The patch is against a 2.6.11 kernel tree. I am running this with a
32-bit X server (compiled up from X.org CVS as of a couple of weeks
ago) and 32-bit DRI libraries and clients. All the userland stuff is
identical to what I am using under a 32-bit kernel on my G4 powerbook
(which is a 32-bit machine of course). I haven't tried compiling up a
64-bit X server or clients yet.
In the compatibility routines I have assumed that the kernel can
safely access user addresses after set_fs(KERNEL_DS). That is, where
an ioctl argument structure contains pointers to other structures, and
those other structures are already compatible between the 32-bit and
64-bit ABIs (i.e. they only contain things like chars, shorts or
ints), I just check the address with access_ok() and then pass it
through to the 64-bit ioctl code. I believe this approach may not
work on sparc64, but it does work on ppc64 and x86_64 at least.
One tricky area which may need to be revisited is the question of how
to handle the handles which we pass back to userspace to identify
mappings. These handles are generated in the ADDMAP ioctl and then
passed in as the offset value to mmap. However, offset values for
mmap seem to be generated in other ways as well, particularly for AGP
mappings.
The approach I have ended up with is to generate a fake 32-bit handle
only for _DRM_SHM mappings. The handles for other mappings (AGP, REG,
FB) are physical addresses which are already limited to 32 bits, and
generating fake handles for them created all sorts of problems in the
mmap/nopage code.
This patch has been updated to use the new compatibility ioctls.
From: Paul Mackerras <paulus@samba.org>
Signed-off-by: Dave Airlie <airlied@linux.ie>
9a18664506