Re: [PATCH RFC] Fixes to Linus' 'remove set_fs patch'

[Michael Schmitz <schmitzmic@gmail.com>]

Hi Linus,

On 7/07/21 6:47 am, Linus Torvalds wrote:
> On Mon, Jul 5, 2021 at 4:20 PM Michael Schmitz <schmitzmic@gmail.com> wrote:
>> Linus' patch didn't compile out of the box. I'm a little
>> worried that I had to patch mm/maccess.c,
> Yeah, that's not the right way to do it.
>
> See the other thread where Christoph rightly points out that the
> problem is that m68k needs __{get,put}_kernel_nofault() and then just
> enable CONFIG_HAVE_GET_KERNEL_NOFAULT.
Yep, saw that late yesterday and started implementing those.
>
> __{get,put}_kernel_nofault() is basically the similar to
> __get_user/__put_user, except for
>
>   (a) it uses kernel addresses (so in m68k parlance, "move" rather than "moves")
Thanks for reminding me ... after backing out the mm/maccess.c changes, 
new __get/put_kernel_asm() macros and basically copying the 
__get/put_kernel_nofault() framework from mips, I still have a working 
kernel. I'd like to test those changes on actual hardware though (and 
will need help from someone with a 040 or 060 machine for that).
>
>   (b) it uses the "goto err_label" approach instead of being an
> expression that returns an error value
>
> That (b) is what allows the use of "asm goto" to implement the
> exception case much more efficiently (but you can avoid having to
> convert to that model by just doing something like
>
>          if (unlikely(__gu_err)) goto err_label;
>
> inside the macro instead - even architectures that use the "asm goto"
> do that for old compilers that don't support "asm goto" with output
> values.
>
> It shouldn't be too hard to implement on m68k - you'd likely want to
> make the helper macros in <asm/uaccess.h> take that "MOVES" thing as a
> new argument, so that you can generate code for the "move" vs "moves"
> cases using the same helper macros, just with different arguments.

True - let's get some test coverage first though.

Cheers,

     Michael


>
>              Linus