Merge branch 'origin' into devel-stable

Conflicts:
	arch/arm/mach-mx2/devices.c
	arch/arm/mach-mx2/devices.h
	sound/soc/pxa/pxa-ssp.c

58 files changed, 1638 insertions, 721 deletions

diff --git a/arch/sh/kernel/cpu/Makefile b/arch/sh/kernel/cpu/Makefile
index d97c803719e..0e48bc61c27 100644
--- a/arch/sh/kernel/cpu/Makefile
+++ b/arch/sh/kernel/cpu/Makefile
@@ -17,5 +17,7 @@ obj-$(CONFIG_ARCH_SHMOBILE)	+= shmobile/
 
 obj-$(CONFIG_SH_ADC)		+= adc.o
 obj-$(CONFIG_SH_CLK_CPG)	+= clock-cpg.o
+obj-$(CONFIG_SH_FPU)		+= fpu.o
+obj-$(CONFIG_SH_FPU_EMU)	+= fpu.o
 
 obj-y	+= irq/ init.o clock.o hwblk.o
diff --git a/arch/sh/kernel/cpu/adc.c b/arch/sh/kernel/cpu/adc.c
index da3d6877f93..d307571d54b 100644
--- a/arch/sh/kernel/cpu/adc.c
+++ b/arch/sh/kernel/cpu/adc.c
@@ -18,19 +18,19 @@ int adc_single(unsigned int channel)
 
 	off = (channel & 0x03) << 2;
 
-	csr = ctrl_inb(ADCSR);
+	csr = __raw_readb(ADCSR);
 	csr = channel | ADCSR_ADST | ADCSR_CKS;
-	ctrl_outb(csr, ADCSR);
+	__raw_writeb(csr, ADCSR);
 
 	do {
-		csr = ctrl_inb(ADCSR);
+		csr = __raw_readb(ADCSR);
 	} while ((csr & ADCSR_ADF) == 0);
 
 	csr &= ~(ADCSR_ADF | ADCSR_ADST);
-	ctrl_outb(csr, ADCSR);
+	__raw_writeb(csr, ADCSR);
 
-	return (((ctrl_inb(ADDRAH + off) << 8) |
-		ctrl_inb(ADDRAL + off)) >> 6);
+	return (((__raw_readb(ADDRAH + off) << 8) |
+		__raw_readb(ADDRAL + off)) >> 6);
 }
 
 EXPORT_SYMBOL(adc_single);
diff --git a/arch/sh/kernel/cpu/clock-cpg.c b/arch/sh/kernel/cpu/clock-cpg.c
index 6dfe2cced3f..eed5eaff96b 100644
--- a/arch/sh/kernel/cpu/clock-cpg.c
+++ b/arch/sh/kernel/cpu/clock-cpg.c
@@ -149,7 +149,8 @@ int __init sh_clk_div6_register(struct clk *clks, int nr)
 
 static unsigned long sh_clk_div4_recalc(struct clk *clk)
 {
-	struct clk_div_mult_table *table = clk->priv;
+	struct clk_div4_table *d4t = clk->priv;
+	struct clk_div_mult_table *table = d4t->div_mult_table;
 	unsigned int idx;
 
 	clk_rate_table_build(clk, clk->freq_table, table->nr_divisors,
@@ -160,17 +161,90 @@ static unsigned long sh_clk_div4_recalc(struct clk *clk)
 	return clk->freq_table[idx].frequency;
 }
 
+static int sh_clk_div4_set_parent(struct clk *clk, struct clk *parent)
+{
+	struct clk_div4_table *d4t = clk->priv;
+	struct clk_div_mult_table *table = d4t->div_mult_table;
+	u32 value;
+	int ret;
+
+	if (!strcmp("pll_clk", parent->name))
+		value = __raw_readl(clk->enable_reg) & ~(1 << 7);
+	else
+		value = __raw_readl(clk->enable_reg) | (1 << 7);
+
+	ret = clk_reparent(clk, parent);
+	if (ret < 0)
+		return ret;
+
+	__raw_writel(value, clk->enable_reg);
+
+	/* Rebiuld the frequency table */
+	clk_rate_table_build(clk, clk->freq_table, table->nr_divisors,
+			     table, &clk->arch_flags);
+
+	return 0;
+}
+
+static int sh_clk_div4_set_rate(struct clk *clk, unsigned long rate, int algo_id)
+{
+	struct clk_div4_table *d4t = clk->priv;
+	unsigned long value;
+	int idx = clk_rate_table_find(clk, clk->freq_table, rate);
+	if (idx < 0)
+		return idx;
+
+	value = __raw_readl(clk->enable_reg);
+	value &= ~(0xf << clk->enable_bit);
+	value |= (idx << clk->enable_bit);
+	__raw_writel(value, clk->enable_reg);
+
+	if (d4t->kick)
+		d4t->kick(clk);
+
+	return 0;
+}
+
+static int sh_clk_div4_enable(struct clk *clk)
+{
+	__raw_writel(__raw_readl(clk->enable_reg) & ~(1 << 8), clk->enable_reg);
+	return 0;
+}
+
+static void sh_clk_div4_disable(struct clk *clk)
+{
+	__raw_writel(__raw_readl(clk->enable_reg) | (1 << 8), clk->enable_reg);
+}
+
 static struct clk_ops sh_clk_div4_clk_ops = {
 	.recalc		= sh_clk_div4_recalc,
+	.set_rate	= sh_clk_div4_set_rate,
 	.round_rate	= sh_clk_div_round_rate,
 };
 
-int __init sh_clk_div4_register(struct clk *clks, int nr,
-				struct clk_div_mult_table *table)
+static struct clk_ops sh_clk_div4_enable_clk_ops = {
+	.recalc		= sh_clk_div4_recalc,
+	.set_rate	= sh_clk_div4_set_rate,
+	.round_rate	= sh_clk_div_round_rate,
+	.enable		= sh_clk_div4_enable,
+	.disable	= sh_clk_div4_disable,
+};
+
+static struct clk_ops sh_clk_div4_reparent_clk_ops = {
+	.recalc		= sh_clk_div4_recalc,
+	.set_rate	= sh_clk_div4_set_rate,
+	.round_rate	= sh_clk_div_round_rate,
+	.enable		= sh_clk_div4_enable,
+	.disable	= sh_clk_div4_disable,
+	.set_parent	= sh_clk_div4_set_parent,
+};
+
+static int __init sh_clk_div4_register_ops(struct clk *clks, int nr,
+			struct clk_div4_table *table, struct clk_ops *ops)
 {
 	struct clk *clkp;
 	void *freq_table;
-	int nr_divs = table->nr_divisors;
+	int nr_divs = table->div_mult_table->nr_divisors;
 	int freq_table_size = sizeof(struct cpufreq_frequency_table);
 	int ret = 0;
 	int k;
@@ -185,7 +259,7 @@ int __init sh_clk_div4_register(struct clk *clks, int nr,
 	for (k = 0; !ret && (k < nr); k++) {
 		clkp = clks + k;
 
-		clkp->ops = &sh_clk_div4_clk_ops;
+		clkp->ops = ops;
 		clkp->id = -1;
 		clkp->priv = table;
 
@@ -198,6 +272,26 @@ int __init sh_clk_div4_register(struct clk *clks, int nr,
 	return ret;
 }
 
+int __init sh_clk_div4_register(struct clk *clks, int nr,
+				struct clk_div4_table *table)
+{
+	return sh_clk_div4_register_ops(clks, nr, table, &sh_clk_div4_clk_ops);
+}
+
+int __init sh_clk_div4_enable_register(struct clk *clks, int nr,
+				struct clk_div4_table *table)
+{
+	return sh_clk_div4_register_ops(clks, nr, table,
+					&sh_clk_div4_enable_clk_ops);
+}
+
+int __init sh_clk_div4_reparent_register(struct clk *clks, int nr,
+				struct clk_div4_table *table)
+{
+	return sh_clk_div4_register_ops(clks, nr, table,
+					&sh_clk_div4_reparent_clk_ops);
+}
+
 #ifdef CONFIG_SH_CLK_CPG_LEGACY
 static struct clk master_clk = {
 	.name		= "master_clk",
diff --git a/arch/sh/kernel/cpu/fpu.c b/arch/sh/kernel/cpu/fpu.c
new file mode 100644
index 00000000000..f059ed62cf5
--- /dev/null
+++ b/arch/sh/kernel/cpu/fpu.c
@@ -0,0 +1,84 @@
+#include <linux/sched.h>
+#include <asm/processor.h>
+#include <asm/fpu.h>
+
+int init_fpu(struct task_struct *tsk)
+{
+	if (tsk_used_math(tsk)) {
+		if ((boot_cpu_data.flags & CPU_HAS_FPU) && tsk == current)
+			unlazy_fpu(tsk, task_pt_regs(tsk));
+		return 0;
+	}
+
+	/*
+	 * Memory allocation at the first usage of the FPU and other state.
+	 */
+	if (!tsk->thread.xstate) {
+		tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
+						      GFP_KERNEL);
+		if (!tsk->thread.xstate)
+			return -ENOMEM;
+	}
+
+	if (boot_cpu_data.flags & CPU_HAS_FPU) {
+		struct sh_fpu_hard_struct *fp = &tsk->thread.xstate->hardfpu;
+		memset(fp, 0, xstate_size);
+		fp->fpscr = FPSCR_INIT;
+	} else {
+		struct sh_fpu_soft_struct *fp = &tsk->thread.xstate->softfpu;
+		memset(fp, 0, xstate_size);
+		fp->fpscr = FPSCR_INIT;
+	}
+
+	set_stopped_child_used_math(tsk);
+	return 0;
+}
+
+#ifdef CONFIG_SH_FPU
+void __fpu_state_restore(void)
+{
+	struct task_struct *tsk = current;
+
+	restore_fpu(tsk);
+
+	task_thread_info(tsk)->status |= TS_USEDFPU;
+	tsk->fpu_counter++;
+}
+
+void fpu_state_restore(struct pt_regs *regs)
+{
+	struct task_struct *tsk = current;
+
+	if (unlikely(!user_mode(regs))) {
+		printk(KERN_ERR "BUG: FPU is used in kernel mode.\n");
+		BUG();
+		return;
+	}
+
+	if (!tsk_used_math(tsk)) {
+		local_irq_enable();
+		/*
+		 * does a slab alloc which can sleep
+		 */
+		if (init_fpu(tsk)) {
+			/*
+			 * ran out of memory!
+			 */
+			do_group_exit(SIGKILL);
+			return;
+		}
+		local_irq_disable();
+	}
+
+	grab_fpu(regs);
+
+	__fpu_state_restore();
+}
+
+BUILD_TRAP_HANDLER(fpu_state_restore)
+{
+	TRAP_HANDLER_DECL;
+
+	fpu_state_restore(regs);
+}
+#endif /* CONFIG_SH_FPU */
diff --git a/arch/sh/kernel/cpu/init.c b/arch/sh/kernel/cpu/init.c
index 89b4b76c0d7..c736422344e 100644
--- a/arch/sh/kernel/cpu/init.c
+++ b/arch/sh/kernel/cpu/init.c
@@ -24,22 +24,32 @@
 #include <asm/elf.h>
 #include <asm/io.h>
 #include <asm/smp.h>
-#ifdef CONFIG_SUPERH32
-#include <asm/ubc.h>
+#include <asm/sh_bios.h>
+
+#ifdef CONFIG_SH_FPU
+#define cpu_has_fpu	1
+#else
+#define cpu_has_fpu	0
+#endif
+
+#ifdef CONFIG_SH_DSP
+#define cpu_has_dsp	1
+#else
+#define cpu_has_dsp	0
 #endif
 
 /*
  * Generic wrapper for command line arguments to disable on-chip
  * peripherals (nofpu, nodsp, and so forth).
  */
-#define onchip_setup(x)				\
-static int x##_disabled __initdata = 0;		\
-						\
-static int __init x##_setup(char *opts)		\
-{						\
-	x##_disabled = 1;			\
-	return 1;				\
-}						\
+#define onchip_setup(x)					\
+static int x##_disabled __initdata = !cpu_has_##x;	\
+							\
+static int __init x##_setup(char *opts)			\
+{							\
+	x##_disabled = 1;				\
+	return 1;					\
+}							\
 __setup("no" __stringify(x), x##_setup);
 
 onchip_setup(fpu);
@@ -52,10 +62,10 @@ onchip_setup(dsp);
 static void __init speculative_execution_init(void)
 {
 	/* Clear RABD */
-	ctrl_outl(ctrl_inl(CPUOPM) & ~CPUOPM_RABD, CPUOPM);
+	__raw_writel(__raw_readl(CPUOPM) & ~CPUOPM_RABD, CPUOPM);
 
 	/* Flush the update */
-	(void)ctrl_inl(CPUOPM);
+	(void)__raw_readl(CPUOPM);
 	ctrl_barrier();
 }
 #else
@@ -89,7 +99,7 @@ static void __init expmask_init(void)
 #endif
 
 /* 2nd-level cache init */
-void __uses_jump_to_uncached __attribute__ ((weak)) l2_cache_init(void)
+void __attribute__ ((weak)) l2_cache_init(void)
 {
 }
 
@@ -97,12 +107,12 @@ void __uses_jump_to_uncached __attribute__ ((weak)) l2_cache_init(void)
  * Generic first-level cache init
  */
 #ifdef CONFIG_SUPERH32
-static void __uses_jump_to_uncached cache_init(void)
+static void cache_init(void)
 {
 	unsigned long ccr, flags;
 
 	jump_to_uncached();
-	ccr = ctrl_inl(CCR);
+	ccr = __raw_readl(CCR);
 
 	/*
 	 * At this point we don't know whether the cache is enabled or not - a
@@ -146,7 +156,7 @@ static void __uses_jump_to_uncached cache_init(void)
 			for (addr = addrstart;
 			     addr < addrstart + waysize;
 			     addr += current_cpu_data.dcache.linesz)
-				ctrl_outl(0, addr);
+				__raw_writel(0, addr);
 
 			addrstart += current_cpu_data.dcache.way_incr;
 		} while (--ways);
@@ -179,7 +189,7 @@ static void __uses_jump_to_uncached cache_init(void)
 
 	l2_cache_init();
 
-	ctrl_outl(flags, CCR);
+	__raw_writel(flags, CCR);
 	back_to_cached();
 }
 #else
@@ -207,6 +217,18 @@ static void detect_cache_shape(void)
 		l2_cache_shape = -1; /* No S-cache */
 }
 
+static void __init fpu_init(void)
+{
+	/* Disable the FPU */
+	if (fpu_disabled && (current_cpu_data.flags & CPU_HAS_FPU)) {
+		printk("FPU Disabled\n");
+		current_cpu_data.flags &= ~CPU_HAS_FPU;
+	}
+
+	disable_fpu();
+	clear_used_math();
+}
+
 #ifdef CONFIG_SH_DSP
 static void __init release_dsp(void)
 {
@@ -244,28 +266,35 @@ static void __init dsp_init(void)
 	if (sr & SR_DSP)
 		current_cpu_data.flags |= CPU_HAS_DSP;
 
+	/* Disable the DSP */
+	if (dsp_disabled && (current_cpu_data.flags & CPU_HAS_DSP)) {
+		printk("DSP Disabled\n");
+		current_cpu_data.flags &= ~CPU_HAS_DSP;
+	}
+
 	/* Now that we've determined the DSP status, clear the DSP bit. */
 	release_dsp();
 }
+#else
+static inline void __init dsp_init(void) { }
 #endif /* CONFIG_SH_DSP */
 
 /**
  * sh_cpu_init
  *
- * This is our initial entry point for each CPU, and is invoked on the boot
- * CPU prior to calling start_kernel(). For SMP, a combination of this and
- * start_secondary() will bring up each processor to a ready state prior
- * to hand forking the idle loop.
+ * This is our initial entry point for each CPU, and is invoked on the
+ * boot CPU prior to calling start_kernel(). For SMP, a combination of
+ * this and start_secondary() will bring up each processor to a ready
+ * state prior to hand forking the idle loop.
  *
- * We do all of the basic processor init here, including setting up the
- * caches, FPU, DSP, kicking the UBC, etc. By the time start_kernel() is
- * hit (and subsequently platform_setup()) things like determining the
- * CPU subtype and initial configuration will all be done.
+ * We do all of the basic processor init here, including setting up
+ * the caches, FPU, DSP, etc. By the time start_kernel() is hit (and
+ * subsequently platform_setup()) things like determining the CPU
+ * subtype and initial configuration will all be done.
  *
  * Each processor family is still responsible for doing its own probing
  * and cache configuration in detect_cpu_and_cache_system().
  */
-
 asmlinkage void __init sh_cpu_init(void)
 {
 	current_thread_info()->cpu = hard_smp_processor_id();
@@ -302,18 +331,8 @@ asmlinkage void __init sh_cpu_init(void)
 		detect_cache_shape();
 	}
 
-	/* Disable the FPU */
-	if (fpu_disabled) {
-		printk("FPU Disabled\n");
-		current_cpu_data.flags &= ~CPU_HAS_FPU;
-	}
-
-	/* FPU initialization */
-	disable_fpu();
-	if ((current_cpu_data.flags & CPU_HAS_FPU)) {
-		current_thread_info()->status &= ~TS_USEDFPU;
-		clear_used_math();
-	}
+	fpu_init();
+	dsp_init();
 
 	/*
 	 * Initialize the per-CPU ASID cache very early, since the
@@ -321,18 +340,24 @@ asmlinkage void __init sh_cpu_init(void)
 	 */
 	current_cpu_data.asid_cache = NO_CONTEXT;
 
-#ifdef CONFIG_SH_DSP
-	/* Probe for DSP */
-	dsp_init();
-
-	/* Disable the DSP */
-	if (dsp_disabled) {
-		printk("DSP Disabled\n");
-		current_cpu_data.flags &= ~CPU_HAS_DSP;
-		release_dsp();
-	}
-#endif
-
 	speculative_execution_init();
 	expmask_init();
+
+	/* Do the rest of the boot processor setup */
+	if (raw_smp_processor_id() == 0) {
+		/* Save off the BIOS VBR, if there is one */
+		sh_bios_vbr_init();
+
+		/*
+		 * Setup VBR for boot CPU. Secondary CPUs do this through
+		 * start_secondary().
+		 */
+		per_cpu_trap_init();
+
+		/*
+		 * Boot processor to setup the FP and extended state
+		 * context info.
+		 */
+		init_thread_xstate();
+	}
 }
diff --git a/arch/sh/kernel/cpu/irq/intc-sh5.c b/arch/sh/kernel/cpu/irq/intc-sh5.c
index 06e7e2959b5..96a23958394 100644
--- a/arch/sh/kernel/cpu/irq/intc-sh5.c
+++ b/arch/sh/kernel/cpu/irq/intc-sh5.c
@@ -123,7 +123,7 @@ static void enable_intc_irq(unsigned int irq)
 		bitmask = 1 << (irq - 32);
 	}
 
-	ctrl_outl(bitmask, reg);
+	__raw_writel(bitmask, reg);
 }
 
 static void disable_intc_irq(unsigned int irq)
@@ -139,7 +139,7 @@ static void disable_intc_irq(unsigned int irq)
 		bitmask = 1 << (irq - 32);
 	}
 
-	ctrl_outl(bitmask, reg);
+	__raw_writel(bitmask, reg);
 }
 
 static void mask_and_ack_intc(unsigned int irq)
@@ -170,11 +170,11 @@ void __init plat_irq_setup(void)
 
 
 	/* Disable all interrupts and set all priorities to 0 to avoid trouble */
-	ctrl_outl(-1, INTC_INTDSB_0);
-	ctrl_outl(-1, INTC_INTDSB_1);
+	__raw_writel(-1, INTC_INTDSB_0);
+	__raw_writel(-1, INTC_INTDSB_1);
 
 	for (reg = INTC_INTPRI_0, i = 0; i < INTC_INTPRI_PREGS; i++, reg += 8)
-		ctrl_outl( NO_PRIORITY, reg);
+		__raw_writel( NO_PRIORITY, reg);
 
 
 #ifdef CONFIG_SH_CAYMAN
@@ -199,7 +199,7 @@ void __init plat_irq_setup(void)
 			reg = INTC_ICR_SET;
 			i = IRQ_IRL0;
 		}
-		ctrl_outl(INTC_ICR_IRLM, reg);
+		__raw_writel(INTC_ICR_IRLM, reg);
 
