs to do to keep up is implement the memory manager > part. I think I've got a reasonable design going that should take about > a month of reimplementing for the FreeBSD kernel. If someone wanted to > get us closer to doing that, > > git clone anongit.freedesktop.org:/git/mesa/drm > > and make everything but TTM (the *_HAVE_FENCE and *_HAVE_BUFFER defines) > work on -current. > > To see what we're working on for what we're calling the "graphics > execution manager" now (memory management plus caching management plus > plans for execution scheduling), > > git-remote add anholt people.freedesktop.org:~anholt/drm > git-fetch anholt > git-checkout anholt/drm-gem > > The interesting things we're needing from the kernel: > - Private storage associated with file descriptors > - Callback to driver on destruction of file descriptor > - Ability to allocate a swap-backed pile of memory > - Ability to pin pages from that object and get addresses for them to > stuff into the GTT. > - Ability to map pages in the kernel with the uncached bits set > (non-intel) > - Ability to map pages to userland with the uncached bits set > (non-intel, likely not required, but may come at a performance > cost). > > Interesting things we're likely going to need from the kernel: > - Ability to create fds above 1024 from our driver, associated with our > own set of file operations (read/write/mmap/ioctl/close). > - Ability to look up those fds and get our private storage associated > with them. > > If above goes through, then another likely adventure: > - Creating a filesystem exposing those objects we've been making fds > for. > > email message attachment, "Forwarded message - Re: Forward: Updated > FreeBSD kernel feature requests (NVIDIA graphics)" > > -------- Forwarded Message -------- > > From: Eric Anholt <eric@anholt.net> > > To: Marcel Moolenaar <xcllnt@mac.com> > > Cc: John Baldwin <jhb@freebsd.org>, Coleman Kane > > <cokane@freebsd.org>, Dag-Erling Smørgrav <des@des.no>, Martin > > Cracauer <cracauer@cons.org>, gnn@freebsd.org, > > developers@freebsd.org > > Subject: Re: Forward: Updated FreeBSD kernel feature requests > > (NVIDIA graphics) > > Date: Thu, 08 May 2008 15:54:17 -0700 > > Mailer: Evolution 2.22.0 FreeBSD GNOME Team Port > > > > On Wed, 2008-05-07 at 10:37 -0700, Marcel Moolenaar wrote: > > > On May 7, 2008, at 8:29 AM, John Baldwin wrote: > > > > > > > Quite, and this has been covered several times already. In fact, > > > > they engage > > > > in several hacks to support Linux. Other OS's such as Solaris and > > > > OS X have > > > > cleaner interfaces for many of the things they wish to do. > > > > > > I think this is where I normally say that we need kernel drivers > > > for graphics hardware. I'm not going to do that anymore; I've been > > > stating the obvious too often already... > > > > It's OK, we're finally listening. By the end of the year, major Linux > > distributions should be shipping "DRM modesetting" -- the DRM device > > takes over your graphics card and manages memory, execution, and > > suspend/resume. Your kernel console and X Server then sit on top of > > that, submitting video mode setting and command execution requests to > > the DRM. The chips I would expect to be supported by then are all > > Intel, pre-r500 ATI at least, and a subset of nvidia through nouveau. > > > > What FreeBSD needs to do to keep up is implement the memory manager > > part. I think I've got a reasonable design going that should take about > > a month of reimplementing for the FreeBSD kernel. If someone wanted to > > get us closer to doing that, > > > > git clone anongit.freedesktop.org:/git/mesa/drm > > > > and make everything but TTM (the *_HAVE_FENCE and *_HAVE_BUFFER defines) > > work on -current. > > > > To see what we're working on for what we're calling the "graphics > > execution manager" now (memory management plus caching management plus > > plans for execution scheduling), > > > > git-remote add anholt people.freedesktop.org:~anholt/drm > > git-fetch anholt > > git-checkout anholt/drm-gem > > > > The interesting things we're needing from the kernel: > > - Private storage associated with file descriptors > > - Callback to driver on destruction of file descriptor > > - Ability to allocate a swap-backed pile of memory > > - Ability to pin pages from that object and get addresses for them to > > stuff into the GTT. > > - Ability to map pages in the kernel with the uncached bits set > > (non-intel) > > - Ability to map pages to userland with the uncached bits set > > (non-intel, likely not required, but may come at a performance > > cost). > > > > Interesting things we're considering needing from the kernel: > > - Ability to create fds above 1024 from our driver, associated with our > > own set of file operations (read/write/mmap/ioctl/close). > > - Ability to look up those fds and get our private storage associated > > with them. > > > > Down the line, likely to happen: > > - Creating a filesystem exposing those objects we've been making fds > > for. Eric, I mentioned earlier that I would get my local git repo's changes to the mesa/drm tree available from fd.o up somewhere. I have them here: * http://www.cokane.org/cgi-bin/gitweb.cgi?p=drm.git My custom branch is 'cokane-master' The fd.o branch is tracked on 'fd-master' You can ignore the 'master' branch for now... I need to re-org my git a bit. Right now, this gets my RS690 notebook to almost work with the in-development radeonhd DRI code, but it causes Xorg to run in a busyloop when I try using the xf86-video-ati driver with it. I did a run at trying to get the vblank stuff ported over, but I got busy and haven't touched it in a bit. I also tried to tweak anything else that kept the radeon code from compiling under FreeBSD. If I have another set of eyes on this it might help me front-burner it more often to get patches in here-and-there. I've tried shoving some bug reports up to the DRI project, but they haven't been acted upon yet... -- Coleman Kane [-- Attachment #2 --] -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.9 (FreeBSD) iEYEABECAAYFAkgrXmwACgkQcMSxQcXat5dpKACfcxjNSh8P0aBxFxu1M+FJBxpf srMAnAkJRaFXh2fRnWdW7YZbiumPDITq =xxUW -----END PGP SIGNATURE-----
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1210801773.2466.12.camel>