clk: bcm281xx: add initial clock framework support

Add code for device tree support of clocks in the BCM281xx family of
SoCs.  Machines in this family use peripheral clocks implemented by
"Kona" clock control units (CCUs).  (Other Broadcom SoC families use
Kona style CCUs as well, but support for them is not yet upstream.)

A BCM281xx SoC has multiple CCUs, each of which manages a set of
clocks on the SoC.  A Kona peripheral clock is composite clock that
may include a gate, a parent clock multiplexor, and zero, one
or two dividers.  There is a variety of gate types, and many gates
implement hardware-managed gating (often called "auto-gating").
Most dividers divide their input clock signal by an integer value
(one or more).  There are also "fractional" dividers which allow
division by non-integer values.  To accomodate such dividers,
clock rates and dividers are generally maintained by the code in
"scaled" form, which allows integer and fractional dividers to
be handled in a uniform way.

If present, the gate for a Kona peripheral clock must be enabled
when a change is made to its multiplexor or one of its dividers.
Additionally, dividers and multiplexors have trigger registers which
must be used whenever the divider value or selected parent clock is
changed.  The same trigger is often used for a divider and
multiplexor, and a BCM281xx peripheral clock occasionally has two
triggers.

The gate, dividers, and parent clock selector are treated in this
code as "components" of a peripheral clock.  Their functionality is
implemented directly--e.g. the common clock framework gate
implementation is not used for a Kona peripheral clock gate.  (This
has being considered though, and the intention is to evolve this
code to leverage common code as much as possible.)

The source code is divided into three general portions:

    drivers/clk/bcm/clk-kona.h
    drivers/clk/bcm/clk-kona.c
        These implement the basic Kona clock functionality,
        including the clk_ops methods and various routines to
        manipulate registers and interpret their values.  This
        includes some functions used to set clocks to a desired
        initial state (though this feature is only partially
        implemented here).

    drivers/clk/bcm/clk-kona-setup.c
        This contains generic run-time initialization code for
        data structures representing Kona CCUs and clocks.  This
        encapsulates the clock structure initialization that can't
        be done statically.  Note that there is a great deal of
        validity-checking code here, making explicit certain
        assumptions in the code.   This is mostly useful for adding
        new clock definitions and could possibly be disabled for
        production use.

    drivers/clk/bcm/clk-bcm281xx.c
        This file defines the specific CCUs used by BCM281XX family
        SoCs, as well as the specific clocks implemented by each.
        It declares a device tree clock match entry for each CCU
        defined.

    include/dt-bindings/clock/bcm281xx.h
        This file defines the selector (index) values used to
        identify a particular clock provided by a CCU.  It consists
        entirely of C preprocessor constants, to be used by both the
        C source and device tree source files.

Signed-off-by: Alex Elder <elder@linaro.org>
Reviewed-by: Tim Kryger <tim.kryger@linaro.org>
Reviewed-by: Matt Porter <mporter@linaro.org>
Acked-by: Mike Turquette <mturquette@linaro.org>
Signed-off-by: Matt Porter <mporter@linaro.org>

1 files changed, 1033 insertions, 0 deletions

diff --git a/drivers/clk/bcm/clk-kona.c b/drivers/clk/bcm/clk-kona.c
new file mode 100644
index 000000000000..e3d339e08309
--- /dev/null
+++ b/drivers/clk/bcm/clk-kona.c
@@ -0,0 +1,1033 @@
+/*
+ * Copyright (C) 2013 Broadcom Corporation
+ * Copyright 2013 Linaro Limited
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation version 2.
+ *
+ * This program is distributed "as is" WITHOUT ANY WARRANTY of any
+ * kind, whether express or implied; without even the implied warranty
+ * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ */
+
+#include "clk-kona.h"
+
+#include <linux/delay.h>
+
+#define CCU_ACCESS_PASSWORD      0xA5A500
+#define CLK_GATE_DELAY_LOOP      2000
+
+/* Bitfield operations */
+
+/* Produces a mask of set bits covering a range of a 32-bit value */
+static inline u32 bitfield_mask(u32 shift, u32 width)
+{
+	return ((1 << width) - 1) << shift;
+}
+
+/* Extract the value of a bitfield found within a given register value */
+static inline u32 bitfield_extract(u32 reg_val, u32 shift, u32 width)
+{
+	return (reg_val & bitfield_mask(shift, width)) >> shift;
+}
+
+/* Replace the value of a bitfield found within a given register value */
+static inline u32 bitfield_replace(u32 reg_val, u32 shift, u32 width, u32 val)
+{
+	u32 mask = bitfield_mask(shift, width);
+
+	return (reg_val & ~mask) | (val << shift);
+}
+
+/* Divider and scaling helpers */
+
+/*
+ * Implement DIV_ROUND_CLOSEST() for 64-bit dividend and both values
+ * unsigned.  Note that unlike do_div(), the remainder is discarded
+ * and the return value is the quotient (not the remainder).
