Re: [PATCH 8/9] drm/bridge: put the bridge returned by drm_bridge_get_next_bridge()

[Luca Ceresoli <luca.ceresoli@bootlin.com>]

Hi Maxime,

On Thu, 10 Jul 2025 09:27:09 +0200
Maxime Ripard <mripard@kernel.org> wrote:

> Hi,
> 
> On Wed, Jul 09, 2025 at 06:48:07PM +0200, Luca Ceresoli wrote:
> > The bridge returned by drm_bridge_get_next_bridge() is refcounted. Put it
> > when done.
> > 
> > Signed-off-by: Luca Ceresoli <luca.ceresoli@bootlin.com>  
> 
> You should really expand a bit more your commit logs, and provide the
> context of why you think putting drm_bridge_put where you do is a good idea.
> 
> > ---
> >  drivers/gpu/drm/drm_bridge.c | 2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/drm_bridge.c b/drivers/gpu/drm/drm_bridge.c
> > index 0b450b334afd82e0460f18fdd248f79d0a2b153d..05e85457099ab1e0a23ea7842c9654c9a6881dfb 100644
> > --- a/drivers/gpu/drm/drm_bridge.c
> > +++ b/drivers/gpu/drm/drm_bridge.c
> > @@ -1147,6 +1147,8 @@ drm_atomic_bridge_propagate_bus_flags(struct drm_bridge *bridge,
> >  	} else {
> >  		next_bridge_state = drm_atomic_get_new_bridge_state(state,
> >  								next_bridge);
> > +		drm_bridge_put(next_bridge);
> > +
> >  		/*
> >  		 * No bridge state attached to the next bridge, just leave the
> >  		 * flags to 0.  
> 
> In particular, I don't think it is here.
> 
> You still have a variable in scope after that branch that you would have
> given up the reference for, which is pretty dangerous.
> 
> Also, the bridge state lifetime is shorter than the bridge lifetime
> itself, so we probably want to have the drm_bridge_put after we're done
> with next_bridge_state too.

Totally agree about this.

I theory moving the _put just after the last usage of next_bridge_state
would be enough. However...

> Overall, I think using __free here is probably the most robust solution.

...I'm OK with using use __free here, even though it doesn't look
strictly necessary. However for patch 9 the code path is slightly more
complex, so I'll use __free for both.

With this change, this patch would become:

@@ -1121,7 +1121,6 @@ drm_atomic_bridge_propagate_bus_flags(struct drm_bridge *bridge,
                                      struct drm_atomic_state *state)
 {
        struct drm_bridge_state *bridge_state, *next_bridge_state;
-       struct drm_bridge *next_bridge;
        u32 output_flags = 0;
 
        bridge_state = drm_atomic_get_new_bridge_state(state, bridge);
@@ -1130,7 +1129,7 @@ drm_atomic_bridge_propagate_bus_flags(struct drm_bridge *bridge,
        if (!bridge_state)
                return;
 
-       next_bridge = drm_bridge_get_next_bridge(bridge);
+       struct drm_bridge *next_bridge __free(drm_bridge_put) = drm_bridge_get_next_bridge(bridge);
 
        /*
         * Let's try to apply the most common case here, that is, propagate

And a tentative commit message body is:

  The bridge returned by drm_bridge_get_next_bridge() is refcounted. Put
  it when done. We need to ensure it is not put before either
  next_bridge or next_bridge_state is in use, thus for simplicity use a
  cleanup action.

I'll resend with the above changes (unless you have more improvements to
suggest) in a few days, to wait for any feedback on patch 1.

Luca

-- 
Luca Ceresoli, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com