Re: [PATCH v2] kernel.h: handle pointers to arrays better in container_of()

[Ian Abbott]

On 23/05/17 14:23, Ian Abbott wrote:
> On 23/05/17 13:02, Ian Abbott wrote:
> > On 23/05/17 12:24, Peter Zijlstra wrote:
> > > On Tue, May 23, 2017 at 11:32:02AM +0100, Ian Abbott wrote:
> > >
> > > > > #define container_of(ptr, type, member) (                \
> > > > >     _Static_assert(__builtin_types_compatible_p(            \
> > > > >             typeof(*ptr), typeof( ((type *)0)->member )), "WUT"),    \
> > > > >     ((type *)((char *)(ptr) - offsetof(type, member)));        \
> > > > > )
> > > >
> > > > It's a fine suggestion (if more parentheses are added), but
> > > > _Static_assert
> > > > is a C11 feature, and I thought the kernel was using gnu89 (unless
> > > > it's been
> > > > updated since).
> > >
> > > We have BUILD_BUG_ON() that should be similar in functionality.
> >
> > Yes, I think BUILD_BUG_ON_ZERO() will be better in this case to avoid
> > some baggage.  Also, __same_type() can be used for type checking.
>
> BUILD_BUG_ON_ZERO() in the left-hand side of a comma expression produced
> lots of warnings about it having no effect, so I tried using
> BUILD_BUG_ON() instead as follows:
>
> #define container_of(ptr, type, member) ({            \
>     BUILD_BUG_ON(!__same_type((ptr), &((type *)0)->member));\
>     ((type *)((char *)(ptr) - offsetof(type, member)));})
>
> Unfortunately, this does result in compiler failures in places where the
> types do not in fact match, for example `slab_show()` in
> "mm/slab_common.c" (`list_entry()` uses `container_of()`):
>
> static int slab_show(struct seq_file *m, void *p)
> {
>     struct kmem_cache *s = list_entry(p, struct kmem_cache,
> root_caches_node);
>
>
> I then changed it to allow `ptr` to be a `void *` too:
>
> #define container_of(ptr, type, member) ({            \
>     BUILD_BUG_ON(!__same_type((ptr), &((type *)0)->member) &&\
>              !__same_type((ptr), void *));        \
>     ((type *)((char *)(ptr) - offsetof(type, member)));})
>
> Now it fails on `external_name()` in "fs/dcache.c":
>
> static inline struct external_name *external_name(struct dentry *dentry)
> {
>     return container_of(dentry->d_name.name, struct external_name,
> name[0]);
> }
>
> Here, `dentry->d_name.name` is of type `const unsigned char *` and
> `name[0]` in `struct external_name` is of type `unsigned char`.  The
> error is:
>
> error: call to ‘__compiletime_assert_258’ declared with attribute error:
> BUILD_BUG_ON failed: !__same_type((dentry->d_name.name), &((struct
> external_name *)0)->name[0]) && !__same_type((dentry->d_name.name), void *)
>   _compiletime_assert(condition, msg, __compiletime_assert_, __LINE__)
>                                       ^
>
> I'm not quite sure why that is failing yet.

It fails because `const unsigned char *` is not compatible with 
`unsigned char *` as fast as `__builtin_types_compatible_p()` is 
concerned (quite rightly so).  However, a simple change fixes that:

#define container_of(ptr, type, member) ({				\
	BUILD_BUG_ON(!__same_type(*(ptr), ((type *)0)->member) &&	\
		     !__same_type((ptr), void *));			\
	((type *)((char *)(ptr) - offsetof(type, member)));})

That one seems to work fine (after including <linux/bug.h> in kernel.h 
and fixing up include/asm-generic/bug.h a little bit, which will be in a 
separate patch.

Thanks for the help guys.  v3 will be a series of 2 patches.

--
-=( Ian Abbott @ MEV Ltd.    E-mail: <abbotti@xxxxxxxxx> )=-
-=(                          Web: http://www.mev.co.uk/  )=-