[ARM] tegra: clock: Add dvfs support, bug fixes, and cleanups

- Add drivers to clock lookup table
- Add new pll_m entries
- Support I2C U16 divider
- Fix rate reporting on 32.768kHz clock
- Call propagate rate only if set_rate succeeds
- Add support for audio_sync clock
- Add 24MHz to PLLA frequency list
- Correct i2s1/2/spdifout mux
- Add suspend support
- Fix enable/disable parent clocks in set_parent
- Add max_rate parameter to all clocks
- DVFS support
- Add virtual cpu clock with dvfs
- Support clk_round_rate
- Fix requesting very high periph frequencies
- Add quirks for PLLU:
   PLLU is slightly different from the rest of the PLLs.  The
   lock enable bit is at bit 22 instead of 18 in the MISC
   register, and the post divider field is a single bit with
   reversed values from other PLLs.
- Simplify recalculating clock rates
- Fix UART divider flags
- Remove unused clock ops

Signed-off-by: Colin Cross <ccross@android.com>

7 files changed, 1005 insertions, 210 deletions

diff --git a/arch/arm/mach-tegra/Makefile b/arch/arm/mach-tegra/Makefile
index 62034752527..3642e58e507 100644
--- a/arch/arm/mach-tegra/Makefile
+++ b/arch/arm/mach-tegra/Makefile
@@ -8,6 +8,7 @@ obj-y                                   += pinmux.o
 obj-y					+= fuse.o
 obj-$(CONFIG_ARCH_TEGRA_2x_SOC)         += clock.o
 obj-$(CONFIG_ARCH_TEGRA_2x_SOC)         += tegra2_clocks.o
+obj-$(CONFIG_ARCH_TEGRA_2x_SOC)         += tegra2_dvfs.o
 obj-$(CONFIG_ARCH_TEGRA_2x_SOC)		+= pinmux-t2-tables.o
 obj-$(CONFIG_SMP)                       += platsmp.o localtimer.o headsmp.o
 obj-$(CONFIG_HOTPLUG_CPU)               += hotplug.o
diff --git a/arch/arm/mach-tegra/clock.c b/arch/arm/mach-tegra/clock.c
index 03ad578349b..ae19f95585b 100644
--- a/arch/arm/mach-tegra/clock.c
+++ b/arch/arm/mach-tegra/clock.c
@@ -24,13 +24,80 @@
 #include <linux/debugfs.h>
 #include <linux/slab.h>
 #include <linux/seq_file.h>
+#include <linux/regulator/consumer.h>
 #include <asm/clkdev.h>
 
 #include "clock.h"
+#include "board.h"
+#include "fuse.h"
 
 static LIST_HEAD(clocks);
 
 static DEFINE_SPINLOCK(clock_lock);
+static DEFINE_MUTEX(dvfs_lock);
+
+static int clk_is_dvfs(struct clk *c)
+{
+	return (c->dvfs != NULL);
+};
+
+static int dvfs_set_rate(struct dvfs *d, unsigned long rate)
+{
+	struct dvfs_table *t;
+
+	if (d->table == NULL)
+		return -ENODEV;
+
+	for (t = d->table; t->rate != 0; t++) {
+		if (rate <= t->rate) {
+			if (!d->reg)
+				return 0;
+
+			return regulator_set_voltage(d->reg,
+				t->millivolts * 1000,
+				d->max_millivolts * 1000);
+		}
+	}
+
+	return -EINVAL;
+}
+
+static void dvfs_init(struct clk *c)
+{
+	int process_id;
+	int i;
+	struct dvfs_table *table;
+
+	process_id = c->dvfs->cpu ? tegra_core_process_id() :
+		tegra_cpu_process_id();
+
+	for (i = 0; i < c->dvfs->process_id_table_length; i++)
+		if (process_id == c->dvfs->process_id_table[i].process_id)
+			c->dvfs->table = c->dvfs->process_id_table[i].table;
+
+	if (c->dvfs->table == NULL) {
+		pr_err("Failed to find dvfs table for clock %s process %d\n",
+			c->name, process_id);
+		return;
+	}
+
+	c->dvfs->max_millivolts = 0;
+	for (table = c->dvfs->table; table->rate != 0; table++)
+		if (c->dvfs->max_millivolts < table->millivolts)
+			c->dvfs->max_millivolts = table->millivolts;
+
+	c->dvfs->reg = regulator_get(NULL, c->dvfs->reg_id);
+
+	if (IS_ERR(c->dvfs->reg)) {
+		pr_err("Failed to get regulator %s for clock %s\n",
+			c->dvfs->reg_id, c->name);
+		c->dvfs->reg = NULL;
+		return;
+	}
+
+	if (c->refcnt > 0)
+		dvfs_set_rate(c->dvfs, c->rate);
+}
 
 struct clk *tegra_get_clock_by_name(const char *name)
 {
@@ -48,14 +115,31 @@ struct clk *tegra_get_clock_by_name(const char *name)
 	return ret;
 }
 
