A framebuffer reconstructed from a hardware state readout describes a
buffer that the firmware allocated, outside of the kernel memory
allocator, so its planes have no GEM object attached.

drm_gem_plane_helper_prepare_fb() retrieves the GEM object of every
plane through drm_gem_fb_get_obj() to extract the implicit fences from
its reservation. For a readout framebuffer, this triggers
drm_WARN_ON_ONCE(!fb->obj[plane]), returns NULL, and fails the whole
commit with -EINVAL.

This can be hit as soon as the state readout installs such a
framebuffer on a plane: any commit that includes the plane without
userspace providing a new framebuffer will duplicate its current
state, readout framebuffer included, and run prepare_fb() on it.
Blanking the display through DPMS right after boot is such a case,
since it pulls the plane in through drm_atomic_add_affected_planes().

Access the framebuffer obj[] array directly and skip the planes
without a GEM object: there is no reservation to extract implicit
fences from anyway. This mirrors what drm_fb_dma_get_scanout_buffer()
does for such framebuffers.

Signed-off-by: Maxime Ripard <[email protected]>
---
 drivers/gpu/drm/drm_gem_atomic_helper.c | 14 +++++++++-----
 1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/drm_gem_atomic_helper.c 
b/drivers/gpu/drm/drm_gem_atomic_helper.c
index c952cfed84c1..bab354d12fe7 100644
--- a/drivers/gpu/drm/drm_gem_atomic_helper.c
+++ b/drivers/gpu/drm/drm_gem_atomic_helper.c
@@ -138,11 +138,15 @@ int drm_gem_plane_helper_prepare_fb(struct drm_plane 
*plane,
 
        for (i = 0; i < state->fb->format->num_planes; ++i) {
-               struct drm_gem_object *obj = drm_gem_fb_get_obj(state->fb, i);
+               struct drm_gem_object *obj = state->fb->obj[i];
                struct dma_fence *new;
 
-               if (!obj) {
-                       ret = -EINVAL;
-                       goto error;
-               }
+               /*
+                * A framebuffer reconstructed from a hardware state
+                * readout has no GEM object backing its planes, as it
+                * describes a buffer allocated outside of the kernel:
+                * there is no implicit fencing to wait on.
+                */
+               if (!obj)
+                       continue;
 
                ret = dma_resv_get_singleton(obj->resv, usage, &new);

-- 
2.55.0

Reply via email to