 		/* Set interrupt priorities according to platform description */
 		for (data = 0, reg = INTC_INTPRI_0; i < NR_INTC_IRQS; i++) {
@@ -207,7 +207,7 @@ void __init plat_irq_setup(void)
 				((i % INTC_INTPRI_PPREG) * 4);
 			if ((i % INTC_INTPRI_PPREG) == (INTC_INTPRI_PPREG - 1)) {
 				/* Upon the 7th, set Priority Register */
-				ctrl_outl(data, reg);
+				__raw_writel(data, reg);
 				data = 0;
 				reg += 8;
 			}
diff --git a/arch/sh/kernel/cpu/sh2/clock-sh7619.c b/arch/sh/kernel/cpu/sh2/clock-sh7619.c
index 4fe863170e3..0c9f24d7a02 100644
--- a/arch/sh/kernel/cpu/sh2/clock-sh7619.c
+++ b/arch/sh/kernel/cpu/sh2/clock-sh7619.c
@@ -31,7 +31,7 @@ static const int pfc_divisors[] = {1,2,0,4};
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= PLL2 * pll1rate[(ctrl_inw(FREQCR) >> 8) & 7];
+	clk->rate *= PLL2 * pll1rate[(__raw_readw(FREQCR) >> 8) & 7];
 }
 
 static struct clk_ops sh7619_master_clk_ops = {
@@ -40,7 +40,7 @@ static struct clk_ops sh7619_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -50,7 +50,7 @@ static struct clk_ops sh7619_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	return clk->parent->rate / pll1rate[(ctrl_inw(FREQCR) >> 8) & 7];
+	return clk->parent->rate / pll1rate[(__raw_readw(FREQCR) >> 8) & 7];
 }
 
 static struct clk_ops sh7619_bus_clk_ops = {
diff --git a/arch/sh/kernel/cpu/sh2a/clock-sh7201.c b/arch/sh/kernel/cpu/sh2a/clock-sh7201.c
index 7814c76159a..b26264dc2ae 100644
--- a/arch/sh/kernel/cpu/sh2a/clock-sh7201.c
+++ b/arch/sh/kernel/cpu/sh2a/clock-sh7201.c
@@ -34,7 +34,7 @@ static const int pfc_divisors[]={1,2,3,4,6,8,12};
 
 static void master_clk_init(struct clk *clk)
 {
-	return 10000000 * PLL2 * pll1rate[(ctrl_inw(FREQCR) >> 8) & 0x0007];
+	return 10000000 * PLL2 * pll1rate[(__raw_readw(FREQCR) >> 8) & 0x0007];
 }
 
 static struct clk_ops sh7201_master_clk_ops = {
@@ -43,7 +43,7 @@ static struct clk_ops sh7201_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -53,7 +53,7 @@ static struct clk_ops sh7201_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -63,7 +63,7 @@ static struct clk_ops sh7201_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inw(FREQCR) >> 4) & 0x0007);
+	int idx = ((__raw_readw(FREQCR) >> 4) & 0x0007);
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh2a/clock-sh7203.c b/arch/sh/kernel/cpu/sh2a/clock-sh7203.c
index 94098696510..7e75d8f7950 100644
--- a/arch/sh/kernel/cpu/sh2a/clock-sh7203.c
+++ b/arch/sh/kernel/cpu/sh2a/clock-sh7203.c
@@ -39,7 +39,7 @@ static const int pfc_divisors[]={1,2,3,4,6,8,12};
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= pll1rate[(ctrl_inw(FREQCR) >> 8) & 0x0003] * PLL2 ;
+	clk->rate *= pll1rate[(__raw_readw(FREQCR) >> 8) & 0x0003] * PLL2 ;
 }
 
 static struct clk_ops sh7203_master_clk_ops = {
@@ -48,7 +48,7 @@ static struct clk_ops sh7203_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -58,7 +58,7 @@ static struct clk_ops sh7203_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx-2];
 }
 
diff --git a/arch/sh/kernel/cpu/sh2a/clock-sh7206.c b/arch/sh/kernel/cpu/sh2a/clock-sh7206.c
index c2268bdecee..b27a5e2687a 100644
--- a/arch/sh/kernel/cpu/sh2a/clock-sh7206.c
+++ b/arch/sh/kernel/cpu/sh2a/clock-sh7206.c
@@ -34,7 +34,7 @@ static const int pfc_divisors[]={1,2,3,4,6,8,12};
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= PLL2 * pll1rate[(ctrl_inw(FREQCR) >> 8) & 0x0007];
+	clk->rate *= PLL2 * pll1rate[(__raw_readw(FREQCR) >> 8) & 0x0007];
 }
 
 static struct clk_ops sh7206_master_clk_ops = {
@@ -43,7 +43,7 @@ static struct clk_ops sh7206_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -53,7 +53,7 @@ static struct clk_ops sh7206_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	return clk->parent->rate / pll1rate[(ctrl_inw(FREQCR) >> 8) & 0x0007];
+	return clk->parent->rate / pll1rate[(__raw_readw(FREQCR) >> 8) & 0x0007];
 }
 
 static struct clk_ops sh7206_bus_clk_ops = {
@@ -62,7 +62,7 @@ static struct clk_ops sh7206_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FREQCR) & 0x0007);
+	int idx = (__raw_readw(FREQCR) & 0x0007);
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh2a/fpu.c b/arch/sh/kernel/cpu/sh2a/fpu.c
index d395ce5740e..488d24e0cdf 100644
--- a/arch/sh/kernel/cpu/sh2a/fpu.c
+++ b/arch/sh/kernel/cpu/sh2a/fpu.c
@@ -26,8 +26,7 @@
 /*
  * Save FPU registers onto task structure.
  */
-void
-save_fpu(struct task_struct *tsk)
+void save_fpu(struct task_struct *tsk)
 {
 	unsigned long dummy;
 
@@ -52,7 +51,7 @@ save_fpu(struct task_struct *tsk)
 		     "fmov.s	fr0, @-%0\n\t"
 		     "lds	%3, fpscr\n\t"
 		     : "=r" (dummy)
-		     : "0" ((char *)(&tsk->thread.fpu.hard.status)),
+		     : "0" ((char *)(&tsk->thread.xstate->hardfpu.status)),
 		       "r" (FPSCR_RCHG),
 		       "r" (FPSCR_INIT)
 		     : "memory");
@@ -60,8 +59,7 @@ save_fpu(struct task_struct *tsk)
 	disable_fpu();
 }
 
-static void
-restore_fpu(struct task_struct *tsk)
+void restore_fpu(struct task_struct *tsk)
 {
 	unsigned long dummy;
 
@@ -85,45 +83,12 @@ restore_fpu(struct task_struct *tsk)
 		     "lds.l	@%0+, fpscr\n\t"
 		     "lds.l	@%0+, fpul\n\t"
 		     : "=r" (dummy)
-		     : "0" (&tsk->thread.fpu), "r" (FPSCR_RCHG)
+		     : "0" (tsk->thread.xstate), "r" (FPSCR_RCHG)
 		     : "memory");
 	disable_fpu();
 }
 
 /*
- * Load the FPU with signalling NANS.  This bit pattern we're using
- * has the property that no matter wether considered as single or as
- * double precission represents signaling NANS.
- */
-
-static void
-fpu_init(void)
-{
-	enable_fpu();
-	asm volatile("lds	%0, fpul\n\t"
-		     "fsts	fpul, fr0\n\t"
-		     "fsts	fpul, fr1\n\t"
-		     "fsts	fpul, fr2\n\t"
-		     "fsts	fpul, fr3\n\t"
-		     "fsts	fpul, fr4\n\t"
-		     "fsts	fpul, fr5\n\t"
-		     "fsts	fpul, fr6\n\t"
-		     "fsts	fpul, fr7\n\t"
-		     "fsts	fpul, fr8\n\t"
-		     "fsts	fpul, fr9\n\t"
-		     "fsts	fpul, fr10\n\t"
-		     "fsts	fpul, fr11\n\t"
-		     "fsts	fpul, fr12\n\t"
-		     "fsts	fpul, fr13\n\t"
-		     "fsts	fpul, fr14\n\t"
-		     "fsts	fpul, fr15\n\t"
-		     "lds	%2, fpscr\n\t"
-		     : /* no output */
-		     : "r" (0), "r" (FPSCR_RCHG), "r" (FPSCR_INIT));
-	disable_fpu();
-}
-
-/*
  *	Emulate arithmetic ops on denormalized number for some FPU insns.
  */
 
@@ -490,9 +455,9 @@ ieee_fpe_handler (struct pt_regs *regs)
 	if ((finsn & 0xf1ff) == 0xf0ad) { /* fcnvsd */
 		struct task_struct *tsk = current;
 
-		if ((tsk->thread.fpu.hard.fpscr & FPSCR_FPU_ERROR)) {
+		if ((tsk->thread.xstate->hardfpu.fpscr & FPSCR_FPU_ERROR)) {
 			/* FPU error */
-			denormal_to_double (&tsk->thread.fpu.hard,
+			denormal_to_double (&tsk->thread.xstate->hardfpu,
 					    (finsn >> 8) & 0xf);
 		} else
 			return 0;
@@ -507,9 +472,9 @@ ieee_fpe_handler (struct pt_regs *regs)
 
 		n = (finsn >> 8) & 0xf;
 		m = (finsn >> 4) & 0xf;
-		hx = tsk->thread.fpu.hard.fp_regs[n];
-		hy = tsk->thread.fpu.hard.fp_regs[m];
-		fpscr = tsk->thread.fpu.hard.fpscr;
+		hx = tsk->thread.xstate->hardfpu.fp_regs[n];
+		hy = tsk->thread.xstate->hardfpu.fp_regs[m];
+		fpscr = tsk->thread.xstate->hardfpu.fpscr;
 		prec = fpscr & (1 << 19);
 
 		if ((fpscr & FPSCR_FPU_ERROR)
@@ -519,15 +484,15 @@ ieee_fpe_handler (struct pt_regs *regs)
 
 			/* FPU error because of denormal */
 			llx = ((long long) hx << 32)
-			       | tsk->thread.fpu.hard.fp_regs[n+1];
+			       | tsk->thread.xstate->hardfpu.fp_regs[n+1];
 			lly = ((long long) hy << 32)
-			       | tsk->thread.fpu.hard.fp_regs[m+1];
+			       | tsk->thread.xstate->hardfpu.fp_regs[m+1];
 			if ((hx & 0x7fffffff) >= 0x00100000)
 				llx = denormal_muld(lly, llx);
 			else
 				llx = denormal_muld(llx, lly);
-			tsk->thread.fpu.hard.fp_regs[n] = llx >> 32;
-			tsk->thread.fpu.hard.fp_regs[n+1] = llx & 0xffffffff;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
+			tsk->thread.xstate->hardfpu.fp_regs[n+1] = llx & 0xffffffff;
 		} else if ((fpscr & FPSCR_FPU_ERROR)
 		     && (!prec && ((hx & 0x7fffffff) < 0x00800000
 				   || (hy & 0x7fffffff) < 0x00800000))) {
@@ -536,7 +501,7 @@ ieee_fpe_handler (struct pt_regs *regs)
 				hx = denormal_mulf(hy, hx);
 			else
 				hx = denormal_mulf(hx, hy);
-			tsk->thread.fpu.hard.fp_regs[n] = hx;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
 		} else
 			return 0;
 
@@ -550,9 +515,9 @@ ieee_fpe_handler (struct pt_regs *regs)
 
 		n = (finsn >> 8) & 0xf;
 		m = (finsn >> 4) & 0xf;
-		hx = tsk->thread.fpu.hard.fp_regs[n];
-		hy = tsk->thread.fpu.hard.fp_regs[m];
-		fpscr = tsk->thread.fpu.hard.fpscr;
+		hx = tsk->thread.xstate->hardfpu.fp_regs[n];
+		hy = tsk->thread.xstate->hardfpu.fp_regs[m];
+		fpscr = tsk->thread.xstate->hardfpu.fpscr;
 		prec = fpscr & (1 << 19);
 
 		if ((fpscr & FPSCR_FPU_ERROR)
@@ -562,15 +527,15 @@ ieee_fpe_handler (struct pt_regs *regs)
 
 			/* FPU error because of denormal */
 			llx = ((long long) hx << 32)
-			       | tsk->thread.fpu.hard.fp_regs[n+1];
+			       | tsk->thread.xstate->hardfpu.fp_regs[n+1];
 			lly = ((long long) hy << 32)
-			       | tsk->thread.fpu.hard.fp_regs[m+1];
+			       | tsk->thread.xstate->hardfpu.fp_regs[m+1];
 			if ((finsn & 0xf00f) == 0xf000)
 				llx = denormal_addd(llx, lly);
 			else
 				llx = denormal_addd(llx, lly ^ (1LL << 63));
-			tsk->thread.fpu.hard.fp_regs[n] = llx >> 32;
-			tsk->thread.fpu.hard.fp_regs[n+1] = llx & 0xffffffff;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
+			tsk->thread.xstate->hardfpu.fp_regs[n+1] = llx & 0xffffffff;
 		} else if ((fpscr & FPSCR_FPU_ERROR)
 		     && (!prec && ((hx & 0x7fffffff) < 0x00800000
 				   || (hy & 0x7fffffff) < 0x00800000))) {
@@ -579,7 +544,7 @@ ieee_fpe_handler (struct pt_regs *regs)
 				hx = denormal_addf(hx, hy);
 			else
 				hx = denormal_addf(hx, hy ^ 0x80000000);
-			tsk->thread.fpu.hard.fp_regs[n] = hx;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
 		} else
 			return 0;
 
@@ -597,7 +562,7 @@ BUILD_TRAP_HANDLER(fpu_error)
 
 	__unlazy_fpu(tsk, regs);
 	if (ieee_fpe_handler(regs)) {
-		tsk->thread.fpu.hard.fpscr &=
+		tsk->thread.xstate->hardfpu.fpscr &=
 			~(FPSCR_CAUSE_MASK | FPSCR_FLAG_MASK);
 		grab_fpu(regs);
 		restore_fpu(tsk);
@@ -607,33 +572,3 @@ BUILD_TRAP_HANDLER(fpu_error)
 
 	force_sig(SIGFPE, tsk);
 }
-
-void fpu_state_restore(struct pt_regs *regs)
-{
-	struct task_struct *tsk = current;
-
-	grab_fpu(regs);
-	if (unlikely(!user_mode(regs))) {
-		printk(KERN_ERR "BUG: FPU is used in kernel mode.\n");
-		BUG();
-		return;
-	}
-
-	if (likely(used_math())) {
-		/* Using the FPU again.  */
-		restore_fpu(tsk);
-	} else	{
-		/* First time FPU user.  */
-		fpu_init();
-		set_used_math();
-	}
-	task_thread_info(tsk)->status |= TS_USEDFPU;
-	tsk->fpu_counter++;
-}
-
-BUILD_TRAP_HANDLER(fpu_state_restore)
-{
-	TRAP_HANDLER_DECL;
-
-	fpu_state_restore(regs);
-}
diff --git a/arch/sh/kernel/cpu/sh3/clock-sh3.c b/arch/sh/kernel/cpu/sh3/clock-sh3.c
index 27b8738f0b0..b78384afac0 100644
--- a/arch/sh/kernel/cpu/sh3/clock-sh3.c
+++ b/arch/sh/kernel/cpu/sh3/clock-sh3.c
@@ -28,7 +28,7 @@ static int pfc_divisors[]    = { 1, 2, 3, 4, 6, 1, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x2000) >> 11) | (frqcr & 0x0003);
 
 	clk->rate *= pfc_divisors[idx];
@@ -40,7 +40,7 @@ static struct clk_ops sh3_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x2000) >> 11) | (frqcr & 0x0003);
 
 	return clk->parent->rate / pfc_divisors[idx];
@@ -52,7 +52,7 @@ static struct clk_ops sh3_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x8000) >> 13) | ((frqcr & 0x0030) >> 4);
 
 	return clk->parent->rate / stc_multipliers[idx];
@@ -64,7 +64,7 @@ static struct clk_ops sh3_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x4000) >> 12) | ((frqcr & 0x000c) >> 2);
 
 	return clk->parent->rate / ifc_divisors[idx];
diff --git a/arch/sh/kernel/cpu/sh3/clock-sh7705.c b/arch/sh/kernel/cpu/sh3/clock-sh7705.c
index 0ca8f2c3646..0ecea1451c6 100644
--- a/arch/sh/kernel/cpu/sh3/clock-sh7705.c
+++ b/arch/sh/kernel/cpu/sh3/clock-sh7705.c
@@ -32,7 +32,7 @@ static int pfc_divisors[]    = { 1, 2, 3, 4, 6, 1, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= pfc_divisors[ctrl_inw(FRQCR) & 0x0003];
+	clk->rate *= pfc_divisors[__raw_readw(FRQCR) & 0x0003];
 }
 
 static struct clk_ops sh7705_master_clk_ops = {
@@ -41,7 +41,7 @@ static struct clk_ops sh7705_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = ctrl_inw(FRQCR) & 0x0003;
+	int idx = __raw_readw(FRQCR) & 0x0003;
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -51,7 +51,7 @@ static struct clk_ops sh7705_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) & 0x0300) >> 8;
+	int idx = (__raw_readw(FRQCR) & 0x0300) >> 8;
 	return clk->parent->rate / stc_multipliers[idx];
 }
 
@@ -61,7 +61,7 @@ static struct clk_ops sh7705_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) & 0x0030) >> 4;
+	int idx = (__raw_readw(FRQCR) & 0x0030) >> 4;
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh3/clock-sh7706.c b/arch/sh/kernel/cpu/sh3/clock-sh7706.c
index 4bf7887d310..6f9ff8b57dd 100644
--- a/arch/sh/kernel/cpu/sh3/clock-sh7706.c
+++ b/arch/sh/kernel/cpu/sh3/clock-sh7706.c
@@ -24,7 +24,7 @@ static int pfc_divisors[]    = { 1, 2, 4, 1, 3, 6, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x2000) >> 11) | (frqcr & 0x0003);
 
 	clk->rate *= pfc_divisors[idx];
@@ -36,7 +36,7 @@ static struct clk_ops sh7706_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x2000) >> 11) | (frqcr & 0x0003);
 
 	return clk->parent->rate / pfc_divisors[idx];
@@ -48,7 +48,7 @@ static struct clk_ops sh7706_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x8000) >> 13) | ((frqcr & 0x0030) >> 4);
 
 	return clk->parent->rate / stc_multipliers[idx];
@@ -60,7 +60,7 @@ static struct clk_ops sh7706_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x4000) >> 12) | ((frqcr & 0x000c) >> 2);
 
 	return clk->parent->rate / ifc_divisors[idx];
diff --git a/arch/sh/kernel/cpu/sh3/clock-sh7709.c b/arch/sh/kernel/cpu/sh3/clock-sh7709.c
index e8749505bd2..f302ba09e68 100644
--- a/arch/sh/kernel/cpu/sh3/clock-sh7709.c
+++ b/arch/sh/kernel/cpu/sh3/clock-sh7709.c
@@ -24,7 +24,7 @@ static int pfc_divisors[]    = { 1, 2, 4, 1, 3, 6, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x2000) >> 11) | (frqcr & 0x0003);
 
 	clk->rate *= pfc_divisors[idx];
@@ -36,7 +36,7 @@ static struct clk_ops sh7709_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x2000) >> 11) | (frqcr & 0x0003);
 
 	return clk->parent->rate / pfc_divisors[idx];
@@ -48,7 +48,7 @@ static struct clk_ops sh7709_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = (frqcr & 0x0080) ?
 		((frqcr & 0x8000) >> 13) | ((frqcr & 0x0030) >> 4) : 1;
 
@@ -61,7 +61,7 @@ static struct clk_ops sh7709_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = ((frqcr & 0x4000) >> 12) | ((frqcr & 0x000c) >> 2);
 
 	return clk->parent->rate / ifc_divisors[idx];
diff --git a/arch/sh/kernel/cpu/sh3/clock-sh7710.c b/arch/sh/kernel/cpu/sh3/clock-sh7710.c
index 030a58ba18a..29a87d8946a 100644
--- a/arch/sh/kernel/cpu/sh3/clock-sh7710.c
+++ b/arch/sh/kernel/cpu/sh3/clock-sh7710.c
@@ -26,7 +26,7 @@ static int md_table[] = { 1, 2, 3, 4, 6, 8, 12 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= md_table[ctrl_inw(FRQCR) & 0x0007];
+	clk->rate *= md_table[__raw_readw(FRQCR) & 0x0007];
 }
 
 static struct clk_ops sh7710_master_clk_ops = {
@@ -35,7 +35,7 @@ static struct clk_ops sh7710_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) & 0x0007);
+	int idx = (__raw_readw(FRQCR) & 0x0007);
 	return clk->parent->rate / md_table[idx];
 }
 
@@ -45,7 +45,7 @@ static struct clk_ops sh7710_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) & 0x0700) >> 8;
+	int idx = (__raw_readw(FRQCR) & 0x0700) >> 8;
 	return clk->parent->rate / md_table[idx];
 }
 
@@ -55,7 +55,7 @@ static struct clk_ops sh7710_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) & 0x0070) >> 4;
+	int idx = (__raw_readw(FRQCR) & 0x0070) >> 4;
 	return clk->parent->rate / md_table[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh3/clock-sh7712.c b/arch/sh/kernel/cpu/sh3/clock-sh7712.c
index 6428ee6c77e..b0d0c520399 100644
--- a/arch/sh/kernel/cpu/sh3/clock-sh7712.c
+++ b/arch/sh/kernel/cpu/sh3/clock-sh7712.c
@@ -23,7 +23,7 @@ static int divisors[]    = { 1, 2, 3, 4, 6 };
 
 static void master_clk_init(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = (frqcr & 0x0300) >> 8;
 
 	clk->rate *= multipliers[idx];
@@ -35,7 +35,7 @@ static struct clk_ops sh7712_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = frqcr & 0x0007;
 
 	return clk->parent->rate / divisors[idx];
@@ -47,7 +47,7 @@ static struct clk_ops sh7712_module_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int frqcr = ctrl_inw(FRQCR);
+	int frqcr = __raw_readw(FRQCR);
 	int idx = (frqcr & 0x0030) >> 4;
 
 	return clk->parent->rate / divisors[idx];
diff --git a/arch/sh/kernel/cpu/sh3/ex.S b/arch/sh/kernel/cpu/sh3/ex.S
index 46610c35c23..99b4d020179 100644
--- a/arch/sh/kernel/cpu/sh3/ex.S
+++ b/arch/sh/kernel/cpu/sh3/ex.S
@@ -49,7 +49,7 @@ ENTRY(exception_handling_table)
 	.long	exception_error	! reserved_instruction (filled by trap_init) /* 180 */
 	.long	exception_error	! illegal_slot_instruction (filled by trap_init) /*1A0*/
 	.long	nmi_trap_handler	/* 1C0 */	! Allow trap to debugger
-	.long	break_point_trap	/* 1E0 */
+	.long	breakpoint_trap_handler	/* 1E0 */
 
 	/*
 	 * Pad the remainder of the table out, exceptions residing in far
diff --git a/arch/sh/kernel/cpu/sh3/probe.c b/arch/sh/kernel/cpu/sh3/probe.c
index f9c7df64eb0..295ec4c99e9 100644
--- a/arch/sh/kernel/cpu/sh3/probe.c
+++ b/arch/sh/kernel/cpu/sh3/probe.c
@@ -16,7 +16,7 @@
 #include <asm/cache.h>
 #include <asm/io.h>
 
-int __uses_jump_to_uncached detect_cpu_and_cache_system(void)
+int detect_cpu_and_cache_system(void)
 {
 	unsigned long addr0, addr1, data0, data1, data2, data3;
 
@@ -30,23 +30,23 @@ int __uses_jump_to_uncached detect_cpu_and_cache_system(void)
 	addr1 = CACHE_OC_ADDRESS_ARRAY + (1 << 12);
 
 	/* First, write back & invalidate */
-	data0  = ctrl_inl(addr0);
-	ctrl_outl(data0&~(SH_CACHE_VALID|SH_CACHE_UPDATED), addr0);
-	data1  = ctrl_inl(addr1);
-	ctrl_outl(data1&~(SH_CACHE_VALID|SH_CACHE_UPDATED), addr1);
+	data0  = __raw_readl(addr0);
+	__raw_writel(data0&~(SH_CACHE_VALID|SH_CACHE_UPDATED), addr0);
+	data1  = __raw_readl(addr1);
+	__raw_writel(data1&~(SH_CACHE_VALID|SH_CACHE_UPDATED), addr1);
 
 	/* Next, check if there's shadow or not */
-	data0 = ctrl_inl(addr0);
+	data0 = __raw_readl(addr0);
 	data0 ^= SH_CACHE_VALID;
-	ctrl_outl(data0, addr0);
-	data1 = ctrl_inl(addr1);
+	__raw_writel(data0, addr0);
+	data1 = __raw_readl(addr1);
 	data2 = data1 ^ SH_CACHE_VALID;
-	ctrl_outl(data2, addr1);
-	data3 = ctrl_inl(addr0);
+	__raw_writel(data2, addr1);
+	data3 = __raw_readl(addr0);
 
