diff --git a/arch/sh/mm/cache-sh4.c b/arch/sh/mm/cache-sh4.c
index 524cea5b47f9..94c05d09c3f7 100644
--- a/arch/sh/mm/cache-sh4.c
+++ b/arch/sh/mm/cache-sh4.c
@@ -2,7 +2,7 @@
  * arch/sh/mm/cache-sh4.c
  *
  * Copyright (C) 1999, 2000, 2002  Niibe Yutaka
- * Copyright (C) 2001, 2002, 2003, 2004  Paul Mundt
+ * Copyright (C) 2001, 2002, 2003, 2004, 2005  Paul Mundt
  * Copyright (C) 2003  Richard Curnow
  *
  * This file is subject to the terms and conditions of the GNU General Public
@@ -25,6 +25,8 @@
 #include <asm/mmu_context.h>
 #include <asm/cacheflush.h>
 
+extern void __flush_cache_4096(unsigned long addr, unsigned long phys,
+			       unsigned long exec_offset);
 extern void __flush_cache_4096_all(unsigned long start);
 static void __flush_cache_4096_all_ex(unsigned long start);
 extern void __flush_dcache_all(void);
@@ -112,9 +114,14 @@ static void __flush_dcache_all_ex(void)
 {
 	unsigned long addr, end_addr, entry_offset;
 
-	end_addr = CACHE_OC_ADDRESS_ARRAY + (cpu_data->dcache.sets << cpu_data->dcache.entry_shift) * cpu_data->dcache.ways;
+	end_addr = CACHE_OC_ADDRESS_ARRAY +
+		(cpu_data->dcache.sets << cpu_data->dcache.entry_shift) *
+		 cpu_data->dcache.ways;
+
 	entry_offset = 1 << cpu_data->dcache.entry_shift;
-	for (addr = CACHE_OC_ADDRESS_ARRAY; addr < end_addr; addr += entry_offset) {
+	for (addr = CACHE_OC_ADDRESS_ARRAY;
+	     addr < end_addr;
+	     addr += entry_offset) {
 		ctrl_outl(0, addr);
 	}
 }
@@ -125,7 +132,8 @@ static void __flush_cache_4096_all_ex(unsigned long start)
 	int i;
 
 	entry_offset = 1 << cpu_data->dcache.entry_shift;
-	for (i = 0; i < cpu_data->dcache.ways; i++, start += cpu_data->dcache.way_incr) {
+	for (i = 0; i < cpu_data->dcache.ways;
+	     i++, start += cpu_data->dcache.way_incr) {
 		for (addr = CACHE_OC_ADDRESS_ARRAY + start;
 		     addr < CACHE_OC_ADDRESS_ARRAY + 4096 + start;
 		     addr += entry_offset) {
@@ -153,14 +161,14 @@ void flush_icache_range(unsigned long start, unsigned long end)
 }
 
 /*
- * Write back the D-cache and purge the I-cache for signal trampoline. 
+ * Write back the D-cache and purge the I-cache for signal trampoline.
  * .. which happens to be the same behavior as flush_icache_range().
  * So, we simply flush out a line.
  */
 void flush_cache_sigtramp(unsigned long addr)
 {
 	unsigned long v, index;
-	unsigned long flags; 
+	unsigned long flags;
 	int i;
 
 	v = addr & ~(L1_CACHE_BYTES-1);
@@ -172,7 +180,8 @@ void flush_cache_sigtramp(unsigned long addr)
 
 	local_irq_save(flags);
 	jump_to_P2();
-	for(i = 0; i < cpu_data->icache.ways; i++, index += cpu_data->icache.way_incr)
+	for (i = 0; i < cpu_data->icache.ways;
+	     i++, index += cpu_data->icache.way_incr)
 		ctrl_outl(0, index);	/* Clear out Valid-bit */
 	back_to_P1();
 	local_irq_restore(flags);
@@ -181,8 +190,7 @@ void flush_cache_sigtramp(unsigned long addr)
 static inline void flush_cache_4096(unsigned long start,
 				    unsigned long phys)
 {
-	unsigned long flags; 
-	extern void __flush_cache_4096(unsigned long addr, unsigned long phys, unsigned long exec_offset);
+	unsigned long flags;
 
 	/*
 	 * SH7751, SH7751R, and ST40 have no restriction to handle cache.
@@ -191,10 +199,12 @@ static inline void flush_cache_4096(unsigned long start,
 	if ((cpu_data->flags & CPU_HAS_P2_FLUSH_BUG)
 	   || start < CACHE_OC_ADDRESS_ARRAY) {
 		local_irq_save(flags);
-		__flush_cache_4096(start | SH_CACHE_ASSOC, P1SEGADDR(phys), 0x20000000);
+		__flush_cache_4096(start | SH_CACHE_ASSOC,
+				   P1SEGADDR(phys), 0x20000000);
 		local_irq_restore(flags);
 	} else {
-		__flush_cache_4096(start | SH_CACHE_ASSOC, P1SEGADDR(phys), 0);
+		__flush_cache_4096(start | SH_CACHE_ASSOC,
+				   P1SEGADDR(phys), 0);
 	}
 }
 
@@ -231,29 +241,22 @@ static inline void flush_icache_all(void)
 	local_irq_restore(flags);
 }
 
-void flush_cache_all(void)
+void flush_dcache_all(void)
 {
 	if (cpu_data->dcache.ways == 1)
 		__flush_dcache_all();
 	else
 		__flush_dcache_all_ex();
+}
+
+void flush_cache_all(void)
+{
+	flush_dcache_all();
 	flush_icache_all();
 }
 
 void flush_cache_mm(struct mm_struct *mm)
 {
-	/* Is there any good way? */
-	/* XXX: possibly call flush_cache_range for each vm area */
-	/* 
-	 * FIXME: Really, the optimal solution here would be able to flush out
-	 * individual lines created by the specified context, but this isn't
-	 * feasible for a number of architectures (such as MIPS, and some
-	 * SPARC) .. is this possible for SuperH?
-	 *
-	 * In the meantime, we'll just flush all of the caches.. this
-	 * seems to be the simplest way to avoid at least a few wasted
-	 * cache flushes. -Lethal
-	 */
 	flush_cache_all();
 }
 
@@ -301,13 +304,30 @@ void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 	unsigned long p = start & PAGE_MASK;
 	pgd_t *dir;
 	pmd_t *pmd;
+	pud_t *pud;
 	pte_t *pte;
 	pte_t entry;
 	unsigned long phys;
 	unsigned long d = 0;
 
+	/*
+	 * Don't bother with the lookup and alias check if we have a
+	 * wide range to cover, just blow away the dcache in its
+	 * entirety instead. -- PFM.
+	 */
+	if (((end - start) >> PAGE_SHIFT) >= 64) {
+		flush_dcache_all();
+
+		if (vma->vm_flags & VM_EXEC)
+			flush_icache_all();
+
+		return;
+	}
+
 	dir = pgd_offset(vma->vm_mm, p);
-	pmd = pmd_offset(dir, p);
+	pud = pud_offset(dir, p);
+	pmd = pmd_offset(pud, p);
+	end = PAGE_ALIGN(end);
 
 	do {
 		if (pmd_none(*pmd) || pmd_bad(*pmd)) {
@@ -322,7 +342,7 @@ void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 			if ((pte_val(entry) & _PAGE_PRESENT)) {
 				phys = pte_val(entry)&PTE_PHYS_MASK;
 				if ((p^phys) & CACHE_ALIAS) {
-					d |= 1 << ((p & CACHE_ALIAS)>>12); 
+					d |= 1 << ((p & CACHE_ALIAS)>>12);
 					d |= 1 << ((phys & CACHE_ALIAS)>>12);
 					if (d == 0x0f)
 						goto loop_exit;