+ */
+u64 do_div_round_closest(u64 dividend, unsigned long divisor)
+{
+	u64 result;
+
+	result = dividend + ((u64)divisor >> 1);
+	(void)do_div(result, divisor);
+
+	return result;
+}
+
+/* Convert a divider into the scaled divisor value it represents. */
+static inline u64 scaled_div_value(struct bcm_clk_div *div, u32 reg_div)
+{
+	return (u64)reg_div + ((u64)1 << div->frac_width);
+}
+
+/*
+ * Build a scaled divider value as close as possible to the
+ * given whole part (div_value) and fractional part (expressed
+ * in billionths).
+ */
+u64 scaled_div_build(struct bcm_clk_div *div, u32 div_value, u32 billionths)
+{
+	u64 combined;
+
+	BUG_ON(!div_value);
+	BUG_ON(billionths >= BILLION);
+
+	combined = (u64)div_value * BILLION + billionths;
+	combined <<= div->frac_width;
+
+	return do_div_round_closest(combined, BILLION);
+}
+
+/* The scaled minimum divisor representable by a divider */
+static inline u64
+scaled_div_min(struct bcm_clk_div *div)
+{
+	if (divider_is_fixed(div))
+		return (u64)div->fixed;
+
+	return scaled_div_value(div, 0);
+}
+
+/* The scaled maximum divisor representable by a divider */
+u64 scaled_div_max(struct bcm_clk_div *div)
+{
+	u32 reg_div;
+
+	if (divider_is_fixed(div))
+		return (u64)div->fixed;
+
+	reg_div = ((u32)1 << div->width) - 1;
+
+	return scaled_div_value(div, reg_div);
+}
+
+/*
+ * Convert a scaled divisor into its divider representation as
+ * stored in a divider register field.
+ */
+static inline u32
+divider(struct bcm_clk_div *div, u64 scaled_div)
+{
+	BUG_ON(scaled_div < scaled_div_min(div));
+	BUG_ON(scaled_div > scaled_div_max(div));
+
+	return (u32)(scaled_div - ((u64)1 << div->frac_width));
+}
+
+/* Return a rate scaled for use when dividing by a scaled divisor. */
+static inline u64
+scale_rate(struct bcm_clk_div *div, u32 rate)
+{
+	if (divider_is_fixed(div))
+		return (u64)rate;
+
+	return (u64)rate << div->frac_width;
+}
+
+/* CCU access */
+
+/* Read a 32-bit register value from a CCU's address space. */
+static inline u32 __ccu_read(struct ccu_data *ccu, u32 reg_offset)
+{
+	return readl(ccu->base + reg_offset);
+}
+
+/* Write a 32-bit register value into a CCU's address space. */
+static inline void
+__ccu_write(struct ccu_data *ccu, u32 reg_offset, u32 reg_val)
+{
+	writel(reg_val, ccu->base + reg_offset);
+}
+
+static inline unsigned long ccu_lock(struct ccu_data *ccu)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&ccu->lock, flags);
+
+	return flags;
+}
+static inline void ccu_unlock(struct ccu_data *ccu, unsigned long flags)
+{
+	spin_unlock_irqrestore(&ccu->lock, flags);
+}
+
+/*
+ * Enable/disable write access to CCU protected registers.  The
+ * WR_ACCESS register for all CCUs is at offset 0.
+ */
+static inline void __ccu_write_enable(struct ccu_data *ccu)
+{
+	if (ccu->write_enabled) {
+		pr_err("%s: access already enabled for %s\n", __func__,
+			ccu->name);
+		return;
+	}
+	ccu->write_enabled = true;
+	__ccu_write(ccu, 0, CCU_ACCESS_PASSWORD | 1);
+}
+
+static inline void __ccu_write_disable(struct ccu_data *ccu)
+{
+	if (!ccu->write_enabled) {
+		pr_err("%s: access wasn't enabled for %s\n", __func__,
+			ccu->name);
+		return;
+	}
+
+	__ccu_write(ccu, 0, CCU_ACCESS_PASSWORD);
+	ccu->write_enabled = false;
+}
+
+/*
+ * Poll a register in a CCU's address space, returning when the
+ * specified bit in that register's value is set (or clear).  Delay
+ * a microsecond after each read of the register.  Returns true if
+ * successful, or false if we gave up trying.
+ *
+ * Caller must ensure the CCU lock is held.