 	/* Lastly, invaliate them. */
-	ctrl_outl(data0&~SH_CACHE_VALID, addr0);
-	ctrl_outl(data2&~SH_CACHE_VALID, addr1);
+	__raw_writel(data0&~SH_CACHE_VALID, addr0);
+	__raw_writel(data2&~SH_CACHE_VALID, addr1);
 
 	back_to_cached();
 
@@ -94,9 +94,9 @@ int __uses_jump_to_uncached detect_cpu_and_cache_system(void)
 		boot_cpu_data.dcache.way_incr	= (1 << 13);
 		boot_cpu_data.dcache.entry_mask	= 0x1ff0;
 		boot_cpu_data.dcache.sets	= 512;
-		ctrl_outl(CCR_CACHE_32KB, CCR3_REG);
+		__raw_writel(CCR_CACHE_32KB, CCR3_REG);
 #else
-		ctrl_outl(CCR_CACHE_16KB, CCR3_REG);
+		__raw_writel(CCR_CACHE_16KB, CCR3_REG);
 #endif
 #endif
 	}
diff --git a/arch/sh/kernel/cpu/sh3/setup-sh3.c b/arch/sh/kernel/cpu/sh3/setup-sh3.c
index c9884685785..53be70b9811 100644
--- a/arch/sh/kernel/cpu/sh3/setup-sh3.c
+++ b/arch/sh/kernel/cpu/sh3/setup-sh3.c
@@ -58,7 +58,7 @@ static DECLARE_INTC_DESC_ACK(intc_desc_irq45, "sh3-irq45",
 void __init plat_irq_setup_pins(int mode)
 {
 	if (mode == IRQ_MODE_IRQ) {
-		ctrl_outw(ctrl_inw(INTC_ICR1) & ~INTC_ICR1_IRQLVL, INTC_ICR1);
+		__raw_writew(__raw_readw(INTC_ICR1) & ~INTC_ICR1_IRQLVL, INTC_ICR1);
 		register_intc_controller(&intc_desc_irq0123);
 		return;
 	}
diff --git a/arch/sh/kernel/cpu/sh4/clock-sh4-202.c b/arch/sh/kernel/cpu/sh4/clock-sh4-202.c
index 21421e34e7d..6b80850294d 100644
--- a/arch/sh/kernel/cpu/sh4/clock-sh4-202.c
+++ b/arch/sh/kernel/cpu/sh4/clock-sh4-202.c
@@ -23,7 +23,7 @@ static int frqcr3_values[]   = { 0, 1, 2, 3, 4, 5, 6  };
 
 static unsigned long emi_clk_recalc(struct clk *clk)
 {
-	int idx = ctrl_inl(CPG2_FRQCR3) & 0x0007;
+	int idx = __raw_readl(CPG2_FRQCR3) & 0x0007;
 	return clk->parent->rate / frqcr3_divisors[idx];
 }
 
@@ -52,7 +52,7 @@ static struct clk sh4202_emi_clk = {
 
 static unsigned long femi_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(CPG2_FRQCR3) >> 3) & 0x0007;
+	int idx = (__raw_readl(CPG2_FRQCR3) >> 3) & 0x0007;
 	return clk->parent->rate / frqcr3_divisors[idx];
 }
 
@@ -92,7 +92,7 @@ static void shoc_clk_init(struct clk *clk)
 
 static unsigned long shoc_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(CPG2_FRQCR3) >> 6) & 0x0007;
+	int idx = (__raw_readl(CPG2_FRQCR3) >> 6) & 0x0007;
 	return clk->parent->rate / frqcr3_divisors[idx];
 }
 
@@ -122,10 +122,10 @@ static int shoc_clk_set_rate(struct clk *clk, unsigned long rate, int algo_id)
 
 	tmp = frqcr3_lookup(clk, rate);
 
-	frqcr3 = ctrl_inl(CPG2_FRQCR3);
+	frqcr3 = __raw_readl(CPG2_FRQCR3);
 	frqcr3 &= ~(0x0007 << 6);
 	frqcr3 |= tmp << 6;
-	ctrl_outl(frqcr3, CPG2_FRQCR3);
+	__raw_writel(frqcr3, CPG2_FRQCR3);
 
 	clk->rate = clk->parent->rate / frqcr3_divisors[tmp];
 
diff --git a/arch/sh/kernel/cpu/sh4/clock-sh4.c b/arch/sh/kernel/cpu/sh4/clock-sh4.c
index 73294d9cd04..5add75c1f53 100644
--- a/arch/sh/kernel/cpu/sh4/clock-sh4.c
+++ b/arch/sh/kernel/cpu/sh4/clock-sh4.c
@@ -28,7 +28,7 @@ static int pfc_divisors[] = { 2, 3, 4, 6, 8, 2, 2, 2 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= pfc_divisors[ctrl_inw(FRQCR) & 0x0007];
+	clk->rate *= pfc_divisors[__raw_readw(FRQCR) & 0x0007];
 }
 
 static struct clk_ops sh4_master_clk_ops = {
@@ -37,7 +37,7 @@ static struct clk_ops sh4_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) & 0x0007);
+	int idx = (__raw_readw(FRQCR) & 0x0007);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -47,7 +47,7 @@ static struct clk_ops sh4_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) >> 3) & 0x0007;
+	int idx = (__raw_readw(FRQCR) >> 3) & 0x0007;
 	return clk->parent->rate / bfc_divisors[idx];
 }
 
@@ -57,7 +57,7 @@ static struct clk_ops sh4_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(FRQCR) >> 6) & 0x0007;
+	int idx = (__raw_readw(FRQCR) >> 6) & 0x0007;
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh4/fpu.c b/arch/sh/kernel/cpu/sh4/fpu.c
index e97857aec8a..447482d7f65 100644
--- a/arch/sh/kernel/cpu/sh4/fpu.c
+++ b/arch/sh/kernel/cpu/sh4/fpu.c
@@ -85,14 +85,14 @@ void save_fpu(struct task_struct *tsk)
 		      "fmov.s	fr1, @-%0\n\t"
 		      "fmov.s	fr0, @-%0\n\t"
 		      "lds	%3, fpscr\n\t":"=r" (dummy)
-		      :"0"((char *)(&tsk->thread.fpu.hard.status)),
+		      :"0"((char *)(&tsk->thread.xstate->hardfpu.status)),
 		      "r"(FPSCR_RCHG), "r"(FPSCR_INIT)
 		      :"memory");
 
 	disable_fpu();
 }
 
-static void restore_fpu(struct task_struct *tsk)
+void restore_fpu(struct task_struct *tsk)
 {
 	unsigned long dummy;
 
@@ -135,62 +135,11 @@ static void restore_fpu(struct task_struct *tsk)
 		      "lds.l	@%0+, fpscr\n\t"
 		      "lds.l	@%0+, fpul\n\t"
 		      :"=r" (dummy)
-		      :"0"(&tsk->thread.fpu), "r"(FPSCR_RCHG)
+		      :"0" (tsk->thread.xstate), "r" (FPSCR_RCHG)
 		      :"memory");
 	disable_fpu();
 }
 
-/*
- * Load the FPU with signalling NANS.  This bit pattern we're using
- * has the property that no matter wether considered as single or as
- * double precision represents signaling NANS.
- */
-
-static void fpu_init(void)
-{
-	enable_fpu();
-	asm volatile (	"lds	%0, fpul\n\t"
-			"lds	%1, fpscr\n\t"
-			"fsts	fpul, fr0\n\t"
-			"fsts	fpul, fr1\n\t"
-			"fsts	fpul, fr2\n\t"
-			"fsts	fpul, fr3\n\t"
-			"fsts	fpul, fr4\n\t"
-			"fsts	fpul, fr5\n\t"
-			"fsts	fpul, fr6\n\t"
-			"fsts	fpul, fr7\n\t"
-			"fsts	fpul, fr8\n\t"
-			"fsts	fpul, fr9\n\t"
-			"fsts	fpul, fr10\n\t"
-			"fsts	fpul, fr11\n\t"
-			"fsts	fpul, fr12\n\t"
-			"fsts	fpul, fr13\n\t"
-			"fsts	fpul, fr14\n\t"
-			"fsts	fpul, fr15\n\t"
-			"frchg\n\t"
-			"fsts	fpul, fr0\n\t"
-			"fsts	fpul, fr1\n\t"
-			"fsts	fpul, fr2\n\t"
-			"fsts	fpul, fr3\n\t"
-			"fsts	fpul, fr4\n\t"
-			"fsts	fpul, fr5\n\t"
-			"fsts	fpul, fr6\n\t"
-			"fsts	fpul, fr7\n\t"
-			"fsts	fpul, fr8\n\t"
-			"fsts	fpul, fr9\n\t"
-			"fsts	fpul, fr10\n\t"
-			"fsts	fpul, fr11\n\t"
-			"fsts	fpul, fr12\n\t"
-			"fsts	fpul, fr13\n\t"
-			"fsts	fpul, fr14\n\t"
-			"fsts	fpul, fr15\n\t"
-			"frchg\n\t"
-			"lds	%2, fpscr\n\t"
-			:	/* no output */
-			:"r" (0), "r"(FPSCR_RCHG), "r"(FPSCR_INIT));
-	disable_fpu();
-}
-
 /**
  *      denormal_to_double - Given denormalized float number,
  *                           store double float
@@ -282,9 +231,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 		/* fcnvsd */
 		struct task_struct *tsk = current;
 
-		if ((tsk->thread.fpu.hard.fpscr & FPSCR_CAUSE_ERROR))
+		if ((tsk->thread.xstate->hardfpu.fpscr & FPSCR_CAUSE_ERROR))
 			/* FPU error */
-			denormal_to_double(&tsk->thread.fpu.hard,
+			denormal_to_double(&tsk->thread.xstate->hardfpu,
 					   (finsn >> 8) & 0xf);
 		else
 			return 0;
@@ -300,9 +249,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 
 		n = (finsn >> 8) & 0xf;
 		m = (finsn >> 4) & 0xf;
-		hx = tsk->thread.fpu.hard.fp_regs[n];
-		hy = tsk->thread.fpu.hard.fp_regs[m];
-		fpscr = tsk->thread.fpu.hard.fpscr;
+		hx = tsk->thread.xstate->hardfpu.fp_regs[n];
+		hy = tsk->thread.xstate->hardfpu.fp_regs[m];
+		fpscr = tsk->thread.xstate->hardfpu.fpscr;
 		prec = fpscr & FPSCR_DBL_PRECISION;
 
 		if ((fpscr & FPSCR_CAUSE_ERROR)
@@ -312,18 +261,18 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 
 			/* FPU error because of denormal (doubles) */
 			llx = ((long long)hx << 32)
-			    | tsk->thread.fpu.hard.fp_regs[n + 1];
+			    | tsk->thread.xstate->hardfpu.fp_regs[n + 1];
 			lly = ((long long)hy << 32)
-			    | tsk->thread.fpu.hard.fp_regs[m + 1];
+			    | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
 			llx = float64_mul(llx, lly);
-			tsk->thread.fpu.hard.fp_regs[n] = llx >> 32;
-			tsk->thread.fpu.hard.fp_regs[n + 1] = llx & 0xffffffff;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
+			tsk->thread.xstate->hardfpu.fp_regs[n + 1] = llx & 0xffffffff;
 		} else if ((fpscr & FPSCR_CAUSE_ERROR)
 			   && (!prec && ((hx & 0x7fffffff) < 0x00800000
 					 || (hy & 0x7fffffff) < 0x00800000))) {
 			/* FPU error because of denormal (floats) */
 			hx = float32_mul(hx, hy);
-			tsk->thread.fpu.hard.fp_regs[n] = hx;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
 		} else
 			return 0;
 
@@ -338,9 +287,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 
 		n = (finsn >> 8) & 0xf;
 		m = (finsn >> 4) & 0xf;
-		hx = tsk->thread.fpu.hard.fp_regs[n];
-		hy = tsk->thread.fpu.hard.fp_regs[m];
-		fpscr = tsk->thread.fpu.hard.fpscr;
+		hx = tsk->thread.xstate->hardfpu.fp_regs[n];
+		hy = tsk->thread.xstate->hardfpu.fp_regs[m];
+		fpscr = tsk->thread.xstate->hardfpu.fpscr;
 		prec = fpscr & FPSCR_DBL_PRECISION;
 
 		if ((fpscr & FPSCR_CAUSE_ERROR)
@@ -350,15 +299,15 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 
 			/* FPU error because of denormal (doubles) */
 			llx = ((long long)hx << 32)
-			    | tsk->thread.fpu.hard.fp_regs[n + 1];
+			    | tsk->thread.xstate->hardfpu.fp_regs[n + 1];
 			lly = ((long long)hy << 32)
-			    | tsk->thread.fpu.hard.fp_regs[m + 1];
+			    | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
 			if ((finsn & 0xf00f) == 0xf000)
 				llx = float64_add(llx, lly);
 			else
 				llx = float64_sub(llx, lly);
-			tsk->thread.fpu.hard.fp_regs[n] = llx >> 32;
-			tsk->thread.fpu.hard.fp_regs[n + 1] = llx & 0xffffffff;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
+			tsk->thread.xstate->hardfpu.fp_regs[n + 1] = llx & 0xffffffff;
 		} else if ((fpscr & FPSCR_CAUSE_ERROR)
 			   && (!prec && ((hx & 0x7fffffff) < 0x00800000
 					 || (hy & 0x7fffffff) < 0x00800000))) {
@@ -367,7 +316,7 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 				hx = float32_add(hx, hy);
 			else
 				hx = float32_sub(hx, hy);
-			tsk->thread.fpu.hard.fp_regs[n] = hx;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
 		} else
 			return 0;
 
@@ -382,9 +331,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 
 		n = (finsn >> 8) & 0xf;
 		m = (finsn >> 4) & 0xf;
-		hx = tsk->thread.fpu.hard.fp_regs[n];
-		hy = tsk->thread.fpu.hard.fp_regs[m];
-		fpscr = tsk->thread.fpu.hard.fpscr;
+		hx = tsk->thread.xstate->hardfpu.fp_regs[n];
+		hy = tsk->thread.xstate->hardfpu.fp_regs[m];
+		fpscr = tsk->thread.xstate->hardfpu.fpscr;
 		prec = fpscr & FPSCR_DBL_PRECISION;
 
 		if ((fpscr & FPSCR_CAUSE_ERROR)
@@ -394,20 +343,20 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 
 			/* FPU error because of denormal (doubles) */
 			llx = ((long long)hx << 32)
-			    | tsk->thread.fpu.hard.fp_regs[n + 1];
+			    | tsk->thread.xstate->hardfpu.fp_regs[n + 1];
 			lly = ((long long)hy << 32)
-			    | tsk->thread.fpu.hard.fp_regs[m + 1];
+			    | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
 
 			llx = float64_div(llx, lly);
 
-			tsk->thread.fpu.hard.fp_regs[n] = llx >> 32;
-			tsk->thread.fpu.hard.fp_regs[n + 1] = llx & 0xffffffff;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
+			tsk->thread.xstate->hardfpu.fp_regs[n + 1] = llx & 0xffffffff;
 		} else if ((fpscr & FPSCR_CAUSE_ERROR)
 			   && (!prec && ((hx & 0x7fffffff) < 0x00800000
 					 || (hy & 0x7fffffff) < 0x00800000))) {
 			/* FPU error because of denormal (floats) */
 			hx = float32_div(hx, hy);
-			tsk->thread.fpu.hard.fp_regs[n] = hx;
+			tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
 		} else
 			return 0;
 
@@ -420,17 +369,17 @@ static int ieee_fpe_handler(struct pt_regs *regs)
 		unsigned int hx;
 
 		m = (finsn >> 8) & 0x7;
-		hx = tsk->thread.fpu.hard.fp_regs[m];
+		hx = tsk->thread.xstate->hardfpu.fp_regs[m];
 
-		if ((tsk->thread.fpu.hard.fpscr & FPSCR_CAUSE_ERROR)
+		if ((tsk->thread.xstate->hardfpu.fpscr & FPSCR_CAUSE_ERROR)
 			&& ((hx & 0x7fffffff) < 0x00100000)) {
 			/* subnormal double to float conversion */
 			long long llx;
 
-			llx = ((long long)tsk->thread.fpu.hard.fp_regs[m] << 32)
-			    | tsk->thread.fpu.hard.fp_regs[m + 1];
+			llx = ((long long)tsk->thread.xstate->hardfpu.fp_regs[m] << 32)
+			    | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
 
-			tsk->thread.fpu.hard.fpul = float64_to_float32(llx);
+			tsk->thread.xstate->hardfpu.fpul = float64_to_float32(llx);
 		} else
 			return 0;
 
@@ -449,7 +398,7 @@ void float_raise(unsigned int flags)
 int float_rounding_mode(void)
 {
 	struct task_struct *tsk = current;
-	int roundingMode = FPSCR_ROUNDING_MODE(tsk->thread.fpu.hard.fpscr);
+	int roundingMode = FPSCR_ROUNDING_MODE(tsk->thread.xstate->hardfpu.fpscr);
 	return roundingMode;
 }
 
@@ -461,16 +410,16 @@ BUILD_TRAP_HANDLER(fpu_error)
 	__unlazy_fpu(tsk, regs);
 	fpu_exception_flags = 0;
 	if (ieee_fpe_handler(regs)) {
-		tsk->thread.fpu.hard.fpscr &=
+		tsk->thread.xstate->hardfpu.fpscr &=
 		    ~(FPSCR_CAUSE_MASK | FPSCR_FLAG_MASK);
-		tsk->thread.fpu.hard.fpscr |= fpu_exception_flags;
+		tsk->thread.xstate->hardfpu.fpscr |= fpu_exception_flags;
 		/* Set the FPSCR flag as well as cause bits - simply
 		 * replicate the cause */
-		tsk->thread.fpu.hard.fpscr |= (fpu_exception_flags >> 10);
+		tsk->thread.xstate->hardfpu.fpscr |= (fpu_exception_flags >> 10);
 		grab_fpu(regs);
 		restore_fpu(tsk);
 		task_thread_info(tsk)->status |= TS_USEDFPU;
-		if ((((tsk->thread.fpu.hard.fpscr & FPSCR_ENABLE_MASK) >> 7) &
+		if ((((tsk->thread.xstate->hardfpu.fpscr & FPSCR_ENABLE_MASK) >> 7) &
 		     (fpu_exception_flags >> 2)) == 0) {
 			return;
 		}
@@ -478,33 +427,3 @@ BUILD_TRAP_HANDLER(fpu_error)
 
 	force_sig(SIGFPE, tsk);
 }
-
-void fpu_state_restore(struct pt_regs *regs)
-{
-	struct task_struct *tsk = current;
-
-	grab_fpu(regs);
-	if (unlikely(!user_mode(regs))) {
-		printk(KERN_ERR "BUG: FPU is used in kernel mode.\n");
-		BUG();
-		return;
-	}
-
-	if (likely(used_math())) {
-		/* Using the FPU again.  */
-		restore_fpu(tsk);
-	} else {
-		/* First time FPU user.  */
-		fpu_init();
-		set_used_math();
-	}
-	task_thread_info(tsk)->status |= TS_USEDFPU;
-	tsk->fpu_counter++;
-}
-
-BUILD_TRAP_HANDLER(fpu_state_restore)
-{
-	TRAP_HANDLER_DECL;
-
-	fpu_state_restore(regs);
-}
diff --git a/arch/sh/kernel/cpu/sh4/probe.c b/arch/sh/kernel/cpu/sh4/probe.c
index d36f0c45f55..822977a06d8 100644
--- a/arch/sh/kernel/cpu/sh4/probe.c
+++ b/arch/sh/kernel/cpu/sh4/probe.c
@@ -28,9 +28,9 @@ int __init detect_cpu_and_cache_system(void)
 		[9] = (1 << 16)
 	};
 
-	pvr = (ctrl_inl(CCN_PVR) >> 8) & 0xffffff;
-	prr = (ctrl_inl(CCN_PRR) >> 4) & 0xff;
-	cvr = (ctrl_inl(CCN_CVR));
+	pvr = (__raw_readl(CCN_PVR) >> 8) & 0xffffff;
+	prr = (__raw_readl(CCN_PRR) >> 4) & 0xff;
+	cvr = (__raw_readl(CCN_CVR));
 
 	/*
 	 * Setup some sane SH-4 defaults for the icache
@@ -71,11 +71,11 @@ int __init detect_cpu_and_cache_system(void)
 		boot_cpu_data.dcache.ways = 4;
 	} else {
 		/* And some SH-4 defaults.. */
-		boot_cpu_data.flags |= CPU_HAS_PTEA;
+		boot_cpu_data.flags |= CPU_HAS_PTEA | CPU_HAS_FPU;
 		boot_cpu_data.family = CPU_FAMILY_SH4;
 	}
 
-	/* FPU detection works for everyone */
+	/* FPU detection works for almost everyone */
 	if ((cvr & 0x20000000))
 		boot_cpu_data.flags |= CPU_HAS_FPU;
 
