[git pull] drm fixes

[Linus Torvalds]

On Mon, Mar 2, 2015 at 1:04 AM, Daniel Vetter <daniel at ffwll.ch> wrote:
>
> And can you please attach a bactrace of the WARN in your patch, just to
> double-check you blow up at the same spot?

So the dmesg I attached had a backtrace for the new WARN_ONCE() (in
addition to an unrelated(?) one from i915_gem_free_object()).

Or did you mean a backtrace of the oops when things go wrong, when my
patch is *not* applied? My first email had that with the kref.h
warning from drm_framebuffer_reference, which is otherwise the same
thing.

And after things go wrong, and the plane handling thing uses a
framebuffer that has already been freed, then the resulting oopses end
up being kind of random. Which was partly why it ended up beign so
hard to finally bisect, and I actually eventually gave up on it -
because sometimes things would just happen to work, sometimes things
would blow up and oops when X started (resulting in just a dead
machine), sometimes things would oops at bootup.

The most common oops was something going wrong when the framebuffer
was free'd for the second time, and it would cause a NULL pointer
dereference in drm_framebuffer_free() or one of the routines it called
(so often a NULL pointer dereference in
"mutex_lock(&dev->mode_config.fb_lock)" because "dev" was NULL, or the
call to "fb->funcs->destroy(fb)" would oops because "fb->funcs" was
NULL.

                         Linus