Re: [PATCH] m68knommu: remove set_fs()

On Mon, Jul 05, 2021 at 01:39:03PM -0700, Linus Torvalds wrote:
> On Mon, Jul 5, 2021 at 1:46 AM Geert Uytterhoeven <[email protected]> wrote:
> >
> > Probably this should be
> >
> >     select SET_FS if CPU_HAS_ADDRESS_SPACES
> 
> Actually, I don't think m68k has a single real "set_fs()" at all, and
> it should just be converted as-is to not use CONFIG_SET_FS.
> 
> Yes, there is a "set_fs()" function, but none of the remaining uses
> actually are the traditional kernel style of "use kernel addresses as
> user addresses". So as far as the *kernel* is concerned, m68k already
> looks like a no-SET_FS architecture, and "set-fs()" is purely a
> syntactic thing.

It still needs "real" kernel-style set_fs for the mm/maccess.c routines,
but adding __{get,put}_kernel_nofault should not be too hard.

> So I think the right thing to do looks something like this:
> 
>  - make the rule be that SFC/DFC is always normally USER_DATA
> 
>  - the special m68k sequences that need to play with special segments
> will always do
> 
>         preempt_disable();
>         set_segment(..whatever segment they need..);
>         .. do the special operation ..
>         set_segment(USER_DATA);
>         preempt_enable();
> 
>  - set_fs() goes away entirely, because the user access functions
> always work on USER_DATA and SFC/DFC is always right for them.

Yes, that's what I mean with needing a more work.