sh: Get SH-5 caches working again post-unification.

A number of cleanups to get the SH-5 cache management code in line with
the rest of the SH backend.

Signed-off-by: Paul Mundt <lethal@linux-sh.org>

5 files changed, 448 insertions, 648 deletions

diff --git a/arch/sh/kernel/cpu/sh5/probe.c b/arch/sh/kernel/cpu/sh5/probe.c
index 15d167f..31f8cb0 100644
--- a/arch/sh/kernel/cpu/sh5/probe.c
+++ b/arch/sh/kernel/cpu/sh5/probe.c
@@ -20,19 +20,18 @@ int __init detect_cpu_and_cache_system(void)
 {
 	unsigned long long cir;
 
-	/* Do peeks in real mode to avoid having to set up a mapping for the
-	   WPC registers. On SH5-101 cut2, such a mapping would be exposed to
-	   an address translation erratum which would make it hard to set up
-	   correctly. */
+	/*
+	 * Do peeks in real mode to avoid having to set up a mapping for
+	 * the WPC registers. On SH5-101 cut2, such a mapping would be
+	 * exposed to an address translation erratum which would make it
+	 * hard to set up correctly.
+	 */
 	cir = peek_real_address_q(0x0d000008);
-	if ((cir & 0xffff) == 0x5103) {
+	if ((cir & 0xffff) == 0x5103)
 		boot_cpu_data.type = CPU_SH5_103;
-	} else if (((cir >> 32) & 0xffff) == 0x51e2) {
+	else if (((cir >> 32) & 0xffff) == 0x51e2)
 		/* CPU.VCR aliased at CIR address on SH5-101 */
 		boot_cpu_data.type = CPU_SH5_101;
-	} else {
-		boot_cpu_data.type = CPU_SH_NONE;
-	}
 
 	/*
 	 * First, setup some sane values for the I-cache.
@@ -40,37 +39,33 @@ int __init detect_cpu_and_cache_system(void)
 	boot_cpu_data.icache.ways		= 4;
 	boot_cpu_data.icache.sets		= 256;
 	boot_cpu_data.icache.linesz		= L1_CACHE_BYTES;
+	boot_cpu_data.icache.way_incr		= (1 << 13);
+	boot_cpu_data.icache.entry_shift	= 5;
+	boot_cpu_data.icache.way_size		= boot_cpu_data.icache.sets *
+						  boot_cpu_data.icache.linesz;
+	boot_cpu_data.icache.entry_mask		= 0x1fe0;
+	boot_cpu_data.icache.flags		= 0;
 
-#if 0
 	/*
-	 * FIXME: This can probably be cleaned up a bit as well.. for example,
-	 * do we really need the way shift _and_ the way_step_shift ?? Judging
-	 * by the existing code, I would guess no.. is there any valid reason
-	 * why we need to be tracking this around?
+	 * Next, setup some sane values for the D-cache.
+	 *
+	 * On the SH5, these are pretty consistent with the I-cache settings,
+	 * so we just copy over the existing definitions.. these can be fixed
+	 * up later, especially if we add runtime CPU probing.
+	 *
+	 * Though in the meantime it saves us from having to duplicate all of
+	 * the above definitions..
 	 */
-	boot_cpu_data.icache.way_shift		= 13;
-	boot_cpu_data.icache.entry_shift	= 5;
-	boot_cpu_data.icache.set_shift		= 4;
-	boot_cpu_data.icache.way_step_shift	= 16;
-	boot_cpu_data.icache.asid_shift		= 2;
+	boot_cpu_data.dcache		= boot_cpu_data.icache;
 
 	/*
-	 * way offset = cache size / associativity, so just don't factor in
-	 * associativity in the first place..
+	 * Setup any cache-related flags here
 	 */
-	boot_cpu_data.icache.way_ofs	= boot_cpu_data.icache.sets *
-					  boot_cpu_data.icache.linesz;
-
-	boot_cpu_data.icache.asid_mask		= 0x3fc;
-	boot_cpu_data.icache.idx_mask		= 0x1fe0;
-	boot_cpu_data.icache.epn_mask		= 0xffffe000;
+#if defined(CONFIG_CACHE_WRITETHROUGH)
+	set_bit(SH_CACHE_MODE_WT, &(boot_cpu_data.dcache.flags));
+#elif defined(CONFIG_CACHE_WRITEBACK)
+	set_bit(SH_CACHE_MODE_WB, &(boot_cpu_data.dcache.flags));
 #endif
 
-	boot_cpu_data.icache.flags		= 0;
-
-	/* A trivial starting point.. */
-	memcpy(&boot_cpu_data.dcache,
-	       &boot_cpu_data.icache, sizeof(struct cache_info));
-
 	return 0;
 }
diff --git a/arch/sh/mm/cache-sh5.c b/arch/sh/mm/cache-sh5.c
index 5d1f615..3877321 100644
--- a/arch/sh/mm/cache-sh5.c
+++ b/arch/sh/mm/cache-sh5.c
@@ -1,10 +1,10 @@
 /*
  * arch/sh/mm/cache-sh5.c
  *
- * Original version Copyright (C) 2000, 2001  Paolo Alberelli
- * Second version Copyright (C) benedict.gaster@superh.com 2002
- * Third version Copyright Richard.Curnow@superh.com 2003
- * Hacks to third version Copyright (C) 2003 Paul Mundt
+ * Copyright (C) 2000, 2001  Paolo Alberelli
+ * Copyright (C) 2002  Benedict Gaster
+ * Copyright (C) 2003  Richard Curnow
+ * Copyright (C) 2003 - 2008  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
@@ -13,101 +13,20 @@
 #include <linux/init.h>
 #include <linux/mman.h>
 #include <linux/mm.h>
-#include <linux/threads.h>
-#include <asm/page.h>
-#include <asm/pgtable.h>
+#include <asm/tlb.h>
 #include <asm/processor.h>
 #include <asm/cache.h>
-#include <asm/tlb.h>
-#include <asm/io.h>
+#include <asm/pgalloc.h>
 #include <asm/uaccess.h>
 #include <asm/mmu_context.h>
-#include <asm/pgalloc.h> /* for flush_itlb_range */
-
-#include <linux/proc_fs.h>
-
-/* This function is in entry.S */
-extern unsigned long switch_and_save_asid(unsigned long new_asid);
 
 /* Wired TLB entry for the D-cache */
 static unsigned long long dtlb_cache_slot;
 
-/**
- * sh64_cache_init()
- *
- * This is pretty much just a straightforward clone of the SH
- * detect_cpu_and_cache_system().
- *
- * This function is responsible for setting up all of the cache
- * info dynamically as well as taking care of CPU probing and
- * setting up the relevant subtype data.
- *
- * FIXME: For the time being, we only really support the SH5-101
- * out of the box, and don't support dynamic probing for things
- * like the SH5-103 or even cut2 of the SH5-101. Implement this
- * later!
- */
-int __init sh64_cache_init(void)
+void __init p3_cache_init(void)
 {
-	/*
-	 * First, setup some sane values for the I-cache.
-	 */
-	cpu_data->icache.ways		= 4;
-	cpu_data->icache.sets		= 256;
-	cpu_data->icache.linesz		= L1_CACHE_BYTES;
-
-	/*
-	 * FIXME: This can probably be cleaned up a bit as well.. for example,
-	 * do we really need the way shift _and_ the way_step_shift ?? Judging
-	 * by the existing code, I would guess no.. is there any valid reason
-	 * why we need to be tracking this around?
-	 */
-	cpu_data->icache.way_shift	= 13;
-	cpu_data->icache.entry_shift	= 5;
-	cpu_data->icache.set_shift	= 4;
-	cpu_data->icache.way_step_shift	= 16;
-	cpu_data->icache.asid_shift	= 2;
-
-	/*
-	 * way offset = cache size / associativity, so just don't factor in
-	 * associativity in the first place..
-	 */
-	cpu_data->icache.way_ofs	= cpu_data->icache.sets *
-					  cpu_data->icache.linesz;
-
-	cpu_data->icache.asid_mask	= 0x3fc;
-	cpu_data->icache.idx_mask	= 0x1fe0;
-	cpu_data->icache.epn_mask	= 0xffffe000;
-	cpu_data->icache.flags		= 0;
-
-	/*
-	 * Next, setup some sane values for the D-cache.
-	 *
-	 * On the SH5, these are pretty consistent with the I-cache settings,
-	 * so we just copy over the existing definitions.. these can be fixed
-	 * up later, especially if we add runtime CPU probing.
-	 *
-	 * Though in the meantime it saves us from having to duplicate all of
-	 * the above definitions..
-	 */
-	cpu_data->dcache		= cpu_data->icache;
-
-	/*
-	 * Setup any cache-related flags here
-	 */
-#if defined(CONFIG_DCACHE_WRITE_THROUGH)
-	set_bit(SH_CACHE_MODE_WT, &(cpu_data->dcache.flags));
-#elif defined(CONFIG_DCACHE_WRITE_BACK)
-	set_bit(SH_CACHE_MODE_WB, &(cpu_data->dcache.flags));
-#endif
-
-	/*
-	 * We also need to reserve a slot for the D-cache in the DTLB, so we
-	 * do this now ..
-	 */
-	dtlb_cache_slot			= sh64_get_wired_dtlb_entry();
-
-	return 0;
+	/* Reserve a slot for dcache colouring in the DTLB */
+	dtlb_cache_slot	= sh64_get_wired_dtlb_entry();
 }
 
 #ifdef CONFIG_DCACHE_DISABLED
@@ -116,73 +35,48 @@ int __init sh64_cache_init(void)
 #define sh64_dcache_purge_user_range(mm, start, end)		do { } while (0)
 #define sh64_dcache_purge_phy_page(paddr)			do { } while (0)
 #define sh64_dcache_purge_virt_page(mm, eaddr)			do { } while (0)
-#define sh64_dcache_purge_kernel_range(start, end)		do { } while (0)
-#define sh64_dcache_wback_current_user_range(start, end)	do { } while (0)
 #endif
 
-/*##########################################################################*/
-
-/* From here onwards, a rewrite of the implementation,
-   by Richard.Curnow@superh.com.
-
-   The major changes in this compared to the old version are;
-   1. use more selective purging through OCBP instead of using ALLOCO to purge
-      by natural replacement.  This avoids purging out unrelated cache lines
-      that happen to be in the same set.
-   2. exploit the APIs copy_user_page and clear_user_page better
-   3. be more selective about I-cache purging, in particular use invalidate_all
-      more sparingly.
-
-   */
-
-/*##########################################################################
-			       SUPPORT FUNCTIONS
-  ##########################################################################*/
-
-/****************************************************************************/
-/* The following group of functions deal with mapping and unmapping a temporary
-   page into the DTLB slot that have been set aside for our exclusive use. */
-/* In order to accomplish this, we use the generic interface for adding and
-   removing a wired slot entry as defined in arch/sh/mm/tlb-sh5.c */
-/****************************************************************************/
-
-static unsigned long slot_own_flags;
-
-static inline void sh64_setup_dtlb_cache_slot(unsigned long eaddr, unsigned long asid, unsigned long paddr)
+/*
+ * The following group of functions deal with mapping and unmapping a
+ * temporary page into a DTLB slot that has been set aside for exclusive
+ * use.
+ */
+static inline void
+sh64_setup_dtlb_cache_slot(unsigned long eaddr, unsigned long asid,
+			   unsigned long paddr)
 {
-	local_irq_save(slot_own_flags);
+	local_irq_disable();
 	sh64_setup_tlb_slot(dtlb_cache_slot, eaddr, asid, paddr);
 }
 
 static inline void sh64_teardown_dtlb_cache_slot(void)
 {
 	sh64_teardown_tlb_slot(dtlb_cache_slot);
-	local_irq_restore(slot_own_flags);
+	local_irq_enable();
 }
 
-/****************************************************************************/
-
 #ifndef CONFIG_ICACHE_DISABLED
-
-static void __inline__ sh64_icache_inv_all(void)
+static inline void sh64_icache_inv_all(void)
 {
 	unsigned long long addr, flag, data;
 	unsigned int flags;
 
-	addr=ICCR0;
-	flag=ICCR0_ICI;
-	data=0;
+	addr = ICCR0;
+	flag = ICCR0_ICI;
+	data = 0;
 
 	/* Make this a critical section for safety (probably not strictly necessary.) */
 	local_irq_save(flags);
 
 	/* Without %1 it gets unexplicably wrong */
-	asm volatile("getcfg	%3, 0, %0\n\t"
-			"or	%0, %2, %0\n\t"
-			"putcfg	%3, 0, %0\n\t"
-			"synci"
-			: "=&r" (data)
-			: "0" (data), "r" (flag), "r" (addr));
+	__asm__ __volatile__ (
+		"getcfg	%3, 0, %0\n\t"
+		"or	%0, %2, %0\n\t"
+		"putcfg	%3, 0, %0\n\t"
+		"synci"
+		: "=&r" (data)
+		: "0" (data), "r" (flag), "r" (addr));
 
 	local_irq_restore(flags);
 }
@@ -193,20 +87,12 @@ static void sh64_icache_inv_kernel_range(unsigned long start, unsigned long end)
 	 * the addresses lie in the kernel superpage. */
 
 	unsigned long long ullend, addr, aligned_start;
-#if (NEFF == 32)
 	aligned_start = (unsigned long long)(signed long long)(signed long) start;
-#else
-#error "NEFF != 32"
-#endif
-	aligned_start &= L1_CACHE_ALIGN_MASK;
-	addr = aligned_start;
-#if (NEFF == 32)
+	addr = L1_CACHE_ALIGN(aligned_start);
 	ullend = (unsigned long long) (signed long long) (signed long) end;
-#else
-#error "NEFF != 32"
-#endif
+
 	while (addr <= ullend) {
-		asm __volatile__ ("icbi %0, 0" : : "r" (addr));
+		__asm__ __volatile__ ("icbi %0, 0" : : "r" (addr));
 		addr += L1_CACHE_BYTES;
 	}
 }
@@ -215,7 +101,7 @@ static void sh64_icache_inv_user_page(struct vm_area_struct *vma, unsigned long
 {
 	/* If we get called, we know that vma->vm_flags contains VM_EXEC.
 	   Also, eaddr is page-aligned. */
-
+	unsigned int cpu = smp_processor_id();
 	unsigned long long addr, end_addr;
 	unsigned long flags = 0;
 	unsigned long running_asid, vma_asid;
@@ -237,17 +123,17 @@ static void sh64_icache_inv_user_page(struct vm_area_struct *vma, unsigned long
 	*/
 
 	running_asid = get_asid();
-	vma_asid = (vma->vm_mm->context & MMU_CONTEXT_ASID_MASK);
+	vma_asid = cpu_asid(cpu, vma->vm_mm);
 	if (running_asid != vma_asid) {
 		local_irq_save(flags);
 		switch_and_save_asid(vma_asid);
 	}
 	while (addr < end_addr) {
 		/* Worth unrolling a little */
-		asm __volatile__("icbi %0,  0" : : "r" (addr));
-		asm __volatile__("icbi %0, 32" : : "r" (addr));
-		asm __volatile__("icbi %0, 64" : : "r" (addr));
-		asm __volatile__("icbi %0, 96" : : "r" (addr));
+		__asm__ __volatile__("icbi %0,  0" : : "r" (addr));
+		__asm__ __volatile__("icbi %0, 32" : : "r" (addr));
+		__asm__ __volatile__("icbi %0, 64" : : "r" (addr));
+		__asm__ __volatile__("icbi %0, 96" : : "r" (addr));
 		addr += 128;
 	}
 	if (running_asid != vma_asid) {
@@ -256,8 +142,6 @@ static void sh64_icache_inv_user_page(struct vm_area_struct *vma, unsigned long
 	}
 }
 
-/****************************************************************************/
-
 static void sh64_icache_inv_user_page_range(struct mm_struct *mm,
 			  unsigned long start, unsigned long end)
 {
@@ -275,10 +159,10 @@ static void sh64_icache_inv_user_page_range(struct mm_struct *mm,
 	   possible with the D-cache.  Just assume 64 for now as a working
 	   figure.
 	   */
-
 	int n_pages;
 
-	if (!mm) return;
+	if (!mm)
+		return;
 
 	n_pages = ((end - start) >> PAGE_SHIFT);
 	if (n_pages >= 64) {
@@ -290,7 +174,7 @@ static void sh64_icache_inv_user_page_range(struct mm_struct *mm,
 		unsigned long mm_asid, current_asid;
 		unsigned long long flags = 0ULL;
 
-		mm_asid = mm->context & MMU_CONTEXT_ASID_MASK;
+		mm_asid = cpu_asid(smp_processor_id(), mm);
 		current_asid = get_asid();
 
 		if (mm_asid != current_asid) {
@@ -322,6 +206,7 @@ static void sh64_icache_inv_user_page_range(struct mm_struct *mm,
 			}
 			aligned_start = vma->vm_end; /* Skip to start of next region */
 		}
+
 		if (mm_asid != current_asid) {
 			switch_and_save_asid(current_asid);
 			local_irq_restore(flags);
@@ -329,47 +214,46 @@ static void sh64_icache_inv_user_page_range(struct mm_struct *mm,
 	}
 }
 
+/*
+ * Invalidate a small range of user context I-cache, not necessarily page
+ * (or even cache-line) aligned.
+ *
+ * Since this is used inside ptrace, the ASID in the mm context typically
+ * won't match current_asid.  We'll have to switch ASID to do this.  For
+ * safety, and given that the range will be small, do all this under cli.
+ *
+ * Note, there is a hazard that the ASID in mm->context is no longer
+ * actually associated with mm, i.e. if the mm->context has started a new
+ * cycle since mm was last active.  However, this is just a performance
+ * issue: all that happens is that we invalidate lines belonging to
+ * another mm, so the owning process has to refill them when that mm goes
+ * live again.  mm itself can't have any cache entries because there will
+ * have been a flush_cache_all when the new mm->context cycle started.
+ */
 static void sh64_icache_inv_user_small_range(struct mm_struct *mm,
 						unsigned long start, int len)
 {
-
-	/* Invalidate a small range of user context I-cache, not necessarily
-	   page (or even cache-line) aligned. */
-
 	unsigned long long eaddr = start;
 	unsigned long long eaddr_end = start + len;
 	unsigned long current_asid, mm_asid;
 	unsigned long long flags;
 	unsigned long long epage_start;
 
-	/* Since this is used inside ptrace, the ASID in the mm context
-	   typically won't match current_asid.  We'll have to switch ASID to do
-	   this.  For safety, and given that the range will be small, do all
-	   this under cli.
-
-	   Note, there is a hazard that the ASID in mm->context is no longer
-	   actually associated with mm, i.e. if the mm->context has started a
-	   new cycle since mm was last active.  However, this is just a
-	   performance issue: all that happens is that we invalidate lines
-	   belonging to another mm, so the owning process has to refill them
-	   when that mm goes live again.  mm itself can't have any cache
-	   entries because there will have been a flush_cache_all when the new
-	   mm->context cycle started. */
-
-	/* Align to start of cache line.  Otherwise, suppose len==8 and start
-	   was at 32N+28 : the last 4 bytes wouldn't get invalidated. */
-	eaddr = start & L1_CACHE_ALIGN_MASK;
+	/*
+	 * Align to start of cache line.  Otherwise, suppose len==8 and
+	 * start was at 32N+28 : the last 4 bytes wouldn't get invalidated.
+	 */
+	eaddr = L1_CACHE_ALIGN(start);
 	eaddr_end = start + len;
 
+	mm_asid = cpu_asid(smp_processor_id(), mm);
 	local_irq_save(flags);
-	mm_asid = mm->context & MMU_CONTEXT_ASID_MASK;
 	current_asid = switch_and_save_asid(mm_asid);
 
 	epage_start = eaddr & PAGE_MASK;
 
-	while (eaddr < eaddr_end)
-	{
-		asm __volatile__("icbi %0, 0" : : "r" (eaddr));
+	while (eaddr < eaddr_end) {
+		__asm__ __volatile__("icbi %0, 0" : : "r" (eaddr));
 		eaddr += L1_CACHE_BYTES;
 	}
 	switch_and_save_asid(current_asid);
@@ -394,30 +278,24 @@ static void sh64_icache_inv_current_user_range(unsigned long start, unsigned lon
 	   been recycled since we were last active in which case we might just
 	   invalidate another processes I-cache entries : no worries, just a
 	   performance drop for him. */
-	aligned_start = start & L1_CACHE_ALIGN_MASK;
+	aligned_start = L1_CACHE_ALIGN(start);
 	addr = aligned_start;
 	while (addr < ull_end) {
-		asm __volatile__ ("icbi %0, 0" : : "r" (addr));
-		asm __volatile__ ("nop");
-		asm __volatile__ ("nop");
+		__asm__ __volatile__ ("icbi %0, 0" : : "r" (addr));
+		__asm__ __volatile__ ("nop");
+		__asm__ __volatile__ ("nop");
 		addr += L1_CACHE_BYTES;
 	}
 }
-
 #endif /* !CONFIG_ICACHE_DISABLED */
 
-/****************************************************************************/
-
 #ifndef CONFIG_DCACHE_DISABLED
-
 /* Buffer used as the target of alloco instructions to purge data from cache
    sets by natural eviction. -- RPC */
-#define DUMMY_ALLOCO_AREA_SIZE L1_CACHE_SIZE_BYTES + (1024 * 4)
+#define DUMMY_ALLOCO_AREA_SIZE ((L1_CACHE_BYTES << 10) + (1024 * 4))
 static unsigned char dummy_alloco_area[DUMMY_ALLOCO_AREA_SIZE] __cacheline_aligned = { 0, };
 
-/****************************************************************************/
-
-static void __inline__ sh64_dcache_purge_sets(int sets_to_purge_base, int n_sets)
+static void inline sh64_dcache_purge_sets(int sets_to_purge_base, int n_sets)
 {
 	/* Purge all ways in a particular block of sets, specified by the base
 	   set number and number of sets.  Can handle wrap-around, if that's
@@ -428,102 +306,86 @@ static void __inline__ sh64_dcache_purge_sets(int sets_to_purge_base, int n_sets
 	int j;
 	int set_offset;
 
-	dummy_buffer_base_set = ((int)&dummy_alloco_area & cpu_data->dcache.idx_mask) >> cpu_data->dcache.entry_shift;
+	dummy_buffer_base_set = ((int)&dummy_alloco_area &
+				 cpu_data->dcache.entry_mask) >>
+				 cpu_data->dcache.entry_shift;
 	set_offset = sets_to_purge_base - dummy_buffer_base_set;
 
-	for (j=0; j<n_sets; j++, set_offset++) {
+	for (j = 0; j < n_sets; j++, set_offset++) {
 		set_offset &= (cpu_data->dcache.sets - 1);
-		eaddr0 = (unsigned long long)dummy_alloco_area + (set_offset << cpu_data->dcache.entry_shift);
-
-		/* Do one alloco which hits the required set per cache way.  For
-		   write-back mode, this will purge the #ways resident lines.   There's
-		   little point unrolling this loop because the allocos stall more if
-		   they're too close together. */
-		eaddr1 = eaddr0 + cpu_data->dcache.way_ofs * cpu_data->dcache.ways;
-		for (eaddr=eaddr0; eaddr<eaddr1; eaddr+=cpu_data->dcache.way_ofs) {
-			asm __volatile__ ("alloco %0, 0" : : "r" (eaddr));
-			asm __volatile__ ("synco"); /* TAKum03020 */
+		eaddr0 = (unsigned long long)dummy_alloco_area +
+			(set_offset << cpu_data->dcache.entry_shift);
+
+		/*
+		 * Do one alloco which hits the required set per cache
+		 * way.  For write-back mode, this will purge the #ways
+		 * resident lines.  There's little point unrolling this
+		 * loop because the allocos stall more if they're too
+		 * close together.
+		 */
+		eaddr1 = eaddr0 + cpu_data->dcache.way_size *
+				  cpu_data->dcache.ways;
+
+		for (eaddr = eaddr0; eaddr < eaddr1;
+		     eaddr += cpu_data->dcache.way_size) {
+			__asm__ __volatile__ ("alloco %0, 0" : : "r" (eaddr));
+			__asm__ __volatile__ ("synco"); /* TAKum03020 */
 		}
 
-		eaddr1 = eaddr0 + cpu_data->dcache.way_ofs * cpu_data->dcache.ways;
-		for (eaddr=eaddr0; eaddr<eaddr1; eaddr+=cpu_data->dcache.way_ofs) {
-			/* Load from each address.  Required because alloco is a NOP if
-			   the cache is write-through.  Write-through is a config option. */
+		eaddr1 = eaddr0 + cpu_data->dcache.way_size *
+				  cpu_data->dcache.ways;
+
+		for (eaddr = eaddr0; eaddr < eaddr1;
+		     eaddr += cpu_data->dcache.way_size) {
+			/*
+			 * Load from each address.  Required because
+			 * alloco is a NOP if the cache is write-through.
+			 */
 			if (test_bit(SH_CACHE_MODE_WT, &(cpu_data->dcache.flags)))
-				*(volatile unsigned char *)(int)eaddr;
+				ctrl_inb(eaddr);
 		}
 	}
 
-	/* Don't use OCBI to invalidate the lines.  That costs cycles directly.
-	   If the dummy block is just left resident, it will naturally get
-	   evicted as required.  */
-
-	return;
+	/*
+	 * Don't use OCBI to invalidate the lines.  That costs cycles
+	 * directly.  If the dummy block is just left resident, it will
+	 * naturally get evicted as required.
+	 */
 }
 
-/****************************************************************************/
-
+/*
+ * Purge the entire contents of the dcache.  The most efficient way to
+ * achieve this is to use alloco instructions on a region of unused
+ * memory equal in size to the cache, thereby causing the current
+ * contents to be discarded by natural eviction.  The alternative, namely
+ * reading every tag, setting up a mapping for the corresponding page and
+ * doing an OCBP for the line, would be much more expensive.
+ */
 static void sh64_dcache_purge_all(void)
 {
-	/* Purge the entire contents of the dcache.  The most efficient way to
-	   achieve this is to use alloco instructions on a region of unused
-	   memory equal in size to the cache, thereby causing the current
-	   contents to be discarded by natural eviction.  The alternative,
-	   namely reading every tag, setting up a mapping for the corresponding
-	   page and doing an OCBP for the line, would be much more expensive.
-	   */
 
 	sh64_dcache_purge_sets(0, cpu_data->dcache.sets);
-
-	return;
-
 }
 
-/****************************************************************************/
-
-static void sh64_dcache_purge_kernel_range(unsigned long start, unsigned long end)
-{
-	/* Purge the range of addresses [start,end] from the D-cache.  The
-	   addresses lie in the superpage mapping.  There's no harm if we
-	   overpurge at either end - just a small performance loss. */
-	unsigned long long ullend, addr, aligned_start;
-#if (NEFF == 32)
-	aligned_start = (unsigned long long)(signed long long)(signed long) start;
-#else
-#error "NEFF != 32"
-#endif
-	aligned_start &= L1_CACHE_ALIGN_MASK;
-	addr = aligned_start;
-#if (NEFF == 32)
-	ullend = (unsigned long long) (signed long long) (signed long) end;
-#else
-#error "NEFF != 32"
-#endif
-	while (addr <= ullend) {
-		asm __volatile__ ("ocbp %0, 0" : : "r" (addr));
-		addr += L1_CACHE_BYTES;
-	}
-	return;
-}
 
 /* Assumes this address (+ (2**n_synbits) pages up from it) aren't used for
    anything else in the kernel */
 #define MAGIC_PAGE0_START 0xffffffffec000000ULL
 
-static void sh64_dcache_purge_coloured_phy_page(unsigned long paddr, unsigned long eaddr)
+/* Purge the physical page 'paddr' from the cache.  It's known that any
+ * cache lines requiring attention have the same page colour as the the
+ * address 'eaddr'.
+ *
+ * This relies on the fact that the D-cache matches on physical tags when
+ * no virtual tag matches.  So we create an alias for the original page
+ * and purge through that.  (Alternatively, we could have done this by
+ * switching ASID to match the original mapping and purged through that,
+ * but that involves ASID switching cost + probably a TLBMISS + refill
+ * anyway.)
+ */
+static void sh64_dcache_purge_coloured_phy_page(unsigned long paddr,
+					        unsigned long eaddr)
 {
-	/* Purge the physical page 'paddr' from the cache.  It's known that any
-	   cache lines requiring attention have the same page colour as the the
-	   address 'eaddr'.
-
-	   This relies on the fact that the D-cache matches on physical tags
-	   when no virtual tag matches.  So we create an alias for the original
-	   page and purge through that.  (Alternatively, we could have done
-	   this by switching ASID to match the original mapping and purged
-	   through that, but that involves ASID switching cost + probably a
-	   TLBMISS + refill anyway.)
-	   */
-
 	unsigned long long magic_page_start;
 	unsigned long long magic_eaddr, magic_eaddr_end;
 
@@ -531,47 +393,45 @@ static void sh64_dcache_purge_coloured_phy_page(unsigned long paddr, unsigned lo
 
 	/* As long as the kernel is not pre-emptible, this doesn't need to be
 	   under cli/sti. */
-
 	sh64_setup_dtlb_cache_slot(magic_page_start, get_asid(), paddr);
 
 	magic_eaddr = magic_page_start;
 	magic_eaddr_end = magic_eaddr + PAGE_SIZE;
+
 	while (magic_eaddr < magic_eaddr_end) {
 		/* Little point in unrolling this loop - the OCBPs are blocking
 		   and won't go any quicker (i.e. the loop overhead is parallel
 		   to part of the OCBP execution.) */
-		asm __volatile__ ("ocbp %0, 0" : : "r" (magic_eaddr));
+		__asm__ __volatile__ ("ocbp %0, 0" : : "r" (magic_eaddr));
 		magic_eaddr += L1_CACHE_BYTES;
 	}
 
 	sh64_teardown_dtlb_cache_slot();
 }
 
-/****************************************************************************/
-
+/*
+ * Purge a page given its physical start address, by creating a temporary
+ * 1 page mapping and purging across that.  Even if we know the virtual
+ * address (& vma or mm) of the page, the method here is more elegant
+ * because it avoids issues of coping with page faults on the purge
+ * instructions (i.e. no special-case code required in the critical path
+ * in the TLB miss handling).
+ */
 static void sh64_dcache_purge_phy_page(unsigned long paddr)
 {
-	/* Pure a page given its physical start address, by creating a
-	   temporary 1 page mapping and purging across that.  Even if we know
-	   the virtual address (& vma or mm) of the page, the method here is
-	   more elegant because it avoids issues of coping with page faults on
-	   the purge instructions (i.e. no special-case code required in the
-	   critical path in the TLB miss handling). */
-
 	unsigned long long eaddr_start, eaddr, eaddr_end;
 	int i;
 
 	/* As long as the kernel is not pre-emptible, this doesn't need to be
 	   under cli/sti. */
-
 	eaddr_start = MAGIC_PAGE0_START;
-	for (i=0; i < (1 << CACHE_OC_N_SYNBITS); i++) {
+	for (i = 0; i < (1 << CACHE_OC_N_SYNBITS); i++) {
 		sh64_setup_dtlb_cache_slot(eaddr_start, get_asid(), paddr);
 
 		eaddr = eaddr_start;
 		eaddr_end = eaddr + PAGE_SIZE;
 		while (eaddr < eaddr_end) {
-			asm __volatile__ ("ocbp %0, 0" : : "r" (eaddr));
+			__asm__ __volatile__ ("ocbp %0, 0" : : "r" (eaddr));
 			eaddr += L1_CACHE_BYTES;
 		}
 
@@ -584,6 +444,7 @@ static void sh64_dcache_purge_user_pages(struct mm_struct *mm,
 				unsigned long addr, unsigned long end)
 {
 	pgd_t *pgd;
+	pud_t *pud;
 	pmd_t *pmd;
 	pte_t *pte;
 	pte_t entry;
@@ -597,7 +458,11 @@ static void sh64_dcache_purge_user_pages(struct mm_struct *mm,
 	if (pgd_bad(*pgd))
 		return;
 
-	pmd = pmd_offset(pgd, addr);
+	pud = pud_offset(pgd, addr);
+	if (pud_none(*pud) || pud_bad(*pud))
+		return;
+
+	pmd = pmd_offset(pud, addr);
 	if (pmd_none(*pmd) || pmd_bad(*pmd))
 		return;
 
@@ -611,421 +476,357 @@ static void sh64_dcache_purge_user_pages(struct mm_struct *mm,
 	} while (pte++, addr += PAGE_SIZE, addr != end);
 	pte_unmap_unlock(pte - 1, ptl);
 }
-/****************************************************************************/
 
+/*
+ * There are at least 5 choices for the implementation of this, with
+ * pros (+), cons(-), comments(*):
+ *
+ * 1. ocbp each line in the range through the original user's ASID
+ *    + no lines spuriously evicted
+ *    - tlbmiss handling (must either handle faults on demand => extra
+ *	special-case code in tlbmiss critical path), or map the page in
+ *	advance (=> flush_tlb_range in advance to avoid multiple hits)
+ *    - ASID switching
+ *    - expensive for large ranges
+ *
+ * 2. temporarily map each page in the range to a special effective
+ *    address and ocbp through the temporary mapping; relies on the
+ *    fact that SH-5 OCB* always do TLB lookup and match on ptags (they
+ *    never look at the etags)
+ *    + no spurious evictions
+ *    - expensive for large ranges
+ *    * surely cheaper than (1)
+ *
+ * 3. walk all the lines in the cache, check the tags, if a match
+ *    occurs create a page mapping to ocbp the line through
+ *    + no spurious evictions
+ *    - tag inspection overhead
+ *    - (especially for small ranges)
+ *    - potential cost of setting up/tearing down page mapping for
+ *	every line that matches the range
+ *    * cost partly independent of range size
+ *
+ * 4. walk all the lines in the cache, check the tags, if a match
+ *    occurs use 4 * alloco to purge the line (+3 other probably
+ *    innocent victims) by natural eviction
+ *    + no tlb mapping overheads
+ *    - spurious evictions
+ *    - tag inspection overhead
+ *
+ * 5. implement like flush_cache_all
+ *    + no tag inspection overhead
+ *    - spurious evictions
+ *    - bad for small ranges
+ *
+ * (1) can be ruled out as more expensive than (2).  (2) appears best
+ * for small ranges.  The choice between (3), (4) and (5) for large
+ * ranges and the range size for the large/small boundary need
+ * benchmarking to determine.
+ *
+ * For now use approach (2) for small ranges and (5) for large ones.
+ */
 static void sh64_dcache_purge_user_range(struct mm_struct *mm,
 			  unsigned long start, unsigned long end)
 {
-	/* There are at least 5 choices for the implementation of this, with
-	   pros (+), cons(-), comments(*):
-
-	   1. ocbp each line in the range through the original user's ASID
-	      + no lines spuriously evicted
-	      - tlbmiss handling (must either handle faults on demand => extra
-		special-case code in tlbmiss critical path), or map the page in
-		advance (=> flush_tlb_range in advance to avoid multiple hits)
-	      - ASID switching
-	      - expensive for large ranges
-
-	   2. temporarily map each page in the range to a special effective
-	      address and ocbp through the temporary mapping; relies on the
-	      fact that SH-5 OCB* always do TLB lookup and match on ptags (they
-	      never look at the etags)
-	      + no spurious evictions
-	      - expensive for large ranges
-	      * surely cheaper than (1)
-
-	   3. walk all the lines in the cache, check the tags, if a match
-	      occurs create a page mapping to ocbp the line through
-	      + no spurious evictions
-	      - tag inspection overhead
-	      - (especially for small ranges)
-	      - potential cost of setting up/tearing down page mapping for
-		every line that matches the range
-	      * cost partly independent of range size
-
-	   4. walk all the lines in the cache, check the tags, if a match
-	      occurs use 4 * alloco to purge the line (+3 other probably
-	      innocent victims) by natural eviction
-	      + no tlb mapping overheads
-	      - spurious evictions
-	      - tag inspection overhead
-
-	   5. implement like flush_cache_all
-	      + no tag inspection overhead
-	      - spurious evictions
-	      - bad for small ranges
-
-	   (1) can be ruled out as more expensive than (2).  (2) appears best
-	   for small ranges.  The choice between (3), (4) and (5) for large
-	   ranges and the range size for the large/small boundary need
-	   benchmarking to determine.
-
-	   For now use approach (2) for small ranges and (5) for large ones.
-
-	   */
+	int n_pages = ((end - start) >> PAGE_SHIFT);
 
-	int n_pages;
-
-	n_pages = ((end - start) >> PAGE_SHIFT);
 	if (n_pages >= 64 || ((start ^ (end - 1)) & PMD_MASK)) {
-#if 1
 		sh64_dcache_purge_all();
-#else
-		unsigned long long set, way;
-		unsigned long mm_asid = mm->context & MMU_CONTEXT_ASID_MASK;
-		for (set = 0; set < cpu_data->dcache.sets; set++) {
-			unsigned long long set_base_config_addr = CACHE_OC_ADDRESS_ARRAY + (set << cpu_data->dcache.set_shift);
-			for (way = 0; way < cpu_data->dcache.ways; way++) {
-				unsigned long long config_addr = set_base_config_addr + (way << cpu_data->dcache.way_step_shift);
-				unsigned long long tag0;
-				unsigned long line_valid;
-
-				asm __volatile__("getcfg %1, 0, %0" : "=r" (tag0) : "r" (config_addr));
-				line_valid = tag0 & SH_CACHE_VALID;
-				if (line_valid) {
-					unsigned long cache_asid;
-					unsigned long epn;
-
-					cache_asid = (tag0 & cpu_data->dcache.asid_mask) >> cpu_data->dcache.asid_shift;
-					/* The next line needs some
-					   explanation.  The virtual tags
-					   encode bits [31:13] of the virtual
-					   address, bit [12] of the 'tag' being
-					   implied by the cache set index. */
-					epn = (tag0 & cpu_data->dcache.epn_mask) | ((set & 0x80) << cpu_data->dcache.entry_shift);
-
-					if ((cache_asid == mm_asid) && (start <= epn) && (epn < end)) {
-						/* TODO : could optimise this
-						   call by batching multiple
-						   adjacent sets together. */
-						sh64_dcache_purge_sets(set, 1);
-						break; /* Don't waste time inspecting other ways for this set */
-					}
-				}
-			}
-		}
-#endif
 	} else {
 		/* Small range, covered by a single page table page */
 		start &= PAGE_MASK;	/* should already be so */
 		end = PAGE_ALIGN(end);	/* should already be so */
 		sh64_dcache_purge_user_pages(mm, start, end);
 	}
-	return;
 }
 
-static void sh64_dcache_wback_current_user_range(unsigned long start, unsigned long end)
+/*
+ * Purge the range of addresses from the D-cache.
+ *
+ * The addresses lie in the superpage mapping. There's no harm if we
+ * overpurge at either end - just a small performance loss.
+ */
+void __flush_purge_region(void *start, int size)
 {
-	unsigned long long aligned_start;
-	unsigned long long ull_end;
-	unsigned long long addr;
-
-	ull_end = end;
+	unsigned long long ullend, addr, aligned_start;
 
-	/* Just wback over the range using the natural addresses.  TLB miss
-	   handling will be OK (TBC) : the range has just been written to by
-	   the signal frame setup code, so the PTEs must exist.
+	aligned_start = (unsigned long long)(signed long long)(signed long) start;
+	addr = L1_CACHE_ALIGN(aligned_start);
+	ullend = (unsigned long long) (signed long long) (signed long) start + size;
 
-	   Note, if we have CONFIG_PREEMPT and get preempted inside this loop,
-	   it doesn't matter, even if the pid->ASID mapping changes whilst
-	   we're away.  In that case the cache will have been flushed when the
-	   mapping was renewed.  So the writebacks below will be nugatory (and
-	   we'll doubtless have to fault the TLB entry/ies in again with the
-	   new ASID), but it's a rare case.
-	   */
-	aligned_start = start & L1_CACHE_ALIGN_MASK;
-	addr = aligned_start;
-	while (addr < ull_end) {
-		asm __volatile__ ("ocbwb %0, 0" : : "r" (addr));
+	while (addr <= ullend) {
+		__asm__ __volatile__ ("ocbp %0, 0" : : "r" (addr));
 		addr += L1_CACHE_BYTES;
 	}
 }
 
-/****************************************************************************/
-
-/* These *MUST* lie in an area of virtual address space that's otherwise unused. */
-#define UNIQUE_EADDR_START 0xe0000000UL
-#define UNIQUE_EADDR_END   0xe8000000UL
-
-static unsigned long sh64_make_unique_eaddr(unsigned long user_eaddr, unsigned long paddr)
+void __flush_wback_region(void *start, int size)
 {
-	/* Given a physical address paddr, and a user virtual address
-	   user_eaddr which will eventually be mapped to it, create a one-off
-	   kernel-private eaddr mapped to the same paddr.  This is used for
-	   creating special destination pages for copy_user_page and
-	   clear_user_page */
+	unsigned long long ullend, addr, aligned_start;
 
-	static unsigned long current_pointer = UNIQUE_EADDR_START;
-	unsigned long coloured_pointer;
+	aligned_start = (unsigned long long)(signed long long)(signed long) start;
+	addr = L1_CACHE_ALIGN(aligned_start);
+	ullend = (unsigned long long) (signed long long) (signed long) start + size;
 
-	if (current_pointer == UNIQUE_EADDR_END) {
-		sh64_dcache_purge_all();
-		current_pointer = UNIQUE_EADDR_START;
+	while (addr < ullend) {
+		__asm__ __volatile__ ("ocbwb %0, 0" : : "r" (addr));
+		addr += L1_CACHE_BYTES;
 	}
-
-	coloured_pointer = (current_pointer & ~CACHE_OC_SYN_MASK) | (user_eaddr & CACHE_OC_SYN_MASK);
-	sh64_setup_dtlb_cache_slot(coloured_pointer, get_asid(), paddr);
-
-	current_pointer += (PAGE_SIZE << CACHE_OC_N_SYNBITS);
-
-	return coloured_pointer;
 }
 
-/****************************************************************************/
-
-static void sh64_copy_user_page_coloured(void *to, void *from, unsigned long address)
+void __flush_invalidate_region(void *start, int size)
 {
-	void *coloured_to;
-
-	/* Discard any existing cache entries of the wrong colour.  These are
-	   present quite often, if the kernel has recently used the page
-	   internally, then given it up, then it's been allocated to the user.
-	   */
-	sh64_dcache_purge_coloured_phy_page(__pa(to), (unsigned long) to);
-
-	coloured_to = (void *) sh64_make_unique_eaddr(address, __pa(to));
-	sh64_page_copy(from, coloured_to);
-
-	sh64_teardown_dtlb_cache_slot();
-}
-
-static void sh64_clear_user_page_coloured(void *to, unsigned long address)
-{
-	void *coloured_to;
-
-	/* Discard any existing kernel-originated lines of the wrong colour (as
-	   above) */
-	sh64_dcache_purge_coloured_phy_page(__pa(to), (unsigned long) to);
+	unsigned long long ullend, addr, aligned_start;
 
-	coloured_to = (void *) sh64_make_unique_eaddr(address, __pa(to));
-	sh64_page_clear(coloured_to);
+	aligned_start = (unsigned long long)(signed long long)(signed long) start;
+	addr = L1_CACHE_ALIGN(aligned_start);
+	ullend = (unsigned long long) (signed long long) (signed long) start + size;
 
-	sh64_teardown_dtlb_cache_slot();
+	while (addr < ullend) {
+		__asm__ __volatile__ ("ocbi %0, 0" : : "r" (addr));
+		addr += L1_CACHE_BYTES;
+	}
 }
-
 #endif /* !CONFIG_DCACHE_DISABLED */
 
-/****************************************************************************/
-
-/*##########################################################################
-			    EXTERNALLY CALLABLE API.
-  ##########################################################################*/
-
-/* These functions are described in Documentation/cachetlb.txt.
-   Each one of these functions varies in behaviour depending on whether the
-   I-cache and/or D-cache are configured out.
-
-   Note that the Linux term 'flush' corresponds to what is termed 'purge' in
-   the sh/sh64 jargon for the D-cache, i.e. write back dirty data then
-   invalidate the cache lines, and 'invalidate' for the I-cache.
-   */
-
-#undef FLUSH_TRACE
-
+/*
+ * Invalidate the entire contents of both caches, after writing back to
+ * memory any dirty data from the D-cache.
+ */
 void flush_cache_all(void)
 {
-	/* Invalidate the entire contents of both caches, after writing back to
-	   memory any dirty data from the D-cache. */
 	sh64_dcache_purge_all();
 	sh64_icache_inv_all();
 }
 
-/****************************************************************************/
-
+/*
+ * Invalidate an entire user-address space from both caches, after
+ * writing back dirty data (e.g. for shared mmap etc).
+ *
+ * This could be coded selectively by inspecting all the tags then
+ * doing 4*alloco on any set containing a match (as for
+ * flush_cache_range), but fork/exit/execve (where this is called from)
+ * are expensive anyway.
+ *
+ * Have to do a purge here, despite the comments re I-cache below.
+ * There could be odd-coloured dirty data associated with the mm still
+ * in the cache - if this gets written out through natural eviction
+ * after the kernel has reused the page there will be chaos.
+ *
+ * The mm being torn down won't ever be active again, so any Icache
+ * lines tagged with its ASID won't be visible for the rest of the
+ * lifetime of this ASID cycle.  Before the ASID gets reused, there
+ * will be a flush_cache_all.  Hence we don't need to touch the
+ * I-cache.  This is similar to the lack of action needed in
+ * flush_tlb_mm - see fault.c.
+ */
 void flush_cache_mm(struct mm_struct *mm)
 {
-	/* Invalidate an entire user-address space from both caches, after
-	   writing back dirty data (e.g. for shared mmap etc). */
-
-	/* This could be coded selectively by inspecting all the tags then
-	   doing 4*alloco on any set containing a match (as for
-	   flush_cache_range), but fork/exit/execve (where this is called from)
-	   are expensive anyway. */
-
-	/* Have to do a purge here, despite the comments re I-cache below.
-	   There could be odd-coloured dirty data associated with the mm still
-	   in the cache - if this gets written out through natural eviction
-	   after the kernel has reused the page there will be chaos.
-	   */
-
 	sh64_dcache_purge_all();
-
-	/* The mm being torn down won't ever be active again, so any Icache
-	   lines tagged with its ASID won't be visible for the rest of the
-	   lifetime of this ASID cycle.  Before the ASID gets reused, there
-	   will be a flush_cache_all.  Hence we don't need to touch the
-	   I-cache.  This is similar to the lack of action needed in
-	   flush_tlb_mm - see fault.c. */
 }
 
-/****************************************************************************/
-
+/*
+ * Invalidate (from both caches) the range [start,end) of virtual
+ * addresses from the user address space specified by mm, after writing
+ * back any dirty data.
+ *
+ * Note, 'end' is 1 byte beyond the end of the range to flush.
+ */
 void flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 		       unsigned long end)
 {
 	struct mm_struct *mm = vma->vm_mm;
 
-	/* Invalidate (from both caches) the range [start,end) of virtual
-	   addresses from the user address space specified by mm, after writing
-	   back any dirty data.
-
-	   Note, 'end' is 1 byte beyond the end of the range to flush. */
-
 	sh64_dcache_purge_user_range(mm, start, end);
 	sh64_icache_inv_user_page_range(mm, start, end);
 }
 
-/****************************************************************************/
-
-void flush_cache_page(struct vm_area_struct *vma, unsigned long eaddr, unsigned long pfn)
+/*
+ * Invalidate any entries in either cache for the vma within the user
+ * address space vma->vm_mm for the page starting at virtual address
+ * 'eaddr'.   This seems to be used primarily in breaking COW.  Note,
+ * the I-cache must be searched too in case the page in question is
+ * both writable and being executed from (e.g. stack trampolines.)
+ *
+ * Note, this is called with pte lock held.
+ */
+void flush_cache_page(struct vm_area_struct *vma, unsigned long eaddr,
+		      unsigned long pfn)
 {
-	/* Invalidate any entries in either cache for the vma within the user
-	   address space vma->vm_mm for the page starting at virtual address
-	   'eaddr'.   This seems to be used primarily in breaking COW.  Note,
-	   the I-cache must be searched too in case the page in question is
-	   both writable and being executed from (e.g. stack trampolines.)
-
-	   Note, this is called with pte lock held.
-	   */
-
 	sh64_dcache_purge_phy_page(pfn << PAGE_SHIFT);
 
-	if (vma->vm_flags & VM_EXEC) {
+	if (vma->vm_flags & VM_EXEC)
 		sh64_icache_inv_user_page(vma, eaddr);
-	}
 }
 
-/****************************************************************************/
+void flush_dcache_page(struct page *page)
+{
+	sh64_dcache_purge_phy_page(page_to_phys(page));
+	wmb();
+}
 
-#ifndef CONFIG_DCACHE_DISABLED
+/*
+ * Flush the range [start,end] of kernel virtual adddress space from
+ * the I-cache.  The corresponding range must be purged from the
+ * D-cache also because the SH-5 doesn't have cache snooping between
+ * the caches.  The addresses will be visible through the superpage
+ * mapping, therefore it's guaranteed that there no cache entries for
+ * the range in cache sets of the wrong colour.
+ */
+void flush_icache_range(unsigned long start, unsigned long end)
+{
+	__flush_purge_region((void *)start, end);
+	wmb();
+	sh64_icache_inv_kernel_range(start, end);
+}
 