+static void clk_recalculate_rate(struct clk *c)
+{
+	u64 rate;
+
+	if (!c->parent)
+		return;
+
+	rate = c->parent->rate;
+
+	if (c->mul != 0 && c->div != 0) {
+		rate = rate * c->mul;
+		do_div(rate, c->div);
+	}
+
+	if (rate > c->max_rate)
+		pr_warn("clocks: Set clock %s to rate %llu, max is %lu\n",
+			c->name, rate, c->max_rate);
+
+	c->rate = rate;
+}
+
 int clk_reparent(struct clk *c, struct clk *parent)
 {
 	pr_debug("%s: %s\n", __func__, c->name);
-	if (c->refcnt && c->parent)
-		clk_disable_locked(c->parent);
 	c->parent = parent;
-	if (c->refcnt && c->parent)
-		clk_enable_locked(c->parent);
 	list_del(&c->sibling);
 	list_add_tail(&c->sibling, &parent->children);
 	return 0;
@@ -67,8 +151,7 @@ static void propagate_rate(struct clk *c)
 	pr_debug("%s: %s\n", __func__, c->name);
 	list_for_each_entry(clkp, &c->children, sibling) {
 		pr_debug("   %s\n", clkp->name);
-		if (clkp->ops->recalculate_rate)
-			clkp->ops->recalculate_rate(clkp);
+		clk_recalculate_rate(clkp);
 		propagate_rate(clkp);
 	}
 }
@@ -77,6 +160,8 @@ void clk_init(struct clk *c)
 {
 	unsigned long flags;
 
+	pr_debug("%s: %s\n", __func__, c->name);
+
 	spin_lock_irqsave(&clock_lock, flags);
 
 	INIT_LIST_HEAD(&c->children);
@@ -85,6 +170,8 @@ void clk_init(struct clk *c)
 	if (c->ops && c->ops->init)
 		c->ops->init(c);
 
+	clk_recalculate_rate(c);
+
 	list_add(&c->node, &clocks);
 
 	if (c->parent)
@@ -122,13 +209,38 @@ int clk_enable_locked(struct clk *c)
 	return 0;
 }
 
+int clk_enable_cansleep(struct clk *c)
+{
+	int ret;
+	unsigned long flags;
+
+	mutex_lock(&dvfs_lock);
+
+	if (clk_is_dvfs(c) && c->refcnt > 0)
+		dvfs_set_rate(c->dvfs, c->rate);
+
+	spin_lock_irqsave(&clock_lock, flags);
+	ret = clk_enable_locked(c);
+	spin_unlock_irqrestore(&clock_lock, flags);
+
+	mutex_unlock(&dvfs_lock);
+
+	return ret;
+}
+EXPORT_SYMBOL(clk_enable_cansleep);
+
 int clk_enable(struct clk *c)
 {
 	int ret;
 	unsigned long flags;
+
+	if (clk_is_dvfs(c))
+		BUG();
+
 	spin_lock_irqsave(&clock_lock, flags);
 	ret = clk_enable_locked(c);
 	spin_unlock_irqrestore(&clock_lock, flags);
+
 	return ret;
 }
 EXPORT_SYMBOL(clk_enable);
@@ -152,9 +264,30 @@ void clk_disable_locked(struct clk *c)
 	c->refcnt--;
 }
 
+void clk_disable_cansleep(struct clk *c)
+{
+	unsigned long flags;
+
+	mutex_lock(&dvfs_lock);
+
+	spin_lock_irqsave(&clock_lock, flags);
+	clk_disable_locked(c);
+	spin_unlock_irqrestore(&clock_lock, flags);
+
+	if (clk_is_dvfs(c) && c->refcnt == 0)
+		dvfs_set_rate(c->dvfs, c->rate);
+
+	mutex_unlock(&dvfs_lock);
+}
+EXPORT_SYMBOL(clk_disable_cansleep);
+
 void clk_disable(struct clk *c)
 {
 	unsigned long flags;
+
+	if (clk_is_dvfs(c))
+		BUG();
+
 	spin_lock_irqsave(&clock_lock, flags);
 	clk_disable_locked(c);
 	spin_unlock_irqrestore(&clock_lock, flags);
@@ -175,6 +308,8 @@ int clk_set_parent_locked(struct clk *c, struct clk *parent)
 	if (ret)
 		return ret;
 
+	clk_recalculate_rate(c);
+
 	propagate_rate(c);
 
 	return 0;
@@ -197,22 +332,69 @@ struct clk *clk_get_parent(struct clk *c)
 }
 EXPORT_SYMBOL(clk_get_parent);
 
-int clk_set_rate(struct clk *c, unsigned long rate)
+int clk_set_rate_locked(struct clk *c, unsigned long rate)
+{
+	int ret;
+
+	if (rate > c->max_rate)
+		rate = c->max_rate;
+
+	if (!c->ops || !c->ops->set_rate)
+		return -ENOSYS;
+
+	ret = c->ops->set_rate(c, rate);
+
+	if (ret)
+		return ret;
+
+	clk_recalculate_rate(c);
+
+	propagate_rate(c);
+
+	return 0;
+}
+
+int clk_set_rate_cansleep(struct clk *c, unsigned long rate)
 {
 	int ret = 0;
 	unsigned long flags;
 
+	pr_debug("%s: %s\n", __func__, c->name);
+
+	mutex_lock(&dvfs_lock);
+
+	if (rate > c->rate)
+		ret = dvfs_set_rate(c->dvfs, rate);
+	if (ret)
+		goto out;
+
 	spin_lock_irqsave(&clock_lock, flags);
+	ret = clk_set_rate_locked(c, rate);
+	spin_unlock_irqrestore(&clock_lock, flags);
 
-	pr_debug("%s: %s\n", __func__, c->name);
+	if (ret)
+		goto out;
 
-	if (c->ops && c->ops->set_rate)
-		ret = c->ops->set_rate(c, rate);
-	else
-		ret = -ENOSYS;
+	ret = dvfs_set_rate(c->dvfs, rate);
 
