Hi Michael,
 
 On ven., juil. 08 2016, Michael Turquette [off-list ref] wrote:

Quoting Gregory CLEMENT (2016-07-07 15:37:51)

quoted

+#include <linux/clk-provider.h>
+#include <linux/clk.h>

Same question as my previous email. Is clk.h necessary? Is this driver
also a clk consumer?

I think I can remove it indeed.

quoted

+static int armada_3700_add_composite_clk(const struct clk_periph_data *data,
+                                        const char * const *parent_name,
+                                        void __iomem *reg, spinlock_t *lock,
+                                        struct device *dev, struct clk_hw *hw)
+{
+       const struct clk_ops *mux_ops = NULL, *gate_ops = NULL,
+               *div_ops = NULL;
+       struct clk_hw *mux_hw = NULL, *gate_hw = NULL, *div_hw = NULL;
+       const char * const *names;
+       struct clk_mux *mux = NULL;
+       struct clk_gate *gate = NULL;
+       struct clk_divider *div = NULL;
+       struct clk_double_div *double_div = NULL;
+       int num_parent;
+       int ret = 0;
+
+       if (data->gate_shift != UNUSED) {
+               gate = devm_kzalloc(dev, sizeof(*gate), GFP_KERNEL);
+
+               if (!gate)
+                       return -ENOMEM;
+
+               gate->reg = reg + CLK_DIS;
+               gate->bit_idx = data->gate_shift;
+               gate->lock = lock;
+               gate_ops = &clk_gate_ops;
+               gate_hw = &gate->hw;
+       }
+
+       if (data->mux_shift != UNUSED) {
+               mux = devm_kzalloc(dev, sizeof(*mux), GFP_KERNEL);
+
+               if (!mux) {
+                       ret = -ENOMEM;
+                       goto free_gate;
+               }
+
+               mux->reg = reg + TBG_SEL;
+               mux->shift = data->mux_shift;
+               mux->mask = 0x3;
+               mux->lock = lock;
+               mux_ops = &clk_mux_ro_ops;
+               mux_hw = &mux->hw;
+       }
+
+       if (data->div_reg1 != UNUSED) {
+               if (data->div_reg2 == UNUSED) {
+                       const struct clk_div_table *clkt;
+                       int table_size = 0;
+
+                       div = devm_kzalloc(dev, sizeof(*div), GFP_KERNEL);
+                       if (!div) {
+                               ret = -ENOMEM;
+                               goto free_mux;
+                       }
+
+                       div->reg = reg + data->div_reg1;
+                       div->table = data->table;
+                       for (clkt = div->table; clkt->div; clkt++)
+                               table_size++;
+                       div->width = order_base_2(table_size);
+                       div->lock = lock;
+                       div_ops = &clk_divider_ro_ops;
+                       div_hw = &div->hw;
+               } else {
+                       double_div = devm_kzalloc(dev, sizeof(*double_div),
+                                                 GFP_KERNEL);
+                       if (!double_div) {
+                               ret = -ENOMEM;
+                               goto free_mux;
+                       }
+
+                       double_div->reg1 = reg + data->div_reg1;
+                       double_div->shift1 = data->div_shift1;
+                       double_div->reg2 = reg + data->div_reg1;
+                       double_div->shift2 = data->div_shift2;
+                       div_ops = &clk_double_div_ops;
+                       div_hw = &double_div->hw;
+               }
+       }
+
+       switch (data->flags) {
+       case XTAL_CHILD:
+               /* the xtal clock is the 5th clock */
+               names = &parent_name[4];
+               num_parent = 1;
+               break;
+       case TBGA_S_CHILD:
+               /* the TBG A S clock is the 3rd clock */
+               names = &parent_name[2];
+               num_parent = 1;
+               break;
+       case GBE_CORE_CHILD:
+               names = &gbe_name[1];
+               num_parent = 1;
+               break;
+       case  GBE_50_CHILD:
+               names = &gbe_name[0];
+               num_parent = 1;
+               break;
+       case  GBE_125_CHILD:
+               names = &gbe_name[2];
+               num_parent = 1;
+               break;
+       default:
+               names = parent_name;
+               num_parent = 4;
+       }
+       hw = clk_hw_register_composite(dev, data->name,
+                                    names, num_parent,
+                                    mux_hw, mux_ops,
+                                    div_hw, div_ops,
+                                    gate_hw, gate_ops,
+                                    CLK_IGNORE_UNUSED);
+       if (IS_ERR(hw)) {
+               ret = PTR_ERR(hw);
+               goto free_div;
+       }
+
+       return 0;
+free_div:
+       devm_kfree(dev, div);
+       devm_kfree(dev, double_div);
+free_mux:
+       devm_kfree(dev, mux);
+free_gate:
+       devm_kfree(dev, gate);
+       return ret;
+}

Can this "add" function (aka registration function) be replaced with
static data instead? I think that all of the static data exists already,
this function can be removed and your probe can call clk_hw_register
directly.

I see your point and indeed we can remove some allocation. However using
clk_hw_register with composite clock is not straight forward. Indeed the
clk_ops is filled in the clk_hw_register_composite function and none of
these operations are exported outside the clk-composite.c file.

We can't directly point a clk_gate_ops structure or a clk_divider_ops as
you did in the drivers/clk/meson/gxbb.c file for example. For clk
composite it would need modify the framework to either export all the
operation or to create set of operations directly usable and I would
like to avoid doing it when we are close to the merge window.

However I can use static data for the clk_mux, clk_gate and clk_divider
struct.

Is it OK for you?

Thanks,

Gregory

It might need a macro though, since composite clock structures are
rather messy. This avoids a lot of unnecessary allocations and time
populating data that we already have access to. In general I am trying
to encourage clk drivers to use only clk_hw_register() in their probe
instead of the helper registration functions.

Similarly I am discouraging drivers from populating hw.init at run-time,
since we already have that data for that at compile-time.

Regards,
Mike

-- 
Gregory Clement, Free Electrons
Kernel, drivers, real-time and embedded Linux
development, consulting, training and support.
http://free-electrons.com