Re: [PATCH v3 09/16] reset: starfive-jh7100: Add StarFive JH7100 reset driver

[Yury Norov <yury.norov@xxxxxxxxx>]

On Tue, Nov 2, 2021 at 12:59 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:
>
> On Tue, 2 Nov 2021 at 20:43, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> > +Cc: Yury (bitmap expert)
> >
> > On Tue, Nov 2, 2021 at 6:50 PM Emil Renner Berthing <kernel@xxxxxxxx> wrote:
> > >
> > > Add a driver for the StarFive JH7100 reset controller.
> >
> > ...
> >
> > > +#define BIT_MASK32(x) BIT((x) % 32)
> >
> > Possible namespace collision.
> >
> > ...
> >
> > > +/*
> > > + * the registers work like a 32bit bitmap, so writing a 1 to the m'th bit of
> > > + * the n'th ASSERT register asserts line 32n + m, and writing a 0 deasserts the
> > > + * same line.

We don't have 32-bit bitmaps. Bitmaps are always arrays of unsigned longs. On a
64-bit system this '32-bit bitmap' may be broken due to endianness issues.

> > > + * most reset lines have their status inverted so a 0 in the STATUS register
> > > + * means the line is asserted and a 1 means it's deasserted. a few lines don't
> > > + * though, so store the expected value of the status registers when all lines
> > > + * are asserted.
> > > + */
> >
> > Besides missing capitalization,
>
> I'm confused. it was you who wanted all comments to capitalized the same..
> 64bi
> if it sounds like bitmap, use bitmap.
> > I have checked DT definitions and it seems you don't even need the
> > BIT_MASK() macro,
> >
> > > +static const u32 jh7100_reset_asserted[4] = {
> > > +       /* STATUS0 register */
> > > +       BIT_MASK32(JH7100_RST_U74) |

I think we have no BIT_MASK32() for a good reason. Natural alignment is
always preferable.

> > > +       BIT_MASK32(JH7100_RST_VP6_DRESET) |
> > > +       BIT_MASK32(JH7100_RST_VP6_BRESET),
> > > +       /* STATUS1 register */
> > > +       BIT_MASK32(JH7100_RST_HIFI4_DRESET) |
> > > +       BIT_MASK32(JH7100_RST_HIFI4_BRESET),
> > > +       /* STATUS2 register */
> > > +       BIT_MASK32(JH7100_RST_E24),
> > > +       /* STATUS3 register */
> > > +       0,
> > > +};
> >
> > Yury, do we have any clever (clean) way to initialize a bitmap with
> > particular bits so that it will be a constant from the beginning? If
> > no, any suggestion what we can provide to such users?

If you want your array to be a true bitmap, ie, all bitmap functions should
work with it correctly, you'd initialize it like this:

static const unsigned long jh7100_reset_asserted[] = {
        BITMAP_FROM_U64(BIT_MASK(JH7100_RST_VP6_DRESET) |
                          BIT_MASK(JH7100_RST_VP6_BRESET) |
                          BIT_MASK(JH7100_RST_HIFI4_DRESET) |
                          BIT_MASK(JH7100_RST_HIFI4_BRESET)),
        BITMAP_FROM_U64(BIT_MASK(JH7100_RST_E24)),
}

Look at lib/test_bitmap.c for example, and comment to BITMAP_FROM_U64()
for internal details.

On the other hand, if you hardware has tricky requirements for bit
positions, and they should depend on endianness and/or size of
long in a way not compatible with bitmaps, you probably know better
how to handle this. Just don't refer to your structure as a bitmap.

Thanks,
Yury

> The problem is, that even if we could initialize this without the
> monstrosity in our last conversation a 64bit bitmap would still
> produce worse code. As it is now it's simply a 32bit load and mask
> with index and mask already calculated for the registers. In the
> status callback the mask can even be folded into the register read
> mask. With a 64bit bitmap you'd need to calculate new 64bit index and
> masks, and then conditionally shift the bits into position.
>
> If this reflection of the 32bit registers bothers you that much we
> could alternatively do
>
> static bool jh7100_reset_inverted(unsigned int idx)
> {
>   switch (idx) {
>   case JH7100_RST_U74:
>   case JH7100_RST_VP6_DRESET:
>   ..
>     return false;
>   }
>   return true;
> }
>
> It'd still produce worse code, but at least it would be readable.
>
> /Emil
>
> > ...
> >
> > > +       dev_dbg(rcdev->dev, "reset(%lu)\n", id);
> >
> > These debug messages are useless since one should use ftrace facility instead,
> >
> > --
> > With Best Regards,
> > Andy Shevchenko