On Wed, Jan 16, 2019 at 10:59 AM Luc Van Oostenryck
<luc.vanoostenryck@xxxxxxxxx> wrote:
> *) I've always considered __noderef as really meaning "must not be
> dereferenced except by some special purpose operation which
> effectively do the real deref". In the present case (percpu &
> uaccess (on x86)), it's clear that the special operation is
> the asm memory op but isn't this essentially incidental?
I agree that it's somewhat incidental rather than anything hugely by design.
That said, the *design* of __noderef was literally to make sure that
the *compiler* never derefences things, and the "special thing" that
does dereference was literally meant to be an inline asm. The
"{get,put}_user()" case was exactly what __noderef was designed for.
So it *is* very much by design, but then the implementation not
*checking* inline asm arguments was kind of an incidental thing.
Which is why I very much admit that it's annoying.
The good news is that the asm constraints aren't really all that
architecture-specific. Yes, all the various register classes etc
obviously are, but within the concept of "the compiler never
dereferences it", the only asm constraints that matter are "m" and
"p", I think.
"m" means "memory", and "p" means "address". They are kind of the
same, except for a gcc memory access ordering situation, I think.
> Thinking a bit more about it and looking at some dumps I made:
> It seems that most of these these 'noderef in asm operand'
> warnings (99.5%) come from percpu operations (which have an API
> using an lvalue and not a pointer, unlike the uaccess API).
>
> I begin to suspect that the warnings occuring with an uaccess
> operation come from some other kind of error (possibly a missing
> call to degenerate()).
Hmm. The uaccess cases definitely pass the lvalue in some cases. Not
the cases where we just end up doing a special "call" sequence (the
normal get_user/put_user() case). But look closer: the cases where we
actaully inline the access itself, we generate a "LKmode" lvalue with
that "__m(addr)" thing, and the actual access is done by the inline
asm.
See __get_user_asm(), __get_user_asm_nozero(), __get_user_asm_ex() and
__put_user_goto().
(Some of this has changed recently, so __put_user_goto() is a new
version of __put_user_asm(), so if you don't look at current kernel
-git, that's what you'll see instead).
Linus
