Re: [PATCH] sh: Implement __get_user_u64() required for 64-bit get_user()

[Geert Uytterhoeven]

Hi Rich,

On Mon, Jun 1, 2020 at 5:03 AM Rich Felker <dalias@xxxxxxxx> wrote:
> On Sun, May 31, 2020 at 12:43:11PM +0200, Geert Uytterhoeven wrote:
> > On Sun, May 31, 2020 at 11:59 AM John Paul Adrian Glaubitz
> > <glaubitz@xxxxxxxxxxxxxxxxxxx> wrote:
> > > On 5/31/20 11:54 AM, John Paul Adrian Glaubitz wrote:
> > > > On 5/31/20 11:52 AM, Geert Uytterhoeven wrote:
> > > >> As this is the 64-bit variant, I think this single move should be
> > > >> replaced by a double move:
> > > >>
> > > >>        "mov  #0,%R1\n\t" \
> > > >>        "mov  #0,%S1\n\t" \
> > > >>
> > > >> Same for the big endian version below.
> > > >>
> > > >> Disclaimer: uncompiled, untested, no SH assembler expert.
> > > >
> > > > Right, this makes sense. I'll send a new patch shortly.
> > >
> > > Hmm, this change is not the case for __put_user_asm() vs. __put_user_u64().
> > > But I have to admit, I don't know what the part below "3:\n\t" is for.
> >
> > It's part of the exception handling, in case the passed (userspace) pointer
> > points to an inaccessible address, and triggers an exception.
> >
> > For an invalid store, nothing is done, besides returning -EFAULT.
> > Hence there's no "mov #0, %1\n\t" in the put_user case.
> > For an invalid load, the data is replaced by zero, and -EFAULT is returned.
> >
> > > +__asm__ __volatile__( \
> > > +       "1:\n\t" \
> > > +       "mov.l  %2,%R1\n\t" \
> > > +       "mov.l  %T2,%S1\n\t" \
> > > +       "2:\n" \
> >
> > (reordering the two sections for easier explanation)
> >
> > > +       ".section       __ex_table,\"a\"\n\t" \
> > > +       ".long  1b, 3b\n\t" \
> >
> > In case an exception happens for the instruction at 1b, jump to 3b.
> >
> > Note that the m68k version has two entries here: one for each half of
> > the 64-bit access[*].
> > I don't know if that is really needed (and thus SH needs it, too), or if
> > the exception code handles subsequent instructions automatically.
>
> Can I propose a different solution? For archs where there isn't
> actually any 64-bit load or store instruction, does it make sense to
> be writing asm just to do two 32-bit loads/stores, especially when
> this code is not in a hot path?
>
> What about just having the 64-bit versions call the corresponding
> 32-bit version twice? (Ideally this would even be arch-generic and
> could replace the m68k asm.) It would return EFAULT if either of the
> 32-bit calls did.

Yes, that's an option, too.

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds