[PATCH v2] drm/panfrost: Document base field location constraint in panfrost_gem_object

[Boris Brezillon]

On Fri, 29 Nov 2019 21:43:45 +0200
Ville Syrjälä <ville.syrjala at linux.intel.com> wrote:

> On Fri, Nov 29, 2019 at 08:24:37PM +0100, Boris Brezillon wrote:
> > On Fri, 29 Nov 2019 19:40:38 +0100
> > Daniel Vetter <daniel at ffwll.ch> wrote:
> >   
> > > On Fri, Nov 29, 2019 at 03:19:36PM +0100, Boris Brezillon wrote:  
> > > > On Fri, 29 Nov 2019 14:13:33 +0000
> > > > Steven Price <steven.price at arm.com> wrote:
> > > >     
> > > > > On 29/11/2019 13:56, Boris Brezillon wrote:    
> > > > > > I've spent hours chasing a memory corruption that was caused by
> > > > > > insertion of an extra field field before ->base. Let's document the
> > > > > > fact that base has to be the first field in panfrost_gem_object.
> > > > > > 
> > > > > > Signed-off-by: Boris Brezillon <boris.brezillon at collabora.com>      
> > > > > 
> > > > > This seems to be a limitation imposed by the gem_create_object()
> > > > > callback - e.g. it's assumed that kfree() can be directly called on the
> > > > > result. Useful to have the documentation though.    
> > > > 
> > > > Oh, you're right, I didn't catch that one.    
> > > 
> > > As a general rule of thumb, never insert anything before a struct member
> > > called base. Even more so if it's of the same kind of $thing, but less
> > > spezialed. This pattern is so common it's fairly often not documented
> > > anywhere.  
> > 
> > I could argue that anything using container_of() in its to_xxx() helper
> > is misleading the user into thinking the position of the base field
> > doesn't matter (which is exactly what happened here), but I feel like I
> > already lost this battle, so I'll simply drop the patch.  
> 
> I agree that this is a bit annoying. Apart from the "let's pass
> the wrong thing to kfree()"

This is already a disputable design choice IMHO. Why should you
delegate the free() to someone who didn't allocate the thing in the
first place.

> issue the other problem is NULL
> pointers. If those don't get preserved when going between the
> base and derived class the code will turn to bad spaghetti.
> 
> Not a problem for the typical to_foo_crtc() since we could
> hide an explicit NULL check in there. But the other direction
> generally just uses &foo_crtc->base so not going to work.

Well, if you inherit from the base struct, you'll inevitably have
NULL checks because of pointer dereferences you have when accessing
other fields, so I'd expects most of them to be already present (not
saying finding the remaining ones is an easy task).

The real problem I see here is that we re-use helpers manipulating base
objects without providing wrappers, and that can only work if base is
the first field in the child struct.

> Shouldn't be an impossible task to wrap all of those as well,
> but I guess no one's motivated enough to actually do it.
> 

Actually, I never suggested to do any of that, because I know how
invasive/risky such a change would be. All I was trying to do was make
an implicit requirement more explicit with a comment stating the fact
and explaining why (BTW, I like your suggestion to use BUILD_BUG_ON()
to catch those problems at compilation time).

Anyway, enough on that matter. I'll try to remember that the use of
container_of() does not necessarily means fields can be shuffled.