Quoting Michael Tretter (2020-12-15 03:38:09)
> On Sat, 12 Dec 2020 21:55:34 -0800, Stephen Boyd wrote:
> > Quoting Michael Tretter (2020-11-15 23:55:28)
> > > diff --git a/drivers/soc/xilinx/xlnx_vcu.c b/drivers/soc/xilinx/xlnx_vcu.c
> > > index 725e646aa726..cedc8b7859f7 100644
> > > --- a/drivers/soc/xilinx/xlnx_vcu.c
> > > +++ b/drivers/soc/xilinx/xlnx_vcu.c
> > > @@ -545,6 +512,146 @@ static int xvcu_set_pll(struct xvcu_device *xvcu)
> > > return xvcu_pll_enable(xvcu);
> > > }
> > >
> > > +static struct clk_hw *xvcu_clk_hw_register_leaf(struct device *dev,
> > > + const char *name,
> > > + const char * const *parent_names,
> > > + u8 num_parents,
> > > + struct clk_hw *parent_default,
> > > + void __iomem *reg,
> > > + spinlock_t *lock)
> > > +{
> > > + unsigned long flags = CLK_SET_RATE_NO_REPARENT | CLK_SET_RATE_PARENT;
> > > + u8 divider_flags = CLK_DIVIDER_ONE_BASED | CLK_DIVIDER_ALLOW_ZERO |
> >
> > Why u8?
>
> __clk_hw_register_divider expects u8 as divider_flags.
Ok.
>
> >
> > > + CLK_DIVIDER_ROUND_CLOSEST;
> > > + struct clk_hw *mux = NULL;
> > > + struct clk_hw *divider = NULL;
> > > + struct clk_hw *gate = NULL;
> > > + char *name_mux;
> > > + char *name_div;
> > > + int err;
> > > +
> > > + name_mux = devm_kasprintf(dev, GFP_KERNEL, "%s%s", name, "_mux");
> > > + if (!name_mux) {
> > > + err = -ENOMEM;
> > > + goto err;
> > > + }
> > > + mux = clk_hw_register_mux(dev, name_mux, parent_names, num_parents,
> > > + flags, reg, 0, 1, 0, lock);
> > > + if (IS_ERR(mux)) {
> > > + err = PTR_ERR(divider);
> > > + goto err;
> > > + }
> > > + clk_hw_set_parent(mux, parent_default);
> >
> > Can this be done from assigned-clock-parents binding?
>
> Could be done, if the driver provides at least the PLL and the mux in addition
> to the actual output clocks. Otherwise, it is not possible to reference the
> PLL post divider and the mux from the device tree. I wanted to avoid to expose
> the complexity to the device tree. Would you prefer, if all clocks are
> provided instead of only the output clocks?
It is fine to do this in software too so not a big deal and no need to
expose it from the device.
>
> >
> > > +
> > > +static int xvcu_register_clock_provider(struct xvcu_device *xvcu)
> > > +{
> > > + struct device *dev = xvcu->dev;
> > > + const char *parent_names[2];
> > > + struct clk_hw *parent_default;
> > > + struct clk_hw_onecell_data *data;
> > > + struct clk_hw **hws;
> > > + void __iomem *reg_base = xvcu->vcu_slcr_ba;
> > > +
> > > + data = devm_kzalloc(dev, struct_size(data, hws, CLK_XVCU_NUM_CLOCKS), GFP_KERNEL);
> > > + if (!data)
> > > + return -ENOMEM;
> > > + data->num = CLK_XVCU_NUM_CLOCKS;
> > > + hws = data->hws;
> > > +
> > > + xvcu->clk_data = data;
> > > +
> > > + parent_default = xvcu->pll;
> > > + parent_names[0] = "dummy";
> >
> > What is "dummy"?
>
> According to the register reference [0], the output clocks can be driven by an
> external clock instead of the PLL, but the VCU Product Guide [1] does not
> mention any ports for actually driving the clock. From my understanding of the
> IP core, this is a clock mux which has a not-connected first parent. Maybe
> someone at Xilinx can clarify, what is happening here.
>
> [0] https://www.xilinx.com/html_docs/registers/ug1087/ug1087-zynq-ultrascale-registers.html
> [1] https://www.xilinx.com/support/documentation-navigation/see-all-versions.html?xlnxproducttypes=IP%20Cores&xlnxipcoresname=v-vcu
>
> What would be a better way to handle this?
>
> >
> > > + parent_names[1] = clk_hw_get_name(parent_default);
> >
> > Can we use the new way of specifying clk parents from DT or by using
> > direct pointers instead of using string names? The idea is to get rid of
> > clk_hw_get_name() eventually.
>
> It should be possible to use the direct pointers, but this really depends on
> how the "dummy" clock is handled.
>
I think if clk_parent_data is used then we can have the binding list the
external clk as a 'clocks' property and then if it's not present in DT
we will know that it can't ever be a parent. So it hopefully "just
works" if clk_parent_data is used.
