On Wed, 29 May 2019 15:20:52 +0200
Philipp Zabel <p.zabel@xxxxxxxxxxxxxx> wrote:
> Hi Boris,
>
> thank you for the review.
>
> On Wed, 2019-05-29 at 13:46 +0200, Boris Brezillon wrote:
> > On Wed, 29 May 2019 11:54:20 +0200
> > Philipp Zabel <p.zabel@xxxxxxxxxxxxxx> wrote:
> >
> > > Add support for multiple register ranges with SoC specific names.
> > >
> > > Signed-off-by: Philipp Zabel <p.zabel@xxxxxxxxxxxxxx>
> > > ---
> > > drivers/staging/media/hantro/hantro.h | 7 ++++++-
> > > drivers/staging/media/hantro/hantro_drv.c | 25 +++++++++++++++++------
> > > 2 files changed, 25 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/drivers/staging/media/hantro/hantro.h b/drivers/staging/media/hantro/hantro.h
> > > index 6b90fe48bcdf..b796867808d5 100644
> > > --- a/drivers/staging/media/hantro/hantro.h
> > > +++ b/drivers/staging/media/hantro/hantro.h
> > > @@ -27,6 +27,7 @@
> > >
> > > #define HANTRO_MAX_CLOCKS 4
> > > #define HANTRO_MAX_IRQS 3
> > > +#define HANTRO_MAX_REG_RANGES 4
> > >
> > > #define MPEG2_MB_DIM 16
> > > #define MPEG2_MB_WIDTH(w) DIV_ROUND_UP(w, MPEG2_MB_DIM)
> > > @@ -63,6 +64,8 @@ struct hantro_codec_ops;
> > > * @num_irqs: number of irqs in the arrays
> > > * @clk_names: array of clock names
> > > * @num_clocks: number of clocks in the array
> > > + * @reg_names: array of register range names
> > > + * @num_regs: number of register range names in the array
> > > */
> > > struct hantro_variant {
> > > unsigned int enc_offset;
> > > @@ -80,6 +83,8 @@ struct hantro_variant {
> > > int num_irqs;
> > > const char *clk_names[HANTRO_MAX_CLOCKS];
> > > int num_clocks;
> > > + const char *reg_names[HANTRO_MAX_REG_RANGES];
> > > + int num_regs;
>
> Do you suggest
> const char * const *reg_names;
> ...
Yes.
>
> > > };
> > >
> > > /**
> > > @@ -170,7 +175,7 @@ struct hantro_dev {
> > > struct platform_device *pdev;
> > > struct device *dev;
> > > struct clk_bulk_data clocks[HANTRO_MAX_CLOCKS];
> > > - void __iomem *base;
> > > + void __iomem *base[HANTRO_MAX_REG_RANGES];
> >
> > Same comment as for the irq stuff.
>
> ... and
> void __iomem **base;
> to get rid of HANTRO_MAX_REG_RANGES?
This one would have to be dynamically allocated, but yes.
>
> Would you like to see the same for clk_names?
It'd be better, indeed.
>
> > > void __iomem *enc_base;
> > > void __iomem *dec_base;
> > >
> > > diff --git a/drivers/staging/media/hantro/hantro_drv.c b/drivers/staging/media/hantro/hantro_drv.c
> > > index f677b40bcd2d..bd02b27258e3 100644
> > > --- a/drivers/staging/media/hantro/hantro_drv.c
> > > +++ b/drivers/staging/media/hantro/hantro_drv.c
> > > @@ -692,12 +692,25 @@ static int hantro_probe(struct platform_device *pdev)
> > > if (ret)
> > > return ret;
> > >
> > > - res = platform_get_resource(vpu->pdev, IORESOURCE_MEM, 0);
> > > - vpu->base = devm_ioremap_resource(vpu->dev, res);
> > > - if (IS_ERR(vpu->base))
> > > - return PTR_ERR(vpu->base);
> > > - vpu->enc_base = vpu->base + vpu->variant->enc_offset;
> > > - vpu->dec_base = vpu->base + vpu->variant->dec_offset;
> > > + if (vpu->variant->num_regs) {
> > > + for (i = 0; i < vpu->variant->num_regs; i++) {
> > > + const char *reg_name = vpu->variant->reg_names[i];
> > > +
> > > + res = platform_get_resource_byname(vpu->pdev,
> > > + IORESOURCE_MEM,
> > > + reg_name);
> > > + vpu->base[i] = devm_ioremap_resource(vpu->dev, res);
> > > + if (IS_ERR(vpu->base[i]))
> > > + return PTR_ERR(vpu->base[i]);
> > > + }
> > > + } else {
> > > + res = platform_get_resource(vpu->pdev, IORESOURCE_MEM, 0);
> > > + vpu->base[0] = devm_ioremap_resource(vpu->dev, res);
> > > + if (IS_ERR(vpu->base[0]))
> > > + return PTR_ERR(vpu->base[0]);
> > > + vpu->enc_base = vpu->base[0] + vpu->variant->enc_offset;
> > > + vpu->dec_base = vpu->base[0] + vpu->variant->dec_offset;
> >
> > I see ->dec_based is assigned in ->hw_init() in patch 8, so maybe it's
> > better to have the same workflow for rk variants: assign
> > vpu->{dec,enc}_base in ->hw_init()
>
> I didn't want to change this around too much, as dec_base is just needed
> for the vdpu_read/write functions, and I expect we'll have to somehow
> replace these anyway when adding G2 support.
If G1 and G2 blocks are completely independent I think they should be
represented as separate instances, and we can then re-use the same
accessors. Am I missing something?
> Adding yet another set of register accessors for g1_read/write vs
> g2_read/write isn't very convenient. Maybe it woudl be better to call
> the register accessors with the base as a parameter instead of
> hantro_dev.
Looks like the reg base is actually per functionality (decoder or
encoder) not per device. Maybe we should stop passing hantro_dev
directly and instead expose hantro_func objects that would have a
pointer to the underlying hantro_dev plus extra attributes like a base
address for regs, ....
That might help cope with the single-instance+multi-func vs
single-instance-single-func difference we have between rk and imx
integration, and we would also have a single set of read/write
accessors.
>
> Also the kerneldoc comment says .init() should "initialize hardware".
> Should that be changed to "variant specific initialization" if the
> enc/dec_base are set there?
>
> > and set ->num_regs to 1 (plus a
> > fallback to platform_get_resource() instead of
> > platform_get_resource_byname() when ->reg_names[0] == NULL).
>
> I suppose we could do that, but
>
> static const char * const rk3288_regs[] = {
> NULL
> }
>
> const struct hantro_variant rk3288_vpu_variant = {
> .reg_n
> ames = rk3288_regs,
> .num_regs = ARRAY_SIZE(rk3288_regs)
> };
>
> would look a bit strange if we were to get rid of
> HANTRO_MAX_REG_RANGES...
You're right, we can probably stay with num_regs = 0 for the unnamed
mem-resource case.
