drm/i915/gen9: Set PIN_ZONE_4G end to 4GB - 1 page

Kernel and userspace are able to handle 4GB (1<<32) address space range,
but "A32 Stateless Model" is not. According to documentation, A32 accesses
are based on General State Base Address and bound checking is in place.
Because size field (instruction State Base Address) limitation, it is not
possible to address full 4GB memory region.

A32 Stateless Model is used by some libraries and without this patch, the
last page of 4GB address space is not accessible in 32bit processes.

Reported-by: Artur Harasimiuk <artur.harasimiuk@intel.com>
Signed-off-by: Michel Thierry <michel.thierry@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1452512367-23614-1-git-send-email-michel.thierry@intel.com
Cc: drm-intel-fixes@lists.freedesktop.org
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
(cherry picked from commit 1892faa9ec5d51b07d646cbd5597cd30e049aa51)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>

diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index d8efc9d..2c24ff3 100644 (file)
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -3484,7 +3484,7 @@ i915_gem_object_bind_to_vm(struct drm_i915_gem_object *obj,
        if (flags & PIN_MAPPABLE)
                end = min_t(u64, end, dev_priv->gtt.mappable_end);
        if (flags & PIN_ZONE_4G)
-               end = min_t(u64, end, (1ULL << 32));
+               end = min_t(u64, end, (1ULL << 32) - PAGE_SIZE);
 
        if (alignment == 0)
                alignment = flags & PIN_MAPPABLE ? fence_alignment :