-	propagate_rate(c);
+out:
+	mutex_unlock(&dvfs_lock);
+	return ret;
+}
+EXPORT_SYMBOL(clk_set_rate_cansleep);
+
+int clk_set_rate(struct clk *c, unsigned long rate)
+{
+	int ret = 0;
+	unsigned long flags;
+
+	pr_debug("%s: %s\n", __func__, c->name);
+
+	if (clk_is_dvfs(c))
+		BUG();
 
+	spin_lock_irqsave(&clock_lock, flags);
+	ret = clk_set_rate_locked(c, rate);
 	spin_unlock_irqrestore(&clock_lock, flags);
 
 	return ret;
@@ -235,6 +417,20 @@ unsigned long clk_get_rate(struct clk *c)
 }
 EXPORT_SYMBOL(clk_get_rate);
 
+long clk_round_rate(struct clk *c, unsigned long rate)
+{
+	pr_debug("%s: %s\n", __func__, c->name);
+
+	if (!c->ops || !c->ops->round_rate)
+		return -ENOSYS;
+
+	if (rate > c->max_rate)
+		rate = c->max_rate;
+
+	return c->ops->round_rate(c, rate);
+}
+EXPORT_SYMBOL(clk_round_rate);
+
 static int tegra_clk_init_one_from_table(struct tegra_clk_init_table *table)
 {
 	struct clk *c;
@@ -308,13 +504,28 @@ void tegra_periph_reset_assert(struct clk *c)
 }
 EXPORT_SYMBOL(tegra_periph_reset_assert);
 
-int __init tegra_init_clock(void)
+void __init tegra_init_clock(void)
 {
 	tegra2_init_clocks();
+}
+
+int __init tegra_init_dvfs(void)
+{
+	struct clk *c, *safe;
+
+	mutex_lock(&dvfs_lock);
+
+	list_for_each_entry_safe(c, safe, &clocks, node)
+		if (c->dvfs)
+			dvfs_init(c);
+
+	mutex_unlock(&dvfs_lock);
 
 	return 0;
 }
 
+late_initcall(tegra_init_dvfs);
+
 #ifdef CONFIG_DEBUG_FS
 static struct dentry *clk_debugfs_root;
 
@@ -324,7 +535,7 @@ static void clock_tree_show_one(struct seq_file *s, struct clk *c, int level)
 	struct clk *child;
 	struct clk *safe;
 	const char *state = "uninit";
-	char div[5] = {0};
+	char div[8] = {0};
 
 	if (c->state == ON)
 		state = "on";
@@ -332,16 +543,26 @@ static void clock_tree_show_one(struct seq_file *s, struct clk *c, int level)
 		state = "off";
 
 	if (c->mul != 0 && c->div != 0) {
-		BUG_ON(c->mul > 2);
-		if (c->mul > c->div)
-			snprintf(div, sizeof(div), "x%d", c->mul / c->div);
-		else
+		if (c->mul > c->div) {
+			int mul = c->mul / c->div;
+			int mul2 = (c->mul * 10 / c->div) % 10;
+			int mul3 = (c->mul * 10) % c->div;
+			if (mul2 == 0 && mul3 == 0)
+				snprintf(div, sizeof(div), "x%d", mul);
+			else if (mul3 == 0)
+				snprintf(div, sizeof(div), "x%d.%d", mul, mul2);
+			else
+				snprintf(div, sizeof(div), "x%d.%d..", mul, mul2);
+		} else {
 			snprintf(div, sizeof(div), "%d%s", c->div / c->mul,
 				(c->div % c->mul) ? ".5" : "");
+		}
 	}
 
-	seq_printf(s, "%*s%-*s %-6s %-3d %-5s %-10lu\n",
-		level * 3 + 1, c->set ? "" : "*",
+	seq_printf(s, "%*s%c%c%-*s %-6s %-3d %-8s %-10lu\n",
+		level * 3 + 1, "",
+		c->rate > c->max_rate ? '!' : ' ',
+		!c->set ? '*' : ' ',
 		30 - level * 3, c->name,
 		state, c->refcnt, div, c->rate);
 	list_for_each_entry_safe(child, safe, &c->children, sibling) {
@@ -353,8 +574,8 @@ static int clock_tree_show(struct seq_file *s, void *data)
 {
 	struct clk *c;
 	unsigned long flags;
-	seq_printf(s, " clock                          state  ref div   rate      \n");
-	seq_printf(s, "-----------------------------------------------------------\n");
+	seq_printf(s, "   clock                          state  ref div      rate\n");
+	seq_printf(s, "--------------------------------------------------------------\n");
 	spin_lock_irqsave(&clock_lock, flags);
 	list_for_each_entry(c, &clocks, node)
 		if (c->parent == NULL)
diff --git a/arch/arm/mach-tegra/clock.h b/arch/arm/mach-tegra/clock.h
index af7c70e2a3b..94fd859770f 100644
--- a/arch/arm/mach-tegra/clock.h
+++ b/arch/arm/mach-tegra/clock.h
@@ -27,18 +27,43 @@
 #define DIV_U71			(1 << 1)
 #define DIV_U71_FIXED		(1 << 2)
 #define DIV_2			(1 << 3)
-#define PLL_FIXED		(1 << 4)
-#define PLL_HAS_CPCON		(1 << 5)
-#define MUX			(1 << 6)
-#define PLLD			(1 << 7)
-#define PERIPH_NO_RESET		(1 << 8)
-#define PERIPH_NO_ENB		(1 << 9)
-#define PERIPH_EMC_ENB		(1 << 10)
-#define PERIPH_MANUAL_RESET	(1 << 11)
-#define PLL_ALT_MISC_REG	(1 << 12)
+#define DIV_U16			(1 << 4)
+#define PLL_FIXED		(1 << 5)
+#define PLL_HAS_CPCON		(1 << 6)
+#define MUX			(1 << 7)
+#define PLLD			(1 << 8)
+#define PERIPH_NO_RESET		(1 << 9)
+#define PERIPH_NO_ENB		(1 << 10)
+#define PERIPH_EMC_ENB		(1 << 11)
+#define PERIPH_MANUAL_RESET	(1 << 12)
+#define PLL_ALT_MISC_REG	(1 << 13)
+#define PLLU			(1 << 14)
 #define ENABLE_ON_INIT		(1 << 28)
 
