[Intel-gfx] [PATCH] drm/i915: Move drm_framebuffer_unreference out of struct_mutex for takeover

[Ville Syrjälä]

On Mon, Apr 13, 2015 at 02:37:41PM +0100, Tvrtko Ursulin wrote:
> 
> On 04/13/2015 01:09 PM, Jani Nikula wrote:
> > On Thu, 26 Mar 2015, Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com> wrote:
> >> On 03/26/2015 01:30 PM, Ville Syrjälä wrote:
> >>> On Thu, Mar 26, 2015 at 12:39:40PM +0000, Tvrtko Ursulin wrote:
> >>>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >>>>
> >>>> intel_user_framebuffer_destroy() requires the struct_mutex for its
> >>>> object bookkeeping, so this means that all calls to
> >>>> drm_framebuffer_unreference must be held without that lock.
> >>>>
> >>>> This is a simplified version of the identically named patch by Chris Wilson.
> >>>>
> >>>> References: https://bugs.freedesktop.org/show_bug.cgi?id=89166
> >>>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> >>>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> >>>> ---
> >>>>    drivers/gpu/drm/i915/intel_display.c | 10 ++++++++++
> >>>>    1 file changed, 10 insertions(+)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> >>>> index cb50854..0788507 100644
> >>>> --- a/drivers/gpu/drm/i915/intel_display.c
> >>>> +++ b/drivers/gpu/drm/i915/intel_display.c
> >>>> @@ -14020,11 +14020,21 @@ void intel_modeset_gem_init(struct drm_device *dev)
> >>>>    					       c->primary->fb,
> >>>>    					       c->primary->state,
> >>>>    					       NULL)) {
> >>>> +			/*
> >>>> +			 * We must drop struct_mutex when calling
> >>>> +			 * drm_framebuffer_unreference and it is safe to do so
> >>>> +			 * because it is not needed at this point anyway.
> >>>> +			 * At this stage the driver is still single-threaded and
> >>>> +			 * we are taking it only to silence a warning in
> >>>> +			 * intel_pin_and_fence_fb_obj.
> >>>> +			 */
> >>>> +			mutex_unlock(&dev->struct_mutex);
> >>>>    			DRM_ERROR("failed to pin boot fb on pipe %d\n",
> >>>>    				  to_intel_crtc(c)->pipe);
> >>>>    			drm_framebuffer_unreference(c->primary->fb);
> >>>>    			c->primary->fb = NULL;
> >>>>    			update_state_fb(c->primary);
> >>>> +			mutex_lock(&dev->struct_mutex);
> >>>>    		}
> >>>>    	}
> >>>>    	mutex_unlock(&dev->struct_mutex);
> >>>
> >>> Just grab the mutex around the pin_and_fence inside the loop. It doesn't
> >>> protect anything else.
> >>
> >> Well the comment says so, but this way it only grabs and releases it
> >> once if there are multiple active crtcs and nothing fails. So I was
> >> hoping the comment was enough to explain the reality, even though the
> >> other option would be more obvious code strictly speaking.
> >
> > Tvrtko & Ville, can you reach a solution on this one? Or is there a
> > new patch that I may have missed?
> 
> It was pretty much bike shedding - I am happy with this version since it 
> has a single lock/unlock on the normal path, compared to one pair per 
> active display with what Ville wanted.
> 
> Either approach makes for unclear code so needs a big comment anyway. 
> Which leaves only the exact placement of mutex_lock/unlock under discussion.

I don't see what's unclear about locking only around the call that needs
the lock.

> 
> If we want to spend this much time on this that is.
> 
> Regards,
> 
> Tvrtko

-- 
Ville Syrjälä
Intel OTC