Re: [PATCH 7/8] percpu: add __percpu sparse annotations to hw_breakpoint

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Jan 26, 2010 at 03:01:14AM +0100, Frederic Weisbecker wrote:
> On Tue, Jan 26, 2010 at 10:19:04AM +0900, Tejun Heo wrote:
> > Hello,
> > 
> > On 01/26/2010 10:02 AM, Frederic Weisbecker wrote:
> > > Well, sorry I must be missing something obvious, but is it impossible
> > > to make per_cpu(var, cpu) returning something cast in:
> > > 
> > > 	(typeof(var) __force)
> > > 
> > > Or I guess you did that already and it is not working with static
> > > arrays, or?
> > 
> > Yeap, the definition looks like
> > 
> >  #define SHIFT_PERCPU_PTR(__p, __offset)	({			\
> > 	__verify_pcpu_ptr((__p));					\
> > 	RELOC_HIDE((typeof(*(__p)) __kernel __force *)(__p), (__offset)); \
> >  })
> > 
> >  #define per_cpu(var, cpu) \
> > 	(*SHIFT_PERCPU_PTR(&(var), per_cpu_offset(cpu)))
> > 
> > but it just ends up putting the __force at the wrong layer.  It seems
> > that (typeof(var) __kernel __force) tell sparse var is in the kernel
> > address space but not its members.
> 
> 
> So, may be it considers you are applying the address space overriding
> to the pointer to the type and not to the type itself.
> 
> Consider:
> 
> 	int __percpu i;
> 
> What you do above *might* be considered as if SHIFT_PERCPU_PTR
> returns something of a type:
> 
> 	int * __percpu i;
> 
> So the pointer is in the normal address space, but its content is in
> __percpu address space.
> 
> What if you do this:
> 
> 
> #define SHIFT_PERCPU_PTR(__p, __offset)      ({                      \
>       __verify_pcpu_ptr((__p));                                       \
>       RELOC_HIDE((__p), (__offset)); \
> })
> 
> #define per_cpu(var, cpu) \
>       (typeof(var) __kernel __force)(*SHIFT_PERCPU_PTR(&(var), per_cpu_offset(cpu)))
> 
> This should work because &(var) should be dereferencable directly, since
> it is not of type "__force t" but of type "*__force t"
> 
> And you're not doing anymore this:
> 
> 	*(int * __kernel __force) i;
> but
> 	*(int __kernel __force *) i;



The above is perhaps a bit confusing.
To be more clear, in the first case you only cast the pointer
to the type, which gives you a pointer valid in kernel space
to data valid in percpu space.

The second case gives you something valid in kernel space for both.