 struct clk;
+struct regulator;
+
+struct dvfs_table {
+	unsigned long rate;
+	int millivolts;
+};
+
+struct dvfs_process_id_table {
+	int process_id;
+	struct dvfs_table *table;
+};
+
+
+struct dvfs {
+	struct regulator *reg;
+	struct dvfs_table *table;
+	int max_millivolts;
+
+	int process_id_table_length;
+	const char *reg_id;
+	bool cpu;
+	struct dvfs_process_id_table process_id_table[];
+};
 
 struct clk_mux_sel {
 	struct clk	*input;
@@ -58,12 +83,9 @@ struct clk_ops {
 	void		(*init)(struct clk *);
 	int		(*enable)(struct clk *);
 	void		(*disable)(struct clk *);
-	void		(*recalc)(struct clk *);
 	int		(*set_parent)(struct clk *, struct clk *);
 	int		(*set_rate)(struct clk *, unsigned long);
-	unsigned long	(*get_rate)(struct clk *);
 	long		(*round_rate)(struct clk *, unsigned long);
-	unsigned long	(*recalculate_rate)(struct clk *);
 };
 
 enum clk_state {
@@ -85,6 +107,7 @@ struct clk {
 	struct clk			*parent;
 	struct clk_lookup		lookup;
 	unsigned long			rate;
+	unsigned long			max_rate;
 	u32				flags;
 	u32				refcnt;
 	const char			*name;
@@ -103,10 +126,6 @@ struct clk {
 	unsigned long			cf_max;
 	unsigned long			vco_min;
 	unsigned long			vco_max;
-	u32				m;
-	u32				n;
-	u32				p;
-	u32				cpcon;
 	const struct clk_pll_table	*pll_table;
 
 	/* DIV */
@@ -117,6 +136,12 @@ struct clk {
 	const struct clk_mux_sel	*inputs;
 	u32				sel;
 	u32				reg_mask;
+
+	/* Virtual cpu clock */
+	struct clk			*main;
+	struct clk			*backup;
+
+	struct dvfs			*dvfs;
 };
 
 
@@ -141,6 +166,7 @@ unsigned long clk_measure_input_freq(void);
 void clk_disable_locked(struct clk *c);
 int clk_enable_locked(struct clk *c);
 int clk_set_parent_locked(struct clk *c, struct clk *parent);
+int clk_set_rate_locked(struct clk *c, unsigned long rate);
 int clk_reparent(struct clk *c, struct clk *parent);
 void tegra_clk_init_from_table(struct tegra_clk_init_table *table);
 
diff --git a/arch/arm/mach-tegra/include/mach/clk.h b/arch/arm/mach-tegra/include/mach/clk.h
index 2896f25ebfb..d7723955dac 100644
--- a/arch/arm/mach-tegra/include/mach/clk.h
+++ b/arch/arm/mach-tegra/include/mach/clk.h
@@ -23,4 +23,9 @@
 void tegra_periph_reset_deassert(struct clk *c);
 void tegra_periph_reset_assert(struct clk *c);
 
+int clk_enable_cansleep(struct clk *clk);
+void clk_disable_cansleep(struct clk *clk);
+int clk_set_rate_cansleep(struct clk *clk, unsigned long rate);
+int clk_set_parent_cansleep(struct clk *clk, struct clk *parent);
+
 #endif
diff --git a/arch/arm/mach-tegra/tegra2_clocks.c b/arch/arm/mach-tegra/tegra2_clocks.c
index 426163231ff..cf1c9d0ef7e 100644
--- a/arch/arm/mach-tegra/tegra2_clocks.c
+++ b/arch/arm/mach-tegra/tegra2_clocks.c
@@ -30,14 +30,21 @@
 #include <mach/iomap.h>
 
 #include "clock.h"
+#include "fuse.h"
+#include "tegra2_dvfs.h"
 
 #define RST_DEVICES			0x004
 #define RST_DEVICES_SET			0x300
 #define RST_DEVICES_CLR			0x304
+#define RST_DEVICES_NUM			3
 
 #define CLK_OUT_ENB			0x010
 #define CLK_OUT_ENB_SET			0x320
 #define CLK_OUT_ENB_CLR			0x324
+#define CLK_OUT_ENB_NUM			3
+
+#define CLK_MASK_ARM			0x44
+#define MISC_CLK_ENB			0x48
 
 #define OSC_CTRL			0x50
 #define OSC_CTRL_OSC_FREQ_MASK		(3<<30)
@@ -45,6 +52,7 @@
 #define OSC_CTRL_OSC_FREQ_19_2MHZ	(1<<30)
 #define OSC_CTRL_OSC_FREQ_12MHZ		(2<<30)
 #define OSC_CTRL_OSC_FREQ_26MHZ		(3<<30)
+#define OSC_CTRL_MASK			0x3f2
 
 #define OSC_FREQ_DET			0x58
 #define OSC_FREQ_DET_TRIG		(1<<31)
@@ -53,10 +61,17 @@
 #define OSC_FREQ_DET_BUSY		(1<<31)
 #define OSC_FREQ_DET_CNT_MASK		0xFFFF
 
+#define PERIPH_CLK_SOURCE_I2S1		0x100
+#define PERIPH_CLK_SOURCE_EMC		0x19c
+#define PERIPH_CLK_SOURCE_OSC		0x1fc
+#define PERIPH_CLK_SOURCE_NUM \
+	((PERIPH_CLK_SOURCE_OSC - PERIPH_CLK_SOURCE_I2S1) / 4)
+
 #define PERIPH_CLK_SOURCE_MASK		(3<<30)
 #define PERIPH_CLK_SOURCE_SHIFT		30
 #define PERIPH_CLK_SOURCE_ENABLE	(1<<28)
-#define PERIPH_CLK_SOURCE_DIV_MASK	0xFF
+#define PERIPH_CLK_SOURCE_DIVU71_MASK	0xFF
+#define PERIPH_CLK_SOURCE_DIVU16_MASK	0xFFFF
 #define PERIPH_CLK_SOURCE_DIV_SHIFT	0
 
 #define PLL_BASE			0x0
@@ -79,8 +94,9 @@
 #define PLL_OUT_RESET_DISABLE		(1<<0)
 
 #define PLL_MISC(c)			(((c)->flags & PLL_ALT_MISC_REG) ? 0x4 : 0xc)
+#define PLL_MISC_LOCK_ENABLE(c)		(((c)->flags & PLLU) ? (1<<22) : (1<<18))
+
 #define PLL_MISC_DCCON_SHIFT		20
-#define PLL_MISC_LOCK_ENABLE		(1<<18)
 #define PLL_MISC_CPCON_SHIFT		8
 #define PLL_MISC_CPCON_MASK		(0xF<<PLL_MISC_CPCON_SHIFT)
 #define PLL_MISC_LFCON_SHIFT		4
@@ -88,6 +104,8 @@
 #define PLL_MISC_VCOCON_SHIFT		0
 #define PLL_MISC_VCOCON_MASK		(0xF<<PLL_MISC_VCOCON_SHIFT)
 
+#define PLLU_BASE_POST_DIV		(1<<20)
+
 #define PLLD_MISC_CLKENABLE		(1<<30)
 #define PLLD_MISC_DIV_RST		(1<<23)
 #define PLLD_MISC_DCCON_SHIFT		12
@@ -143,30 +161,37 @@ unsigned long clk_measure_input_freq(void)
 	}
 }
 
-static int clk_div71_get_divider(struct clk *c, unsigned long rate)
+static int clk_div71_get_divider(unsigned long parent_rate, unsigned long rate)
 {
-	unsigned long divider_u71;
+	s64 divider_u71 = parent_rate * 2;
+	divider_u71 += rate - 1;
+	do_div(divider_u71, rate);
 
-	divider_u71 = DIV_ROUND_UP(c->rate * 2, rate);
+	if (divider_u71 - 2 < 0)
+		return 0;
 
-	if (divider_u71 - 2 > 255 || divider_u71 - 2 < 0)
+	if (divider_u71 - 2 > 255)
 		return -EINVAL;
 
 	return divider_u71 - 2;
 }
 
-static unsigned long tegra2_clk_recalculate_rate(struct clk *c)
+static int clk_div16_get_divider(unsigned long parent_rate, unsigned long rate)
 {
-	unsigned long rate;
-	rate = c->parent->rate;
+	s64 divider_u16;
 
-	if (c->mul != 0 && c->div != 0)
-		c->rate = rate * c->mul / c->div;
-	else
-		c->rate = rate;
-	return c->rate;
-}
+	divider_u16 = parent_rate;
+	divider_u16 += rate - 1;
+	do_div(divider_u16, rate);
 
+	if (divider_u16 - 1 < 0)
+		return 0;
+
+	if (divider_u16 - 1 > 255)
+		return -EINVAL;
+
+	return divider_u16 - 1;
+}
 
 /* clk_m functions */
 static unsigned long tegra2_clk_m_autodetect_rate(struct clk *c)
@@ -244,7 +269,6 @@ static void tegra2_super_clk_init(struct clk *c)
 	}
 	BUG_ON(sel->input == NULL);
 	c->parent = sel->input;
-	tegra2_clk_recalculate_rate(c);
 }
 
