From: Tormod Volden debian.tormod@gmail.com
Drivers using multiple framebuffers got broken by commit 41c2e75e60200a860a74b7c84a6375c105e7437f which ignored the framebuffer (or register) map offset when looking for existing maps. The rationale was that the kernel-userspace ABI is fixed at a 32-bit offset, so the real offsets could not always be handed over for comparison.
Instead of ignoring the offset we will compare the lower 32 bit. Drivers using multiple framebuffers should just make sure that the lower 32 bit are different. The existing drivers in question are practically limited to 32-bit systems so that should be fine for them.
It is assumed that current drivers always specify a correct framebuffer map offset, even if this offset was ignored since above commit. So this patch should not change anything for drivers using only one framebuffer.
Drivers needing multiple framebuffers with 64-bit map offsets will need to cook up something, for instance keeping an ID in the lower bit which is to be aligned away when it comes to using the offset.
All of above applies to _DRM_REGISTERS as well.
Signed-off-by: Tormod Volden debian.tormod@gmail.com ---
On Mon, May 30, 2011 at 1:29 AM, Dave Airlie wrote:
If you test it and it works I like it best. Simple and clear, and pretty close to what I was thinking was a good idea.
As you say if someone needs this functionality in a new driver they can fix it, but really new drivers shouldn't be doing anything in this area.
Dave.
Whoops, there was a less obvious fallthrough from the _DRM_SHM case above, where we do not want to compare offsets at all if it contains a lock(*).
This patch has been tested on savage, and for verification also on radeon with DRI1 and DRI2.
Tormod
(*) It actually checks if _DRM_CONTAINS_LOCK is the /only/ flag set. I suppose this is intentional. My v2 patch does not change anything in the case of _DRM_SHM: If it contains a lock, it returns a match without comparing offsets. If no lock, it compares the full offsets. Is this because the only _DRM_SHM used by userspace is the one with a lock, so there is never a need to check a userspace-provided offset, or are those always within 32 bit so a full check is ok?
drivers/gpu/drm/drm_bufs.c | 14 +++++++++----- 1 files changed, 9 insertions(+), 5 deletions(-)
diff --git a/drivers/gpu/drm/drm_bufs.c b/drivers/gpu/drm/drm_bufs.c index 3e257a5..40ccfbc 100644 --- a/drivers/gpu/drm/drm_bufs.c +++ b/drivers/gpu/drm/drm_bufs.c @@ -46,10 +46,11 @@ static struct drm_map_list *drm_find_matching_map(struct drm_device *dev, list_for_each_entry(entry, &dev->maplist, head) { /* * Because the kernel-userspace ABI is fixed at a 32-bit offset - * while PCI resources may live above that, we ignore the map - * offset for maps of type _DRM_FRAMEBUFFER or _DRM_REGISTERS. - * It is assumed that each driver will have only one resource of - * each type. + * while PCI resources may live above that, we only compare the + * lower 32 bits of the map offset for maps of type + * _DRM_FRAMEBUFFER or _DRM_REGISTERS. + * It is assumed that if a driver have more than one resource + * of each type, the lower 32 bits are different. */ if (!entry->map || map->type != entry->map->type || @@ -59,9 +60,12 @@ static struct drm_map_list *drm_find_matching_map(struct drm_device *dev, case _DRM_SHM: if (map->flags != _DRM_CONTAINS_LOCK) break; + return entry; case _DRM_REGISTERS: case _DRM_FRAME_BUFFER: - return entry; + if ((entry->map->offset & 0xffffffff) == + (map->offset & 0xffffffff)) + return entry; default: /* Make gcc happy */ ; }