On Thu, Sep 06, 2018 at 04:01:28PM -0700, Linus Torvalds wrote:
> On Thu, Sep 6, 2018 at 3:52 PM Luc Van Oostenryck
> <luc.vanoostenryck@xxxxxxxxx> wrote:
> >
> > I suppose that hard-coding these names wouldn't be too bad.
> > Otherwise, I only see something like:
>
> I was thinking that we could do something like actually change sparse
> to track a token instead of a number for the address space, and then
> we'd just make the kernel eventually switch to
>
> #define __user __attribute__((noderef, address_space(user)))
>
> and the address space would just magically change from '1' to 'user'.
>
> The sparse validation tests would then just use a token name, and you
> coulkd just have a warning like
>
> warning: cast removes 'user' address space of expression
>
> and yeah, the kernel warnings would look odd for a while ("cast
> removes '1' address space of expression"), but we'd know that they
> eventually would become more legible as we start relying on a newer
> version of sparse.
>
> Maybe it's not a big deal, and maybe it's too painful to change the
> asn from an integer to a token pointer. I didn't even look at that.
Yes, that would also be possible, of course.
I think it would be a good thing to have a name in the message.
It shouldn't be a big problem to deal with a pointer (most uses would
be unchanged, a few places use things like 'as1 |= sym->ctype.as;'
which will need a bit more care).
OTOH, this would destroy one of my patch to shrink the size of
struct ctype (worth maybe 0.5% of perf), though. I'm also a bit
annoyed about this transition period. I'll think about it.
-- Luc