 static int tegra2_super_clk_enable(struct clk *c)
@@ -266,6 +290,7 @@ static int tegra2_super_clk_set_parent(struct clk *c, struct clk *p)
 	u32 val;
 	const struct clk_mux_sel *sel;
 	int shift;
+
 	val = clk_readl(c->reg + SUPER_CLK_MUX);;
 	BUG_ON(((val & SUPER_STATE_MASK) != SUPER_STATE_RUN) &&
 		((val & SUPER_STATE_MASK) != SUPER_STATE_IDLE));
@@ -273,11 +298,18 @@ static int tegra2_super_clk_set_parent(struct clk *c, struct clk *p)
 		SUPER_IDLE_SOURCE_SHIFT : SUPER_RUN_SOURCE_SHIFT;
 	for (sel = c->inputs; sel->input != NULL; sel++) {
 		if (sel->input == p) {
-			clk_reparent(c, p);
 			val &= ~(SUPER_SOURCE_MASK << shift);
 			val |= sel->value << shift;
+
+			if (c->refcnt)
+				clk_enable_locked(p);
+
 			clk_writel(val, c->reg);
-			c->rate = c->parent->rate;
+
+			if (c->refcnt && c->parent)
+				clk_disable_locked(c->parent);
+
+			clk_reparent(c, p);
 			return 0;
 		}
 	}
@@ -289,7 +321,61 @@ static struct clk_ops tegra_super_ops = {
 	.enable			= tegra2_super_clk_enable,
 	.disable		= tegra2_super_clk_disable,
 	.set_parent		= tegra2_super_clk_set_parent,
-	.recalculate_rate	= tegra2_clk_recalculate_rate,
+};
+
+/* virtual cpu clock functions */
+/* some clocks can not be stopped (cpu, memory bus) while the SoC is running.
+   To change the frequency of these clocks, the parent pll may need to be
+   reprogrammed, so the clock must be moved off the pll, the pll reprogrammed,
+   and then the clock moved back to the pll.  To hide this sequence, a virtual
+   clock handles it.
+ */
+static void tegra2_cpu_clk_init(struct clk *c)
+{
+}
+
+static int tegra2_cpu_clk_enable(struct clk *c)
+{
+	return 0;
+}
+
+static void tegra2_cpu_clk_disable(struct clk *c)
+{
+	pr_debug("%s on clock %s\n", __func__, c->name);
+
+	/* oops - don't disable the CPU clock! */
+	BUG();
+}
+
+static int tegra2_cpu_clk_set_rate(struct clk *c, unsigned long rate)
+{
+	int ret;
+	ret = clk_set_parent_locked(c->parent, c->backup);
+	if (ret) {
+		pr_err("Failed to switch cpu to clock %s\n", c->backup->name);
+		return ret;
+	}
+
+	ret = clk_set_rate_locked(c->main, rate);
+	if (ret) {
+		pr_err("Failed to change cpu pll to %lu\n", rate);
+		return ret;
+	}
+
+	ret = clk_set_parent_locked(c->parent, c->main);
+	if (ret) {
+		pr_err("Failed to switch cpu to clock %s\n", c->main->name);
+		return ret;
+	}
+
+	return 0;
+}
+
+static struct clk_ops tegra_cpu_ops = {
+	.init     = tegra2_cpu_clk_init,
+	.enable   = tegra2_cpu_clk_enable,
+	.disable  = tegra2_cpu_clk_disable,
+	.set_rate = tegra2_cpu_clk_set_rate,
 };
 
 /* bus clock functions */
@@ -299,7 +385,6 @@ static void tegra2_bus_clk_init(struct clk *c)
 	c->state = ((val >> c->reg_shift) & BUS_CLK_DISABLE) ? OFF : ON;
 	c->div = ((val >> c->reg_shift) & BUS_CLK_DIV_MASK) + 1;
 	c->mul = 1;
-	tegra2_clk_recalculate_rate(c);
 }
 
 static int tegra2_bus_clk_enable(struct clk *c)
@@ -340,27 +425,15 @@ static struct clk_ops tegra_bus_ops = {
 	.enable			= tegra2_bus_clk_enable,
 	.disable		= tegra2_bus_clk_disable,
 	.set_rate		= tegra2_bus_clk_set_rate,
-	.recalculate_rate	= tegra2_clk_recalculate_rate,
 };
 
 /* PLL Functions */
-static unsigned long tegra2_pll_clk_recalculate_rate(struct clk *c)
-{
-	u64 rate;
-	rate = c->parent->rate;
-	rate *= c->n;
-	do_div(rate, c->m);
-	if (c->p == 2)
-		rate >>= 1;
-	c->rate = rate;
-	return c->rate;
-}
-
 static int tegra2_pll_clk_wait_for_lock(struct clk *c)
 {
 	ktime_t before;
 
 	before = ktime_get();
+
 	while (!(clk_readl(c->reg + PLL_BASE) & PLL_BASE_LOCK)) {
 		if (ktime_us_delta(ktime_get(), before) > 5000) {
 			pr_err("Timed out waiting for lock bit on pll %s",
@@ -380,24 +453,19 @@ static void tegra2_pll_clk_init(struct clk *c)
 
 	if (c->flags & PLL_FIXED && !(val & PLL_BASE_OVERRIDE)) {
 		pr_warning("Clock %s has unknown fixed frequency\n", c->name);
-		c->n = 1;
-		c->m = 0;
-		c->p = 1;
+		c->mul = 1;
+		c->div = 1;
 	} else if (val & PLL_BASE_BYPASS) {
-		c->n = 1;
-		c->m = 1;
-		c->p = 1;
+		c->mul = 1;
+		c->div = 1;
 	} else {
-		c->n = (val & PLL_BASE_DIVN_MASK) >> PLL_BASE_DIVN_SHIFT;
-		c->m = (val & PLL_BASE_DIVM_MASK) >> PLL_BASE_DIVM_SHIFT;
-		c->p = (val & PLL_BASE_DIVP_MASK) ? 2 : 1;
+		c->mul = (val & PLL_BASE_DIVN_MASK) >> PLL_BASE_DIVN_SHIFT;
+		c->div = (val & PLL_BASE_DIVM_MASK) >> PLL_BASE_DIVM_SHIFT;
+		if (c->flags & PLLU)
+			c->div *= (val & PLLU_BASE_POST_DIV) ? 1 : 2;
+		else
+			c->div *= (val & PLL_BASE_DIVP_MASK) ? 2 : 1;
 	}
-
-	val = clk_readl(c->reg + PLL_MISC(c));
-	if (c->flags & PLL_HAS_CPCON)
-		c->cpcon = (val & PLL_MISC_CPCON_MASK) >> PLL_MISC_CPCON_SHIFT;
-
-	tegra2_pll_clk_recalculate_rate(c);
 }
 
 static int tegra2_pll_clk_enable(struct clk *c)
@@ -411,7 +479,7 @@ static int tegra2_pll_clk_enable(struct clk *c)
 	clk_writel(val, c->reg + PLL_BASE);
 
 	val = clk_readl(c->reg + PLL_MISC(c));
-	val |= PLL_MISC_LOCK_ENABLE;
+	val |= PLL_MISC_LOCK_ENABLE(c);
 	clk_writel(val, c->reg + PLL_MISC(c));
 
 	tegra2_pll_clk_wait_for_lock(c);
@@ -441,33 +509,36 @@ static int tegra2_pll_clk_set_rate(struct clk *c, unsigned long rate)
 	input_rate = c->parent->rate;
 	for (sel = c->pll_table; sel->input_rate != 0; sel++) {
 		if (sel->input_rate == input_rate && sel->output_rate == rate) {
-			c->n = sel->n;
-			c->m = sel->m;
-			c->p = sel->p;
-			c->cpcon = sel->cpcon;
+			c->mul = sel->n;
+			c->div = sel->m * sel->p;
 
 			val = clk_readl(c->reg + PLL_BASE);
 			if (c->flags & PLL_FIXED)
 				val |= PLL_BASE_OVERRIDE;
 			val &= ~(PLL_BASE_DIVP_MASK | PLL_BASE_DIVN_MASK |
 				 PLL_BASE_DIVM_MASK);
-			val |= (c->m << PLL_BASE_DIVM_SHIFT) |
-				(c->n << PLL_BASE_DIVN_SHIFT);
-			BUG_ON(c->p > 2);
-			if (c->p == 2)
-				val |= 1 << PLL_BASE_DIVP_SHIFT;
+			val |= (sel->m << PLL_BASE_DIVM_SHIFT) |
+				(sel->n << PLL_BASE_DIVN_SHIFT);
+			BUG_ON(sel->p < 1 || sel->p > 2);
+			if (c->flags & PLLU) {
+				if (sel->p == 1)
+					val |= PLLU_BASE_POST_DIV;
+			} else {
+				if (sel->p == 2)
+					val |= 1 << PLL_BASE_DIVP_SHIFT;
+			}
 			clk_writel(val, c->reg + PLL_BASE);
 
 			if (c->flags & PLL_HAS_CPCON) {
-				val = c->cpcon << PLL_MISC_CPCON_SHIFT;
-				val |= PLL_MISC_LOCK_ENABLE;
+				val = clk_readl(c->reg + PLL_MISC(c));
+				val &= ~PLL_MISC_CPCON_MASK;
+				val |= sel->cpcon << PLL_MISC_CPCON_SHIFT;
 				clk_writel(val, c->reg + PLL_MISC(c));
 			}
 
 			if (c->state == ON)
 				tegra2_pll_clk_enable(c);
 
-			c->rate = rate;
 			return 0;
 		}
 	}
@@ -479,7 +550,21 @@ static struct clk_ops tegra_pll_ops = {
 	.enable			= tegra2_pll_clk_enable,
 	.disable		= tegra2_pll_clk_disable,
 	.set_rate		= tegra2_pll_clk_set_rate,
-	.recalculate_rate	= tegra2_pll_clk_recalculate_rate,
+};
+
+static void tegra2_pllx_clk_init(struct clk *c)
+{
+	tegra2_pll_clk_init(c);
+
+	if (tegra_sku_id() == 7)
+		c->max_rate = 750000000;
+}
+
+static struct clk_ops tegra_pllx_ops = {
+	.init     = tegra2_pllx_clk_init,
+	.enable   = tegra2_pll_clk_enable,
+	.disable  = tegra2_pll_clk_disable,
+	.set_rate = tegra2_pll_clk_set_rate,
 };
 
 /* Clock divider ops */
@@ -503,8 +588,6 @@ static void tegra2_pll_div_clk_init(struct clk *c)
 		c->div = 1;
 		c->mul = 1;
 	}
-
-	tegra2_clk_recalculate_rate(c);
 }
 
 static int tegra2_pll_div_clk_enable(struct clk *c)
@@ -565,7 +648,7 @@ static int tegra2_pll_div_clk_set_rate(struct clk *c, unsigned long rate)
 	int divider_u71;
 	pr_debug("%s: %s %lu\n", __func__, c->name, rate);
 	if (c->flags & DIV_U71) {
-		divider_u71 = clk_div71_get_divider(c->parent, rate);
+		divider_u71 = clk_div71_get_divider(c->parent->rate, rate);
 		if (divider_u71 >= 0) {
 			val = clk_readl(c->reg);
 			new_val = val >> c->reg_shift;
@@ -580,25 +663,37 @@ static int tegra2_pll_div_clk_set_rate(struct clk *c, unsigned long rate)
 			clk_writel(val, c->reg);
 			c->div = divider_u71 + 2;
 			c->mul = 2;
-			tegra2_clk_recalculate_rate(c);
 			return 0;
 		}
 	} else if (c->flags & DIV_2) {
-		if (c->parent->rate == rate * 2) {
-			c->rate = rate;
+		if (c->parent->rate == rate * 2)
 			return 0;
-		}
 	}
 	return -EINVAL;
 }
 
+static long tegra2_pll_div_clk_round_rate(struct clk *c, unsigned long rate)
+{
+	int divider;
+	pr_debug("%s: %s %lu\n", __func__, c->name, rate);
+
+	if (c->flags & DIV_U71) {
+		divider = clk_div71_get_divider(c->parent->rate, rate);
+		if (divider < 0)
+			return divider;
+		return c->parent->rate * 2 / (divider + 2);
+	} else if (c->flags & DIV_2) {
+		return c->parent->rate / 2;
+	}
+	return -EINVAL;
+}
 
 static struct clk_ops tegra_pll_div_ops = {
 	.init			= tegra2_pll_div_clk_init,
 	.enable			= tegra2_pll_div_clk_enable,
 	.disable		= tegra2_pll_div_clk_disable,
 	.set_rate		= tegra2_pll_div_clk_set_rate,
-	.recalculate_rate	= tegra2_clk_recalculate_rate,
+	.round_rate		= tegra2_pll_div_clk_round_rate,
 };
 
 /* Periph clk ops */
@@ -621,9 +716,13 @@ static void tegra2_periph_clk_init(struct clk *c)
 	}
 
 	if (c->flags & DIV_U71) {
-		u32 divu71 = val & PERIPH_CLK_SOURCE_DIV_MASK;
+		u32 divu71 = val & PERIPH_CLK_SOURCE_DIVU71_MASK;
 		c->div = divu71 + 2;
 		c->mul = 2;
+	} else if (c->flags & DIV_U16) {
+		u32 divu16 = val & PERIPH_CLK_SOURCE_DIVU16_MASK;
+		c->div = divu16 + 1;
+		c->mul = 1;
 	} else {
 		c->div = 1;
 		c->mul = 1;
@@ -637,7 +736,6 @@ static void tegra2_periph_clk_init(struct clk *c)
 		if (clk_readl(RST_DEVICES + PERIPH_CLK_TO_ENB_REG(c)) &
 				PERIPH_CLK_TO_ENB_BIT(c))
 			c->state = OFF;
-	tegra2_clk_recalculate_rate(c);
 }
 
 static int tegra2_periph_clk_enable(struct clk *c)
@@ -692,12 +790,19 @@ static int tegra2_periph_clk_set_parent(struct clk *c, struct clk *p)
 	pr_debug("%s: %s %s\n", __func__, c->name, p->name);
 	for (sel = c->inputs; sel->input != NULL; sel++) {
 		if (sel->input == p) {
-			clk_reparent(c, p);
 			val = clk_readl(c->reg);
 			val &= ~PERIPH_CLK_SOURCE_MASK;
 			val |= (sel->value) << PERIPH_CLK_SOURCE_SHIFT;
+
+			if (c->refcnt)
+				clk_enable_locked(p);
+
 			clk_writel(val, c->reg);
-			c->rate = c->parent->rate;
+
+			if (c->refcnt && c->parent)
+				clk_disable_locked(c->parent);
+
+			clk_reparent(c, p);
 			return 0;
 		}
 	}
@@ -708,20 +813,55 @@ static int tegra2_periph_clk_set_parent(struct clk *c, struct clk *p)
 static int tegra2_periph_clk_set_rate(struct clk *c, unsigned long rate)
 {
 	u32 val;
-	int divider_u71;
+	int divider;
 	pr_debug("%s: %lu\n", __func__, rate);
 	if (c->flags & DIV_U71) {
-		divider_u71 = clk_div71_get_divider(c->parent, rate);
-		if (divider_u71 >= 0) {
+		divider = clk_div71_get_divider(c->parent->rate, rate);
+		if (divider >= 0) {
 			val = clk_readl(c->reg);
-			val &= ~PERIPH_CLK_SOURCE_DIV_MASK;
-			val |= divider_u71;
+			val &= ~PERIPH_CLK_SOURCE_DIVU71_MASK;
+			val |= divider;
 			clk_writel(val, c->reg);
-			c->div = divider_u71 + 2;
+			c->div = divider + 2;
 			c->mul = 2;
-			tegra2_clk_recalculate_rate(c);
 			return 0;
 		}
+	} else if (c->flags & DIV_U16) {
+		divider = clk_div16_get_divider(c->parent->rate, rate);
+		if (divider >= 0) {
+			val = clk_readl(c->reg);
+			val &= ~PERIPH_CLK_SOURCE_DIVU16_MASK;
+			val |= divider;
+			clk_writel(val, c->reg);
+			c->div = divider + 1;
+			c->mul = 1;
+			return 0;
+		}
+	} else if (c->parent->rate <= rate) {
+		c->div = 1;
+		c->mul = 1;
+		return 0;
+	}
+	return -EINVAL;
+}
+
+static long tegra2_periph_clk_round_rate(struct clk *c,
+	unsigned long rate)
+{
+	int divider;
+	pr_debug("%s: %s %lu\n", __func__, c->name, rate);
+
+	if (c->flags & DIV_U71) {
+		divider = clk_div71_get_divider(c->parent->rate, rate);
+		if (divider < 0)
+			return divider;
+
+		return c->parent->rate * 2 / (divider + 2);
+	} else if (c->flags & DIV_U16) {
+		divider = clk_div16_get_divider(c->parent->rate, rate);
+		if (divider < 0)
+			return divider;
+		return c->parent->rate / (divider + 1);
 	}
 	return -EINV