Hi Linus,
Please pull the 'agp-fixes' branch from
git://master.kernel.org/pub/scm/linux/kernel/git/airlied/agp-2.6.git agp-fixes
This contains two fixes,
oops fix: intel g33 fix for 1MB aperture setting
PAE kernel + AGP + GEM fix. The patch contains more explaination, but
since GEM uses shmem to allocate pages it can pass highmem pages into AGP,
AGP just needs to pass those addressess around cleanly in dma_addr_t. The
correct answer is to use struct page *, but that is way too invasive for
this stage of a kernel release.
Dave.
drivers/char/agp/agp.h | 8 ++++----
drivers/char/agp/efficeon-agp.c | 2 +-
drivers/char/agp/generic.c | 15 ++++++---------
drivers/char/agp/hp-agp.c | 2 +-
drivers/char/agp/i460-agp.c | 2 +-
drivers/char/agp/intel-agp.c | 11 +++++++++--
drivers/char/agp/parisc-agp.c | 2 +-
drivers/char/agp/sgi-agp.c | 2 +-
drivers/gpu/drm/drm_memory.c | 3 ++-
drivers/gpu/drm/drm_vm.c | 4 ++--
include/linux/agp_backend.h | 4 ++--
11 files changed, 30 insertions(+), 25 deletions(-)
commit d5e4cc00c911a72709a8d012179980a6157ba552
Author: Brandon Philips <[email protected]>
Date: Tue Oct 28 22:22:18 2008 -0700
intel-agp: Avoid oops for G33 on 1MB stolen case
This is similar to f443675affe3f16dd428e46f0f7fd3f4d703eeab which was
reverted because it broke older X.org driver. This patch only fixes
the 1MB stolen case since it causes an oops.
Xorg will not work without the accompanying patch[1] but avoiding an
oops and making it possible to work with patched xorg driver is
reasonable.
[1] http://ifup.org/~philips/review/xf86-video-intel-G33-1mb.patch
Explanation of the oops:
> static void intel_i830_init_gtt_entries(void)
...
> } else if (IS_G33) {
> /* G33's GTT size defined in gmch_ctrl */
> switch (gmch_ctrl & G33_PGETBL_SIZE_MASK) {
> case G33_PGETBL_SIZE_1M:
> size = 1024;
> break;
...
> size += 4;
size = 1028
Then since we have the BIOS setting 1MB for the device in the GMCH
control we get to here:
> } else {
> switch (gmch_ctrl & I855_GMCH_GMS_MASK) {
> case I855_GMCH_GMS_STOLEN_1M:
> gtt_entries = MB(1) - KB(size);
> break;
MB(1) = 1 * 1024 * 1024
KB(1028) = 1028 * 1024
MB(1) - KB(1028) = -4096
> gtt_entries /= KB(4);
> intel_private.gtt_entries = gtt_entries;
We end up with -1 in gtt_entries.
This leads to intel_i915_configure reading/writing to areas outside of
mapped memory and the oops.
Bugzilla: https://bugzilla.novell.com/show_bug.cgi?id=391261
Signed-off-by: Brandon Philips <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
commit b4f5dfa4b12efeea816a6f16ddb4c8851adbd39a
Author: Pierre Willenbrock <[email protected]>
Date: Mon Nov 3 21:14:56 2008 -0800
agp: change physical address type from unsigned long to dma_addr_t
This fixes enough of the interface for i915 DRM with CONFIG_HIGHMEM64. It
doesn't fix the physical addresses when passed to intelfb, mtrr or userspace.
airlied: This patch is sufficent to fix GEM use cases, we should overhaul AGP
later to make it use struct page *, instead of this. AGP allocated pages are
safe because we never use highmem pages, GEM allocated pages never use
virt_to_*.
Signed-off-by: Pierre Willenbrock <[email protected]>
Signed-off-by: Eric Anholt <[email protected]>
Signed-off-by: Dave Airlie <[email protected]>
On Thu, Nov 6, 2008 at 4:46 PM, Dave Airlie <[email protected]> wrote:
> Hi Linus,
>
> Please pull the 'agp-fixes' branch from
> git://master.kernel.org/pub/scm/linux/kernel/git/airlied/agp-2.6.git agp-fixes
Actually Linus please skip this, we really want this PAE patch in but
I suspect I will need
to do more widespread changes than this patch does :(
Dave.