+ */
+static inline bool
+__ccu_wait_bit(struct ccu_data *ccu, u32 reg_offset, u32 bit, bool want)
+{
+	unsigned int tries;
+	u32 bit_mask = 1 << bit;
+
+	for (tries = 0; tries < CLK_GATE_DELAY_LOOP; tries++) {
+		u32 val;
+		bool bit_val;
+
+		val = __ccu_read(ccu, reg_offset);
+		bit_val = (val & bit_mask) != 0;
+		if (bit_val == want)
+			return true;
+		udelay(1);
+	}
+	return false;
+}
+
+/* Gate operations */
+
+/* Determine whether a clock is gated.  CCU lock must be held.  */
+static bool
+__is_clk_gate_enabled(struct ccu_data *ccu, struct bcm_clk_gate *gate)
+{
+	u32 bit_mask;
+	u32 reg_val;
+
+	/* If there is no gate we can assume it's enabled. */
+	if (!gate_exists(gate))
+		return true;
+
+	bit_mask = 1 << gate->status_bit;
+	reg_val = __ccu_read(ccu, gate->offset);
+
+	return (reg_val & bit_mask) != 0;
+}
+
+/* Determine whether a clock is gated. */
+static bool
+is_clk_gate_enabled(struct ccu_data *ccu, struct bcm_clk_gate *gate)
+{
+	long flags;
+	bool ret;
+
+	/* Avoid taking the lock if we can */
+	if (!gate_exists(gate))
+		return true;
+
+	flags = ccu_lock(ccu);
+	ret = __is_clk_gate_enabled(ccu, gate);
+	ccu_unlock(ccu, flags);
+
+	return ret;
+}
+
+/*
+ * Commit our desired gate state to the hardware.
+ * Returns true if successful, false otherwise.
+ */
+static bool
+__gate_commit(struct ccu_data *ccu, struct bcm_clk_gate *gate)
+{
+	u32 reg_val;
+	u32 mask;
+	bool enabled = false;
+
+	BUG_ON(!gate_exists(gate));
+	if (!gate_is_sw_controllable(gate))
+		return true;		/* Nothing we can change */
+
+	reg_val = __ccu_read(ccu, gate->offset);
+
+	/* For a hardware/software gate, set which is in control */
+	if (gate_is_hw_controllable(gate)) {
+		mask = (u32)1 << gate->hw_sw_sel_bit;
+		if (gate_is_sw_managed(gate))
+			reg_val |= mask;
+		else
+			reg_val &= ~mask;
+	}
+
+	/*
+	 * If software is in control, enable or disable the gate.
+	 * If hardware is, clear the enabled bit for good measure.
+	 * If a software controlled gate can't be disabled, we're
+	 * required to write a 0 into the enable bit (but the gate
+	 * will be enabled).
+	 */
+	mask = (u32)1 << gate->en_bit;
+	if (gate_is_sw_managed(gate) && (enabled = gate_is_enabled(gate)) &&
+			!gate_is_no_disable(gate))
+		reg_val |= mask;
+	else
+		reg_val &= ~mask;
+
+	__ccu_write(ccu, gate->offset, reg_val);
+
+	/* For a hardware controlled gate, we're done */
+	if (!gate_is_sw_managed(gate))
+		return true;
+
+	/* Otherwise wait for the gate to be in desired state */
+	return __ccu_wait_bit(ccu, gate->offset, gate->status_bit, enabled);
+}
+
+/*
+ * Initialize a gate.  Our desired state (hardware/software select,
+ * and if software, its enable state) is committed to hardware
+ * without the usual checks to see if it's already set up that way.
+ * Returns true if successful, false otherwise.
+ */
+static bool gate_init(struct ccu_data *ccu, struct bcm_clk_gate *gate)
+{
+	if (!gate_exists(gate))
+		return true;
+	return __gate_commit(ccu, gate);
+}
+
+/*
+ * Set a gate to enabled or disabled state.  Does nothing if the
+ * gate is not currently under software control, or if it is already
+ * in the requested state.  Returns true if successful, false
+ * otherwise.  CCU lock must be held.
+ */
+static bool
+__clk_gate(struct ccu_data *ccu, struct bcm_clk_gate *gate, bool enable)
+{
+	bool ret;
+
+	if (!gate_exists(gate) || !gate_is_sw_managed(gate))
+		return true;	/* Nothing to do */
+
+	if (!enable && gate_is_no_disable(gate)) {
+		pr_warn("%s: invalid gate disable request (ignoring)\n",
+			__func__);
+		return true;
+	}
+
+	if (enable == gate_is_enabled(gate))
+		return true;	/* No change */
+
+	gate_flip_enabled(gate);
+	ret = __gate_commit(ccu, gate);
+	if (!ret)
+		gate_flip_enabled(gate);	/* Revert the change */
+
+	return ret;
+}
+
+/* Enable or disable a gate.  Returns 0 if successful, -EIO otherwise */
+static int clk_gate(struct ccu_data *ccu, const char *name,
+			struct bcm_clk_gate *gate, bool enable)
+{
+	unsigned long flags;
+	bool success;
+
+	/*
+	 * Avoid taking the lock if we can.  We quietly ignore
+	 * requests to change state that don't make sense.
+	 */
+	if (!gate_exists(gate) || !gate_is_sw_managed(gate))
+		return 0;
+	if (!enable && gate_is_no_disable(gate))
+		return 0;
+
+	flags = ccu_lock(ccu);
+	__ccu_write_enable(ccu);
+
+	success = __clk_gate(ccu, gate, enable);
+
+	__ccu_write_disable(ccu);
+	ccu_unlock(ccu, flags);
+
+	if (success)
+		return 0;
+
+	pr_err("%s: failed to %s gate for %s\n", __func__,
+		enable ? "enable" : "disable", name);
+
+	return -EIO;
+}
+
+/* Trigger operations */
+
+/*
+ * Caller must ensure CCU lock is held and access is enabled.
+ * Returns true if successful, false otherwise.
+ */
+static bool __clk_trigger(struct ccu_data *ccu, struct bcm_clk_trig *trig)
+{
+	/* Trigger the clock and wait for it to finish */
+	__ccu_write(ccu, trig->offset, 1 << trig->bit);
+
+	return __ccu_wait_bit(ccu, trig->offset, trig->bit, false);
+}
+
+/* Divider operations */
+
+/* Read a divider value and return the scaled divisor it represents. */
+static u64 divider_read_scaled(struct ccu_data *ccu, struct bcm_clk_div *div)
+{
+	unsigned long flags;
+	u32 reg_val;
+	u32 reg_div;
+
+	if (divider_is_fixed(div))
+		return (u64)div->fixed;
+
+	flags = ccu_lock(ccu);
+	reg_val = __ccu_read(ccu, div->offset);
+	ccu_unlock(ccu, flags);
+
+	/* Extract the full divider field from the register value */
+	reg_div = bitfield_extract(reg_val, div->shift, div->width);
+
+	/* Return the scaled divisor value it represents */
+	return scaled_div_value(div, reg_div);
+}
+
+/*
+ * Convert a divider's scaled divisor value into its recorded form
+ * and commit it into the hardware divider register.
+ *
+ * Returns 0 on success.  Returns -EINVAL for invalid arguments.
+ * Returns -ENXIO if gating failed, and -EIO if a trigger failed.
+ */
+static int __div_commit(struct ccu_data *ccu, struct bcm_clk_gate *gate,
+			struct bcm_clk_div *div, struct bcm_clk_trig *trig)
+{
+	bool enabled;
+	u32 reg_div;
+	u32 reg_val;
+	int ret = 0;
+
+	BUG_ON(divider_is_fixed(div));
+
+	/*
+	 * If we're just initializing the divider, and no initial
+	 * state was defined in the device tree, we just find out
+	 * what its current value is rather than updating it.
+	 */
+	if (div->scaled_div == BAD_SCALED_DIV_VALUE) {
+		reg_val = __ccu_read(ccu, div->offset);
+		reg_div = bitfield_extract(reg_val, div->shift, div->width);
+		div->scaled_div = scaled_div_value(div, reg_div);
+
+		return 0;
+	}
+
+	/* Convert the scaled divisor to the value we need to record */
+	reg_div = divider(div, div->scaled_div);
+
+	/* Clock needs to be enabled before changing the rate */
+	enabled = __is_clk_gate_enabled(ccu, gate);
+	if (!enabled && !__clk_gate(ccu, gate, true)) {
+		ret = -ENXIO;
+		goto out;
+	}
+
+	/* Replace the divider value and record the result */
+	reg_val = __ccu_read(ccu, div->offset);
+	reg_val = bitfield_replace(reg_val, div->shift, div->width, reg_div);
+	__ccu_write(ccu, div->offset, reg_val);
+
+	/* If the trigger fails we still want to disable the gate */
+	if (!__clk_trigger(ccu, trig))
+		ret = -EIO;
+
+	/* Disable the clock again if it was disabled to begin with */
+	if (!enabled && !__clk_gate(ccu, gate, false))
+		ret = ret ? ret : -ENXIO;	/* return first error */
+out:
+	return ret;
+}
+
+/*
+ * Initialize a divider by committing our desired state to hardware
+ * without the usual checks to see if it's already set up that way.
+ * Returns true if successful, false otherwise.
+ */
+static bool div_init(struct ccu_data *ccu, struct bcm_clk_gate *gate,
+			struct bcm_clk_div *div, struct bcm_clk_trig *trig)
+{
+	if (!divider_exists(div) || divider_is_fixed(div))
+		return true;
+	return !__div_commit(ccu, gate, div, trig);
+}
+
+static int divider_write(struct ccu_data *ccu, struct bcm_clk_gate *gate,
+			struct bcm_clk_div *div, struct bcm_clk_trig *trig,
+			u64 scaled_div)
+{
+	unsigned long flags;
+	u64 previous;
+	int ret;
+
+	BUG_ON(divider_is_fixed(div));
+
+	previous = div->scaled_div;
+	if (previous == scaled_div)
+		return 0;	/* No change */
+
+	div->scaled_div = scaled_div;
+
+	flags = ccu_lock(ccu);
+	__ccu_write_enable(ccu);
+
+	ret = __div_commit(ccu, gate, div, trig);
+
+	__ccu_write_disable(ccu);
+	ccu_unlock(ccu, flags);
+
+	if (ret)
+		div->scaled_div = previous;		/* Revert the change */
+
+	return ret;
+
+}
+
+/* Common clock rate helpers */
+
+/*
+ * Implement the common clock framework recalc_rate method, taking
+ * into account a divider and an optional pre-divider.  The
+ * pre-divider register pointer may be NULL.
+ */
+static unsigned long clk_recalc_rate(struct ccu_data *ccu,
+			struct bcm_clk_div *div, struct bcm_clk_div *pre_div,
+			unsigned long parent_rate)
+{
+	u64 scaled_parent_rate;
+	u64 scaled_div;
+	u64 result;
+
+	if (!divider_exists(div))
+		return parent_rate;
+
+	if (parent_rate > (unsigned long)LONG_MAX)
+		return 0;	/* actually this would be a caller bug */
+
+	/*
+	 * If there is a pre-divider, divide the scaled parent rate
+	 * by the pre-divider value first.  In this case--to improve
+	 * accuracy--scale the parent rate by *both* the pre-divider
+	 * value and the divider before actually computing the
+	 * result of the pre-divider.
+	 *
+	 * If there's only one divider, just scale the parent rate.
+	 */
+	if (pre_div && divider_exists(pre_div)) {
+		u64 scaled_rate;
+
+		scaled_rate = scale_rate(pre_div, parent_rate);
+		scaled_rate = scale_rate(div, scaled_rate);
+		scaled_div = divider_read_scaled(ccu, pre_div);
+		scaled_parent_rate = do_div_round_closest(scaled_rate,
+							scaled_div);
+	} else  {
+		scaled_parent_rate = scale_rate(div, parent_rate);
+	}
+
+	/*
+	 * Get the scaled divisor value, and divide the scaled
+	 * parent rate by that to determine this clock's resulting
+	 * rate.
+	 */
+	scaled_div = divider_read_scaled(ccu, div);
+	result = do_div_round_closest(scaled_parent_rate, scaled_div);
+
+	return (unsigned long)result;
+}
+
+/*
+ * Compute the output rate produced when a given parent rate is fed
+ * into two dividers.  The pre-divider can be NULL, and even if it's
+ * non-null it may be nonexistent.  It's also OK for the divider to
+ * be nonexistent, and in that case the pre-divider is also ignored.
+ *
+ * If scaled_div is non-null, it is used to return the scaled divisor
+ * value used by the (downstream) divider to produce that rate.
+ */
+static long round_rate(struct ccu_data *ccu, struct bcm_clk_div *div,
+				struct bcm_clk_div *pre_div,
+				unsigned long rate, unsigned long parent_rate,
+				u64 *scaled_div)
+{
+	u64 scaled_parent_rate;
+	u64 min_scaled_div;
+	u64 max_scaled_div;
+	u64 best_scaled_div;
+	u64 result;
+
+	BUG_ON(!divider_exists(div));
+	BUG_ON(!rate);
+	BUG_ON(parent_rate > (u64)LONG_MAX);
+
+	/*
+	 * If there is a pre-divider, divide the scaled parent rate
+	 * by the pre-divider value first.  In this case--to improve
+	 * accuracy--scale the parent rate by *both* the pre-divider
+	 * value and the divider before actually computing the
+	 * result of the pre-divider.
+	 *
+	 * If there's only one divider, just scale the parent rate.
+	 *
+	 * For simplicity we treat the pre-divider as fixed (for now).
+	 */
+	if (divider_exists(pre_div)) {
+		u64 scaled_rate;
+		u64 scaled_pre_div;
+
+		scaled_rate = scale_rate(pre_div, parent_rate);
+		scaled_rate = scale_rate(div, scaled_rate);
+		scaled_pre_div = divider_read_scaled(ccu, pre_div);
+		scaled_parent_rate = do_div_round_closest(scaled_rate,
+							scaled_pre_div);
+	} else {
+		scaled_parent_rate = scale_rate(div, parent_rate);
+	}
+
+	/*
+	 * Compute the best possible divider and ensure it is in
+	 * range.  A fixed divider can't be changed, so just report
+	 * the best we can do.
+	 */
+	if (!divider_is_fixed(div)) {
+		best_scaled_div = do_div_round_closest(scaled_parent_rate,
+							rate);
+		min_scaled_div = scaled_div_min(div);
+		max_scaled_div = scaled_div_max(div);
+		if (best_scaled_div > max_scaled_div)
+			best_scaled_div = max_scaled_div;
+		else if (best_scaled_div < min_scaled_div)
+			best_scaled_div = min_scaled_div;
+	} else {
+		best_scaled_div = divider_read_scaled(ccu, div);
+	}
+
+	/* OK, figure out the resulting rate */
+	result = do_div_round_closest(scaled_parent_rate, best_scaled_div);
+
+	if (scaled_div)
+		*scaled_div = best_scaled_div;
+
+	return (long)result;
+}
+
+/* Common clock parent helpers */
+
+/*
+ * For a given parent selector (register field) value, find the
+ * index into a selector's parent_sel array that contains it.
+ * Returns the index, or BAD_CLK_INDEX if it's not found.
+ */
+static u8 parent_index(struct bcm_clk_sel *sel, u8 parent_sel)
+{
+	u8 i;
+
+	BUG_ON(sel->parent_count > (u32)U8_MAX);
+	for (i = 0; i < sel->parent_count; i++)
+		if (sel->parent_sel[i] == parent_sel)
+			return i;
+	return BAD_CLK_INDEX;
+}
+
+/*
+ * Fetch the current value of the selector, and translate that into
+ * its corresponding index in the parent array we registered with
+ * the clock framework.
+ *
+ * Returns parent array index that corresponds with the value found,
+ * or BAD_CLK_INDEX if the found value is out of range.
+ */
+static u8 selector_read_index(struct ccu_data *ccu, struct bcm_clk_sel *sel)
+{
+	unsigned long flags;
+	u32 reg_val;
+	u32 parent_sel;
+	u8 index;
+
+	/* If there's no selector, there's only one parent */
+	if (!selector_exists(sel))
+		return 0;
+
+	/* Get the value in the selector register */
+	flags = ccu_lock(ccu);
+	reg_val = __ccu_read(ccu, sel->offset);
+	ccu_unlock(ccu, flags);
+
+	parent_sel = bitfield_extract(reg_val, sel->shift, sel->width);
+
+	/* Look up that selector's parent array index and return it */
+	index = parent_index(sel, parent_sel);
+	if (index == BAD_CLK_INDEX)
+		pr_err("%s: out-of-range parent selector %u (%s 0x%04x)\n",
+			__func__, parent_sel, ccu->name, sel->offset);
+
+	return index;
+}
+
+/*
+ * Commit our desired selector value to the hardware.
+ *
+ * Returns 0 on success.  Returns -EINVAL for invalid arguments.
+ * Returns -ENXIO if gating failed, and -EIO if a trigger failed.
+ */
+static int
+__sel_commit(struct ccu_data *ccu, struct bcm_clk_gate *gate,
+			struct bcm_clk_sel *sel, struct bcm_clk_trig *trig)
+{
+	u32 parent_sel;
+	u32 reg_val;
+	bool enabled;
+	int ret = 0;
+
+	BUG_ON(!selector_exists(sel));
+
+	/*
+	 * If we're just initializing the selector, and no initial
+	 * state was defined in the device tree, we just find out
+	 * what its current value is rather than updating it.
+	 */
+	if (sel->clk_index == BAD_CLK_INDEX) {
+		u8 index;
+
+		reg_val = __ccu_read(ccu, sel->offset);
+		parent_sel = bitfield_extract(reg_val, sel->shift, sel->width);
+		index = parent_index(sel, parent_sel);
+		if (index == BAD_CLK_INDEX)
+			return -EINVAL;
+		sel->clk_index = index;
+
+		return 0;
+	}
+
+	BUG_ON((u32)sel->clk_index >= sel->parent_count);
+	parent_sel = sel->parent_sel[sel->clk_index];
+
+	/* Clock needs to be enabled before changing the parent */
+	enabled = __is_clk_gate_enabled(ccu, gate);
+	if (!enabled && !__clk_gate(ccu, gate, true))
+		return -ENXIO;
+
+	/* Replace the selector value and record the result */
+	reg_val = __ccu_read(ccu, sel->offset);
+	reg_val = bitfield_replace(reg_val, sel->shift, sel->width, parent_sel);
+	__ccu_write(ccu, sel->offset, reg_val);
+
+	/* If the trigger fails we still want to disable the gate */
+	if (!__clk_trigger(ccu, trig))
+		ret = -EIO;
+
+	/* Disable the clock again if it was disabled to begin with */
+	if (!enabled && !__clk_gate(ccu, gate, false))
+		ret = ret ? ret : -ENXIO;	/* return first error */
+
+	return ret;
+}
+
+/*
+ * Initialize a selector by committing our desired state to hardware
+ * without the usual checks to see if it's already set up that way.
+ * Returns true if successful, false otherwise.
+ */
+static bool sel_init(struct ccu_data *ccu, struct bcm_clk_gate *gate,
+			struct bcm_clk_sel *sel, struct bcm_clk_trig *trig)
+{
+	if (!selector_exists(sel))
+		return true;
+	return !__sel_commit(ccu, gate, sel, trig);
+}
+
+/*
+ * Write a new value into a selector register to switch to a
+ * different parent clock.  Returns 0 on success, or an error code
+ * (from __sel_commit()) otherwise.
+ */
+static int selector_write(struct ccu_data *ccu, struct bcm_clk_gate *gate,
+			struct bcm_clk_sel *sel, struct bcm_clk_trig *trig,
+			u8 index)
+{
+	unsigned long flags;
+	u8 previous;
+	int ret;
+
+	previous = sel->clk_index;
+	if (previous == index)
+		return 0;	/* No change */
+
+	sel->clk_index = index;
+
+	flags = ccu_lock(ccu);
+	__ccu_write_enable(ccu);
+
+	ret = __sel_commit(ccu, gate, sel, trig);
+
+	__ccu_write_disable(ccu);
+	ccu_unlock(ccu, flags);
+
+	if (ret)
+		sel->clk_index = previous;	/* Revert the change */
+
+	return ret;
+}
+
+/* Clock operations */
+
+static int kona_peri_clk_enable(struct clk_hw *hw)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct bcm_clk_gate *gate = &bcm_clk->peri->gate;
+
+	return clk_gate(bcm_clk->ccu, bcm_clk->name, gate, true);
+}
+
+static void kona_peri_clk_disable(struct clk_hw *hw)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct bcm_clk_gate *gate = &bcm_clk->peri->gate;
+
+	(void)clk_gate(bcm_clk->ccu, bcm_clk->name, gate, false);
+}
+
+static int kona_peri_clk_is_enabled(struct clk_hw *hw)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct bcm_clk_gate *gate = &bcm_clk->peri->gate;
+
+	return is_clk_gate_enabled(bcm_clk->ccu, gate) ? 1 : 0;
+}
+
+static unsigned long kona_peri_clk_recalc_rate(struct clk_hw *hw,
+			unsigned long parent_rate)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct peri_clk_data *data = bcm_clk->peri;
+
+	return clk_recalc_rate(bcm_clk->ccu, &data->div, &data->pre_div,
+				parent_rate);
+}
+
+static long kona_peri_clk_round_rate(struct clk_hw *hw, unsigned long rate,
+			unsigned long *parent_rate)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct bcm_clk_div *div = &bcm_clk->peri->div;
+
+	if (!divider_exists(div))
+		return __clk_get_rate(hw->clk);
+
+	/* Quietly avoid a zero rate */
+	return round_rate(bcm_clk->ccu, div, &bcm_clk->peri->pre_div,
+				rate ? rate : 1, *parent_rate, NULL);
+}
+
+static int kona_peri_clk_set_parent(struct clk_hw *hw, u8 index)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct peri_clk_data *data = bcm_clk->peri;
+	struct bcm_clk_sel *sel = &data->sel;
+	struct bcm_clk_trig *trig;
+	int ret;
+
+	BUG_ON(index >= sel->parent_count);
+
+	/* If there's only one parent we don't require a selector */
+	if (!selector_exists(sel))
+		return 0;
+
+	/*
+	 * The regular trigger is used by default, but if there's a
+	 * pre-trigger we want to use that instead.
+	 */
+	trig = trigger_exists(&data->pre_trig) ? &data->pre_trig
+					       : &data->trig;
+
+	ret = selector_write(bcm_clk->ccu, &data->gate, sel, trig, index);
+	if (ret == -ENXIO) {
+		pr_err("%s: gating failure for %s\n", __func__, bcm_clk->name);
+		ret = -EIO;	/* Don't proliferate weird errors */
+	} else if (ret == -EIO) {
+		pr_err("%s: %strigger failed for %s\n", __func__,
+			trig == &data->pre_trig ? "pre-" : "",
+			bcm_clk->name);
+	}
+
+	return ret;
+}
+
+static u8 kona_peri_clk_get_parent(struct clk_hw *hw)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct peri_clk_data *data = bcm_clk->peri;
+	u8 index;
+
+	index = selector_read_index(bcm_clk->ccu, &data->sel);
+
+	/* Not all callers would handle an out-of-range value gracefully */
+	return index == BAD_CLK_INDEX ? 0 : index;
+}
+
+static int kona_peri_clk_set_rate(struct clk_hw *hw, unsigned long rate,
+			unsigned long parent_rate)
+{
+	struct kona_clk *bcm_clk = to_kona_clk(hw);
+	struct peri_clk_data *data = bcm_clk->peri;
+	struct bcm_clk_div *div = &data->div;
+	u64 scaled_div = 0;
+	int ret;
+
+	if (parent_rate > (unsigned long)LONG_MAX)
+		return -EINVAL;
+
+	if (rate == __clk_get_rate(hw->clk))
+		return 0;
+
+	if (!divider_exists(div))
+		return rate == parent_rate ? 0 : -EINVAL;
+
+	/*
+	 * A fixed divider can't be changed.  (Nor can a fixed
+	 * pre-divider be, but for now we never actually try to
+	 * change that.)  Tolerate a request for a no-op change.