@@ -124,6 +124,7 @@ int __init detect_cpu_and_cache_system(void)
 		boot_cpu_data.type = CPU_SH7785;
 		break;
 	case 0x4004:
+	case 0x4005:
 		boot_cpu_data.type = CPU_SH7786;
 		boot_cpu_data.flags |= CPU_HAS_PTEAEX | CPU_HAS_L2_CACHE;
 		break;
@@ -160,6 +161,7 @@ int __init detect_cpu_and_cache_system(void)
 		break;
 	case 0x700:
 		boot_cpu_data.type = CPU_SH4_501;
+		boot_cpu_data.flags &= ~CPU_HAS_FPU;
 		boot_cpu_data.icache.ways = 2;
 		boot_cpu_data.dcache.ways = 2;
 		break;
@@ -227,7 +229,7 @@ int __init detect_cpu_and_cache_system(void)
 			 * Size calculation is much more sensible
 			 * than it is for the L1.
 			 *
-			 * Sizes are 128KB, 258KB, 512KB, and 1MB.
+			 * Sizes are 128KB, 256KB, 512KB, and 1MB.
 			 */
 			size = (cvr & 0xf) << 17;
 
diff --git a/arch/sh/kernel/cpu/sh4/setup-sh4-202.c b/arch/sh/kernel/cpu/sh4/setup-sh4-202.c
index 4b733715cdb..b9b7e10ad68 100644
--- a/arch/sh/kernel/cpu/sh4/setup-sh4-202.c
+++ b/arch/sh/kernel/cpu/sh4/setup-sh4-202.c
@@ -198,7 +198,7 @@ void __init plat_irq_setup_pins(int mode)
 {
 	switch (mode) {
 	case IRQ_MODE_IRQ: /* individual interrupt mode for IRL3-0 */
-		ctrl_outw(ctrl_inw(INTC_ICR) | INTC_ICR_IRLM, INTC_ICR);
+		__raw_writew(__raw_readw(INTC_ICR) | INTC_ICR_IRLM, INTC_ICR);
 		register_intc_controller(&intc_desc_irlm);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4/setup-sh7750.c b/arch/sh/kernel/cpu/sh4/setup-sh7750.c
index b2a9df1af64..ffd79e57254 100644
--- a/arch/sh/kernel/cpu/sh4/setup-sh7750.c
+++ b/arch/sh/kernel/cpu/sh4/setup-sh7750.c
@@ -442,7 +442,7 @@ void __init plat_irq_setup_pins(int mode)
 
 	switch (mode) {
 	case IRQ_MODE_IRQ: /* individual interrupt mode for IRL3-0 */
-		ctrl_outw(ctrl_inw(INTC_ICR) | INTC_ICR_IRLM, INTC_ICR);
+		__raw_writew(__raw_readw(INTC_ICR) | INTC_ICR_IRLM, INTC_ICR);
 		register_intc_controller(&intc_desc_irlm);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4/setup-sh7760.c b/arch/sh/kernel/cpu/sh4/setup-sh7760.c
index 5b74cc0b43d..a16eb3656f4 100644
--- a/arch/sh/kernel/cpu/sh4/setup-sh7760.c
+++ b/arch/sh/kernel/cpu/sh4/setup-sh7760.c
@@ -319,7 +319,7 @@ void __init plat_irq_setup_pins(int mode)
 {
 	switch (mode) {
 	case IRQ_MODE_IRQ:
-		ctrl_outw(ctrl_inw(INTC_ICR) | INTC_ICR_IRLM, INTC_ICR);
+		__raw_writew(__raw_readw(INTC_ICR) | INTC_ICR_IRLM, INTC_ICR);
 		register_intc_controller(&intc_desc_irq);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4/sq.c b/arch/sh/kernel/cpu/sh4/sq.c
index 8a8a993f55e..fc065f9da6e 100644
--- a/arch/sh/kernel/cpu/sh4/sq.c
+++ b/arch/sh/kernel/cpu/sh4/sq.c
@@ -43,9 +43,9 @@ static unsigned long *sq_bitmap;
 
 #define store_queue_barrier()			\
 do {						\
-	(void)ctrl_inl(P4SEG_STORE_QUE);	\
-	ctrl_outl(0, P4SEG_STORE_QUE + 0);	\
-	ctrl_outl(0, P4SEG_STORE_QUE + 8);	\
+	(void)__raw_readl(P4SEG_STORE_QUE);	\
+	__raw_writel(0, P4SEG_STORE_QUE + 0);	\
+	__raw_writel(0, P4SEG_STORE_QUE + 8);	\
 } while (0);
 
 /**
@@ -100,7 +100,7 @@ static inline void sq_mapping_list_del(struct sq_mapping *map)
 	spin_unlock_irq(&sq_mapping_lock);
 }
 
-static int __sq_remap(struct sq_mapping *map, unsigned long flags)
+static int __sq_remap(struct sq_mapping *map, pgprot_t prot)
 {
 #if defined(CONFIG_MMU)
 	struct vm_struct *vma;
@@ -113,7 +113,7 @@ static int __sq_remap(struct sq_mapping *map, unsigned long flags)
 
 	if (ioremap_page_range((unsigned long)vma->addr,
 			       (unsigned long)vma->addr + map->size,
-			       vma->phys_addr, __pgprot(flags))) {
+			       vma->phys_addr, prot)) {
 		vunmap(vma->addr);
 		return -EAGAIN;
 	}
@@ -123,8 +123,8 @@ static int __sq_remap(struct sq_mapping *map, unsigned long flags)
 	 * straightforward, as we can just load up each queue's QACR with
 	 * the physical address appropriately masked.
 	 */
-	ctrl_outl(((map->addr >> 26) << 2) & 0x1c, SQ_QACR0);
-	ctrl_outl(((map->addr >> 26) << 2) & 0x1c, SQ_QACR1);
+	__raw_writel(((map->addr >> 26) << 2) & 0x1c, SQ_QACR0);
+	__raw_writel(((map->addr >> 26) << 2) & 0x1c, SQ_QACR1);
 #endif
 
 	return 0;
@@ -135,14 +135,14 @@ static int __sq_remap(struct sq_mapping *map, unsigned long flags)
  * @phys: Physical address of mapping.
  * @size: Length of mapping.
  * @name: User invoking mapping.
- * @flags: Protection flags.
+ * @prot: Protection bits.
  *
  * Remaps the physical address @phys through the next available store queue
  * address of @size length. @name is logged at boot time as well as through
  * the sysfs interface.
  */
 unsigned long sq_remap(unsigned long phys, unsigned int size,
-		       const char *name, unsigned long flags)
+		       const char *name, pgprot_t prot)
 {
 	struct sq_mapping *map;
 	unsigned long end;
@@ -177,7 +177,7 @@ unsigned long sq_remap(unsigned long phys, unsigned int size,
 
 	map->sq_addr = P4SEG_STORE_QUE + (page << PAGE_SHIFT);
 
-	ret = __sq_remap(map, pgprot_val(PAGE_KERNEL_NOCACHE) | flags);
+	ret = __sq_remap(map, prot);
 	if (unlikely(ret != 0))
 		goto out;
 
@@ -309,8 +309,7 @@ static ssize_t mapping_store(const char *buf, size_t count)
 		return -EIO;
 
 	if (likely(len)) {
-		int ret = sq_remap(base, len, "Userspace",
-				   pgprot_val(PAGE_SHARED));
+		int ret = sq_remap(base, len, "Userspace", PAGE_SHARED);
 		if (ret < 0)
 			return ret;
 	} else
diff --git a/arch/sh/kernel/cpu/sh4a/Makefile b/arch/sh/kernel/cpu/sh4a/Makefile
index 33bab477d2e..b144e8af89d 100644
--- a/arch/sh/kernel/cpu/sh4a/Makefile
+++ b/arch/sh/kernel/cpu/sh4a/Makefile
@@ -41,7 +41,8 @@ pinmux-$(CONFIG_CPU_SUBTYPE_SH7757)	:= pinmux-sh7757.o
 pinmux-$(CONFIG_CPU_SUBTYPE_SH7785)	:= pinmux-sh7785.o
 pinmux-$(CONFIG_CPU_SUBTYPE_SH7786)	:= pinmux-sh7786.o
 
-obj-y				+= $(clock-y)
-obj-$(CONFIG_SMP)		+= $(smp-y)
-obj-$(CONFIG_GENERIC_GPIO)	+= $(pinmux-y)
-obj-$(CONFIG_PERF_EVENTS)	+= perf_event.o
+obj-y					+= $(clock-y)
+obj-$(CONFIG_SMP)			+= $(smp-y)
+obj-$(CONFIG_GENERIC_GPIO)		+= $(pinmux-y)
+obj-$(CONFIG_PERF_EVENTS)		+= perf_event.o
+obj-$(CONFIG_HAVE_HW_BREAKPOINT)	+= ubc.o
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7343.c b/arch/sh/kernel/cpu/sh4a/clock-sh7343.c
index 0ee3ee86125..2c16df37eda 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7343.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7343.c
@@ -107,13 +107,17 @@ struct clk *main_clks[] = {
 static int multipliers[] = { 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1 };
 static int divisors[] = { 1, 3, 2, 5, 3, 4, 5, 6, 8, 10, 12, 16, 20 };
 
-static struct clk_div_mult_table div4_table = {
+static struct clk_div_mult_table div4_div_mult_table = {
 	.divisors = divisors,
 	.nr_divisors = ARRAY_SIZE(divisors),
 	.multipliers = multipliers,
 	.nr_multipliers = ARRAY_SIZE(multipliers),
 };
 
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
+};
+
 enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_B3, DIV4_P,
        DIV4_SIUA, DIV4_SIUB, DIV4_NR };
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7366.c b/arch/sh/kernel/cpu/sh4a/clock-sh7366.c
index a95ebaba095..91588d280cd 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7366.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7366.c
@@ -110,13 +110,17 @@ struct clk *main_clks[] = {
 static int multipliers[] = { 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1 };
 static int divisors[] = { 1, 3, 2, 5, 3, 4, 5, 6, 8, 10, 12, 16, 20 };
 
-static struct clk_div_mult_table div4_table = {
+static struct clk_div_mult_table div4_div_mult_table = {
 	.divisors = divisors,
 	.nr_divisors = ARRAY_SIZE(divisors),
 	.multipliers = multipliers,
 	.nr_multipliers = ARRAY_SIZE(multipliers),
 };
 
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
+};
+
 enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_B3, DIV4_P,
        DIV4_SIUA, DIV4_SIUB, DIV4_NR };
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7722.c b/arch/sh/kernel/cpu/sh4a/clock-sh7722.c
index ea38b554dc0..15db6d521c5 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7722.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7722.c
@@ -110,19 +110,22 @@ struct clk *main_clks[] = {
 static int multipliers[] = { 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1 };
 static int divisors[] = { 1, 3, 2, 5, 3, 4, 5, 6, 8, 10, 12, 16, 20 };
 
-static struct clk_div_mult_table div4_table = {
+static struct clk_div_mult_table div4_div_mult_table = {
 	.divisors = divisors,
 	.nr_divisors = ARRAY_SIZE(divisors),
 	.multipliers = multipliers,
 	.nr_multipliers = ARRAY_SIZE(multipliers),
 };
 
-enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_B3, DIV4_P,
-       DIV4_SIUA, DIV4_SIUB, DIV4_IRDA, DIV4_NR };
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
+};
 
 #define DIV4(_str, _reg, _bit, _mask, _flags) \
   SH_CLK_DIV4(_str, &pll_clk, _reg, _bit, _mask, _flags)
 
+enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_B3, DIV4_P, DIV4_NR };
+
 struct clk div4_clks[DIV4_NR] = {
 	[DIV4_I] = DIV4("cpu_clk", FRQCR, 20, 0x1fef, CLK_ENABLE_ON_INIT),
 	[DIV4_U] = DIV4("umem_clk", FRQCR, 16, 0x1fff, CLK_ENABLE_ON_INIT),
@@ -130,9 +133,19 @@ struct clk div4_clks[DIV4_NR] = {
 	[DIV4_B] = DIV4("bus_clk", FRQCR, 8, 0x1fff, CLK_ENABLE_ON_INIT),
 	[DIV4_B3] = DIV4("b3_clk", FRQCR, 4, 0x1fff, CLK_ENABLE_ON_INIT),
 	[DIV4_P] = DIV4("peripheral_clk", FRQCR, 0, 0x1fff, 0),
+};
+
+enum { DIV4_IRDA, DIV4_ENABLE_NR };
+
+struct clk div4_enable_clks[DIV4_ENABLE_NR] = {
+	[DIV4_IRDA] = DIV4("irda_clk", IRDACLKCR, 0, 0x1fff, 0),
+};
+
+enum { DIV4_SIUA, DIV4_SIUB, DIV4_REPARENT_NR };
+
+struct clk div4_reparent_clks[DIV4_REPARENT_NR] = {
 	[DIV4_SIUA] = DIV4("siua_clk", SCLKACR, 0, 0x1fff, 0),
 	[DIV4_SIUB] = DIV4("siub_clk", SCLKBCR, 0, 0x1fff, 0),
-	[DIV4_IRDA] = DIV4("irda_clk", IRDACLKCR, 0, 0x1fff, 0),
 };
 
 struct clk div6_clks[] = {
@@ -189,6 +202,14 @@ int __init arch_clk_init(void)
 		ret = sh_clk_div4_register(div4_clks, DIV4_NR, &div4_table);
 
 	if (!ret)
+		ret = sh_clk_div4_enable_register(div4_enable_clks,
+					DIV4_ENABLE_NR, &div4_table);
+
+	if (!ret)
+		ret = sh_clk_div4_reparent_register(div4_reparent_clks,
+					DIV4_REPARENT_NR, &div4_table);
+
+	if (!ret)
 		ret = sh_clk_div6_register(div6_clks, ARRAY_SIZE(div6_clks));
 
 	if (!ret)
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7723.c b/arch/sh/kernel/cpu/sh4a/clock-sh7723.c
index 20a31c2255a..50babe01fe4 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7723.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7723.c
@@ -110,15 +110,18 @@ struct clk *main_clks[] = {
 static int multipliers[] = { 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1 };
 static int divisors[] = { 1, 3, 2, 5, 3, 4, 5, 6, 8, 10, 12, 16, 20 };
 
-static struct clk_div_mult_table div4_table = {
+static struct clk_div_mult_table div4_div_mult_table = {
 	.divisors = divisors,
 	.nr_divisors = ARRAY_SIZE(divisors),
 	.multipliers = multipliers,
 	.nr_multipliers = ARRAY_SIZE(multipliers),
 };
 
-enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_B3, DIV4_P,
-       DIV4_SIUA, DIV4_SIUB, DIV4_IRDA, DIV4_NR };
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
+};
+
+enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_B3, DIV4_P, DIV4_NR };
 
 #define DIV4(_str, _reg, _bit, _mask, _flags) \
   SH_CLK_DIV4(_str, &pll_clk, _reg, _bit, _mask, _flags)
@@ -130,11 +133,20 @@ struct clk div4_clks[DIV4_NR] = {
 	[DIV4_B] = DIV4("bus_clk", FRQCR, 8, 0x0dbf, CLK_ENABLE_ON_INIT),
 	[DIV4_B3] = DIV4("b3_clk", FRQCR, 4, 0x0db4, CLK_ENABLE_ON_INIT),
 	[DIV4_P] = DIV4("peripheral_clk", FRQCR, 0, 0x0dbf, 0),
-	[DIV4_SIUA] = DIV4("siua_clk", SCLKACR, 0, 0x0dbf, 0),
-	[DIV4_SIUB] = DIV4("siub_clk", SCLKBCR, 0, 0x0dbf, 0),
+};
+
+enum { DIV4_IRDA, DIV4_ENABLE_NR };
+
+struct clk div4_enable_clks[DIV4_ENABLE_NR] = {
 	[DIV4_IRDA] = DIV4("irda_clk", IRDACLKCR, 0, 0x0dbf, 0),
 };
 
+enum { DIV4_SIUA, DIV4_SIUB, DIV4_REPARENT_NR };
+
+struct clk div4_reparent_clks[DIV4_REPARENT_NR] = {
+	[DIV4_SIUA] = DIV4("siua_clk", SCLKACR, 0, 0x0dbf, 0),
+	[DIV4_SIUB] = DIV4("siub_clk", SCLKBCR, 0, 0x0dbf, 0),
+};
 struct clk div6_clks[] = {
 	SH_CLK_DIV6("video_clk", &pll_clk, VCLKCR, 0),
 };
@@ -216,6 +228,14 @@ int __init arch_clk_init(void)
 		ret = sh_clk_div4_register(div4_clks, DIV4_NR, &div4_table);
 
 	if (!ret)
+		ret = sh_clk_div4_enable_register(div4_enable_clks,
+					DIV4_ENABLE_NR, &div4_table);
+
+	if (!ret)
+		ret = sh_clk_div4_reparent_register(div4_reparent_clks,
+					DIV4_REPARENT_NR, &div4_table);
+
+	if (!ret)
 		ret = sh_clk_div6_register(div6_clks, ARRAY_SIZE(div6_clks));
 
 	if (!ret)
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7724.c b/arch/sh/kernel/cpu/sh4a/clock-sh7724.c
index 9db743802f0..6707061fbf5 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7724.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7724.c
@@ -127,13 +127,28 @@ struct clk *main_clks[] = {
 	&div3_clk,
 };
 
+static void div4_kick(struct clk *clk)
+{
+	unsigned long value;
+
+	/* set KICK bit in FRQCRA to update hardware setting */
+	value = __raw_readl(FRQCRA);
+	value |= (1 << 31);
+	__raw_writel(value, FRQCRA);
+}
+
 static int divisors[] = { 2, 3, 4, 6, 8, 12, 16, 0, 24, 32, 36, 48, 0, 72 };
 
-static struct clk_div_mult_table div4_table = {
+static struct clk_div_mult_table div4_div_mult_table = {
 	.divisors = divisors,
 	.nr_divisors = ARRAY_SIZE(divisors),
 };
 
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
+	.kick = div4_kick,
+};
+
 enum { DIV4_I, DIV4_SH, DIV4_B, DIV4_P, DIV4_M1, DIV4_NR };
 
 #define DIV4(_str, _reg, _bit, _mask, _flags) \
@@ -144,7 +159,7 @@ struct clk div4_clks[DIV4_NR] = {
 	[DIV4_SH] = DIV4("shyway_clk", FRQCRA, 12, 0x2f7c, CLK_ENABLE_ON_INIT),
 	[DIV4_B] = DIV4("bus_clk", FRQCRA, 8, 0x2f7c, CLK_ENABLE_ON_INIT),
 	[DIV4_P] = DIV4("peripheral_clk", FRQCRA, 0, 0x2f7c, 0),
-	[DIV4_M1] = DIV4("vpu_clk", FRQCRB, 4, 0x2f7c, 0),
+	[DIV4_M1] = DIV4("vpu_clk", FRQCRB, 4, 0x2f7c, CLK_ENABLE_ON_INIT),
 };
 
 struct clk div6_clks[] = {
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7757.c b/arch/sh/kernel/cpu/sh4a/clock-sh7757.c
index ddc235ca966..86aae60677d 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7757.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7757.c
@@ -35,7 +35,7 @@ static struct clk_ops sh7757_master_clk_ops = {
 
 static void module_clk_recalc(struct clk *clk)
 {
-	int idx = ctrl_inl(FRQCR) & 0x0000000f;
+	int idx = __raw_readl(FRQCR) & 0x0000000f;
 	clk->rate = clk->parent->rate / p1fc_divisors[idx];
 }
 
@@ -45,7 +45,7 @@ static struct clk_ops sh7757_module_clk_ops = {
 
 static void bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(FRQCR) >> 8) & 0x0000000f;
+	int idx = (__raw_readl(FRQCR) >> 8) & 0x0000000f;
 	clk->rate = clk->parent->rate / bfc_divisors[idx];
 }
 
@@ -55,7 +55,7 @@ static struct clk_ops sh7757_bus_clk_ops = {
 
 static void cpu_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(FRQCR) >> 20) & 0x0000000f;
+	int idx = (__raw_readl(FRQCR) >> 20) & 0x0000000f;
 	clk->rate = clk->parent->rate / ifc_divisors[idx];
 }
 
@@ -78,7 +78,7 @@ void __init arch_init_clk_ops(struct clk_ops **ops, int idx)
 
 static void shyway_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(FRQCR) >> 12) & 0x0000000f;
+	int idx = (__raw_readl(FRQCR) >> 12) & 0x0000000f;
 	clk->rate = clk->parent->rate / sfc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7763.c b/arch/sh/kernel/cpu/sh4a/clock-sh7763.c
index 370cd47642e..9f401163e71 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7763.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7763.c
@@ -22,7 +22,7 @@ static int cfc_divisors[] = { 1, 1, 4, 1, 1, 1, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= p0fc_divisors[(ctrl_inl(FRQCR) >> 4) & 0x07];
+	clk->rate *= p0fc_divisors[(__raw_readl(FRQCR) >> 4) & 0x07];
 }
 
 static struct clk_ops sh7763_master_clk_ops = {
@@ -31,7 +31,7 @@ static struct clk_ops sh7763_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 4) & 0x07);
+	int idx = ((__raw_readl(FRQCR) >> 4) & 0x07);
 	return clk->parent->rate / p0fc_divisors[idx];
 }
 
@@ -41,7 +41,7 @@ static struct clk_ops sh7763_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 16) & 0x07);
+	int idx = ((__raw_readl(FRQCR) >> 16) & 0x07);
 	return clk->parent->rate / bfc_divisors[idx];
 }
 
@@ -68,7 +68,7 @@ void __init arch_init_clk_ops(struct clk_ops **ops, int idx)
 
 static unsigned long shyway_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 20) & 0x07);
+	int idx = ((__raw_readl(FRQCR) >> 20) & 0x07);
 	return clk->parent->rate / cfc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7770.c b/arch/sh/kernel/cpu/sh4a/clock-sh7770.c
index e0b89676920..9e3354365d4 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7770.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7770.c
@@ -21,7 +21,7 @@ static int pfc_divisors[] = { 1, 8, 1,10,12,16, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= pfc_divisors[(ctrl_inl(FRQCR) >> 28) & 0x000f];
+	clk->rate *= pfc_divisors[(__raw_readl(FRQCR) >> 28) & 0x000f];
 }
 
 static struct clk_ops sh7770_master_clk_ops = {
@@ -30,7 +30,7 @@ static struct clk_ops sh7770_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 28) & 0x000f);
+	int idx = ((__raw_readl(FRQCR) >> 28) & 0x000f);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -40,7 +40,7 @@ static struct clk_ops sh7770_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(FRQCR) & 0x000f);
+	int idx = (__raw_readl(FRQCR) & 0x000f);
 	return clk->parent->rate / bfc_divisors[idx];
 }
 
@@ -50,7 +50,7 @@ static struct clk_ops sh7770_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 24) & 0x000f);
+	int idx = ((__raw_readl(FRQCR) >> 24) & 0x000f);
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7780.c b/arch/sh/kernel/cpu/sh4a/clock-sh7780.c
index a249d823578..150963a6001 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7780.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7780.c
@@ -22,7 +22,7 @@ static int cfc_divisors[] = { 1, 1, 4, 1, 6, 1, 1, 1 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= pfc_divisors[ctrl_inl(FRQCR) & 0x0003];
+	clk->rate *= pfc_divisors[__raw_readl(FRQCR) & 0x0003];
 }
 
 static struct clk_ops sh7780_master_clk_ops = {
@@ -31,7 +31,7 @@ static struct clk_ops sh7780_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(FRQCR) & 0x0003);
+	int idx = (__raw_readl(FRQCR) & 0x0003);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -41,7 +41,7 @@ static struct clk_ops sh7780_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 16) & 0x0007);
+	int idx = ((__raw_readl(FRQCR) >> 16) & 0x0007);
 	return clk->parent->rate / bfc_divisors[idx];
 }
 
@@ -51,7 +51,7 @@ static struct clk_ops sh7780_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 24) & 0x0001);
+	int idx = ((__raw_readl(FRQCR) >> 24) & 0x0001);
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
@@ -74,7 +74,7 @@ void __init arch_init_clk_ops(struct clk_ops **ops, int idx)
 
 static unsigned long shyway_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> 20) & 0x0007);
+	int idx = ((__raw_readl(FRQCR) >> 20) & 0x0007);
 	return clk->parent->rate / cfc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7785.c b/arch/sh/kernel/cpu/sh4a/clock-sh7785.c
index 73abfbf2f16..d997f0a25b1 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7785.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7785.c
@@ -57,11 +57,15 @@ static struct clk *clks[] = {
 static unsigned int div2[] = { 1, 2, 4, 6, 8, 12, 16, 18,
 			       24, 32, 36, 48 };
 
-static struct clk_div_mult_table div4_table = {
+static struct clk_div_mult_table div4_div_mult_table = {
 	.divisors = div2,
 	.nr_divisors = ARRAY_SIZE(div2),
 };
 
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
+};
+
 enum { DIV4_I, DIV4_U, DIV4_SH, DIV4_B, DIV4_DDR, DIV4_GA,
 	DIV4_DU, DIV4_P, DIV4_NR };
 
diff --git a/arch/sh/kernel/cpu/sh4a/clock-sh7786.c b/arch/sh/kernel/cpu/sh4a/clock-sh7786.c
index a0e8869071a..af69fd46870 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-sh7786.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-sh7786.c
@@ -3,11 +3,7 @@
  *
  * SH7786 support for the clock framework
  *
- * Copyright (C) 2008, 2009  Renesas Solutions Corp.
- * Kuninori Morimoto <morimoto.kuninori@renesas.com>
- *
- * Based on SH7785
- *  Copyright (C) 2007  Paul Mundt
+ *  Copyright (C) 2010  Paul Mundt
  *
  * This file is subject to the terms and conditions of the GNU General Public
  * License.  See the file "COPYING" in the main directory of this archive
@@ -15,127 +11,127 @@
  */
 #include <linux/init.h>
 #include <linux/kernel.h>
+#include <linux/clk.h>
+#include <linux/io.h>
 #include <asm/clock.h>
 #include <asm/freq.h>
-#include <asm/io.h>
-
-static int ifc_divisors[] = { 1, 2, 4, 1 };
-static int sfc_divisors[] = { 1, 1, 4, 1 };
-static int bfc_divisors[] = { 1, 1, 1, 1, 1, 12, 16, 1,
-			     24, 32, 1, 1, 1, 1, 1, 1 };
-static int mfc_divisors[] = { 1, 1, 4, 1 };
-static int pfc_divisors[] = { 1, 1, 1, 1, 1, 1, 16, 1,
-			      24, 32, 1, 48, 1, 1, 1, 1 };
 
-static void master_clk_init(struct clk *clk)
-{
-	clk->rate *= pfc_divisors[ctrl_inl(FRQMR1) & 0x000f];
-}
-
-static struct clk_ops sh7786_master_clk_ops = {
-	.init		= master_clk_init,
+/*
+ * Default rate for the root input clock, reset this with clk_set_rate()
+ * from the platform code.
+ */
+static struct clk extal_clk = {
+	.name		= "extal",
+	.id		= -1,
+	.rate		= 33333333,
 };
 
-static unsigned long module_clk_recalc(struct clk *clk)
+static unsigned long pll_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inl(FRQMR1) & 0x000f);
-	return clk->parent->rate / pfc_divisors[idx];
-}
+	int multiplier;
 
-static struct clk_ops sh7786_module_clk_ops = {
-	.recalc		= module_clk_recalc,
-};
+	/*
+	 * Clock modes 0, 1, and 2 use an x64 multiplier against PLL1,
+	 * while modes 3, 4, and 5 use an x32.
+	 */
+	multiplier = (sh_mv.mv_mode_pins() & 0xf) < 3 ? 64 : 32;
 
-static unsigned long bus_clk_recalc(struct clk *clk)
-{
-	int idx = ((ctrl_inl(FRQMR1) >> 16) & 0x000f);
-	return clk->parent->rate / bfc_divisors[idx];
+	return clk->parent->rate * multiplier;
 }
 
-static struct clk_ops sh7786_bus_clk_ops = {
-	.recalc		= bus_clk_recalc,
+static struct clk_ops pll_clk_ops = {
+	.recalc		= pll_recalc,
 };
 
-static unsigned long cpu_clk_recalc(struct clk *clk)
-{
-	int idx = ((ctrl_inl(FRQMR1) >> 28) & 0x0003);
-	return clk->parent->rate / ifc_divisors[idx];
-}
-
-static struct clk_ops sh7786_cpu_clk_ops = {
-	.recalc		= cpu_clk_recalc,
+static struct clk pll_clk = {
+	.name		= "pll_clk",
+	.id		= -1,
+	.ops		= &pll_clk_ops,
+	.parent		= &extal_clk,
+	.flags		= CLK_ENABLE_ON_INIT,
 };
 
-static struct clk_ops *sh7786_clk_ops[] = {
-	&sh7786_master_clk_ops,
-	&sh7786_module_clk_ops,
-	&sh7786_bus_clk_ops,
-	&sh7786_cpu_clk_ops,
+static struct clk *clks[] = {
+	&extal_clk,
+	&pll_clk,
 };
 
-void __init arch_init_clk_ops(struct clk_ops **ops, int idx)
-{
-	if (idx < ARRAY_SIZE(sh7786_clk_ops))
-		*ops = sh7786_clk_ops[idx];
-}
+static unsigned int div2[] = { 1, 2, 4, 6, 8, 12, 16, 18,
+			       24, 32, 36, 48 };
 
-static unsigned long shyway_clk_recalc(struct clk *clk)
-{
-	int idx = ((ctrl_inl(FRQMR1) >> 20) & 0x0003);
-	return clk->parent->rate / sfc_divisors[idx];
-}
-
-static struct clk_ops sh7786_shyway_clk_ops = {
-	.recalc		= shyway_clk_recalc,
+static struct clk_div_mult_table div4_div_mult_table = {
+	.divisors = div2,
+	.nr_divisors = ARRAY_SIZE(div2),
 };
 
-static struct clk sh7786_shyway_clk = {
-	.name		= "shyway_clk",
-	.flags		= CLK_ENABLE_ON_INIT,
-	.ops		= &sh7786_shyway_clk_ops,
+static struct clk_div4_table div4_table = {
+	.div_mult_table = &div4_div_mult_table,
 };
 
-static unsigned long ddr_clk_recalc(struct clk *clk)
-{
-	int idx = ((ctrl_inl(FRQMR1) >> 12) & 0x0003);
-	return clk->parent->rate / mfc_divisors[idx];
-}
+enum { DIV4_I, DIV4_SH, DIV4_B, DIV4_DDR, DIV4_DU, DIV4_P, DIV4_NR };
 
-static struct clk_ops sh7786_ddr_clk_ops = {
-	.recalc		= ddr_clk_recalc,
-};
+#define DIV4(_str, _bit, _mask, _flags) \
+  SH_CLK_DIV4(_str, &pll_clk, FRQMR1, _bit, _mask, _flags)
 
-static struct clk sh7786_ddr_clk = {
-	.name		= "ddr_clk",
-	.flags		= CLK_ENABLE_ON_INIT,
-	.ops		= &sh7786_ddr_clk_ops,
+struct clk div4_clks[DIV4_NR] = {
+	[DIV4_P] = DIV4("peripheral_clk", 0, 0x0b40, 0),
+	[DIV4_DU] = DIV4("du_clk", 4, 0x0010, 0),
+	[DIV4_DDR] = DIV4("ddr_clk", 12, 0x0002, CLK_ENABLE_ON_INIT),
+	[DIV4_B] = DIV4("bus_clk", 16, 0x0360, CLK_ENABLE_ON_INIT),
+	[DIV4_SH] = DIV4("shyway_clk", 20, 0x0002, CLK_ENABLE_ON_INIT),
+	[DIV4_I] = DIV4("cpu_clk", 28, 0x0006, CLK_ENABLE_ON_INIT),
 };
 
-/*
- * Additional SH7786-specific on-chip clocks that aren't already part of the
- * clock framework
- */
-static struct clk *sh7786_onchip_clocks[] = {
-	&sh7786_shyway_clk,
-	&sh7786_ddr_clk,
+#define MSTPCR0		0xffc40030
+#define MSTPCR1		0xffc40034
+
+static struct clk mstp_clks[] = {
+	/* MSTPCR0 */
+	SH_CLK_MSTP32("scif_fck", 5, &div4_clks[DIV4_P], MSTPCR0, 29, 0),
+	SH_CLK_MSTP32("scif_fck", 4, &div4_clks[DIV4_P], MSTPCR0, 28, 0),
+	SH_CLK_MSTP32("scif_fck", 3, &div4_clks[DIV4_P], MSTPCR0, 27, 0),
+	SH_CLK_MSTP32("scif_fck", 2, &div4_clks[DIV4_P], MSTPCR0, 26, 0),
+	SH_CLK_MSTP32("scif_fck", 1, &div4_clks[DIV4_P], MSTPCR0, 25, 0),
+	SH_CLK_MSTP32("scif_fck", 0, &div4_clks[DIV4_P], MSTPCR0, 24, 0),
+	SH_CLK_MSTP32("ssi_fck", 3, &div4_clks[DIV4_P], MSTPCR0, 23, 0),
+	SH_CLK_MSTP32("ssi_fck", 2, &div4_clks[DIV4_P], MSTPCR0, 22, 0),
+	SH_CLK_MSTP32("ssi_fck", 1, &div4_clks[DIV4_P], MSTPCR0, 21, 0),
+	SH_CLK_MSTP32("ssi_fck", 0, &div4_clks[DIV4_P], MSTPCR0, 20, 0),
+	SH_CLK_MSTP32("hac_fck", 1, &div4_clks[DIV4_P], MSTPCR0, 17, 0),
+	SH_CLK_MSTP32("hac_fck", 0, &div4_clks[DIV4_P], MSTPCR0, 16, 0),
+	SH_CLK_MSTP32("i2c_fck", 1, &div4_clks[DIV4_P], MSTPCR0, 15, 0),
+	SH_CLK_MSTP32("i2c_fck", 0, &div4_clks[DIV4_P], MSTPCR0, 14, 0),
+	SH_CLK_MSTP32("tmu9_11_fck", -1, &div4_clks[DIV4_P], MSTPCR0, 11, 0),
+	SH_CLK_MSTP32("tmu678_fck", -1, &div4_clks[DIV4_P], MSTPCR0, 10, 0),
+	SH_CLK_MSTP32("tmu345_fck", -1, &div4_clks[DIV4_P], MSTPCR0, 9, 0),
+	SH_CLK_MSTP32("tmu012_fck", -1, &div4_clks[DIV4_P], MSTPCR0, 8, 0),
+	SH_CLK_MSTP32("sdif_fck", 1, &div4_clks[DIV4_P], MSTPCR0, 5, 0),
+	SH_CLK_MSTP32("sdif_fck", 0, &div4_clks[DIV4_P], MSTPCR0, 4, 0),
+	SH_CLK_MSTP32("hspi_fck", -1, &div4_clks[DIV4_P], MSTPCR0, 2, 0),
+
+	/* MSTPCR1 */
+	SH_CLK_MSTP32("usb_fck", -1, NULL, MSTPCR1, 12, 0),
+	SH_CLK_MSTP32("pcie_fck", 2, NULL, MSTPCR1, 10, 0),
+	SH_CLK_MSTP32("pcie_fck", 1, NULL, MSTPCR1, 9, 0),
+	SH_CLK_MSTP32("pcie_fck", 0, NULL, MSTPCR1, 8, 0),
+	SH_CLK_MSTP32("dmac_11_6_fck", -1, NULL, MSTPCR1, 5, 0),
+	SH_CLK_MSTP32("dmac_5_0_fck", -1, NULL, MSTPCR1, 4, 0),
+	SH_CLK_MSTP32("du_fck", -1, NULL, MSTPCR1, 3, 0),
+	SH_CLK_MSTP32("ether_fck", -1, NULL, MSTPCR1, 2, 0),
 };
 
 int __init arch_clk_init(void)
 {
-	struct clk *clk;
 	int i, ret = 0;
 
-	cpg_clk_init();
-
-	clk = clk_get(NULL, "master_clk");
-	for (i = 0; i < ARRAY_SIZE(sh7786_onchip_clocks); i++) {
-		struct clk *clkp = sh7786_onchip_clocks[i];
-
-		clkp->parent = clk;
-		ret |= clk_register(clkp);
-	}
+	for (i = 0; i < ARRAY_SIZE(clks); i++)
+		ret |= clk_register(clks[i]);
 
-	clk_put(clk);
+	if (!ret)
+		ret = sh_clk_div4_register(div4_clks, ARRAY_SIZE(div4_clks),
+					   &div4_table);
+	if (!ret)
+		ret = sh_clk_mstp32_register(mstp_clks, ARRAY_SIZE(mstp_clks));
 
 	return ret;
 }
diff --git a/arch/sh/kernel/cpu/sh4a/clock-shx3.c b/arch/sh/kernel/cpu/sh4a/clock-shx3.c
index 23c27d32d98..e75c57bdfa5 100644
--- a/arch/sh/kernel/cpu/sh4a/clock-shx3.c
+++ b/arch/sh/kernel/cpu/sh4a/clock-shx3.c
@@ -33,7 +33,7 @@ static int cfc_divisors[] = { 1, 1, 4, 6 };
 
 static void master_clk_init(struct clk *clk)
 {
-	clk->rate *= pfc_divisors[(ctrl_inl(FRQCR) >> PFC_POS) & PFC_MSK];
+	clk->rate *= pfc_divisors[(__raw_readl(FRQCR) >> PFC_POS) & PFC_MSK];
 }
 
 static struct clk_ops shx3_master_clk_ops = {
@@ -42,7 +42,7 @@ static struct clk_ops shx3_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> PFC_POS) & PFC_MSK);
+	int idx = ((__raw_readl(FRQCR) >> PFC_POS) & PFC_MSK);
 	return clk->parent->rate / pfc_divisors[idx];
 }
 
@@ -52,7 +52,7 @@ static struct clk_ops shx3_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> BFC_POS) & BFC_MSK);
+	int idx = ((__raw_readl(FRQCR) >> BFC_POS) & BFC_MSK);
 	return clk->parent->rate / bfc_divisors[idx];
 }
 
@@ -62,7 +62,7 @@ static struct clk_ops shx3_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> IFC_POS) & IFC_MSK);
+	int idx = ((__raw_readl(FRQCR) >> IFC_POS) & IFC_MSK);
 	return clk->parent->rate / ifc_divisors[idx];
 }
 
@@ -85,7 +85,7 @@ void __init arch_init_clk_ops(struct clk_ops **ops, int idx)
 
 static unsigned long shyway_clk_recalc(struct clk *clk)
 {
-	int idx = ((ctrl_inl(FRQCR) >> CFC_POS) & CFC_MSK);
+	int idx = ((__raw_readl(FRQCR) >> CFC_POS) & CFC_MSK);
 	return clk->parent->rate / cfc_divisors[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh4a/pinmux-sh7722.c b/arch/sh/kernel/cpu/sh4a/pinmux-sh7722.c
index cb9d07bd59f..0688a7502f8 100644
--- a/arch/sh/kernel/cpu/sh4a/pinmux-sh7722.c
+++ b/arch/sh/kernel/cpu/sh4a/pinmux-sh7722.c
@@ -278,6 +278,7 @@ enum {
 	HIZA8_LCDC, HIZA8_HIZ,
 	HIZA7_LCDC, HIZA7_HIZ,
 	HIZA6_LCDC, HIZA6_HIZ,
+	HIZB4_SIUA, HIZB4_HIZ,
 	HIZB1_VIO, HIZB1_HIZ,
 	HIZB0_VIO, HIZB0_HIZ,
 	HIZC15_IRQ7, HIZC15_HIZ,
@@ -546,7 +547,7 @@ static pinmux_enum_t pinmux_data[] = {
 	PINMUX_DATA(VIO_VD2_MARK, PSE3_VIO, MSELB9_VIO2,
 		    HIZB0_VIO, FOE_VIO_VD2),
 	PINMUX_DATA(VIO_HD2_MARK, PSE3_VIO, MSELB9_VIO2,
-		    HIZB1_VIO, HIZB1_VIO, FCE_VIO_HD2),
+		    HIZB1_VIO, FCE_VIO_HD2),
 	PINMUX_DATA(VIO_CLK2_MARK, PSE3_VIO, MSELB9_VIO2,
 		    HIZB1_VIO, FRB_VIO_CLK2),
 
@@ -658,14 +659,14 @@ static pinmux_enum_t pinmux_data[] = {
 	PINMUX_DATA(SDHICLK_MARK, SDHICLK),
 
 	/* SIU - Port A */
-	PINMUX_DATA(SIUAOLR_MARK, PSC13_SIUAOLR, SIUAOLR_SIOF1_SYNC),
-	PINMUX_DATA(SIUAOBT_MARK, PSC14_SIUAOBT, SIUAOBT_SIOF1_SCK),
-	PINMUX_DATA(SIUAISLD_MARK, PSC15_SIUAISLD, SIUAISLD_SIOF1_RXD),
-	PINMUX_DATA(SIUAILR_MARK, PSC11_SIUAILR, SIUAILR_SIOF1_SS2),
-	PINMUX_DATA(SIUAIBT_MARK, PSC12_SIUAIBT, SIUAIBT_SIOF1_SS1),
-	PINMUX_DATA(SIUAOSLD_MARK, PSB0_SIUAOSLD, SIUAOSLD_SIOF1_TXD),
-	PINMUX_DATA(SIUMCKA_MARK, PSE11_SIUMCKA_SIOF1_MCK, PSB1_SIUMCKA, PTK0),
-	PINMUX_DATA(SIUFCKA_MARK, PSE11_SIUFCKA, PTK0),
+	PINMUX_DATA(SIUAOLR_MARK, PSC13_SIUAOLR, HIZB4_SIUA, SIUAOLR_SIOF1_SYNC),
+	PINMUX_DATA(SIUAOBT_MARK, PSC14_SIUAOBT, HIZB4_SIUA, SIUAOBT_SIOF1_SCK),
+	PINMUX_DATA(SIUAISLD_MARK, PSC15_SIUAISLD, HIZB4_SIUA, SIUAISLD_SIOF1_RXD),
+	PINMUX_DATA(SIUAILR_MARK, PSC11_SIUAILR, HIZB4_SIUA, SIUAILR_SIOF1_SS2),
+	PINMUX_DATA(SIUAIBT_MARK, PSC12_SIUAIBT, HIZB4_SIUA, SIUAIBT_SIOF1_SS1),
+	PINMUX_DATA(SIUAOSLD_MARK, PSB0_SIUAOSLD, HIZB4_SIUA, SIUAOSLD_SIOF1_TXD),
+	PINMUX_DATA(SIUMCKA_MARK, PSE11_SIUMCKA_SIOF1_MCK, HIZB4_SIUA, PSB1_SIUMCKA, PTK0),
+	PINMUX_DATA(SIUFCKA_MARK, PSE11_SIUFCKA, HIZB4_SIUA, PTK0),
 
 	/* SIU - Port B */
 	PINMUX_DATA(SIUBOLR_MARK, PSB11_SIUBOLR, SIOSTRB1_SIUBOLR),
@@ -1612,7 +1613,7 @@ static struct pinmux_cfg_reg pinmux_config_regs[] = {
 		0, 0,
 		0, 0,
 		0, 0,
-		0, 0,
+		HIZB4_SIUA, HIZB4_HIZ,
 		0, 0,
 		0, 0,
 		HIZB1_VIO, HIZB1_HIZ,
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7722.c b/arch/sh/kernel/cpu/sh4a/setup-sh7722.c
index b5335b5e309..fd7e3639e84 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7722.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7722.c
@@ -7,19 +7,167 @@
  * License.  See the file "COPYING" in the main directory of this archive
  * for more details.
  */
-#include <linux/platform_device.h>
 #include <linux/init.h>
+#include <linux/mm.h>
+#include <linux/platform_device.h>
 #include <linux/serial.h>
 #include <linux/serial_sci.h>
-#include <linux/mm.h>
+#include <linux/sh_timer.h>
 #include <linux/uio_driver.h>
 #include <linux/usb/m66592.h>
-#include <linux/sh_timer.h>
+
 #include <asm/clock.h>
+#include <asm/dmaengine.h>
 #include <asm/mmzone.h>
-#include <asm/dma-sh.h>
+#include <asm/siu.h>
+
+#include <cpu/dma-register.h>
 #include <cpu/sh7722.h>
 
+static struct sh_dmae_slave_config sh7722_dmae_slaves[] = {
+	{
+		.slave_id	= SHDMA_SLAVE_SCIF0_TX,
+		.addr		= 0xffe0000c,
+		.chcr		= DM_FIX | SM_INC | 0x800 | TS_INDEX2VAL(XMIT_SZ_8BIT),
+		.mid_rid	= 0x21,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SCIF0_RX,
+		.addr		= 0xffe00014,
+		.chcr		= DM_INC | SM_FIX | 0x800 | TS_INDEX2VAL(XMIT_SZ_8BIT),
+		.mid_rid	= 0x22,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SCIF1_TX,
+		.addr		= 0xffe1000c,
+		.chcr		= DM_FIX | SM_INC | 0x800 | TS_INDEX2VAL(XMIT_SZ_8BIT),
+		.mid_rid	= 0x25,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SCIF1_RX,
+		.addr		= 0xffe10014,
+		.chcr		= DM_INC | SM_FIX | 0x800 | TS_INDEX2VAL(XMIT_SZ_8BIT),
+		.mid_rid	= 0x26,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SCIF2_TX,
+		.addr		= 0xffe2000c,
+		.chcr		= DM_FIX | SM_INC | 0x800 | TS_INDEX2VAL(XMIT_SZ_8BIT),
+		.mid_rid	= 0x29,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SCIF2_RX,
+		.addr		= 0xffe20014,
+		.chcr		= DM_INC | SM_FIX | 0x800 | TS_INDEX2VAL(XMIT_SZ_8BIT),
+		.mid_rid	= 0x2a,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SIUA_TX,
+		.addr		= 0xa454c098,
+		.chcr		= DM_FIX | SM_INC | 0x800 | TS_INDEX2VAL(XMIT_SZ_32BIT),
+		.mid_rid	= 0xb1,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SIUA_RX,
+		.addr		= 0xa454c090,
+		.chcr		= DM_INC | SM_FIX | 0x800 | TS_INDEX2VAL(XMIT_SZ_32BIT),
+		.mid_rid	= 0xb2,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SIUB_TX,
+		.addr		= 0xa454c09c,
+		.chcr		= DM_FIX | SM_INC | 0x800 | TS_INDEX2VAL(XMIT_SZ_32BIT),
+		.mid_rid	= 0xb5,
+	}, {
+		.slave_id	= SHDMA_SLAVE_SIUB_RX,
+		.addr		= 0xa454c094,
+		.chcr		= DM_INC | SM_FIX | 0x800 | TS_INDEX2VAL(XMIT_SZ_32BIT),
+		.mid_rid	= 0xb6,
+	},
+};
+
+static struct sh_dmae_channel sh7722_dmae_channels[] = {
+	{
+		.offset = 0,
+		.dmars = 0,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x10,
+		.dmars = 0,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x20,
+		.dmars = 4,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x30,
+		.dmars = 4,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x50,
+		.dmars = 8,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x60,
+		.dmars = 8,
+		.dmars_bit = 8,
+	}
+};
+
+static unsigned int ts_shift[] = TS_SHIFT;
+
+static struct sh_dmae_pdata dma_platform_data = {
+	.slave		= sh7722_dmae_slaves,
+	.slave_num	= ARRAY_SIZE(sh7722_dmae_slaves),
+	.channel	= sh7722_dmae_channels,
+	.channel_num	= ARRAY_SIZE(sh7722_dmae_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
+};
+
+static struct resource sh7722_dmae_resources[] = {
+	[0] = {
+		/* Channel registers and DMAOR */
+		.start	= 0xfe008020,
+		.end	= 0xfe00808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	[1] = {
+		/* DMARSx */
+		.start	= 0xfe009000,
+		.end	= 0xfe00900b,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* DMA error IRQ */
+		.start	= 78,
+		.end	= 78,
+		.flags	= IORESOURCE_IRQ,
+	},
+	{
+		/* IRQ for channels 0-3 */
+		.start	= 48,
+		.end	= 51,
+		.flags	= IORESOURCE_IRQ,
+	},
+	{
+		/* IRQ for channels 4-5 */
+		.start	= 76,
+		.end	= 77,
+		.flags	= IORESOURCE_IRQ,
+	},
+};
+
+struct platform_device dma_device = {
+	.name		= "sh-dma-engine",
+	.id		= -1,
+	.resource	= sh7722_dmae_resources,
+	.num_resources	= ARRAY_SIZE(sh7722_dmae_resources),
+	.dev		= {
+		.platform_data	= &dma_platform_data,
+	},
+	.archdata = {
+		.hwblk_id = HWBLK_DMAC,
+	},
+};
+
 /* Serial */
 static struct plat_sci_port scif0_platform_data = {
 	.mapbase        = 0xffe00000,
@@ -388,15 +536,36 @@ static struct platform_device tmu2_device = {
 	},
 };
 
-static struct sh_dmae_pdata dma_platform_data = {
-	.mode = 0,
+static struct siu_platform siu_platform_data = {
+	.dma_dev	= &dma_device.dev,
+	.dma_slave_tx_a	= SHDMA_SLAVE_SIUA_TX,
+	.dma_slave_rx_a	= SHDMA_SLAVE_SIUA_RX,
+	.dma_slave_tx_b	= SHDMA_SLAVE_SIUB_TX,
+	.dma_slave_rx_b	= SHDMA_SLAVE_SIUB_RX,
 };
 
-static struct platform_device dma_device = {
-	.name		= "sh-dma-engine",
+static struct resource siu_resources[] = {
+	[0] = {
+		.start	= 0xa4540000,
+		.end	= 0xa454c10f,
+		.flags	= IORESOURCE_MEM,
+	},
+	[1] = {
+		.start	= 108,
+		.flags	= IORESOURCE_IRQ,
+	},
+};
+
+static struct platform_device siu_device = {
+	.name		= "sh_siu",
 	.id		= -1,
-	.dev		= {
-		.platform_data	= &dma_platform_data,
+	.dev = {
+		.platform_data	= &siu_platform_data,
+	},
+	.resource	= siu_resources,
+	.num_resources	= ARRAY_SIZE(siu_resources),
+	.archdata = {
+		.hwblk_id = HWBLK_SIU,
 	},
 };
 
@@ -414,6 +583,7 @@ static struct platform_device *sh7722_devices[] __initdata = {
 	&vpu_device,
 	&veu_device,
 	&jpu_device,
+	&siu_device,
 	&dma_device,
 };
 
@@ -446,6 +616,8 @@ void __init plat_early_device_setup(void)
 
 enum {
 	UNUSED=0,
+	ENABLED,
+	DISABLED,
 
 	/* interrupt sources */
 	IRQ0, IRQ1, IRQ2, IRQ3, IRQ4, IRQ5, IRQ6, IRQ7,
@@ -461,7 +633,6 @@ enum {
 	SCIF0, SCIF1, SCIF2, SIOF0, SIOF1, SIO,
 	FLCTL_FLSTEI, FLCTL_FLENDI, FLCTL_FLTREQ0I, FLCTL_FLTREQ1I,
 	I2C_ALI, I2C_TACKI, I2C_WAITI, I2C_DTEI,
-	SDHI0, SDHI1, SDHI2, SDHI3,
 	CMT, TSIF, SIU, TWODG,
 	TMU0, TMU1, TMU2,
 	IRDA, JPU, LCDC,
@@ -494,8 +665,8 @@ static struct intc_vect vectors[] __initdata = {
 	INTC_VECT(FLCTL_FLTREQ0I, 0xdc0), INTC_VECT(FLCTL_FLTREQ1I, 0xde0),
 	INTC_VECT(I2C_ALI, 0xe00), INTC_VECT(I2C_TACKI, 0xe20),
 	INTC_VECT(I2C_WAITI, 0xe40), INTC_VECT(I2C_DTEI, 0xe60),
-	INTC_VECT(SDHI0, 0xe80), INTC_VECT(SDHI1, 0xea0),
-	INTC_VECT(SDHI2, 0xec0), INTC_VECT(SDHI3, 0xee0),
+	INTC_VECT(SDHI, 0xe80), INTC_VECT(SDHI, 0xea0),
+	INTC_VECT(SDHI, 0xec0), INTC_VECT(SDHI, 0xee0),
 	INTC_VECT(CMT, 0xf00), INTC_VECT(TSIF, 0xf20),
 	INTC_VECT(SIU, 0xf80), INTC_VECT(TWODG, 0xfa0),
 	INTC_VECT(TMU0, 0x400), INTC_VECT(TMU1, 0x420),
@@ -513,7 +684,6 @@ static struct intc_group groups[] __initdata = {
 	INTC_GROUP(FLCTL, FLCTL_FLSTEI, FLCTL_FLENDI,
 		   FLCTL_FLTREQ0I, FLCTL_FLTREQ1I),
 	INTC_GROUP(I2C, I2C_ALI, I2C_TACKI, I2C_WAITI, I2C_DTEI),
-	INTC_GROUP(SDHI, SDHI0, SDHI1, SDHI2, SDHI3),
 };
 
 static struct intc_mask_reg mask_registers[] __initdata = {
@@ -535,7 +705,7 @@ static struct intc_mask_reg mask_registers[] __initdata = {
 	  { I2C_DTEI, I2C_WAITI, I2C_TACKI, I2C_ALI,
 	    FLCTL_FLTREQ1I, FLCTL_FLTREQ0I, FLCTL_FLENDI, FLCTL_FLSTEI } },
 	{ 0xa40800a0, 0xa40800e0, 8, /* IMR8 / IMCR8 */
-	  { SDHI3, SDHI2, SDHI1, SDHI0, 0, 0, TWODG, SIU } },
+	  { DISABLED, DISABLED, ENABLED, ENABLED, 0, 0, TWODG, SIU } },
 	{ 0xa40800a4, 0xa40800e4, 8, /* IMR9 / IMCR9 */
 	  { 0, 0, 0, CMT, 0, USB_USBI1, USB_USBI0, } },
 	{ 0xa40800a8, 0xa40800e8, 8, /* IMR10 / IMCR10 */
@@ -573,9 +743,13 @@ static struct intc_mask_reg ack_registers[] __initdata = {
 	  { IRQ0, IRQ1, IRQ2, IRQ3, IRQ4, IRQ5, IRQ6, IRQ7 } },
 };
 
-static DECLARE_INTC_DESC_ACK(intc_desc, "sh7722", vectors, groups,
-			     mask_registers, prio_registers, sense_registers,
-			     ack_registers);
+static struct intc_desc intc_desc __initdata = {
+	.name = "sh7722",
+	.force_enable = ENABLED,
+	.force_disable = DISABLED,
+	.hw = INTC_HW_DESC(vectors, groups, mask_registers,
+			   prio_registers, sense_registers, ack_registers),
+};
 
 void __init plat_irq_setup(void)
 {
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7723.c b/arch/sh/kernel/cpu/sh4a/setup-sh7723.c
index 772b9265d0e..85c61f62470 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7723.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7723.c
@@ -592,14 +592,17 @@ void __init plat_early_device_setup(void)
 #define RAMCR_CACHE_L2FC	0x0002
 #define RAMCR_CACHE_L2E		0x0001
 #define L2_CACHE_ENABLE		(RAMCR_CACHE_L2E|RAMCR_CACHE_L2FC)
-void __uses_jump_to_uncached l2_cache_init(void)
+
+void l2_cache_init(void)
 {
 	/* Enable L2 cache */
-	ctrl_outl(L2_CACHE_ENABLE, RAMCR);
+	__raw_writel(L2_CACHE_ENABLE, RAMCR);
 }
 
 enum {
 	UNUSED=0,
+	ENABLED,
+	DISABLED,
 
 	/* interrupt sources */
 	IRQ0, IRQ1, IRQ2, IRQ3, IRQ4, IRQ5, IRQ6, IRQ7,
@@ -622,7 +625,6 @@ enum {
 	SCIFA_SCIFA1,
 	FLCTL_FLSTEI,FLCTL_FLTENDI,FLCTL_FLTREQ0I,FLCTL_FLTREQ1I,
 	I2C_ALI,I2C_TACKI,I2C_WAITI,I2C_DTEI,
-	SDHI0_SDHII0,SDHI0_SDHII1,SDHI0_SDHII2,
 	CMT_CMTI,
 	TSIF_TSIFI,
 	SIU_SIUI,
@@ -630,7 +632,6 @@ enum {
 	TMU0_TUNI0, TMU0_TUNI1, TMU0_TUNI2,
 	IRDA_IRDAI,
 	ATAPI_ATAPII,
-	SDHI1_SDHII0,SDHI1_SDHII1,SDHI1_SDHII2,
 	VEU2H1_VEU2HI,
 	LCDC_LCDCI,
 	TMU1_TUNI0,TMU1_TUNI1,TMU1_TUNI2,
@@ -701,9 +702,9 @@ static struct intc_vect vectors[] __initdata = {
 	INTC_VECT(I2C_WAITI,0xE40),
 	INTC_VECT(I2C_DTEI,0xE60),
 
-	INTC_VECT(SDHI0_SDHII0,0xE80),
-	INTC_VECT(SDHI0_SDHII1,0xEA0),
-	INTC_VECT(SDHI0_SDHII2,0xEC0),
+	INTC_VECT(SDHI0, 0xE80),
+	INTC_VECT(SDHI0, 0xEA0),
+	INTC_VECT(SDHI0, 0xEC0),
 
 	INTC_VECT(CMT_CMTI,0xF00),
 	INTC_VECT(TSIF_TSIFI,0xF20),
@@ -717,9 +718,9 @@ static struct intc_vect vectors[] __initdata = {
 	INTC_VECT(IRDA_IRDAI,0x480),
 	INTC_VECT(ATAPI_ATAPII,0x4A0),
 
-	INTC_VECT(SDHI1_SDHII0,0x4E0),
-	INTC_VECT(SDHI1_SDHII1,0x500),
-	INTC_VECT(SDHI1_SDHII2,0x520),
+	INTC_VECT(SDHI1, 0x4E0),
+	INTC_VECT(SDHI1, 0x500),
+	INTC_VECT(SDHI1, 0x520),
 
 	INTC_VECT(VEU2H1_VEU2HI,0x560),
 	INTC_VECT(LCDC_LCDCI,0x580),
@@ -738,15 +739,14 @@ static struct intc_group groups[] __initdata = {
 	INTC_GROUP(FLCTL,FLCTL_FLSTEI,FLCTL_FLTENDI,FLCTL_FLTREQ0I,FLCTL_FLTREQ1I),
 	INTC_GROUP(I2C,I2C_ALI,I2C_TACKI,I2C_WAITI,I2C_DTEI),
 	INTC_GROUP(_2DG, _2DG_TRI,_2DG_INI,_2DG_CEI),
-	INTC_GROUP(SDHI1, SDHI1_SDHII0,SDHI1_SDHII1,SDHI1_SDHII2),
 	INTC_GROUP(RTC, RTC_ATI,RTC_PRI,RTC_CUI),
 	INTC_GROUP(DMAC1B, DMAC1B_DEI4,DMAC1B_DEI5,DMAC1B_DADERR),
-	INTC_GROUP(SDHI0,SDHI0_SDHII0,SDHI0_SDHII1,SDHI0_SDHII2),
 };
 
 static struct intc_mask_reg mask_registers[] __initdata = {
 	{ 0xa4080080, 0xa40800c0, 8, /* IMR0 / IMCR0 */
-	  { 0,  TMU1_TUNI2,TMU1_TUNI1,TMU1_TUNI0,0,SDHI1_SDHII2,SDHI1_SDHII1,SDHI1_SDHII0} },
+	  { 0, TMU1_TUNI2, TMU1_TUNI1, TMU1_TUNI0,
+	    0, DISABLED, ENABLED, ENABLED } },
 	{ 0xa4080084, 0xa40800c4, 8, /* IMR1 / IMCR1 */
 	  { VIO_VOUI, VIO_VEU2HI,VIO_BEUI,VIO_CEUI,DMAC0A_DEI3,DMAC0A_DEI2,DMAC0A_DEI1,DMAC0A_DEI0 } },
 	{ 0xa4080088, 0xa40800c8, 8, /* IMR2 / IMCR2 */
@@ -763,7 +763,8 @@ static struct intc_mask_reg mask_registers[] __initdata = {
 	  { I2C_DTEI, I2C_WAITI, I2C_TACKI, I2C_ALI,
 	    FLCTL_FLTREQ1I, FLCTL_FLTREQ0I, FLCTL_FLTENDI, FLCTL_FLSTEI } },
 	{ 0xa40800a0, 0xa40800e0, 8, /* IMR8 / IMCR8 */
-	  { 0,SDHI0_SDHII2,SDHI0_SDHII1,SDHI0_SDHII0,0,0,SCIFA_SCIFA2,SIU_SIUI } },
+	  { 0, DISABLED, ENABLED, ENABLED,
+	    0, 0, SCIFA_SCIFA2, SIU_SIUI } },
 	{ 0xa40800a4, 0xa40800e4, 8, /* IMR9 / IMCR9 */
 	  { 0, 0, 0, CMT_CMTI, 0, 0, USB_USI0,0 } },
 	{ 0xa40800a8, 0xa40800e8, 8, /* IMR10 / IMCR10 */
@@ -803,9 +804,13 @@ static struct intc_mask_reg ack_registers[] __initdata = {
 	  { IRQ0, IRQ1, IRQ2, IRQ3, IRQ4, IRQ5, IRQ6, IRQ7 } },
 };
 
-static DECLARE_INTC_DESC_ACK(intc_desc, "sh7723", vectors, groups,
-			     mask_registers, prio_registers, sense_registers,
-			     ack_registers);
+static struct intc_desc intc_desc __initdata = {
+	.name = "sh7723",
+	.force_enable = ENABLED,
+	.force_disable = DISABLED,
+	.hw = INTC_HW_DESC(vectors, groups, mask_registers,
+			   prio_registers, sense_registers, ack_registers),
+};
 
 void __init plat_irq_setup(void)
 {
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7724.c b/arch/sh/kernel/cpu/sh4a/setup-sh7724.c
index d32f96c1cc1..e7fa2a92fc1 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7724.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7724.c
@@ -21,22 +21,189 @@
 #include <linux/sh_timer.h>
 #include <linux/io.h>
 #include <linux/notifier.h>
+
 #include <asm/suspend.h>
 #include <asm/clock.h>
-#include <asm/dma-sh.h>
+#include <asm/dmaengine.h>
 #include <asm/mmzone.h>
+
+#include <cpu/dma-register.h>
 #include <cpu/sh7724.h>
 
 /* DMA */
-static struct sh_dmae_pdata dma_platform_data = {
-	.mode = SHDMA_DMAOR1,
+static struct sh_dmae_channel sh7724_dmae0_channels[] = {
+	{
+		.offset = 0,
+		.dmars = 0,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x10,
+		.dmars = 0,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x20,
+		.dmars = 4,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x30,
+		.dmars = 4,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x50,
+		.dmars = 8,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x60,
+		.dmars = 8,
+		.dmars_bit = 8,
+	}
+};
+
+static struct sh_dmae_channel sh7724_dmae1_channels[] = {
+	{
+		.offset = 0,
+		.dmars = 0,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x10,
+		.dmars = 0,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x20,
+		.dmars = 4,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x30,
+		.dmars = 4,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x50,
+		.dmars = 8,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x60,
+		.dmars = 8,
+		.dmars_bit = 8,
+	}
+};
+
+static unsigned int ts_shift[] = TS_SHIFT;
+
+static struct sh_dmae_pdata dma0_platform_data = {
+	.channel	= sh7724_dmae0_channels,
+	.channel_num	= ARRAY_SIZE(sh7724_dmae0_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
+};
+
+static struct sh_dmae_pdata dma1_platform_data = {
+	.channel	= sh7724_dmae1_channels,
+	.channel_num	= ARRAY_SIZE(sh7724_dmae1_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
+};
+
+/* Resource order important! */
+static struct resource sh7724_dmae0_resources[] = {
+	{
+		/* Channel registers and DMAOR */
+		.start	= 0xfe008020,
+		.end	= 0xfe00808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* DMARSx */
+		.start	= 0xfe009000,
+		.end	= 0xfe00900b,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* DMA error IRQ */
+		.start	= 78,
+		.end	= 78,
+		.flags	= IORESOURCE_IRQ,
+	},
+	{
+		/* IRQ for channels 0-3 */
+		.start	= 48,
+		.end	= 51,
+		.flags	= IORESOURCE_IRQ,
+	},
+	{
+		/* IRQ for channels 4-5 */
+		.start	= 76,
+		.end	= 77,
+		.flags	= IORESOURCE_IRQ,
+	},
 };
 
-static struct platform_device dma_device = {
-	.name	= "sh-dma-engine",
-	.id		= -1,
-	.dev	= {
-		.platform_data	= &dma_platform_data,
+/* Resource order important! */
+static struct resource sh7724_dmae1_resources[] = {
+	{
+		/* Channel registers and DMAOR */
+		.start	= 0xfdc08020,
+		.end	= 0xfdc0808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* DMARSx */
+		.start	= 0xfdc09000,
+		.end	= 0xfdc0900b,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* DMA error IRQ */
+		.start	= 74,
+		.end	= 74,
+		.flags	= IORESOURCE_IRQ,
+	},
+	{
+		/* IRQ for channels 0-3 */
+		.start	= 40,
+		.end	= 43,
+		.flags	= IORESOURCE_IRQ,
+	},
+	{
+		/* IRQ for channels 4-5 */
+		.start	= 72,
+		.end	= 73,
+		.flags	= IORESOURCE_IRQ,
+	},
+};
+
+static struct platform_device dma0_device = {
+	.name		= "sh-dma-engine",
+	.id		= 0,
+	.resource	= sh7724_dmae0_resources,
+	.num_resources	= ARRAY_SIZE(sh7724_dmae0_resources),
+	.dev		= {
+		.platform_data	= &dma0_platform_data,
+	},
+	.archdata = {
+		.hwblk_id = HWBLK_DMAC0,
+	},
+};
+
+static struct platform_device dma1_device = {
+	.name		= "sh-dma-engine",
+	.id		= 1,
+	.resource	= sh7724_dmae1_resources,
+	.num_resources	= ARRAY_SIZE(sh7724_dmae1_resources),
+	.dev		= {
+		.platform_data	= &dma1_platform_data,
+	},
+	.archdata = {
+		.hwblk_id = HWBLK_DMAC1,
 	},
 };
 
@@ -663,7 +830,8 @@ static struct platform_device *sh7724_devices[] __initdata = {
 	&tmu3_device,
 	&tmu4_device,
 	&tmu5_device,
-	&dma_device,
+	&dma0_device,
+	&dma1_device,
 	&rtc_device,
 	&iic0_device,
 	&iic1_device,
@@ -714,14 +882,17 @@ void __init plat_early_device_setup(void)
 #define RAMCR_CACHE_L2FC	0x0002
 #define RAMCR_CACHE_L2E		0x0001
 #define L2_CACHE_ENABLE		(RAMCR_CACHE_L2E|RAMCR_CACHE_L2FC)
-void __uses_jump_to_uncached l2_cache_init(void)
+
+void l2_cache_init(void)
 {
 	/* Enable L2 cache */
-	ctrl_outl(L2_CACHE_ENABLE, RAMCR);
+	__raw_writel(L2_CACHE_ENABLE, RAMCR);
 }
 
 enum {
 	UNUSED = 0,
+	ENABLED,
+	DISABLED,
 
 	/* interrupt sources */
 	IRQ0, IRQ1, IRQ2, IRQ3, IRQ4, IRQ5, IRQ6, IRQ7,
@@ -750,14 +921,12 @@ enum {
 	ETHI,
 	I2C1_ALI, I2C1_TACKI, I2C1_WAITI, I2C1_DTEI,
 	I2C0_ALI, I2C0_TACKI, I2C0_WAITI, I2C0_DTEI,
-	SDHI0_SDHII0, SDHI0_SDHII1, SDHI0_SDHII2, SDHI0_SDHII3,
 	CMT,
 	TSIF,
 	FSI,
 	SCIFA5,
 	TMU0_TUNI0, TMU0_TUNI1, TMU0_TUNI2,
 	IRDA,
-	SDHI1_SDHII0, SDHI1_SDHII1, SDHI1_SDHII2,
 	JPU,
 	_2DDMAC,
 	MMC_MMC2I, MMC_MMC3I,
@@ -839,10 +1008,10 @@ static struct intc_vect vectors[] __initdata = {
 	INTC_VECT(I2C0_WAITI, 0xE40),
 	INTC_VECT(I2C0_DTEI, 0xE60),
 
-	INTC_VECT(SDHI0_SDHII0, 0xE80),
-	INTC_VECT(SDHI0_SDHII1, 0xEA0),
-	INTC_VECT(SDHI0_SDHII2, 0xEC0),
-	INTC_VECT(SDHI0_SDHII3, 0xEE0),
+	INTC_VECT(SDHI0, 0xE80),
+	INTC_VECT(SDHI0, 0xEA0),
+	INTC_VECT(SDHI0, 0xEC0),
+	INTC_VECT(SDHI0, 0xEE0),
 
 	INTC_VECT(CMT,    0xF00),
 	INTC_VECT(TSIF,   0xF20),
@@ -855,9 +1024,9 @@ static struct intc_vect vectors[] __initdata = {
 
 	INTC_VECT(IRDA,    0x480),
 
-	INTC_VECT(SDHI1_SDHII0, 0x4E0),
-	INTC_VECT(SDHI1_SDHII1, 0x500),
-	INTC_VECT(SDHI1_SDHII2, 0x520),
+	INTC_VECT(SDHI1, 0x4E0),
+	INTC_VECT(SDHI1, 0x500),
+	INTC_VECT(SDHI1, 0x520),
 
 	INTC_VECT(JPU, 0x560),
 	INTC_VECT(_2DDMAC, 0x4A0),
@@ -883,8 +1052,6 @@ static struct intc_group groups[] __initdata = {
 	INTC_GROUP(DMAC0B, DMAC0B_DEI4, DMAC0B_DEI5, DMAC0B_DADERR),
 	INTC_GROUP(I2C0, I2C0_ALI, I2C0_TACKI, I2C0_WAITI, I2C0_DTEI),
 	INTC_GROUP(I2C1, I2C1_ALI, I2C1_TACKI, I2C1_WAITI, I2C1_DTEI),
-	INTC_GROUP(SDHI0, SDHI0_SDHII0, SDHI0_SDHII1, SDHI0_SDHII2, SDHI0_SDHII3),
-	INTC_GROUP(SDHI1, SDHI1_SDHII0, SDHI1_SDHII1, SDHI1_SDHII2),
 	INTC_GROUP(SPU, SPU_SPUI0, SPU_SPUI1),
 	INTC_GROUP(MMCIF, MMC_MMC2I, MMC_MMC3I),
 };
@@ -892,7 +1059,7 @@ static struct intc_group groups[] __initdata = {
 static struct intc_mask_reg mask_registers[] __initdata = {
 	{ 0xa4080080, 0xa40800c0, 8, /* IMR0 / IMCR0 */
 	  { 0, TMU1_TUNI2, TMU1_TUNI1, TMU1_TUNI0,
-	    0, SDHI1_SDHII2, SDHI1_SDHII1, SDHI1_SDHII0 } },
+	    0, DISABLED, ENABLED, ENABLED } },
 	{ 0xa4080084, 0xa40800c4, 8, /* IMR1 / IMCR1 */
 	  { VIO_VOU, VIO_VEU1, VIO_BEU0, VIO_CEU0,
 	    DMAC0A_DEI3, DMAC0A_DEI2, DMAC0A_DEI1, DMAC0A_DEI0 } },
@@ -914,7 +1081,7 @@ static struct intc_mask_reg mask_registers[] __initdata = {
 	  { I2C0_DTEI, I2C0_WAITI, I2C0_TACKI, I2C0_ALI,
 	    I2C1_DTEI, I2C1_WAITI, I2C1_TACKI, I2C1_ALI } },
 	{ 0xa40800a0, 0xa40800e0, 8, /* IMR8 / IMCR8 */
-	  { SDHI0_SDHII3, SDHI0_SDHII2, SDHI0_SDHII1, SDHI0_SDHII0,
+	  { DISABLED, DISABLED, ENABLED, ENABLED,
 	    0, 0, SCIFA5, FSI } },
 	{ 0xa40800a4, 0xa40800e4, 8, /* IMR9 / IMCR9 */
 	  { 0, 0, 0, CMT, 0, USB1, USB0, 0 } },
@@ -961,9 +1128,13 @@ static struct intc_mask_reg ack_registers[] __initdata = {
 	  { IRQ0, IRQ1, IRQ2, IRQ3, IRQ4, IRQ5, IRQ6, IRQ7 } },
 };
 
-static DECLARE_INTC_DESC_ACK(intc_desc, "sh7724", vectors, groups,
-			     mask_registers, prio_registers, sense_registers,
-			     ack_registers);
+static struct intc_desc intc_desc __initdata = {
+	.name = "sh7724",
+	.force_enable = ENABLED,
+	.force_disable = DISABLED,
+	.hw = INTC_HW_DESC(vectors, groups, mask_registers,
+			   prio_registers, sense_registers, ack_registers),
+};
 
 void __init plat_irq_setup(void)
 {
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7757.c b/arch/sh/kernel/cpu/sh4a/setup-sh7757.c
index 37e32efbbaa..e75edf58796 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7757.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7757.c
@@ -487,17 +487,17 @@ static DECLARE_INTC_DESC(intc_desc_irl4567, "sh7757-irl4567", vectors_irl4567,
 void __init plat_irq_setup(void)
 {
 	/* disable IRQ3-0 + IRQ7-4 */
-	ctrl_outl(0xff000000, INTC_INTMSK0);
+	__raw_writel(0xff000000, INTC_INTMSK0);
 
 	/* disable IRL3-0 + IRL7-4 */
-	ctrl_outl(0xc0000000, INTC_INTMSK1);
-	ctrl_outl(0xfffefffe, INTC_INTMSK2);
+	__raw_writel(0xc0000000, INTC_INTMSK1);
+	__raw_writel(0xfffefffe, INTC_INTMSK2);
 
 	/* select IRL mode for IRL3-0 + IRL7-4 */
-	ctrl_outl(ctrl_inl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
 
 	/* disable holding function, ie enable "SH-4 Mode" */
-	ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00200000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) | 0x00200000, INTC_ICR0);
 
 	register_intc_controller(&intc_desc);
 }
@@ -507,32 +507,32 @@ void __init plat_irq_setup_pins(int mode)
 	switch (mode) {
 	case IRQ_MODE_IRQ7654:
 		/* select IRQ mode for IRL7-4 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00400000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00400000, INTC_ICR0);
 		register_intc_controller(&intc_desc_irq4567);
 		break;
 	case IRQ_MODE_IRQ3210:
 		/* select IRQ mode for IRL3-0 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00800000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00800000, INTC_ICR0);
 		register_intc_controller(&intc_desc_irq0123);
 		break;
 	case IRQ_MODE_IRL7654:
 		/* enable IRL7-4 but don't provide any masking */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
-		ctrl_outl(0x0000fffe, INTC_INTMSKCLR2);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x0000fffe, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL3210:
 		/* enable IRL0-3 but don't provide any masking */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
-		ctrl_outl(0xfffe0000, INTC_INTMSKCLR2);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0xfffe0000, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL7654_MASK:
 		/* enable IRL7-4 and mask using cpu intc controller */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_desc_irl4567);
 		break;
 	case IRQ_MODE_IRL3210_MASK:
 		/* enable IRL0-3 and mask using cpu intc controller */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_desc_irl0123);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7763.c b/arch/sh/kernel/cpu/sh4a/setup-sh7763.c
index 6aba26fec41..7f6b0a5f7f8 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7763.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7763.c
@@ -538,11 +538,11 @@ static DECLARE_INTC_DESC(intc_irl3210_desc, "sh7763-irl3210", irl_vectors,
 void __init plat_irq_setup(void)
 {
 	/* disable IRQ7-0 */
-	ctrl_outl(0xff000000, INTC_INTMSK0);
+	__raw_writel(0xff000000, INTC_INTMSK0);
 
 	/* disable IRL3-0 + IRL7-4 */
-	ctrl_outl(0xc0000000, INTC_INTMSK1);
-	ctrl_outl(0xfffefffe, INTC_INTMSK2);
+	__raw_writel(0xc0000000, INTC_INTMSK1);
+	__raw_writel(0xfffefffe, INTC_INTMSK2);
 
 	register_intc_controller(&intc_desc);
 }
@@ -552,27 +552,27 @@ void __init plat_irq_setup_pins(int mode)
 	switch (mode) {
 	case IRQ_MODE_IRQ:
 		/* select IRQ mode for IRL3-0 + IRL7-4 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00c00000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00c00000, INTC_ICR0);
 		register_intc_controller(&intc_irq_desc);
 		break;
 	case IRQ_MODE_IRL7654:
 		/* enable IRL7-4 but don't provide any masking */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
-		ctrl_outl(0x0000fffe, INTC_INTMSKCLR2);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x0000fffe, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL3210:
 		/* enable IRL0-3 but don't provide any masking */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
-		ctrl_outl(0xfffe0000, INTC_INTMSKCLR2);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0xfffe0000, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL7654_MASK:
 		/* enable IRL7-4 and mask using cpu intc controller */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_irl7654_desc);
 		break;
 	case IRQ_MODE_IRL3210_MASK:
 		/* enable IRL0-3 and mask using cpu intc controller */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_irl3210_desc);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7770.c b/arch/sh/kernel/cpu/sh4a/setup-sh7770.c
index c1643bc9590..86d681ecf90 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7770.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7770.c
@@ -694,17 +694,17 @@ static DECLARE_INTC_DESC(intc_irl3210_desc, "sh7780-irl3210", irl_vectors,
 void __init plat_irq_setup(void)
 {
 	/* disable IRQ7-0 */
-	ctrl_outl(0xff000000, INTC_INTMSK0);
+	__raw_writel(0xff000000, INTC_INTMSK0);
 
 	/* disable IRL3-0 + IRL7-4 */
-	ctrl_outl(0xc0000000, INTC_INTMSK1);
-	ctrl_outl(0xfffefffe, INTC_INTMSK2);
+	__raw_writel(0xc0000000, INTC_INTMSK1);
+	__raw_writel(0xfffefffe, INTC_INTMSK2);
 
 	/* select IRL mode for IRL3-0 + IRL7-4 */
-	ctrl_outl(ctrl_inl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
 
 	/* disable holding function, ie enable "SH-4 Mode" */
-	ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00200000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) | 0x00200000, INTC_ICR0);
 
 	register_intc_controller(&intc_desc);
 }
@@ -714,27 +714,27 @@ void __init plat_irq_setup_pins(int mode)
 	switch (mode) {
 	case IRQ_MODE_IRQ:
 		/* select IRQ mode for IRL3-0 + IRL7-4 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00c00000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00c00000, INTC_ICR0);
 		register_intc_controller(&intc_irq_desc);
 		break;
 	case IRQ_MODE_IRL7654:
 		/* enable IRL7-4 but don't provide any masking */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
-		ctrl_outl(0x0000fffe, INTC_INTMSKCLR2);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x0000fffe, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL3210:
 		/* enable IRL0-3 but don't provide any masking */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
-		ctrl_outl(0xfffe0000, INTC_INTMSKCLR2);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0xfffe0000, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL7654_MASK:
 		/* enable IRL7-4 and mask using cpu intc controller */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_irl7654_desc);
 		break;
 	case IRQ_MODE_IRL3210_MASK:
 		/* enable IRL0-3 and mask using cpu intc controller */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_irl3210_desc);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7780.c b/arch/sh/kernel/cpu/sh4a/setup-sh7780.c
index c310558490d..02e792c90de 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7780.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7780.c
@@ -13,7 +13,10 @@
 #include <linux/io.h>
 #include <linux/serial_sci.h>
 #include <linux/sh_timer.h>
-#include <asm/dma-sh.h>
+
+#include <asm/dmaengine.h>
+
+#include <cpu/dma-register.h>
 
 static struct plat_sci_port scif0_platform_data = {
 	.mapbase	= 0xffe00000,
@@ -247,15 +250,131 @@ static struct platform_device rtc_device = {
 	.resource	= rtc_resources,
 };
 
-static struct sh_dmae_pdata dma_platform_data = {
-	.mode = (SHDMA_MIX_IRQ | SHDMA_DMAOR1),
+/* DMA */
+static struct sh_dmae_channel sh7780_dmae0_channels[] = {
+	{
+		.offset = 0,
+		.dmars = 0,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x10,
+		.dmars = 0,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x20,
+		.dmars = 4,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x30,
+		.dmars = 4,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x50,
+		.dmars = 8,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x60,
+		.dmars = 8,
+		.dmars_bit = 8,
+	}
+};
+
+static struct sh_dmae_channel sh7780_dmae1_channels[] = {
+	{
+		.offset = 0,
+	}, {
+		.offset = 0x10,
+	}, {
+		.offset = 0x20,
+	}, {
+		.offset = 0x30,
+	}, {
+		.offset = 0x50,
+	}, {
+		.offset = 0x60,
+	}
+};
+
+static unsigned int ts_shift[] = TS_SHIFT;
+
+static struct sh_dmae_pdata dma0_platform_data = {
+	.channel	= sh7780_dmae0_channels,
+	.channel_num	= ARRAY_SIZE(sh7780_dmae0_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
+};
+
+static struct sh_dmae_pdata dma1_platform_data = {
+	.channel	= sh7780_dmae1_channels,
+	.channel_num	= ARRAY_SIZE(sh7780_dmae1_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
 };
 
-static struct platform_device dma_device = {
+static struct resource sh7780_dmae0_resources[] = {
+	[0] = {
+		/* Channel registers and DMAOR */
+		.start	= 0xfc808020,
+		.end	= 0xfc80808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	[1] = {
+		/* DMARSx */
+		.start	= 0xfc809000,
+		.end	= 0xfc80900b,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* Real DMA error IRQ is 38, and channel IRQs are 34-37, 44-45 */
+		.start	= 34,
+		.end	= 34,
+		.flags	= IORESOURCE_IRQ | IORESOURCE_IRQ_SHAREABLE,
+	},
+};
+
+static struct resource sh7780_dmae1_resources[] = {
+	[0] = {
+		/* Channel registers and DMAOR */
+		.start	= 0xfc818020,
+		.end	= 0xfc81808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	/* DMAC1 has no DMARS */
+	{
+		/* Real DMA error IRQ is 38, and channel IRQs are 46-47, 92-95 */
+		.start	= 46,
+		.end	= 46,
+		.flags	= IORESOURCE_IRQ | IORESOURCE_IRQ_SHAREABLE,
+	},
+};
+
+static struct platform_device dma0_device = {
 	.name           = "sh-dma-engine",
-	.id             = -1,
+	.id             = 0,
+	.resource	= sh7780_dmae0_resources,
+	.num_resources	= ARRAY_SIZE(sh7780_dmae0_resources),
 	.dev            = {
-		.platform_data  = &dma_platform_data,
+		.platform_data	= &dma0_platform_data,
+	},
+};
+
+static struct platform_device dma1_device = {
+	.name		= "sh-dma-engine",
+	.id		= 1,
+	.resource	= sh7780_dmae1_resources,
+	.num_resources	= ARRAY_SIZE(sh7780_dmae1_resources),
+	.dev		= {
+		.platform_data	= &dma1_platform_data,
 	},
 };
 
@@ -269,7 +388,8 @@ static struct platform_device *sh7780_devices[] __initdata = {
 	&tmu4_device,
 	&tmu5_device,
 	&rtc_device,
-	&dma_device,
+	&dma0_device,
+	&dma1_device,
 };
 
 static int __init sh7780_devices_setup(void)
@@ -461,17 +581,17 @@ static DECLARE_INTC_DESC(intc_irl3210_desc, "sh7780-irl3210", irl_vectors,
 void __init plat_irq_setup(void)
 {
 	/* disable IRQ7-0 */
-	ctrl_outl(0xff000000, INTC_INTMSK0);
+	__raw_writel(0xff000000, INTC_INTMSK0);
 
 	/* disable IRL3-0 + IRL7-4 */
-	ctrl_outl(0xc0000000, INTC_INTMSK1);
-	ctrl_outl(0xfffefffe, INTC_INTMSK2);
+	__raw_writel(0xc0000000, INTC_INTMSK1);
+	__raw_writel(0xfffefffe, INTC_INTMSK2);
 
 	/* select IRL mode for IRL3-0 + IRL7-4 */
-	ctrl_outl(ctrl_inl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
 
 	/* disable holding function, ie enable "SH-4 Mode" */
-	ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00200000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) | 0x00200000, INTC_ICR0);
 
 	register_intc_controller(&intc_desc);
 }
@@ -481,27 +601,27 @@ void __init plat_irq_setup_pins(int mode)
 	switch (mode) {
 	case IRQ_MODE_IRQ:
 		/* select IRQ mode for IRL3-0 + IRL7-4 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00c00000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00c00000, INTC_ICR0);
 		register_intc_controller(&intc_irq_desc);
 		break;
 	case IRQ_MODE_IRL7654:
 		/* enable IRL7-4 but don't provide any masking */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
-		ctrl_outl(0x0000fffe, INTC_INTMSKCLR2);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x0000fffe, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL3210:
 		/* enable IRL0-3 but don't provide any masking */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
-		ctrl_outl(0xfffe0000, INTC_INTMSKCLR2);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0xfffe0000, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL7654_MASK:
 		/* enable IRL7-4 and mask using cpu intc controller */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_irl7654_desc);
 		break;
 	case IRQ_MODE_IRL3210_MASK:
 		/* enable IRL0-3 and mask using cpu intc controller */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_irl3210_desc);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7785.c b/arch/sh/kernel/cpu/sh4a/setup-sh7785.c
index f685b9b2199..1fcd88b1671 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7785.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7785.c
@@ -14,9 +14,12 @@
 #include <linux/io.h>
 #include <linux/mm.h>
 #include <linux/sh_timer.h>
-#include <asm/dma-sh.h>
+
+#include <asm/dmaengine.h>
 #include <asm/mmzone.h>
 
+#include <cpu/dma-register.h>
+
 static struct plat_sci_port scif0_platform_data = {
 	.mapbase	= 0xffea0000,
 	.flags		= UPF_BOOT_AUTOCONF,
@@ -295,15 +298,131 @@ static struct platform_device tmu5_device = {
 	.num_resources	= ARRAY_SIZE(tmu5_resources),
 };
 
-static struct sh_dmae_pdata dma_platform_data = {
-	.mode = (SHDMA_MIX_IRQ | SHDMA_DMAOR1),
+/* DMA */
+static struct sh_dmae_channel sh7785_dmae0_channels[] = {
+	{
+		.offset = 0,
+		.dmars = 0,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x10,
+		.dmars = 0,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x20,
+		.dmars = 4,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x30,
+		.dmars = 4,
+		.dmars_bit = 8,
+	}, {
+		.offset = 0x50,
+		.dmars = 8,
+		.dmars_bit = 0,
+	}, {
+		.offset = 0x60,
+		.dmars = 8,
+		.dmars_bit = 8,
+	}
+};
+
+static struct sh_dmae_channel sh7785_dmae1_channels[] = {
+	{
+		.offset = 0,
+	}, {
+		.offset = 0x10,
+	}, {
+		.offset = 0x20,
+	}, {
+		.offset = 0x30,
+	}, {
+		.offset = 0x50,
+	}, {
+		.offset = 0x60,
+	}
+};
+
+static unsigned int ts_shift[] = TS_SHIFT;
+
+static struct sh_dmae_pdata dma0_platform_data = {
+	.channel	= sh7785_dmae0_channels,
+	.channel_num	= ARRAY_SIZE(sh7785_dmae0_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
+};
+
+static struct sh_dmae_pdata dma1_platform_data = {
+	.channel	= sh7785_dmae1_channels,
+	.channel_num	= ARRAY_SIZE(sh7785_dmae1_channels),
+	.ts_low_shift	= CHCR_TS_LOW_SHIFT,
+	.ts_low_mask	= CHCR_TS_LOW_MASK,
+	.ts_high_shift	= CHCR_TS_HIGH_SHIFT,
+	.ts_high_mask	= CHCR_TS_HIGH_MASK,
+	.ts_shift	= ts_shift,
+	.ts_shift_num	= ARRAY_SIZE(ts_shift),
+	.dmaor_init	= DMAOR_INIT,
 };
 
-static struct platform_device dma_device = {
+static struct resource sh7785_dmae0_resources[] = {
+	[0] = {
+		/* Channel registers and DMAOR */
+		.start	= 0xfc808020,
+		.end	= 0xfc80808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	[1] = {
+		/* DMARSx */
+		.start	= 0xfc809000,
+		.end	= 0xfc80900b,
+		.flags	= IORESOURCE_MEM,
+	},
+	{
+		/* Real DMA error IRQ is 39, and channel IRQs are 33-38 */
+		.start	= 33,
+		.end	= 33,
+		.flags	= IORESOURCE_IRQ | IORESOURCE_IRQ_SHAREABLE,
+	},
+};
+
+static struct resource sh7785_dmae1_resources[] = {
+	[0] = {
+		/* Channel registers and DMAOR */
+		.start	= 0xfcc08020,
+		.end	= 0xfcc0808f,
+		.flags	= IORESOURCE_MEM,
+	},
+	/* DMAC1 has no DMARS */
+	{
+		/* Real DMA error IRQ is 58, and channel IRQs are 52-57 */
+		.start	= 52,
+		.end	= 52,
+		.flags	= IORESOURCE_IRQ | IORESOURCE_IRQ_SHAREABLE,
+	},
+};
+
+static struct platform_device dma0_device = {
 	.name           = "sh-dma-engine",
-	.id             = -1,
+	.id             = 0,
+	.resource	= sh7785_dmae0_resources,
+	.num_resources	= ARRAY_SIZE(sh7785_dmae0_resources),
 	.dev            = {
-		.platform_data  = &dma_platform_data,
+		.platform_data	= &dma0_platform_data,
+	},
+};
+
+static struct platform_device dma1_device = {
+	.name		= "sh-dma-engine",
+	.id		= 1,
+	.resource	= sh7785_dmae1_resources,
+	.num_resources	= ARRAY_SIZE(sh7785_dmae1_resources),
+	.dev		= {
+		.platform_data	= &dma1_platform_data,
 	},
 };
 
@@ -320,7 +439,8 @@ static struct platform_device *sh7785_devices[] __initdata = {
 	&tmu3_device,
 	&tmu4_device,
 	&tmu5_device,
-	&dma_device,
+	&dma0_device,
+	&dma1_device,
 };
 
 static int __init sh7785_devices_setup(void)
@@ -541,17 +661,17 @@ static DECLARE_INTC_DESC(intc_desc_irl4567, "sh7785-irl4567", vectors_irl4567,
 void __init plat_irq_setup(void)
 {
 	/* disable IRQ3-0 + IRQ7-4 */
-	ctrl_outl(0xff000000, INTC_INTMSK0);
+	__raw_writel(0xff000000, INTC_INTMSK0);
 
 	/* disable IRL3-0 + IRL7-4 */
-	ctrl_outl(0xc0000000, INTC_INTMSK1);
-	ctrl_outl(0xfffefffe, INTC_INTMSK2);
+	__raw_writel(0xc0000000, INTC_INTMSK1);
+	__raw_writel(0xfffefffe, INTC_INTMSK2);
 
 	/* select IRL mode for IRL3-0 + IRL7-4 */
-	ctrl_outl(ctrl_inl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
 
 	/* disable holding function, ie enable "SH-4 Mode" */
-	ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00200000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) | 0x00200000, INTC_ICR0);
 
 	register_intc_controller(&intc_desc);
 }
@@ -561,32 +681,32 @@ void __init plat_irq_setup_pins(int mode)
 	switch (mode) {
 	case IRQ_MODE_IRQ7654:
 		/* select IRQ mode for IRL7-4 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00400000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00400000, INTC_ICR0);
 		register_intc_controller(&intc_desc_irq4567);
 		break;
 	case IRQ_MODE_IRQ3210:
 		/* select IRQ mode for IRL3-0 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00800000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00800000, INTC_ICR0);
 		register_intc_controller(&intc_desc_irq0123);
 		break;
 	case IRQ_MODE_IRL7654:
 		/* enable IRL7-4 but don't provide any masking */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
-		ctrl_outl(0x0000fffe, INTC_INTMSKCLR2);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x0000fffe, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL3210:
 		/* enable IRL0-3 but don't provide any masking */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
-		ctrl_outl(0xfffe0000, INTC_INTMSKCLR2);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0xfffe0000, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL7654_MASK:
 		/* enable IRL7-4 and mask using cpu intc controller */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_desc_irl4567);
 		break;
 	case IRQ_MODE_IRL3210_MASK:
 		/* enable IRL0-3 and mask using cpu intc controller */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_desc_irl0123);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4a/setup-sh7786.c b/arch/sh/kernel/cpu/sh4a/setup-sh7786.c
index 71673487ace..7e585320710 100644
--- a/arch/sh/kernel/cpu/sh4a/setup-sh7786.c
+++ b/arch/sh/kernel/cpu/sh4a/setup-sh7786.c
@@ -867,14 +867,14 @@ static DECLARE_INTC_DESC(intc_desc_irl4567, "sh7786-irl4567", vectors_irl4567,
 void __init plat_irq_setup(void)
 {
 	/* disable IRQ3-0 + IRQ7-4 */
-	ctrl_outl(0xff000000, INTC_INTMSK0);
+	__raw_writel(0xff000000, INTC_INTMSK0);
 
 	/* disable IRL3-0 + IRL7-4 */
-	ctrl_outl(0xc0000000, INTC_INTMSK1);
-	ctrl_outl(0xfffefffe, INTC_INTMSK2);
+	__raw_writel(0xc0000000, INTC_INTMSK1);
+	__raw_writel(0xfffefffe, INTC_INTMSK2);
 
 	/* select IRL mode for IRL3-0 + IRL7-4 */
-	ctrl_outl(ctrl_inl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
+	__raw_writel(__raw_readl(INTC_ICR0) & ~0x00c00000, INTC_ICR0);
 
 	register_intc_controller(&intc_desc);
 }
@@ -884,32 +884,32 @@ void __init plat_irq_setup_pins(int mode)
 	switch (mode) {
 	case IRQ_MODE_IRQ7654:
 		/* select IRQ mode for IRL7-4 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00400000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00400000, INTC_ICR0);
 		register_intc_controller(&intc_desc_irq4567);
 		break;
 	case IRQ_MODE_IRQ3210:
 		/* select IRQ mode for IRL3-0 */
-		ctrl_outl(ctrl_inl(INTC_ICR0) | 0x00800000, INTC_ICR0);
+		__raw_writel(__raw_readl(INTC_ICR0) | 0x00800000, INTC_ICR0);
 		register_intc_controller(&intc_desc_irq0123);
 		break;
 	case IRQ_MODE_IRL7654:
 		/* enable IRL7-4 but don't provide any masking */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
-		ctrl_outl(0x0000fffe, INTC_INTMSKCLR2);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x0000fffe, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL3210:
 		/* enable IRL0-3 but don't provide any masking */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
-		ctrl_outl(0xfffe0000, INTC_INTMSKCLR2);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0xfffe0000, INTC_INTMSKCLR2);
 		break;
 	case IRQ_MODE_IRL7654_MASK:
 		/* enable IRL7-4 and mask using cpu intc controller */
-		ctrl_outl(0x40000000, INTC_INTMSKCLR1);
+		__raw_writel(0x40000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_desc_irl4567);
 		break;
 	case IRQ_MODE_IRL3210_MASK:
 		/* enable IRL0-3 and mask using cpu intc controller */
-		ctrl_outl(0x80000000, INTC_INTMSKCLR1);
+		__raw_writel(0x80000000, INTC_INTMSKCLR1);
 		register_intc_controller(&intc_desc_irl0123);
 		break;
 	default:
diff --git a/arch/sh/kernel/cpu/sh4a/smp-shx3.c b/arch/sh/kernel/cpu/sh4a/smp-shx3.c
index 5863e0c4d02..11bf4c1e25c 100644
--- a/arch/sh/kernel/cpu/sh4a/smp-shx3.c
+++ b/arch/sh/kernel/cpu/sh4a/smp-shx3.c
@@ -78,7 +78,10 @@ void __init plat_prepare_cpus(unsigned int max_cpus)
 
 void plat_start_cpu(unsigned int cpu, unsigned long entry_point)
 {
-	__raw_writel(entry_point, RESET_REG(cpu));
+	if (__in_29bit_mode())
+		__raw_writel(entry_point, RESET_REG(cpu));
+	else
+		__raw_writel(virt_to_phys(entry_point), RESET_REG(cpu));
 
 	if (!(__raw_readl(STBCR_REG(cpu)) & STBCR_MSTP))
 		__raw_writel(STBCR_MSTP, STBCR_REG(cpu));
diff --git a/arch/sh/kernel/cpu/sh4a/ubc.c b/arch/sh/kernel/cpu/sh4a/ubc.c
new file mode 100644
index 00000000000..efb2745bcb3
--- /dev/null
+++ b/arch/sh/kernel/cpu/sh4a/ubc.c
@@ -0,0 +1,133 @@
+/*
+ * arch/sh/kernel/cpu/sh4a/ubc.c
+ *
+ * On-chip UBC support for SH-4A CPUs.
+ *
+ * Copyright (C) 2009 - 2010  Paul Mundt
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ */
+#include <linux/init.h>
+#include <linux/err.h>
+#include <linux/clk.h>
+#include <linux/io.h>
+#include <asm/hw_breakpoint.h>
+
+#define UBC_CBR(idx)	(0xff200000 + (0x20 * idx))
+#define UBC_CRR(idx)	(0xff200004 + (0x20 * idx))
+#define UBC_CAR(idx)	(0xff200008 + (0x20 * idx))
+#define UBC_CAMR(idx)	(0xff20000c + (0x20 * idx))
+
+#define UBC_CCMFR	0xff200600
+#define UBC_CBCR	0xff200620
+
+/* CRR */
+#define UBC_CRR_PCB	(1 << 1)
+#define UBC_CRR_BIE	(1 << 0)
+
+/* CBR */
+#define UBC_CBR_CE	(1 << 0)
+
+static struct sh_ubc sh4a_ubc;
+
+static void sh4a_ubc_enable(struct arch_hw_breakpoint *info, int idx)
+{
+	__raw_writel(UBC_CBR_CE | info->len | info->type, UBC_CBR(idx));
+	__raw_writel(info->address, UBC_CAR(idx));
+}
+
+static void sh4a_ubc_disable(struct arch_hw_breakpoint *info, int idx)
+{
+	__raw_writel(0, UBC_CBR(idx));
+	__raw_writel(0, UBC_CAR(idx));
+}
+
+static void sh4a_ubc_enable_all(unsigned long mask)
+{
+	int i;
+
+	for (i = 0; i < sh4a_ubc.num_events; i++)
+		if (mask & (1 << i))
+			__raw_writel(__raw_readl(UBC_CBR(i)) | UBC_CBR_CE,
+				     UBC_CBR(i));
+}
+
+static void sh4a_ubc_disable_all(void)
+{
+	int i;
+
+	for (i = 0; i < sh4a_ubc.num_events; i++)
+		__raw_writel(__raw_readl(UBC_CBR(i)) & ~UBC_CBR_CE,
+			     UBC_CBR(i));
+}
+
+static unsigned long sh4a_ubc_active_mask(void)
+{
+	unsigned long active = 0;
+	int i;
+
+	for (i = 0; i < sh4a_ubc.num_events; i++)
+		if (__raw_readl(UBC_CBR(i)) & UBC_CBR_CE)
+			active |= (1 << i);
+
+	return active;
+}
+
+static unsigned long sh4a_ubc_triggered_mask(void)
+{
+	return __raw_readl(UBC_CCMFR);
+}
+
+static void sh4a_ubc_clear_triggered_mask(unsigned long mask)
+{
+	__raw_writel(__raw_readl(UBC_CCMFR) & ~mask, UBC_CCMFR);
+}
+
+static struct sh_ubc sh4a_ubc = {
+	.name			= "SH-4A",
+	.num_events		= 2,
+	.trap_nr		= 0x1e0,
+	.enable			= sh4a_ubc_enable,
+	.disable		= sh4a_ubc_disable,
+	.enable_all		= sh4a_ubc_enable_all,
+	.disable_all		= sh4a_ubc_disable_all,
+	.active_mask		= sh4a_ubc_active_mask,
+	.triggered_mask		= sh4a_ubc_triggered_mask,
+	.clear_triggered_mask	= sh4a_ubc_clear_triggered_mask,
+};
+
+static int __init sh4a_ubc_init(void)
+{
+	struct clk *ubc_iclk = clk_get(NULL, "ubc0");
+	int i;
+
+	/*
+	 * The UBC MSTP bit is optional, as not all platforms will have
+	 * it. Just ignore it if we can't find it.
+	 */
+	if (IS_ERR(ubc_iclk))
+		ubc_iclk = NULL;
+
+	clk_enable(ubc_iclk);
+
+	__raw_writel(0, UBC_CBCR);
+
+	for (i = 0; i < sh4a_ubc.num_events; i++) {
+		__raw_writel(0, UBC_CAMR(i));
+		__raw_writel(0, UBC_CBR(i));
+
+		__raw_writel(UBC_CRR_BIE | UBC_CRR_PCB, UBC_CRR(i));
+
+		/* dummy read for write posting */
+		(void)__raw_readl(UBC_CRR(i));
+	}
+
+	clk_disable(ubc_iclk);
+
+	sh4a_ubc.clk = ubc_iclk;
+
+	return register_sh_ubc(&sh4a_ubc);
+}
+arch_initcall(sh4a_ubc_init);
diff --git a/arch/sh/kernel/cpu/sh5/clock-sh5.c b/arch/sh/kernel/cpu/sh5/clock-sh5.c
index 7f864ebc51d..9cfc19b8dbe 100644
--- a/arch/sh/kernel/cpu/sh5/clock-sh5.c
+++ b/arch/sh/kernel/cpu/sh5/clock-sh5.c
@@ -24,7 +24,7 @@ static unsigned long cprc_base;
 
 static void master_clk_init(struct clk *clk)
 {
-	int idx = (ctrl_inl(cprc_base + 0x00) >> 6) & 0x0007;
+	int idx = (__raw_readl(cprc_base + 0x00) >> 6) & 0x0007;
 	clk->rate *= ifc_table[idx];
 }
 
@@ -34,7 +34,7 @@ static struct clk_ops sh5_master_clk_ops = {
 
 static unsigned long module_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(cprc_base) >> 12) & 0x0007;
+	int idx = (__raw_readw(cprc_base) >> 12) & 0x0007;
 	return clk->parent->rate / ifc_table[idx];
 }
 
@@ -44,7 +44,7 @@ static struct clk_ops sh5_module_clk_ops = {
 
 static unsigned long bus_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(cprc_base) >> 3) & 0x0007;
+	int idx = (__raw_readw(cprc_base) >> 3) & 0x0007;
 	return clk->parent->rate / ifc_table[idx];
 }
 
@@ -54,7 +54,7 @@ static struct clk_ops sh5_bus_clk_ops = {
 
 static unsigned long cpu_clk_recalc(struct clk *clk)
 {
-	int idx = (ctrl_inw(cprc_base) & 0x0007);
+	int idx = (__raw_readw(cprc_base) & 0x0007);
 	return clk->parent->rate / ifc_table[idx];
 }
 
diff --git a/arch/sh/kernel/cpu/sh5/entry.S b/arch/sh/kernel/cpu/sh5/entry.S
index 8f13f73cb2c..6b80295dd7a 100644
--- a/arch/sh/kernel/cpu/sh5/entry.S
+++ b/arch/sh/kernel/cpu/sh5/entry.S
@@ -187,7 +187,7 @@ trap_jtable:
 	.rept 6
 		.long do_exception_error	/* 0x880 - 0x920 */
 	.endr
-	.long	do_software_break_point	/* 0x940 */
+	.long	breakpoint_trap_handler	/* 0x940 */
 	.long	do_exception_error		/* 0x960 */
 	.long	do_single_step		/* 0x980 */
 
@@ -1124,7 +1124,7 @@ fpu_error_or_IRQA:
 	pta	its_IRQ, tr0
 	beqi/l	r4, EVENT_INTERRUPT, tr0
 #ifdef CONFIG_SH_FPU
-	movi	do_fpu_state_restore, r6
+	movi	fpu_state_restore_trap_handler, r6
 #else
 	movi	do_exception_error, r6
 #endif
@@ -1135,7 +1135,7 @@ fpu_error_or_IRQB:
 	pta	its_IRQ, tr0
 	beqi/l	r4, EVENT_INTERRUPT, tr0
 #ifdef CONFIG_SH_FPU
-	movi	do_fpu_state_restore, r6
+	movi	fpu_state_restore_trap_handler, r6
 #else
 	movi	do_exception_error, r6
 #endif
diff --git a/arch/sh/kernel/cpu/sh5/fpu.c b/arch/sh/kernel/cpu/sh5/fpu.c
index 4648ccee6c4..4b3bb35e99f 100644
--- a/arch/sh/kernel/cpu/sh5/fpu.c
+++ b/arch/sh/kernel/cpu/sh5/fpu.c
@@ -15,24 +15,6 @@
 #include <linux/sched.h>
 #include <linux/signal.h>
 #include <asm/processor.h>
-#include <asm/user.h>
-#include <asm/io.h>
-#include <asm/fpu.h>
-
-/*
- * Initially load the FPU with signalling NANS.  This bit pattern
- * has the property that no matter whether considered as single or as
- * double precision, it still represents a signalling NAN.
- */
-#define sNAN64		0xFFFFFFFFFFFFFFFFULL
-#define sNAN32		0xFFFFFFFFUL
-
-static union sh_fpu_union init_fpuregs = {
-	.hard = {
-		.fp_regs = { [0 ... 63] = sNAN32 },
-		.fpscr = FPSCR_INIT
-	}
-};
 
 void save_fpu(struct task_struct *tsk)
 {
@@ -72,12 +54,11 @@ void save_fpu(struct task_struct *tsk)
 		     "fgetscr   fr63\n\t"
 		     "fst.s     %0, (32*8), fr63\n\t"
 		: /* no output */
-		: "r" (&tsk->thread.fpu.hard)
+		: "r" (&tsk->thread.xstate->hardfpu)
 		: "memory");
 }
 
-static inline void
-fpload(struct sh_fpu_hard_struct *fpregs)
+void restore_fpu(struct task_struct *tsk)
 {
 	asm volatile("fld.p     %0, (0*8), fp0\n\t"
 		     "fld.p     %0, (1*8), fp2\n\t"
@@ -116,16 +97,11 @@ fpload(struct sh_fpu_hard_struct *fpregs)
 
 		     "fld.p     %0, (31*8), fp62\n\t"
 		: /* no output */
-		: "r" (fpregs) );
-}
-
-void fpinit(struct sh_fpu_hard_struct *fpregs)
-{
-	*fpregs = init_fpuregs.hard;
+		: "r" (&tsk->thread.xstate->hardfpu)
+		: "memory");
 }
 
-asmlinkage void
-do_fpu_error(unsigned long ex, struct pt_regs *regs)
+asmlinkage void do_fpu_error(unsigned long ex, struct pt_regs *regs)
 {
 	struct task_struct *tsk = current;
 
@@ -133,35 +109,6 @@ do_fpu_error(unsigned long ex, struct pt_regs *regs)
 
 	tsk->thread.trap_no = 11;
 	tsk->thread.error_code = 0;
-	force_sig(SIGFPE, tsk);
-}
-
-
-asmlinkage void
-do_fpu_state_restore(unsigned long ex, struct pt_regs *regs)
-{
-	void die(const char *str, struct pt_regs *regs, long err);
-
-	if (! user_mode(regs))
-		die("FPU used in kernel", regs, ex);
 
-	regs->sr &= ~SR_FD;
-
-	if (last_task_used_math == current)
-		return;
-
-	enable_fpu();
-	if (last_task_used_math != NULL)
-		/* Other processes fpu state, save away */
-		save_fpu(last_task_used_math);
-
-        last_task_used_math = current;
-        if (used_math()) {
-                fpload(&current->thread.fpu.hard);
-        } else {
-		/* First time FPU user.  */
-		fpload(&init_fpuregs.hard);
-                set_used_math();
-        }
-	disable_fpu();
+	force_sig(SIGFPE, tsk);
 }
diff --git a/arch/sh/kernel/cpu/shmobile/pm.c b/arch/sh/kernel/cpu/shmobile/pm.c
index ca029a44743..e5596871270 100644
--- a/arch/sh/kernel/cpu/shmobile/pm.c
+++ b/arch/sh/kernel/cpu/shmobile/pm.c
@@ -33,7 +33,8 @@ ATOMIC_NOTIFIER_HEAD(sh_mobile_post_sleep_notifier_list);
 #define SUSP_MODE_SLEEP		(SUSP_SH_SLEEP)
 #define SUSP_MODE_SLEEP_SF	(SUSP_SH_SLEEP | SUSP_SH_SF)
 #define SUSP_MODE_STANDBY_SF	(SUSP_SH_STANDBY | SUSP_SH_SF)
-#define SUSP_MODE_RSTANDBY	(SUSP_SH_RSTANDBY | SUSP_SH_MMU | SUSP_SH_SF)
+#define SUSP_MODE_RSTANDBY_SF \
+	(SUSP_SH_RSTANDBY | SUSP_SH_MMU | SUSP_SH_REGS | SUSP_SH_SF)
  /*
   * U-standby mode is unsupported since it needs bootloader hacks
   */
diff --git a/arch/sh/kernel/cpu/shmobile/sleep.S b/arch/sh/kernel/cpu/shmobile/sleep.S
index e9dd7fa0abd..e6aac65f575 100644
--- a/arch/sh/kernel/cpu/shmobile/sleep.S
+++ b/arch/sh/kernel/cpu/shmobile/sleep.S
@@ -48,8 +48,48 @@ ENTRY(sh_mobile_sleep_enter_start)
 	stc	sr, r0
 	mov.l	r0, @(SH_SLEEP_SR, r5)
 
-	/* save sp */
+	/* save general purpose registers to stack if needed */
+	mov.l	@(SH_SLEEP_MODE, r5), r0
+	tst	#SUSP_SH_REGS, r0
+	bt	skip_regs_save
+
+	sts.l	pr, @-r15
+	mov.l	r14, @-r15
+	mov.l	r13, @-r15
+	mov.l	r12, @-r15
+	mov.l	r11, @-r15
+	mov.l	r10, @-r15
+	mov.l	r9, @-r15
+	mov.l	r8, @-r15
+
+	/* make sure bank0 is selected, save low registers */
+	mov.l	rb_bit, r9
+	not	r9, r9
+	bsr	set_sr
+	 mov	#0, r10
+
+	bsr	save_low_regs
+	 nop
+
+	/* switch to bank 1, save low registers */
+	mov.l	rb_bit, r10
+	bsr	set_sr
+	 mov	#-1, r9
+
+	bsr	save_low_regs
+	 nop
+
+	/* switch back to bank 0 */
+	mov.l	rb_bit, r9
+	not	r9, r9
+	bsr	set_sr
+	 mov	#0, r10
+
+skip_regs_save:
+
+	/* save sp, also set to internal ram */
 	mov.l	r15, @(SH_SLEEP_SP, r5)
+	mov	r5, r15
 
 	/* save stbcr */
 	bsr     save_register
@@ -60,7 +100,7 @@ ENTRY(sh_mobile_sleep_enter_start)
 	tst	#SUSP_SH_MMU, r0
 	bt	skip_mmu_save_disable
 
-       /* save mmu state */
+	/* save mmu state */
 	bsr	save_register
 	 mov	#SH_SLEEP_REG_PTEH, r0
 
@@ -177,6 +217,29 @@ get_register:
 	mov.l	@(r0, r5), r0
 	rts
 	 nop
+
+set_sr:
+	stc	sr, r8
+	and	r9, r8
+	or	r10, r8
+	ldc	r8, sr
+	rts
+	 nop
+
+save_low_regs:
+	mov.l	r7, @-r15
+	mov.l	r6, @-r15
+	mov.l	r5, @-r15
+	mov.l	r4, @-r15
+	mov.l	r3, @-r15
+	mov.l	r2, @-r15
+	mov.l	r1, @-r15
+	rts
+	 mov.l	r0, @-r15
+
+	.balign 4
+rb_bit:	.long	0x20000000 ! RB=1
+
 ENTRY(sh_mobile_sleep_enter_end)
 
 	.balign 4
@@ -270,6 +333,40 @@ skip_restore_sf:
 	icbi	@r0
 
 skip_restore_mmu:
+
+	/* restore general purpose registers if needed */
+	mov.l	@(SH_SLEEP_MODE, r5), r0
+	tst	#SUSP_SH_REGS, r0
+	bt	skip_restore_regs
+
+	/* switch to bank 1, restore low registers */
+	mov.l	_rb_bit, r10
+	bsr	_set_sr
+	 mov	#-1, r9
+
+	bsr	restore_low_regs
+	 nop
+
+	/* switch to bank0, restore low registers */
+	mov.l	_rb_bit, r9
+	not	r9, r9
+	bsr	_set_sr
+	 mov	#0, r10
+
+	bsr	restore_low_regs
+	 nop
+
+	/* restore the rest of the registers */
+	mov.l	@r15+, r8
+	mov.l	@r15+, r9
+	mov.l	@r15+, r10
+	mov.l	@r15+, r11
+	mov.l	@r15+, r12
+	mov.l	@r15+, r13
+	mov.l	@r15+, r14
+	lds.l	@r15+, pr
+
+skip_restore_regs:
 	rte
 	 nop
 
@@ -283,6 +380,26 @@ restore_register:
 	rts
 	 nop
 
+_set_sr:
+	stc	sr, r8
+	and	r9, r8
+	or	r10, r8
+	ldc	r8, sr
+	rts
+	 nop
+
+restore_low_regs:
+	mov.l	@r15+, r0
+	mov.l	@r15+, r1
+	mov.l	@r15+, r2
+	mov.l	@r15+, r3
+	mov.l	@r15+, r4
+	mov.l	@r15+, r5
+	mov.l	@r15+, r6
+	rts
+	 mov.l	@r15+, r7
+
 	.balign 4
+_rb_bit:	.long	0x20000000 ! RB=1
 1:	.long	~0x7ff
 ENTRY(sh_mobile_sleep_resume_end)