[Intel-gfx] Deadlock in intel_user_framebuffer_destroy()

[Chris Wilson]

On Mon, Jun 15, 2015 at 02:02:23PM +0200, Daniel Vetter wrote:
> On Mon, Jun 15, 2015 at 08:53:02AM +0100, Chris Wilson wrote:
> > On Mon, Jun 15, 2015 at 09:44:15AM +0300, Jani Nikula wrote:
> > > On Wed, 03 Jun 2015, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > > > On Wed, Jun 03, 2015 at 03:43:32PM +0200, Lukas Wunner wrote:
> > > >> Hi,
> > > >> 
> > > >> a deadlock was introduced by commit 60a5ca015ffd2aacfe5674b5a401cd2a37159e07
> > > >> 
> > > >> Author: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > >> Date:   Fri Jun 13 11:10:53 2014 +0300
> > > >> 
> > > >>     drm/i915: Add locking around framebuffer_references--
> > > >> 
> > > >> 
> > > >> The commit amended intel_display.c:intel_user_framebuffer_destroy() with
> > > >> mutex_lock(&dev->struct_mutex).
> > > >> 
> > > >> A few weeks prior Chris Wilson had amended intel_fbdev.c:intelfb_create()
> > > >> with a call to drm_framebuffer_unreference() while &dev->struct_mutex is
> > > >> locked (commit edd586fe705e819bc711b5ed7194a0b6f9f1a7e1, "drm/i915: Discard
> > > >> BIOS framebuffers too small to accommodate chosen mode").
> > > >> 
> > > >
> > > > Just move the mutex_lock down a step.
> > > 
> > > Lucas, did you try this?
> > 
> > There's a goto unlock that also needed to be disabled, such as
> 
> Your previous patch placed the mutex_lock before the goto out_unlock - I
> fail to see what has been broken with that version? Can you resubmit that
> as a proper patch with sob and Lucas' t-b?

There wasn't, I just rewrote it incorrectly. There's also a
drm_framebuffer_remove() called by intelfb_alloc which needs to be moved
out of the mutex. A much larger disentangling of the functions involved
here is required. Tvrstko volunteered himself. Brave, very brave :)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre