Re: [PATCH 2/2] drm/ast: Fix connector leak during driver unload

[Daniel Vetter]

On Fri, Nov 30, 2018 at 11:17:51AM +1100, Sam Bobroff wrote:
> On Thu, Nov 29, 2018 at 09:56:53AM +0100, Daniel Vetter wrote:
> > On Thu, Nov 29, 2018 at 9:05 AM Sam Bobroff <sbobroff@xxxxxxxxxxxxx> wrote:
> > >
> > > On Thu, Nov 29, 2018 at 09:40:53AM +1000, Dave Airlie wrote:
> > > > On Mon, 5 Nov 2018 at 15:59, Sam Bobroff <sbobroff@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > When unloading the ast driver, a warning message is printed by
> > > > > drm_mode_config_cleanup() because a reference is still held to one of
> > > > > the drm_connector structs.
> > > > >
> > > > > Correct this by calling drm_framebuffer_remove() in
> > > > > ast_fbdev_destroy().
> > > > >
> > > > > Signed-off-by: Sam Bobroff <sbobroff@xxxxxxxxxxxxx>
> > > > > ---
> > > > >  drivers/gpu/drm/ast/ast_fb.c | 4 ++++
> > > > >  1 file changed, 4 insertions(+)
> > > > >
> > > > > diff --git a/drivers/gpu/drm/ast/ast_fb.c b/drivers/gpu/drm/ast/ast_fb.c
> > > > > index 0cd827e11fa2..655372ea81e9 100644
> > > > > --- a/drivers/gpu/drm/ast/ast_fb.c
> > > > > +++ b/drivers/gpu/drm/ast/ast_fb.c
> > > > > @@ -263,6 +263,10 @@ static void ast_fbdev_destroy(struct drm_device *dev,
> > > > >  {
> > > > >         struct ast_framebuffer *afb = &afbdev->afb;
> > > > >
> > > > > +       /* drm_framebuffer_remove() expects us to hold a ref, which it
> > > > > +        * will drop, so take one: */
> > > > > +       drm_framebuffer_get(&afb->base);
> > > > > +       drm_framebuffer_remove(&afb->base);
> > > >
> > > > This doesn't seem corret, no other driver does this pattern, and I
> > > > can't believe ast is special here.
> > > >
> > > > The get just doesn't make sense.
> > >
> > > Thanks for having a look at this, as I said in the cover letter I was
> > > concerned that it might not be a good fix.
> > >
> > > But the AST driver does seem to be special (or just old?) because it
> > > embeds the drm_framebuffer directly into ast_fbdev and (almost all)
> > > other drivers dynamically allocate and reference count theirs.
> > >
> > > The drm_framebuffer_get() certainly looks weird but it is there in order
> > > to cause drm_framebuffer_remove() to call legacy_remove_fb(), which it
> > > won't do unless the refcount is at least 2. (And because the
> > > drm_framebuffer isn't dynamically allocated in this case we don't really
> > > care about the reference count anyway.)
> > >
> > > An alternative might be to call legacy_remove_fb() directly, but it's
> > > declared static.  Do you think it would be better to expose it and call
> > > it directly from the AST driver code? Or is there some other better way
> > > to put the drm_connectors?
> > 
> > Your problem isn't the dynamic fb vs. embedded fb for fbdev (you're
> > already using drm_framebuffer_unregister_private to handle that). Your
> > problem is you're not shutting down stuff on driver unload, which
> > means the fb is still in use. drm_atomic_helper_shutdown() takes care
> > of that for atomic drivers.
> > 
> > No idea anymore what to do for legacy code, probably need to open code
> > a shutdown sequence. Definitely not the above.
> > -Daniel
> 
> Well, it looks like drm_crtc_force_disable_all() would also do the job,
> and from looking at nouveau_display_fini() it's used there as an
> alternative to drm_atomic_helper_shutdown().

Ah right, I tried looking for that one but didn't find it with a quick
scan.
 
> Would it be reasonable to call that at the start of
> ast_fbdev_destroy() instead? (Testing shows that it does allow the
> drm_connector to be released. Is it enough/correct though?)

Yes.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch