Merge branch 'next' of https://gitlab.denx.de/u-boot/custodians/u-boot-riscv into next

- Add Sipeed Maix support
- Update clock handler and proper cpu features

5 files changed, 1547 insertions, 0 deletions

diff --git a/drivers/clk/kendryte/Kconfig b/drivers/clk/kendryte/Kconfig
new file mode 100644
index 0000000000..073fca0781
--- /dev/null
+++ b/drivers/clk/kendryte/Kconfig
@@ -0,0 +1,12 @@
+config CLK_K210
+	bool "Clock support for Kendryte K210"
+	depends on CLK && CLK_CCF && CLK_COMPOSITE_CCF
+	help
+	  This enables support clock driver for Kendryte K210 platforms.
+
+config CLK_K210_SET_RATE
+	bool "Enable setting the Kendryte K210 PLL rate"
+	depends on CLK_K210
+	help
+	  Add functionality to calculate new rates for K210 PLLs. Enabling this
+	  feature adds around 1K to U-Boot's final size.
diff --git a/drivers/clk/kendryte/Makefile b/drivers/clk/kendryte/Makefile
new file mode 100644
index 0000000000..6fb68253ae
--- /dev/null
+++ b/drivers/clk/kendryte/Makefile
@@ -0,0 +1 @@
+obj-y += bypass.o clk.o pll.o
diff --git a/drivers/clk/kendryte/bypass.c b/drivers/clk/kendryte/bypass.c
new file mode 100644
index 0000000000..d1fd28175b
--- /dev/null
+++ b/drivers/clk/kendryte/bypass.c
@@ -0,0 +1,270 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Copyright (C) 2020 Sean Anderson <seanga2@gmail.com>
+ */
+
+#define LOG_CATEGORY UCLASS_CLK
+#include <kendryte/bypass.h>
+
+#include <clk-uclass.h>
+#include <linux/clk-provider.h>
+#include <linux/err.h>
+#include <log.h>
+
+#define CLK_K210_BYPASS "k210_clk_bypass"
+
+/*
+ * This is a small driver to do a software bypass of a clock if hardware bypass
+ * is not working. I have tried to write this in a generic fashion, so that it
+ * could be potentially broken out of the kendryte code at some future date.
+ *
+ * Say you have the following clock configuration
+ *
+ * +---+ +---+
+ * |osc| |pll|
+ * +---+ +---+
+ *         ^
+ *        /|
+ *       / |
+ *      /  |
+ *     /   |
+ *    /    |
+ * +---+ +---+
+ * |clk| |clk|
+ * +---+ +---+
+ *
+ * But the pll does not have a bypass, so when you configure the pll, the
+ * configuration needs to change to look like
+ *
+ * +---+ +---+
+ * |osc| |pll|
+ * +---+ +---+
+ *   ^
+ *   |\
+ *   | \
+ *   |  \
+ *   |   \
+ *   |    \
+ * +---+ +---+
+ * |clk| |clk|
+ * +---+ +---+
+ *
+ * To set this up, create a bypass clock with bypassee=pll and alt=osc. When
+ * creating the child clocks, set their parent to the bypass clock. After
+ * creating all the children, call k210_bypass_setchildren().
+ */
+
+static int k210_bypass_dobypass(struct k210_bypass *bypass)
+{
+	int ret, i;
+
+	/*
+	 * If we already have saved parents, then the children are already
+	 * bypassed
+	 */
+	if (bypass->child_count && bypass->saved_parents[0])
+		return 0;
+
+	for (i = 0; i < bypass->child_count; i++) {
+		struct clk *child = bypass->children[i];
+		struct clk *parent = clk_get_parent(child);
+
+		if (IS_ERR(parent)) {
+			for (; i; i--)
+				bypass->saved_parents[i] = NULL;
+			return PTR_ERR(parent);
+		}
+		bypass->saved_parents[i] = parent;
+	}
+
+	for (i = 0; i < bypass->child_count; i++) {
+		struct clk *child = bypass->children[i];
+
+		ret = clk_set_parent(child, bypass->alt);
+		if (ret) {
+			for (; i; i--)
+				clk_set_parent(bypass->children[i],
+					       bypass->saved_parents[i]);
+			for (i = 0; i < bypass->child_count; i++)
+				bypass->saved_parents[i] = NULL;
+			return ret;
+		}
+	}
+
+	return 0;
+}
+
+static int k210_bypass_unbypass(struct k210_bypass *bypass)
+{
+	int err, ret, i;
+
+	if (!bypass->child_count && !bypass->saved_parents[0]) {
+		log_warning("Cannot unbypass children; dobypass not called first\n");
+		return 0;
+	}
+
+	ret = 0;
+	for (i = 0; i < bypass->child_count; i++) {
+		err = clk_set_parent(bypass->children[i],
+				     bypass->saved_parents[i]);
+		if (err)
+			ret = err;
+		bypass->saved_parents[i] = NULL;
+	}
+	return ret;
+}
+
+static ulong k210_bypass_get_rate(struct clk *clk)
+{
+	struct k210_bypass *bypass = to_k210_bypass(clk);
+	const struct clk_ops *ops = bypass->bypassee_ops;
+
+	if (ops->get_rate)
+		return ops->get_rate(bypass->bypassee);
+	else
+		return clk_get_parent_rate(bypass->bypassee);
+}
+
+static ulong k210_bypass_set_rate(struct clk *clk, unsigned long rate)
+{
+	int ret;
+	struct k210_bypass *bypass = to_k210_bypass(clk);
+	const struct clk_ops *ops = bypass->bypassee_ops;
+
+	/* Don't bother bypassing if we aren't going to set the rate */
+	if (!ops->set_rate)
+		return k210_bypass_get_rate(clk);
+
+	ret = k210_bypass_dobypass(bypass);
+	if (ret)
+		return ret;
+
+	ret = ops->set_rate(bypass->bypassee, rate);
+	if (ret < 0)
+		return ret;
+
+	return k210_bypass_unbypass(bypass);
+}
+
+static int k210_bypass_set_parent(struct clk *clk, struct clk *parent)
+{
+	struct k210_bypass *bypass = to_k210_bypass(clk);
+	const struct clk_ops *ops = bypass->bypassee_ops;
+
+	if (ops->set_parent)
+		return ops->set_parent(bypass->bypassee, parent);
+	else
+		return -ENOTSUPP;
+}
+
+/*
+ * For these next two functions, do the bypassing even if there is no
+ * en-/-disable function, since the bypassing itself can be observed in between
+ * calls.
+ */
+static int k210_bypass_enable(struct clk *clk)
+{
+	int ret;
+	struct k210_bypass *bypass = to_k210_bypass(clk);
+	const struct clk_ops *ops = bypass->bypassee_ops;
+
+	ret = k210_bypass_dobypass(bypass);
+	if (ret)
+		return ret;
+
+	if (ops->enable)
+		ret = ops->enable(bypass->bypassee);
+	else
+		ret = 0;
+	if (ret)
+		return ret;
+
+	return k210_bypass_unbypass(bypass);
+}
+
+static int k210_bypass_disable(struct clk *clk)
+{
+	int ret;
+	struct k210_bypass *bypass = to_k210_bypass(clk);
+	const struct clk_ops *ops = bypass->bypassee_ops;
+
+	ret = k210_bypass_dobypass(bypass);
+	if (ret)
+		return ret;
+
+	if (ops->disable)
+		return ops->disable(bypass->bypassee);
+	else
+		return 0;
+}
+
+static const struct clk_ops k210_bypass_ops = {
+	.get_rate = k210_bypass_get_rate,
+	.set_rate = k210_bypass_set_rate,
+	.set_parent = k210_bypass_set_parent,
+	.enable = k210_bypass_enable,
+	.disable = k210_bypass_disable,
+};
+
+int k210_bypass_set_children(struct clk *clk, struct clk **children,
+			     size_t child_count)
+{
+	struct k210_bypass *bypass = to_k210_bypass(clk);
+
+	kfree(bypass->saved_parents);
+	if (child_count) {
+		bypass->saved_parents =
+			kcalloc(child_count, sizeof(struct clk *), GFP_KERNEL);
+		if (!bypass->saved_parents)
+			return -ENOMEM;
+	}
+	bypass->child_count = child_count;
+	bypass->children = children;
+
+	return 0;
+}
+
+struct clk *k210_register_bypass_struct(const char *name,
+					const char *parent_name,
+					struct k210_bypass *bypass)
+{
+	int ret;
+	struct clk *clk;
+
+	clk = &bypass->clk;
+
+	ret = clk_register(clk, CLK_K210_BYPASS, name, parent_name);
+	if (ret)
+		return ERR_PTR(ret);
+
+	bypass->bypassee->dev = clk->dev;
+	return clk;
+}
+
+struct clk *k210_register_bypass(const char *name, const char *parent_name,
+				 struct clk *bypassee,
+				 const struct clk_ops *bypassee_ops,
+				 struct clk *alt)
+{
+	struct clk *clk;
+	struct k210_bypass *bypass;
+
+	bypass = kzalloc(sizeof(*bypass), GFP_KERNEL);
+	if (!bypass)
+		return ERR_PTR(-ENOMEM);
+
+	bypass->bypassee = bypassee;
+	bypass->bypassee_ops = bypassee_ops;
+	bypass->alt = alt;
+
+	clk = k210_register_bypass_struct(name, parent_name, bypass);
+	if (IS_ERR(clk))
+		kfree(bypass);
+	return clk;
+}
+
+U_BOOT_DRIVER(k210_bypass) = {
+	.name	= CLK_K210_BYPASS,
+	.id	= UCLASS_CLK,
+	.ops	= &k210_bypass_ops,
+};
diff --git a/drivers/clk/kendryte/clk.c b/drivers/clk/kendryte/clk.c
new file mode 100644
index 0000000000..981b3b7699
--- /dev/null
+++ b/drivers/clk/kendryte/clk.c
@@ -0,0 +1,663 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Copyright (C) 2019-20 Sean Anderson <seanga2@gmail.com>
+ */
+#include <kendryte/clk.h>
+
+#include <asm/io.h>
+#include <dt-bindings/clock/k210-sysctl.h>
+#include <dt-bindings/mfd/k210-sysctl.h>
+#include <dm.h>
+#include <log.h>
+#include <mapmem.h>
+
+#include <kendryte/bypass.h>
+#include <kendryte/pll.h>
+
+/* All methods are delegated to CCF clocks */
+
+static ulong k210_clk_get_rate(struct clk *clk)
+{
+	struct clk *c;
+	int err = clk_get_by_id(clk->id, &c);
+
+	if (err)
+		return err;
+	return clk_get_rate(c);
+}
+
+static ulong k210_clk_set_rate(struct clk *clk, unsigned long rate)
+{
+	struct clk *c;
+	int err = clk_get_by_id(clk->id, &c);
+
+	if (err)
+		return err;
+	return clk_set_rate(c, rate);
+}
+
+static int k210_clk_set_parent(struct clk *clk, struct clk *parent)
+{
+	struct clk *c, *p;
+	int err = clk_get_by_id(clk->id, &c);
+
+	if (err)
+		return err;
+
+	err = clk_get_by_id(parent->id, &p);
+	if (err)
+		return err;
+
+	return clk_set_parent(c, p);
+}
+
+static int k210_clk_endisable(struct clk *clk, bool enable)
+{
+	struct clk *c;
+	int err = clk_get_by_id(clk->id, &c);
+
+	if (err)
+		return err;
+	return enable ? clk_enable(c) : clk_disable(c);
+}
+
+static int k210_clk_enable(struct clk *clk)
+{
+	return k210_clk_endisable(clk, true);
+}
+
+static int k210_clk_disable(struct clk *clk)
+{
+	return k210_clk_endisable(clk, false);
+}
+
+static const struct clk_ops k210_clk_ops = {
+	.set_rate = k210_clk_set_rate,
+	.get_rate = k210_clk_get_rate,
+	.set_parent = k210_clk_set_parent,
+	.enable = k210_clk_enable,
+	.disable = k210_clk_disable,
+};
+
+/* Parents for muxed clocks */
+static const char * const generic_sels[] = { "in0_half", "pll0_half" };
+/* The first clock is in0, which is filled in by k210_clk_probe */
+static const char *aclk_sels[] = { NULL, "pll0_half" };
+static const char *pll2_sels[] = { NULL, "pll0", "pll1" };
+
+/*
+ * All parameters for different sub-clocks are collected into parameter arrays.
+ * These parameters are then initialized by the clock which uses them during
+ * probe. To save space, ids are automatically generated for each sub-clock by
+ * using an enum. Instead of storing a parameter struct for each clock, even for
+ * those clocks which don't use a particular type of sub-clock, we can just
+ * store the parameters for the clocks which need them.
+ *
+ * So why do it like this? Arranging all the sub-clocks together makes it very
+ * easy to find bugs in the code.
+ */
+
+#define DIV(id, off, shift, width) DIV_FLAGS(id, off, shift, width, 0)
+#define DIV_LIST \
+	DIV_FLAGS(K210_CLK_ACLK, K210_SYSCTL_SEL0, 1, 2, \
+		  CLK_DIVIDER_POWER_OF_TWO) \
+	DIV(K210_CLK_APB0,   K210_SYSCTL_SEL0,  3,  3) \
+	DIV(K210_CLK_APB1,   K210_SYSCTL_SEL0,  6,  3) \
+	DIV(K210_CLK_APB2,   K210_SYSCTL_SEL0,  9,  3) \
+	DIV(K210_CLK_SRAM0,  K210_SYSCTL_THR0,  0,  4) \
+	DIV(K210_CLK_SRAM1,  K210_SYSCTL_THR0,  4,  4) \
+	DIV(K210_CLK_AI,     K210_SYSCTL_THR0,  8,  4) \
+	DIV(K210_CLK_DVP,    K210_SYSCTL_THR0, 12,  4) \
+	DIV(K210_CLK_ROM,    K210_SYSCTL_THR0, 16,  4) \
+	DIV(K210_CLK_SPI0,   K210_SYSCTL_THR1,  0,  8) \
+	DIV(K210_CLK_SPI1,   K210_SYSCTL_THR1,  8,  8) \
+	DIV(K210_CLK_SPI2,   K210_SYSCTL_THR1, 16,  8) \
+	DIV(K210_CLK_SPI3,   K210_SYSCTL_THR1, 24,  8) \
+	DIV(K210_CLK_TIMER0, K210_SYSCTL_THR2,  0,  8) \
+	DIV(K210_CLK_TIMER1, K210_SYSCTL_THR2,  8,  8) \
+	DIV(K210_CLK_TIMER2, K210_SYSCTL_THR2, 16,  8) \
+	DIV(K210_CLK_I2S0,   K210_SYSCTL_THR3,  0, 16) \
+	DIV(K210_CLK_I2S1,   K210_SYSCTL_THR3, 16, 16) \
+	DIV(K210_CLK_I2S2,   K210_SYSCTL_THR4,  0, 16) \
+	DIV(K210_CLK_I2S0_M, K210_SYSCTL_THR4, 16,  8) \
+	DIV(K210_CLK_I2S1_M, K210_SYSCTL_THR4, 24,  8) \
+	DIV(K210_CLK_I2S2_M, K210_SYSCTL_THR4,  0,  8) \
+	DIV(K210_CLK_I2C0,   K210_SYSCTL_THR5,  8,  8) \
+	DIV(K210_CLK_I2C1,   K210_SYSCTL_THR5, 16,  8) \
+	DIV(K210_CLK_I2C2,   K210_SYSCTL_THR5, 24,  8) \
+	DIV(K210_CLK_WDT0,   K210_SYSCTL_THR6,  0,  8) \
+	DIV(K210_CLK_WDT1,   K210_SYSCTL_THR6,  8,  8)
+
+#define _DIVIFY(id) K210_CLK_DIV_##id
+#define DIVIFY(id) _DIVIFY(id)
+
+enum k210_div_ids {
+#define DIV_FLAGS(id, ...) DIVIFY(id),
+	DIV_LIST
+#undef DIV_FLAGS
+};
+
+struct k210_div_params {
+	u8 off;
+	u8 shift;
+	u8 width;
+	u8 flags;
+};
+
+static const struct k210_div_params k210_divs[] = {
+#define DIV_FLAGS(id, _off, _shift, _width, _flags) \
+	[DIVIFY(id)] = { \
+		.off = (_off), \
+		.shift = (_shift), \
+		.width = (_width), \
+		.flags = (_flags), \
+	},
+	DIV_LIST
+#undef DIV_FLAGS
+};
+
+#undef DIV
+#undef DIV_LIST
+
+#define GATE_LIST \
+	GATE(K210_CLK_CPU,    K210_SYSCTL_EN_CENT,  0) \
+	GATE(K210_CLK_SRAM0,  K210_SYSCTL_EN_CENT,  1) \
+	GATE(K210_CLK_SRAM1,  K210_SYSCTL_EN_CENT,  2) \
+	GATE(K210_CLK_APB0,   K210_SYSCTL_EN_CENT,  3) \
+	GATE(K210_CLK_APB1,   K210_SYSCTL_EN_CENT,  4) \
+	GATE(K210_CLK_APB2,   K210_SYSCTL_EN_CENT,  5) \
+	GATE(K210_CLK_ROM,    K210_SYSCTL_EN_PERI,  0) \
+	GATE(K210_CLK_DMA,    K210_SYSCTL_EN_PERI,  1) \
+	GATE(K210_CLK_AI,     K210_SYSCTL_EN_PERI,  2) \
+	GATE(K210_CLK_DVP,    K210_SYSCTL_EN_PERI,  3) \
+	GATE(K210_CLK_FFT,    K210_SYSCTL_EN_PERI,  4) \
+	GATE(K210_CLK_GPIO,   K210_SYSCTL_EN_PERI,  5) \
+	GATE(K210_CLK_SPI0,   K210_SYSCTL_EN_PERI,  6) \
+	GATE(K210_CLK_SPI1,   K210_SYSCTL_EN_PERI,  7) \
+	GATE(K210_CLK_SPI2,   K210_SYSCTL_EN_PERI,  8) \
+	GATE(K210_CLK_SPI3,   K210_SYSCTL_EN_PERI,  9) \
+	GATE(K210_CLK_I2S0,   K210_SYSCTL_EN_PERI, 10) \
+	GATE(K210_CLK_I2S1,   K210_SYSCTL_EN_PERI, 11) \
+	GATE(K210_CLK_I2S2,   K210_SYSCTL_EN_PERI, 12) \
+	GATE(K210_CLK_I2C0,   K210_SYSCTL_EN_PERI, 13) \
+	GATE(K210_CLK_I2C1,   K210_SYSCTL_EN_PERI, 14) \
+	GATE(K210_CLK_I2C2,   K210_SYSCTL_EN_PERI, 15) \
+	GATE(K210_CLK_UART1,  K210_SYSCTL_EN_PERI, 16) \
+	GATE(K210_CLK_UART2,  K210_SYSCTL_EN_PERI, 17) \
+	GATE(K210_CLK_UART3,  K210_SYSCTL_EN_PERI, 18) \
+	GATE(K210_CLK_AES,    K210_SYSCTL_EN_PERI, 19) \
+	GATE(K210_CLK_FPIOA,  K210_SYSCTL_EN_PERI, 20) \
+	GATE(K210_CLK_TIMER0, K210_SYSCTL_EN_PERI, 21) \
+	GATE(K210_CLK_TIMER1, K210_SYSCTL_EN_PERI, 22) \
+	GATE(K210_CLK_TIMER2, K210_SYSCTL_EN_PERI, 23) \
+	GATE(K210_CLK_WDT0,   K210_SYSCTL_EN_PERI, 24) \
+	GATE(K210_CLK_WDT1,   K210_SYSCTL_EN_PERI, 25) \
+	GATE(K210_CLK_SHA,    K210_SYSCTL_EN_PERI, 26) \
+	GATE(K210_CLK_OTP,    K210_SYSCTL_EN_PERI, 27) \
+	GATE(K210_CLK_RTC,    K210_SYSCTL_EN_PERI, 29)
+
+#define _GATEIFY(id) K210_CLK_GATE_##id
+#define GATEIFY(id) _GATEIFY(id)
+
+enum k210_gate_ids {
+#define GATE(id, ...) GATEIFY(id),
+	GATE_LIST
+#undef GATE
+};
+
+struct k210_gate_params {
+	u8 off;
+	u8 bit_idx;
+};
+
+static const struct k210_gate_params k210_gates[] = {
+#define GATE(id, _off, _idx) \
+	[GATEIFY(id)] = { \
+		.off = (_off), \
+		.bit_idx = (_idx), \
+	},
+	GATE_LIST
+#undef GATE
+};
+
+#undef GATE_LIST
+
+#define MUX(id, reg, shift, width) \
+	MUX_PARENTS(id, generic_sels, reg, shift, width)
+#define MUX_LIST \
+	MUX_PARENTS(K210_CLK_PLL2, pll2_sels, K210_SYSCTL_PLL2, 26, 2) \
+	MUX_PARENTS(K210_CLK_ACLK, aclk_sels, K210_SYSCTL_SEL0,  0, 1) \
+	MUX(K210_CLK_SPI3,   K210_SYSCTL_SEL0, 12, 1) \
+	MUX(K210_CLK_TIMER0, K210_SYSCTL_SEL0, 13, 1) \
+	MUX(K210_CLK_TIMER1, K210_SYSCTL_SEL0, 14, 1) \
+	MUX(K210_CLK_TIMER2, K210_SYSCTL_SEL0, 15, 1)
+
+#define _MUXIFY(id) K210_CLK_MUX_##id
+#define MUXIFY(id) _MUXIFY(id)
+
+enum k210_mux_ids {
+#define MUX_PARENTS(id, ...) MUXIFY(id),
+	MUX_LIST
+#undef MUX_PARENTS
+	K210_CLK_MUX_NONE,
+};
+
+struct k210_mux_params {
+	const char *const *parent_names;
+	u8 num_parents;
+	u8 off;
+	u8 shift;
+	u8 width;
+};
+
+static const struct k210_mux_params k210_muxes[] = {
+#define MUX_PARENTS(id, parents, _off, _shift, _width) \
+	[MUXIFY(id)] = { \
+		.parent_names = (const char * const *)(parents), \
+		.num_parents = ARRAY_SIZE(parents), \
+		.off = (_off), \
+		.shift = (_shift), \
+		.width = (_width), \
+	},
+	MUX_LIST
+#undef MUX_PARENTS
+};
+
+#undef MUX
+#undef MUX_LIST
+
+struct k210_pll_params {
+	u8 off;
+	u8 lock_off;
+	u8 shift;
+	u8 width;
+};
+
+static const struct k210_pll_params k210_plls[] = {
+#define PLL(_off, _shift, _width) { \
+	.off = (_off), \
+	.lock_off = K210_SYSCTL_PLL_LOCK, \
+	.shift = (_shift), \
+	.width = (_width), \
+}
+	[0] = PLL(K210_SYSCTL_PLL0,  0, 2),
+	[1] = PLL(K210_SYSCTL_PLL1,  8, 1),
+	[2] = PLL(K210_SYSCTL_PLL2, 16, 1),
+#undef PLL
+};
+
+#define COMP(id) \
+	COMP_FULL(id, MUXIFY(id), DIVIFY(id), GATEIFY(id))
+#define COMP_NOMUX(id) \
+	COMP_FULL(id, K210_CLK_MUX_NONE, DIVIFY(id), GATEIFY(id))
+#define COMP_LIST \
+	COMP(K210_CLK_SPI3) \
+	COMP(K210_CLK_TIMER0) \
+	COMP(K210_CLK_TIMER1) \
+	COMP(K210_CLK_TIMER2) \
+	COMP_NOMUX(K210_CLK_SRAM0) \
+	COMP_NOMUX(K210_CLK_SRAM1) \
+	COMP_NOMUX(K210_CLK_ROM) \
+	COMP_NOMUX(K210_CLK_DVP) \
+	COMP_NOMUX(K210_CLK_APB0) \
+	COMP_NOMUX(K210_CLK_APB1) \
+	COMP_NOMUX(K210_CLK_APB2) \
+	COMP_NOMUX(K210_CLK_AI) \
+	COMP_NOMUX(K210_CLK_I2S0) \
+	COMP_NOMUX(K210_CLK_I2S1) \
+	COMP_NOMUX(K210_CLK_I2S2) \
+	COMP_NOMUX(K210_CLK_WDT0) \
+	COMP_NOMUX(K210_CLK_WDT1) \
+	COMP_NOMUX(K210_CLK_SPI0) \
+	COMP_NOMUX(K210_CLK_SPI1) \
+	COMP_NOMUX(K210_CLK_SPI2) \
+	COMP_NOMUX(K210_CLK_I2C0) \
+	COMP_NOMUX(K210_CLK_I2C1) \
+	COMP_NOMUX(K210_CLK_I2C2)
+
+#define _COMPIFY(id) K210_CLK_COMP_##id
+#define COMPIFY(id) _COMPIFY(id)
+
+enum k210_comp_ids {
+#define COMP_FULL(id, ...) COMPIFY(id),
+	COMP_LIST
+#undef COMP_FULL
+};
+
+struct k210_comp_params {
+	u8 mux;
+	u8 div;
+	u8 gate;
+};
+
+static const struct k210_comp_params k210_comps[] = {
+#define COMP_FULL(id, _mux, _div, _gate) \
+	[COMPIFY(id)] = { \
+		.mux = (_mux), \
+		.div = (_div), \
+		.gate = (_gate), \
+	},
+	COMP_LIST
+#undef COMP_FULL
+};
+
+#undef COMP
+#undef COMP_ID
+#undef COMP_NOMUX
+#undef COMP_NOMUX_ID
+#undef COMP_LIST
+
+static struct clk *k210_bypass_children = {
+	NULL,
+};
+
+/* Helper functions to create sub-clocks */
+static struct clk_mux *k210_create_mux(const struct k210_mux_params *params,
+				       void *base)
+{
+	struct clk_mux *mux = kzalloc(sizeof(*mux), GFP_KERNEL);
+
+	if (!mux)
+		return mux;
+
+	mux->reg = base + params->off;
+	mux->mask = BIT(params->width) - 1;
+	mux->shift = params->shift;
+	mux->parent_names = params->parent_names;
+	mux->num_parents = params->num_parents;
+
+	return mux;
+}
+
+static struct clk_divider *k210_create_div(const struct k210_div_params *params,
+					   void *base)
+{
+	struct clk_divider *div = kzalloc(sizeof(*div), GFP_KERNEL);
+
+	if (!div)
+		return div;
+
+	div->reg = base + params->off;
+	div->shift = params->shift;
+	div->width = params->width;
+	div->flags = params->flags;
+
+	return div;
+}
+
+static struct clk_gate *k210_create_gate(const struct k210_gate_params *params,
+					 void *base)
+{
+	struct clk_gate *gate = kzalloc(sizeof(*gate), GFP_KERNEL);
+
+	if (!gate)
+		return gate;
+
+	gate->reg = base + params->off;
+	gate->bit_idx = params->bit_idx;
+
+	return gate;
+}
+
+static struct k210_pll *k210_create_pll(const struct k210_pll_params *params,
+					void *base)
+{
+	struct k210_pll *pll = kzalloc(sizeof(*pll), GFP_KERNEL);
+
+	if (!pll)
+		return pll;
+
+	pll->reg = base + params->off;
+	pll->lock = base + params->lock_off;
+	pll->shift = params->shift;
+	pll->width = params->width;
+
+	return pll;
+}
+
+/* Create all sub-clocks, and then register the composite clock */
+static struct clk *k210_register_comp(const struct k210_comp_params *params,
+				      void *base, const char *name,
+				      const char *parent)
+{
+	const char *const *parent_names;
+	int num_parents;
+	struct clk *comp;
+	const struct clk_ops *mux_ops;
+	struct clk_mux *mux;
+	struct clk_divider *div;
+	struct clk_gate *gate;
+
+	if (params->mux == K210_CLK_MUX_NONE) {
+		if (!parent)
+			return ERR_PTR(-EINVAL);
+
+		mux_ops = NULL;
+		mux = NULL;
+		parent_names = &parent;
+		num_parents = 1;
+	} else {
+		mux_ops = &clk_mux_ops;
+		mux = k210_create_mux(&k210_muxes[params->mux], base);
+		if (!mux)
+			return ERR_PTR(-ENOMEM);
+
+		parent_names = mux->parent_names;
+		num_parents = mux->num_parents;
+	}
+
+	div = k210_create_div(&k210_divs[params->div], base);
+	if (!div) {
+		comp = ERR_PTR(-ENOMEM);
+		goto cleanup_mux;
+	}
+
+	gate = k210_create_gate(&k210_gates[params->gate], base);
+	if (!gate) {
+		comp = ERR_PTR(-ENOMEM);
+		goto cleanup_div;
+	}
+
+	comp = clk_register_composite(NULL, name, parent_names, num_parents,
+				      &mux->clk, mux_ops,
+				      &div->clk, &clk_divider_ops,
+				      &gate->clk, &clk_gate_ops, 0);
+	if (IS_ERR(comp))
+		goto cleanup_gate;
+	return comp;
+
+cleanup_gate:
+	free(gate);
+cleanup_div:
+	free(div);
+cleanup_mux:
+	if (mux)
+		free(mux);
+	return comp;
+}
+
+static bool probed;
+
+static int k210_clk_probe(struct udevice *dev)
+{
+	int ret;
+	const char *in0;
+	struct clk *in0_clk, *bypass;
+	struct clk_mux *mux;
+	struct clk_divider *div;
+	struct k210_pll *pll;
+	void *base;
+
+	/*
+	 * Only one instance of this driver allowed. This prevents weird bugs
+	 * when the driver fails part-way through probing. Some clocks will
+	 * already have been registered, and re-probing will register them
+	 * again, creating a bunch of duplicates. Better error-handling/cleanup
+	 * could fix this, but it's Probably Not Worth It (TM).
+	 */
+	if (probed)
+		return -ENOTSUPP;
+
+	base = dev_read_addr_ptr(dev_get_parent(dev));
+	if (!base)
+		return -EINVAL;
+
+	in0_clk = kzalloc(sizeof(*in0_clk), GFP_KERNEL);
+	if (!in0_clk)
+		return -ENOMEM;
+
+	ret = clk_get_by_index(dev, 0, in0_clk);
+	if (ret)
+		return ret;
+	in0 = in0_clk->dev->name;
+
+	probed = true;
+
+	aclk_sels[0] = in0;
+	pll2_sels[0] = in0;
+
+	/*
+	 * All PLLs have a broken bypass, but pll0 has the CPU downstream, so we
+	 * need to manually reparent it whenever we configure pll0
+	 */
+	pll = k210_create_pll(&k210_plls[0], base);
+	if (pll) {
+		bypass = k210_register_bypass("pll0", in0, &pll->clk,
+					      &k210_pll_ops, in0_clk);
+		clk_dm(K210_CLK_PLL0, bypass);
+	} else {
+		return -ENOMEM;
+	}
+
+	{
+		const struct k210_pll_params *params = &k210_plls[1];
+
+		clk_dm(K210_CLK_PLL1,
+		       k210_register_pll("pll1", in0, base + params->off,
+					 base + params->lock_off, params->shift,
+					 params->width));
+	}
+
+	/* PLL2 is muxed, so set up a composite clock */
+	mux = k210_create_mux(&k210_muxes[MUXIFY(K210_CLK_PLL2)], base);
+	pll = k210_create_pll(&k210_plls[2], base);
+	if (!mux || !pll) {
+		free(mux);
+		free(pll);
+	} else {
+		clk_dm(K210_CLK_PLL2,
+		       clk_register_composite(NULL, "pll2", pll2_sels,
+					      ARRAY_SIZE(pll2_sels),
+					      &mux->clk, &clk_mux_ops,
+					      &pll->clk, &k210_pll_ops,
+					      &pll->clk, &k210_pll_ops, 0));
+	}
+
+	/* Half-frequency clocks for "even" dividers */
+	clk_dm(K210_CLK_IN0_H,  k210_clk_half("in0_half", in0));
+	clk_dm(K210_CLK_PLL0_H, k210_clk_half("pll0_half", "pll0"));
+	clk_dm(K210_CLK_PLL2_H, k210_clk_half("pll2_half", "pll2"));
+
+	/* ACLK has no gate */
+	mux = k210_create_mux(&k210_muxes[MUXIFY(K210_CLK_ACLK)], base);
+	div = k210_create_div(&k210_divs[DIVIFY(K210_CLK_ACLK)], base);
+	if (!mux || !div) {
+		free(mux);
+		free(div);
+	} else {
+		struct clk *aclk =
+			clk_register_composite(NULL, "aclk", aclk_sels,
+					       ARRAY_SIZE(aclk_sels),
+					       &mux->clk, &clk_mux_ops,
+					       &div->clk, &clk_divider_ops,
+					       NULL, NULL, 0);
+		clk_dm(K210_CLK_ACLK, aclk);
+		if (!IS_ERR(aclk)) {
+			k210_bypass_children = aclk;
+			k210_bypass_set_children(bypass,
+						 &k210_bypass_children, 1);
+		}
+	}
+
+#define REGISTER_COMP(id, name) \
+	clk_dm(id, \
+	       k210_register_comp(&k210_comps[COMPIFY(id)], base, name, NULL))
+	REGISTER_COMP(K210_CLK_SPI3,   "spi3");
+	REGISTER_COMP(K210_CLK_TIMER0, "timer0");
+	REGISTER_COMP(K210_CLK_TIMER1, "timer1");
+	REGISTER_COMP(K210_CLK_TIMER2, "timer2");
+#undef REGISTER_COMP
+
+	/* Dividing clocks, no mux */
+#define REGISTER_COMP_NOMUX(id, name, parent) \
+	clk_dm(id, \
+	       k210_register_comp(&k210_comps[COMPIFY(id)], base, name, parent))
+	REGISTER_COMP_NOMUX(K210_CLK_SRAM0, "sram0", "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_SRAM1, "sram1", "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_ROM,   "rom",   "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_DVP,   "dvp",   "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_APB0,  "apb0",  "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_APB1,  "apb1",  "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_APB2,  "apb2",  "aclk");
+	REGISTER_COMP_NOMUX(K210_CLK_AI,    "ai",    "pll1");
+	REGISTER_COMP_NOMUX(K210_CLK_I2S0,  "i2s0",  "pll2_half");
+	REGISTER_COMP_NOMUX(K210_CLK_I2S1,  "i2s1",  "pll2_half");
+	REGISTER_COMP_NOMUX(K210_CLK_I2S2,  "i2s2",  "pll2_half");
+	REGISTER_COMP_NOMUX(K210_CLK_WDT0,  "wdt0",  "in0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_WDT1,  "wdt1",  "in0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_SPI0,  "spi0",  "pll0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_SPI1,  "spi1",  "pll0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_SPI2,  "spi2",  "pll0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_I2C0,  "i2c0",  "pll0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_I2C1,  "i2c1",  "pll0_half");
+	REGISTER_COMP_NOMUX(K210_CLK_I2C2,  "i2c2",  "pll0_half");
+#undef REGISTER_COMP_NOMUX
+
+	/* Dividing clocks */
+#define REGISTER_DIV(id, name, parent) do {\
+	const struct k210_div_params *params = &k210_divs[DIVIFY(id)]; \
+	clk_dm(id, \
+	       clk_register_divider(NULL, name, parent, 0, base + params->off, \
+				    params->shift, params->width, 0)); \
+} while (false)
+	REGISTER_DIV(K210_CLK_I2S0_M, "i2s0_m", "pll2_half");
+	REGISTER_DIV(K210_CLK_I2S1_M, "i2s1_m", "pll2_half");
+	REGISTER_DIV(K210_CLK_I2S2_M, "i2s2_m", "pll2_half");
+#undef REGISTER_DIV
+
+	/* Gated clocks */
+#define REGISTER_GATE(id, name, parent) do { \
+	const struct k210_gate_params *params = &k210_gates[GATEIFY(id)]; \
+	clk_dm(id, \
+	       clk_register_gate(NULL, name, parent, 0, base + params->off, \
+				 params->bit_idx, 0, NULL)); \
+} while (false)
+	REGISTER_GATE(K210_CLK_CPU,   "cpu",   "aclk");
+	REGISTER_GATE(K210_CLK_DMA,   "dma",   "aclk");
+	REGISTER_GATE(K210_CLK_FFT,   "fft",   "aclk");
+	REGISTER_GATE(K210_CLK_GPIO,  "gpio",  "apb0");
+	REGISTER_GATE(K210_CLK_UART1, "uart1", "apb0");
+	REGISTER_GATE(K210_CLK_UART2, "uart2", "apb0");
+	REGISTER_GATE(K210_CLK_UART3, "uart3", "apb0");
+	REGISTER_GATE(K210_CLK_FPIOA, "fpioa", "apb0");
+	REGISTER_GATE(K210_CLK_SHA,   "sha",   "apb0");
+	REGISTER_GATE(K210_CLK_AES,   "aes",   "apb1");
+	REGISTER_GATE(K210_CLK_OTP,   "otp",   "apb1");
+	REGISTER_GATE(K210_CLK_RTC,   "rtc",   in0);
+#undef REGISTER_GATE
+
+	return 0;
+}
+
+static const struct udevice_id k210_clk_ids[] = {
+	{ .compatible = "kendryte,k210-clk" },
+	{ },
+};
+
+U_BOOT_DRIVER(k210_clk) = {
+	.name = "k210_clk",
+	.id = UCLASS_CLK,
+	.of_match = k210_clk_ids,
+	.ops = &k210_clk_ops,
+	.probe = k210_clk_probe,
+};
diff --git a/drivers/clk/kendryte/pll.c b/drivers/clk/kendryte/pll.c
new file mode 100644
index 0000000000..19e358856a
--- /dev/null
+++ b/drivers/clk/kendryte/pll.c
@@ -0,0 +1,601 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Copyright (C) 2019-20 Sean Anderson <seanga2@gmail.com>
+ */
+#define LOG_CATEGORY UCLASS_CLK
+#include <kendryte/pll.h>
+
+#include <asm/io.h>
+/* For DIV_ROUND_DOWN_ULL, defined in linux/kernel.h */
+#include <div64.h>
+#include <dt-bindings/clock/k210-sysctl.h>
+#include <linux/bitfield.h>
+#include <linux/clk-provider.h>
+#include <linux/delay.h>
+#include <linux/err.h>
+#include <log.h>
+#include <serial.h>
+
+#define CLK_K210_PLL "k210_clk_pll"
+
+#ifdef CONFIG_CLK_K210_SET_RATE
+static int k210_pll_enable(struct clk *clk);
+static int k210_pll_disable(struct clk *clk);
+
+/*
+ * The PLL included with the Kendryte K210 appears to be a True Circuits, Inc.
+ * General-Purpose PLL. The logical layout of the PLL with internal feedback is
+ * approximately the following:
+ *
+ *  +---------------+
+ *  |reference clock|
+ *  +---------------+
+ *          |
+ *          v
+ *        +--+
+ *        |/r|
+ *        +--+
+ *          |
+ *          v
+ *   +-------------+
+ *   |divided clock|
+ *   +-------------+
+ *          |
+ *          v
+ *  +--------------+
+ *  |phase detector|<---+
+ *  +--------------+    |
+ *          |           |
+ *          v   +--------------+
+ *        +---+ |feedback clock|
+ *        |VCO| +--------------+
+ *        +---+         ^
+ *          |    +--+   |
+ *          +--->|/f|---+
+ *          |    +--+
+ *          v
+ *        +---+
+ *        |/od|
+ *        +---+
+ *          |
+ *          v
+ *       +------+
+ *       |output|
+ *       +------+
+ *
+ * The k210 PLLs have three factors: r, f, and od. Because of the feedback mode,
+ * the effect of the division by f is to multiply the input frequency. The
+ * equation for the output rate is
+ *   rate = (rate_in * f) / (r * od).
+ * Moving knowns to one side of the equation, we get
+ *   rate / rate_in = f / (r * od)
+ * Rearranging slightly,
+ *   abs_error = abs((rate / rate_in) - (f / (r * od))).
+ * To get relative, error, we divide by the expected ratio
+ *   error = abs((rate / rate_in) - (f / (r * od))) / (rate / rate_in).
+ * Simplifying,
+ *   error = abs(1 - f / (r * od)) / (rate / rate_in)
+ *   error = abs(1 - (f * rate_in) / (r * od * rate))
+ * Using the constants ratio = rate / rate_in and inv_ratio = rate_in / rate,
+ *   error = abs((f * inv_ratio) / (r * od) - 1)
+ * This is the error used in evaluating parameters.
+ *
+ * r and od are four bits each, while f is six bits. Because r and od are
+ * multiplied together, instead of the full 256 values possible if both bits
+ * were used fully, there are only 97 distinct products. Combined with f, there
+ * are 6208 theoretical settings for the PLL. However, most of these settings
+ * can be ruled out immediately because they do not have the correct ratio.
+ *
+ * In addition to the constraint of approximating the desired ratio, parameters
+ * must also keep internal pll frequencies within acceptable ranges. The divided
+ * clock's minimum and maximum frequencies have a ratio of around 128.  This
+ * leaves fairly substantial room to work with, especially since the only
+ * affected parameter is r. The VCO's minimum and maximum frequency have a ratio
+ * of 5, which is considerably more restrictive.
+ *
+ * The r and od factors are stored in a table. This is to make it easy to find
+ * the next-largest product. Some products have multiple factorizations, but
+ * only when one factor has at least a 2.5x ratio to the factors of the other
+ * factorization. This is because any smaller ratio would not make a difference
+ * when ensuring the VCO's frequency is within spec.
+ *
+ * Throughout the calculation function, fixed point arithmetic is used. Because
+ * the range of rate and rate_in may be up to 1.75 GHz, or around 2^30, 64-bit
+ * 32.32 fixed-point numbers are used to represent ratios. In general, to
+ * implement division, the numerator is first multiplied by 2^32. This gives a
+ * result where the whole number part is in the upper 32 bits, and the fraction
+ * is in the lower 32 bits.
+ *
+ * In general, rounding is done to the closest integer. This helps find the best
+ * approximation for the ratio. Rounding in one direction (e.g down) could cause
+ * the function to miss a better ratio with one of the parameters increased by
+ * one.
+ */
+
+/*
+ * The factors table was generated with the following python code:
+ *
+ * def p(x, y):
+ *    return (1.0*x/y > 2.5) or (1.0*y/x > 2.5)
+ *
+ * factors = {}
+ * for i in range(1, 17):
+ *    for j in range(1, 17):
+ *       fs = factors.get(i*j) or []
+ *       if fs == [] or all([
+ *             (p(i, x) and p(i, y)) or (p(j, x) and p(j, y))
+ *             for (x, y) in fs]):
+ *          fs.append((i, j))
+ *          factors[i*j] = fs
+ *
+ * for k, l in sorted(factors.items()):
+ *    for v in l:
+ *       print("PACK(%s, %s)," % v)
+ */
+#define PACK(r, od) (((((r) - 1) & 0xF) << 4) | (((od) - 1) & 0xF))
+#define UNPACK_R(val) ((((val) >> 4) & 0xF) + 1)
+#define UNPACK_OD(val) (((val) & 0xF) + 1)
+static const u8 factors[] = {
+	PACK(1, 1),
+	PACK(1, 2),
+	PACK(1, 3),
+	PACK(1, 4),
+	PACK(1, 5),
+	PACK(1, 6),
+	PACK(1, 7),
+	PACK(1, 8),
+	PACK(1, 9),
+	PACK(3, 3),
+	PACK(1, 10),
+	PACK(1, 11),
+	PACK(1, 12),
+	PACK(3, 4),
+	PACK(1, 13),
+	PACK(1, 14),
+	PACK(1, 15),
+	PACK(3, 5),
+	PACK(1, 16),
+	PACK(4, 4),
+	PACK(2, 9),
+	PACK(2, 10),
+	PACK(3, 7),
+	PACK(2, 11),
+	PACK(2, 12),
+	PACK(5, 5),
+	PACK(2, 13),
+	PACK(3, 9),
+	PACK(2, 14),
+	PACK(2, 15),
+	PACK(2, 16),
+	PACK(3, 11),
+	PACK(5, 7),
+	PACK(3, 12),
+	PACK(3, 13),
+	PACK(4, 10),
+	PACK(3, 14),
+	PACK(4, 11),
+	PACK(3, 15),
+	PACK(3, 16),
+	PACK(7, 7),
+	PACK(5, 10),
+	PACK(4, 13),
+	PACK(6, 9),
+	PACK(5, 11),
+	PACK(4, 14),
+	PACK(4, 15),
+	PACK(7, 9),
+	PACK(4, 16),
+	PACK(5, 13),
+	PACK(6, 11),
+	PACK(5, 14),
+	PACK(6, 12),
+	PACK(5, 15),
+	PACK(7, 11),
+	PACK(6, 13),
+	PACK(5, 16),
+	PACK(9, 9),
+	PACK(6, 14),
+	PACK(8, 11),
+	PACK(6, 15),
+	PACK(7, 13),
+	PACK(6, 16),
+	PACK(7, 14),
+	PACK(9, 11),
+	PACK(10, 10),
+	PACK(8, 13),
+	PACK(7, 15),
+	PACK(9, 12),
+	PACK(10, 11),
+	PACK(7, 16),
+	PACK(9, 13),
+	PACK(8, 15),
+	PACK(11, 11),
+	PACK(9, 14),
+	PACK(8, 16),
+	PACK(10, 13),
+	PACK(11, 12),
+	PACK(9, 15),
+	PACK(10, 14),
+	PACK(11, 13),
+	PACK(9, 16),
+	PACK(10, 15),
+	PACK(11, 14),
+	PACK(12, 13),
+	PACK(10, 16),
+	PACK(11, 15),
+	PACK(12, 14),
+	PACK(13, 13),
+	PACK(11, 16),
+	PACK(12, 15),
+	PACK(13, 14),
+	PACK(12, 16),
+	PACK(13, 15),
+	PACK(14, 14),
+	PACK(13, 16),
+	PACK(14, 15),
+	PACK(14, 16),
+	PACK(15, 15),
+	PACK(15, 16),
+	PACK(16, 16),
+};
+
+TEST_STATIC int k210_pll_calc_config(u32 rate, u32 rate_in,
+				     struct k210_pll_config *best)
+{
+	int i;
+	s64 error, best_error;
+	u64 ratio, inv_ratio; /* fixed point 32.32 ratio of the rates */
+	u64 max_r;
+	u64 r, f, od;
+
+	/*
+	 * Can't go over 1.75 GHz or under 21.25 MHz due to limitations on the
+	 * VCO frequency. These are not the same limits as below because od can
+	 * reduce the output frequency by 16.
+	 */
+	if (rate > 1750000000 || rate < 21250000)
+		return -EINVAL;
+
+	/* Similar restrictions on the input rate */
+	if (rate_in > 1750000000 || rate_in < 13300000)
+		return -EINVAL;
+
+	ratio = DIV_ROUND_CLOSEST_ULL((u64)rate << 32, rate_in);
+	inv_ratio = DIV_ROUND_CLOSEST_ULL((u64)rate_in << 32, rate);
+	/* Can't increase by more than 64 or reduce by more than 256 */
+	if (rate > rate_in && ratio > (64ULL << 32))
+		return -EINVAL;
+	else if (rate <= rate_in && inv_ratio > (256ULL << 32))
+		return -EINVAL;
+
+	/*
+	 * The divided clock (rate_in / r) must stay between 1.75 GHz and 13.3
+	 * MHz. There is no minimum, since the only way to get a higher input
+	 * clock than 26 MHz is to use a clock generated by a PLL. Because PLLs
+	 * cannot output frequencies greater than 1.75 GHz, the minimum would
+	 * never be greater than one.
+	 */
+	max_r = DIV_ROUND_DOWN_ULL(rate_in, 13300000);
+
+	/* Variables get immediately incremented, so start at -1th iteration */
+	i = -1;
+	f = 0;
+	r = 0;
+	od = 0;
+	best_error = S64_MAX;
+	error = best_error;
+	/* do-while here so we always try at least one ratio */
+	do {
+		/*
+		 * Whether we swapped r and od while enforcing frequency limits
+		 */
+		bool swapped = false;
+		u64 last_od = od;
+		u64 last_r = r;
+
+		/*
+		 * Try the next largest value for f (or r and od) and
+		 * recalculate the other parameters based on that
+		 */
+		if (rate > rate_in) {
+			/*
+			 * Skip factors of the same product if we already tried
+			 * out that product
+			 */
+			do {
+				i++;
+				r = UNPACK_R(factors[i]);
+				od = UNPACK_OD(factors[i]);
+			} while (i + 1 < ARRAY_SIZE(factors) &&
+				 r * od == last_r * last_od);
+
+			/* Round close */
+			f = (r * od * ratio + BIT(31)) >> 32;
+			if (f > 64)
+				f = 64;
+		} else {
+			u64 tmp = ++f * inv_ratio;
+			bool round_up = !!(tmp & BIT(31));
+			u32 goal = (tmp >> 32) + round_up;
+			u32 err, last_err;
+
+			/* Get the next r/od pair in factors */
+			while (r * od < goal && i + 1 < ARRAY_SIZE(factors)) {
+				i++;
+				r = UNPACK_R(factors[i]);
+				od = UNPACK_OD(factors[i]);
+			}
+
+			/*
+			 * This is a case of double rounding. If we rounded up
+			 * above, we need to round down (in cases of ties) here.
+			 * This prevents off-by-one errors resulting from
+			 * choosing X+2 over X when X.Y rounds up to X+1 and
+			 * there is no r * od = X+1. For the converse, when X.Y
+			 * is rounded down to X, we should choose X+1 over X-1.
+			 */
+			err = abs(r * od - goal);
+			last_err = abs(last_r * last_od - goal);
+			if (last_err < err || (round_up && last_err == err)) {
+				i--;
+				r = last_r;
+				od = last_od;
+			}
+		}
+
+		/*
+		 * Enforce limits on internal clock frequencies. If we
+		 * aren't in spec, try swapping r and od. If everything is
+		 * in-spec, calculate the relative error.
+		 */
+		while (true) {
+			/*
+			 * Whether the intermediate frequencies are out-of-spec
+			 */
+			bool out_of_spec = false;
+
+			if (r > max_r) {
+				out_of_spec = true;
+			} else {
+				/*
+				 * There is no way to only divide once; we need
+				 * to examine the frequency with and without the
+				 * effect of od.
+				 */
+				u64 vco = DIV_ROUND_CLOSEST_ULL(rate_in * f, r);
+
+				if (vco > 1750000000 || vco < 340000000)
+					out_of_spec = true;
+			}
+
+			if (out_of_spec) {
+				if (!swapped) {
+					u64 tmp = r;
+
+					r = od;
+					od = tmp;
+					swapped = true;
+					continue;
+				} else {
+					/*
+					 * Try looking ahead to see if there are
+					 * additional factors for the same
+					 * product.
+					 */
+					if (i + 1 < ARRAY_SIZE(factors)) {
+						u64 new_r, new_od;
+
+						i++;
+						new_r = UNPACK_R(factors[i]);
+						new_od = UNPACK_OD(factors[i]);
+						if (r * od == new_r * new_od) {
+							r = new_r;
+							od = new_od;
+							swapped = false;
+							continue;
+						}
+						i--;
+					}
+					break;
+				}
+			}
+
+			error = DIV_ROUND_CLOSEST_ULL(f * inv_ratio, r * od);
+			/* The lower 16 bits are spurious */
+			error = abs((error - BIT(32))) >> 16;
+
+			if (error < best_error) {
+				best->r = r;
+				best->f = f;
+				best->od = od;
+				best_error = error;
+			}
+			break;
+		}
+	} while (f < 64 && i + 1 < ARRAY_SIZE(factors) && error != 0);
+
+	if (best_error == S64_MAX)
+		return -EINVAL;
+
+	log_debug("best error %lld\n", best_error);
+	return 0;
+}
+
+static ulong k210_pll_set_rate(struct clk *clk, ulong rate)
+{
+	int err;
+	long long rate_in = clk_get_parent_rate(clk);
+	struct k210_pll_config config = {};
+	struct k210_pll *pll = to_k210_pll(clk);
+	u32 reg;
+
+	if (rate_in < 0)
+		return rate_in;
+
+	log_debug("Calculating parameters with rate=%lu and rate_in=%lld\n",
+		  rate, rate_in);
+	err = k210_pll_calc_config(rate, rate_in, &config);
+	if (err)
+		return err;
+	log_debug("Got r=%u f=%u od=%u\n", config.r, config.f, config.od);
+
+	/*
+	 * Don't use clk_disable as it might not actually disable the pll due to
+	 * refcounting
+	 */
+	k210_pll_disable(clk);
+
+	reg = readl(pll->reg);
+	reg &= ~K210_PLL_CLKR
+	    &  ~K210_PLL_CLKF
+	    &  ~K210_PLL_CLKOD
+	    &  ~K210_PLL_BWADJ;
+	reg |= FIELD_PREP(K210_PLL_CLKR, config.r - 1)
+	    |  FIELD_PREP(K210_PLL_CLKF, config.f - 1)
+	    |  FIELD_PREP(K210_PLL_CLKOD, config.od - 1)
+	    |  FIELD_PREP(K210_PLL_BWADJ, config.f - 1);
+	writel(reg, pll->reg);
+
+	err = k210_pll_enable(clk);
+	if (err)
+		return err;
+
+	serial_setbrg();
+	return clk_get_rate(clk);
+}
+#endif /* CONFIG_CLK_K210_SET_RATE */
+
+static ulong k210_pll_get_rate(struct clk *clk)
+{
+	long long rate_in = clk_get_parent_rate(clk);
+	struct k210_pll *pll = to_k210_pll(clk);
+	u64 r, f, od;
+	u32 reg = readl(pll->reg);
+
+	if (rate_in < 0 || (reg & K210_PLL_BYPASS))
+		return rate_in;
+
+	if (!(reg & K210_PLL_PWRD))
+		return 0;
+
+	r = FIELD_GET(K210_PLL_CLKR, reg) + 1;
+	f = FIELD_GET(K210_PLL_CLKF, reg) + 1;
+	od = FIELD_GET(K210_PLL_CLKOD, reg) + 1;
+
+	return DIV_ROUND_DOWN_ULL(((u64)rate_in) * f, r * od);
+}
+
+/*
+ * Wait for the PLL to be locked. If the PLL is not locked, try clearing the
+ * slip before retrying
+ */
+static void k210_pll_waitfor_lock(struct k210_pll *pll)
+{
+	u32 mask = GENMASK(pll->width - 1, 0) << pll->shift;
+
+	while (true) {
+		u32 reg = readl(pll->lock);
+
+		if ((reg & mask) == mask)
+			break;
+
+		reg |= BIT(pll->shift + K210_PLL_CLEAR_SLIP);
+		writel(reg, pll->lock);
+	}
+}
+
+/* Adapted from sysctl_pll_enable */
+static int k210_pll_enable(struct clk *clk)
+{
+	struct k210_pll *pll = to_k210_pll(clk);
+	u32 reg = readl(pll->reg);
+
+	if ((reg | K210_PLL_PWRD) && !(reg | K210_PLL_RESET))
+		return 0;
+
+	reg |= K210_PLL_PWRD;
+	writel(reg, pll->reg);
+
+	/* Ensure reset is low before asserting it */
+	reg &= ~K210_PLL_RESET;
+	writel(reg, pll->reg);
+	reg |= K210_PLL_RESET;
+	writel(reg, pll->reg);
+	nop();
+	nop();
+	reg &= ~K210_PLL_RESET;
+	writel(reg, pll->reg);
+
+	k210_pll_waitfor_lock(pll);
+
+	reg &= ~K210_PLL_BYPASS;
+	writel(reg, pll->reg);
+
+	return 0;
+}
+
+static int k210_pll_disable(struct clk *clk)
+{
+	struct k210_pll *pll = to_k210_pll(clk);
+	u32 reg = readl(pll->reg);
+
+	/*
+	 * Bypassing before powering off is important so child clocks don't stop
+	 * working. This is especially important for pll0, the indirect parent
+	 * of the cpu clock.
+	 */
+	reg |= K210_PLL_BYPASS;
+	writel(reg, pll->reg);
+
+	reg &= ~K210_PLL_PWRD;
+	writel(reg, pll->reg);
+	return 0;
+}
+
+const struct clk_ops k210_pll_ops = {
+	.get_rate = k210_pll_get_rate,
+#ifdef CONFIG_CLK_K210_SET_RATE
+	.set_rate = k210_pll_set_rate,
+#endif
+	.enable = k210_pll_enable,
+	.disable = k210_pll_disable,
+};
+
+struct clk *k210_register_pll_struct(const char *name, const char *parent_name,
+				     struct k210_pll *pll)
+{
+	int ret;
+	struct clk *clk = &pll->clk;
+
+	ret = clk_register(clk, CLK_K210_PLL, name, parent_name);
+	if (ret)
+		return ERR_PTR(ret);
+	return clk;
+}
+
+struct clk *k210_register_pll(const char *name, const char *parent_name,
+			      void __iomem *reg, void __iomem *lock, u8 shift,
+			      u8 width)
+{
+	struct clk *clk;
+	struct k210_pll *pll;
+
+	pll = kzalloc(sizeof(*pll), GFP_KERNEL);
+	if (!pll)
+		return ERR_PTR(-ENOMEM);
+	pll->reg = reg;
+	pll->lock = lock;
+	pll->shift = shift;
+	pll->width = width;
+
+	clk = k210_register_pll_struct(name, parent_name, pll);
+	if (IS_ERR(clk))
+		kfree(pll);
+	return clk;
+}
+
+U_BOOT_DRIVER(k210_pll) = {
+	.name	= CLK_K210_PLL,
+	.id	= UCLASS_CLK,
+	.ops	= &k210_pll_ops,
+};