>
> This contains two fixes,
>
> oops fix: intel g33 fix for 1MB aperture setting
>
> PAE kernel + AGP + GEM fix. The patch contains more explaination, but
> since GEM uses shmem to allocate pages it can pass highmem pages into AGP,
> AGP just needs to pass those addressess around cleanly in dma_addr_t. The
> correct answer is to use struct page *, but that is way too invasive for
> this stage of a kernel release.
>
> Dave.
>
> drivers/char/agp/agp.h | 8 ++++----
> drivers/char/agp/efficeon-agp.c | 2 +-
> drivers/char/agp/generic.c | 15 ++++++---------
> drivers/char/agp/hp-agp.c | 2 +-
> drivers/char/agp/i460-agp.c | 2 +-
> drivers/char/agp/intel-agp.c | 11 +++++++++--
> drivers/char/agp/parisc-agp.c | 2 +-
> drivers/char/agp/sgi-agp.c | 2 +-
> drivers/gpu/drm/drm_memory.c | 3 ++-
> drivers/gpu/drm/drm_vm.c | 4 ++--
> include/linux/agp_backend.h | 4 ++--
> 11 files changed, 30 insertions(+), 25 deletions(-)
>
> commit d5e4cc00c911a72709a8d012179980a6157ba552
> Author: Brandon Philips <[email protected]>
> Date: Tue Oct 28 22:22:18 2008 -0700
>
> intel-agp: Avoid oops for G33 on 1MB stolen case
>
> This is similar to f443675affe3f16dd428e46f0f7fd3f4d703eeab which was
> reverted because it broke older X.org driver. This patch only fixes
> the 1MB stolen case since it causes an oops.
>
> Xorg will not work without the accompanying patch[1] but avoiding an
> oops and making it possible to work with patched xorg driver is
> reasonable.
>
> [1] http://ifup.org/~philips/review/xf86-video-intel-G33-1mb.patch
>
> Explanation of the oops:
>
> > static void intel_i830_init_gtt_entries(void)
> ...
> > } else if (IS_G33) {
> > /* G33's GTT size defined in gmch_ctrl */
> > switch (gmch_ctrl & G33_PGETBL_SIZE_MASK) {
> > case G33_PGETBL_SIZE_1M:
> > size = 1024;
> > break;
> ...
> > size += 4;
>
> size = 1028
>
> Then since we have the BIOS setting 1MB for the device in the GMCH
> control we get to here:
>
> > } else {
> > switch (gmch_ctrl & I855_GMCH_GMS_MASK) {
> > case I855_GMCH_GMS_STOLEN_1M:
> > gtt_entries = MB(1) - KB(size);
> > break;
>
> MB(1) = 1 * 1024 * 1024
> KB(1028) = 1028 * 1024
>
> MB(1) - KB(1028) = -4096
>
> > gtt_entries /= KB(4);
> > intel_private.gtt_entries = gtt_entries;
>
> We end up with -1 in gtt_entries.
>
> This leads to intel_i915_configure reading/writing to areas outside of
> mapped memory and the oops.
>
> Bugzilla: https://bugzilla.novell.com/show_bug.cgi?id=391261
>
> Signed-off-by: Brandon Philips <[email protected]>
> Signed-off-by: Dave Airlie <[email protected]>
>
> commit b4f5dfa4b12efeea816a6f16ddb4c8851adbd39a
> Author: Pierre Willenbrock <[email protected]>
> Date: Mon Nov 3 21:14:56 2008 -0800
>
> agp: change physical address type from unsigned long to dma_addr_t
>
> This fixes enough of the interface for i915 DRM with CONFIG_HIGHMEM64. It
> doesn't fix the physical addresses when passed to intelfb, mtrr or userspace.
>
> airlied: This patch is sufficent to fix GEM use cases, we should overhaul AGP
> later to make it use struct page *, instead of this. AGP allocated pages are
> safe because we never use highmem pages, GEM allocated pages never use
> virt_to_*.
>
> Signed-off-by: Pierre Willenbrock <[email protected]>
> Signed-off-by: Eric Anholt <[email protected]>
> Signed-off-by: Dave Airlie <[email protected]>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>