+	 */
+	if (divider_is_fixed(&data->div))
+		return rate == parent_rate ? 0 : -EINVAL;
+
+	/*
+	 * Get the scaled divisor value needed to achieve a clock
+	 * rate as close as possible to what was requested, given
+	 * the parent clock rate supplied.
+	 */
+	(void)round_rate(bcm_clk->ccu, div, &data->pre_div,
+				rate ? rate : 1, parent_rate, &scaled_div);
+
+	/*
+	 * We aren't updating any pre-divider at this point, so
+	 * we'll use the regular trigger.
+	 */
+	ret = divider_write(bcm_clk->ccu, &data->gate, &data->div,
+				&data->trig, scaled_div);
+	if (ret == -ENXIO) {
+		pr_err("%s: gating failure for %s\n", __func__, bcm_clk->name);
+		ret = -EIO;	/* Don't proliferate weird errors */
+	} else if (ret == -EIO) {
+		pr_err("%s: trigger failed for %s\n", __func__, bcm_clk->name);
+	}
+
+	return ret;
+}
+
+struct clk_ops kona_peri_clk_ops = {
+	.enable = kona_peri_clk_enable,
+	.disable = kona_peri_clk_disable,
+	.is_enabled = kona_peri_clk_is_enabled,
+	.recalc_rate = kona_peri_clk_recalc_rate,
+	.round_rate = kona_peri_clk_round_rate,
+	.set_parent = kona_peri_clk_set_parent,
+	.get_parent = kona_peri_clk_get_parent,
+	.set_rate = kona_peri_clk_set_rate,
+};
+
+/* Put a peripheral clock into its initial state */
+static bool __peri_clk_init(struct kona_clk *bcm_clk)
+{
+	struct ccu_data *ccu = bcm_clk->ccu;
+	struct peri_clk_data *peri = bcm_clk->peri;
+	const char *name = bcm_clk->name;
+	struct bcm_clk_trig *trig;
+
+	BUG_ON(bcm_clk->type != bcm_clk_peri);
+
+	if (!gate_init(ccu, &peri->gate)) {
+		pr_err("%s: error initializing gate for %s\n", __func__, name);
+		return false;
+	}
+	if (!div_init(ccu, &peri->gate, &peri->div, &peri->trig)) {
+		pr_err("%s: error initializing divider for %s\n", __func__,
+			name);
+		return false;
+	}
+
+	/*
+	 * For the pre-divider and selector, the pre-trigger is used
+	 * if it's present, otherwise we just use the regular trigger.
+	 */
+	trig = trigger_exists(&peri->pre_trig) ? &peri->pre_trig
+					       : &peri->trig;
+
+	if (!div_init(ccu, &peri->gate, &peri->pre_div, trig)) {
+		pr_err("%s: error initializing pre-divider for %s\n", __func__,
+			name);
+		return false;
+	}
+
+	if (!sel_init(ccu, &peri->gate, &peri->sel, trig)) {
+		pr_err("%s: error initializing selector for %s\n", __func__,
+			name);
+		return false;
+	}
+
+	return true;
+}
+
+static bool __kona_clk_init(struct kona_clk *bcm_clk)
+{
+	switch (bcm_clk->type) {
+	case bcm_clk_peri:
+		return __peri_clk_init(bcm_clk);
+	default:
+		BUG();
+	}
+	return -EINVAL;
+}
+
+/* Set a CCU and all its clocks into their desired initial state */
+bool __init kona_ccu_init(struct ccu_data *ccu)
+{
+	unsigned long flags;
+	unsigned int which;
+	struct clk **clks = ccu->data.clks;
+	bool success = true;
+
+	flags = ccu_lock(ccu);
+	__ccu_write_enable(ccu);
+
+	for (which = 0; which < ccu->data.clk_num; which++) {
+		struct kona_clk *bcm_clk;
+
+		if (!clks[which])
+			continue;
+		bcm_clk = to_kona_clk(__clk_get_hw(clks[which]));
+		success &= __kona_clk_init(bcm_clk);
+	}
+
+	__ccu_write_disable(ccu);
+	ccu_unlock(ccu, flags);
+	return success;
+}