-void copy_user_page(void *to, void *from, unsigned long address, struct page *page)
+/*
+ * Flush the range of user (defined by vma->vm_mm) address space starting
+ * at 'addr' for 'len' bytes from the cache.  The range does not straddle
+ * a page boundary, the unique physical page containing the range is
+ * 'page'.  This seems to be used mainly for invalidating an address
+ * range following a poke into the program text through the ptrace() call
+ * from another process (e.g. for BRK instruction insertion).
+ */
+void flush_icache_user_range(struct vm_area_struct *vma,
+			struct page *page, unsigned long addr, int len)
 {
-	/* 'from' and 'to' are kernel virtual addresses (within the superpage
-	   mapping of the physical RAM).  'address' is the user virtual address
-	   where the copy 'to' will be mapped after.  This allows a custom
-	   mapping to be used to ensure that the new copy is placed in the
-	   right cache sets for the user to see it without having to bounce it
-	   out via memory.  Note however : the call to flush_page_to_ram in
-	   (generic)/mm/memory.c:(break_cow) undoes all this good work in that one
-	   very important case!
-
-	   TBD : can we guarantee that on every call, any cache entries for
-	   'from' are in the same colour sets as 'address' also?  i.e. is this
-	   always used just to deal with COW?  (I suspect not). */
-
-	/* There are two possibilities here for when the page 'from' was last accessed:
-	   * by the kernel : this is OK, no purge required.
-	   * by the/a user (e.g. for break_COW) : need to purge.
-
-	   If the potential user mapping at 'address' is the same colour as
-	   'from' there is no need to purge any cache lines from the 'from'
-	   page mapped into cache sets of colour 'address'.  (The copy will be
-	   accessing the page through 'from').
-	   */
 
-	if (((address ^ (unsigned long) from) & CACHE_OC_SYN_MASK) != 0) {
-		sh64_dcache_purge_coloured_phy_page(__pa(from), address);
-	}
+	sh64_dcache_purge_coloured_phy_page(page_to_phys(page), addr);
+	mb();
 
-	if (((address ^ (unsigned long) to) & CACHE_OC_SYN_MASK) == 0) {
-		/* No synonym problem on destination */
-		sh64_page_copy(from, to);
-	} else {
-		sh64_copy_user_page_coloured(to, from, address);
-	}
+	if (vma->vm_flags & VM_EXEC)
+		sh64_icache_inv_user_small_range(vma->vm_mm, addr, len);
+}
 
-	/* Note, don't need to flush 'from' page from the cache again - it's
-	   done anyway by the generic code */
+/*
+ * For the address range [start,end), write back the data from the
+ * D-cache and invalidate the corresponding region of the I-cache for the
+ * current process.  Used to flush signal trampolines on the stack to
+ * make them executable.
+ */
+void flush_cache_sigtramp(unsigned long vaddr)
+{
+	unsigned long end = vaddr + L1_CACHE_BYTES;
+
+	__flush_wback_region((void *)vaddr, L1_CACHE_BYTES);
+	wmb();
+	sh64_icache_inv_current_user_range(vaddr, end);
 }
 
-void clear_user_page(void *to, unsigned long address, struct page *page)
+/*
+ * These *MUST* lie in an area of virtual address space that's otherwise
+ * unused.
+ */
+#define UNIQUE_EADDR_START 0xe0000000UL
+#define UNIQUE_EADDR_END   0xe8000000UL
+
+/*
+ * Given a physical address paddr, and a user virtual address user_eaddr
+ * which will eventually be mapped to it, create a one-off kernel-private
+ * eaddr mapped to the same paddr.  This is used for creating special
+ * destination pages for copy_user_page and clear_user_page.
+ */
+static unsigned long sh64_make_unique_eaddr(unsigned long user_eaddr,
+					    unsigned long paddr)
 {
-	/* 'to' is a kernel virtual address (within the superpage
-	   mapping of the physical RAM).  'address' is the user virtual address
-	   where the 'to' page will be mapped after.  This allows a custom
-	   mapping to be used to ensure that the new copy is placed in the
-	   right cache sets for the user to see it without having to bounce it
-	   out via memory.
-	*/
+	static unsigned long current_pointer = UNIQUE_EADDR_START;
+	unsigned long coloured_pointer;
 
-	if (((address ^ (unsigned long) to) & CACHE_OC_SYN_MASK) == 0) {
-		/* No synonym problem on destination */
-		sh64_page_clear(to);
-	} else {
-		sh64_clear_user_page_coloured(to, address);
+	if (current_pointer == UNIQUE_EADDR_END) {
+		sh64_dcache_purge_all();
+		current_pointer = UNIQUE_EADDR_START;
 	}
-}
 
-#endif /* !CONFIG_DCACHE_DISABLED */
+	coloured_pointer = (current_pointer & ~CACHE_OC_SYN_MASK) |
+				(user_eaddr & CACHE_OC_SYN_MASK);
+	sh64_setup_dtlb_cache_slot(coloured_pointer, get_asid(), paddr);
 
-/****************************************************************************/
+	current_pointer += (PAGE_SIZE << CACHE_OC_N_SYNBITS);
 
-void flush_dcache_page(struct page *page)
-{
-	sh64_dcache_purge_phy_page(page_to_phys(page));
-	wmb();
+	return coloured_pointer;
 }
 
-/****************************************************************************/
-
-void flush_icache_range(unsigned long start, unsigned long end)
+static void sh64_copy_user_page_coloured(void *to, void *from,
+					 unsigned long address)
 {
-	/* Flush the range [start,end] of kernel virtual adddress space from
-	   the I-cache.  The corresponding range must be purged from the
-	   D-cache also because the SH-5 doesn't have cache snooping between
-	   the caches.  The addresses will be visible through the superpage
-	   mapping, therefore it's guaranteed that there no cache entries for
-	   the range in cache sets of the wrong colour.
+	void *coloured_to;
 
-	   Primarily used for cohering the I-cache after a module has
-	   been loaded.  */
+	/*
+	 * Discard any existing cache entries of the wrong colour.  These are
+	 * present quite often, if the kernel has recently used the page
+	 * internally, then given it up, then it's been allocated to the user.
+	 */
+	sh64_dcache_purge_coloured_phy_page(__pa(to), (unsigned long)to);
 
-	/* We also make sure to purge the same range from the D-cache since
-	   flush_page_to_ram() won't be doing this for us! */
+	coloured_to = (void *)sh64_make_unique_eaddr(address, __pa(to));
+	copy_page(from, coloured_to);
 
-	sh64_dcache_purge_kernel_range(start, end);
-	wmb();
-	sh64_icache_inv_kernel_range(start, end);
+	sh64_teardown_dtlb_cache_slot();
 }
 
-/****************************************************************************/
-
-void flush_icache_user_range(struct vm_area_struct *vma,
-			struct page *page, unsigned long addr, int len)
+static void sh64_clear_user_page_coloured(void *to, unsigned long address)
 {
-	/* Flush the range of user (defined by vma->vm_mm) address space
-	   starting at 'addr' for 'len' bytes from the cache.  The range does
-	   not straddle a page boundary, the unique physical page containing
-	   the range is 'page'.  This seems to be used mainly for invalidating
-	   an address range following a poke into the program text through the
-	   ptrace() call from another process (e.g. for BRK instruction
-	   insertion). */
+	void *coloured_to;
 
-	sh64_dcache_purge_coloured_phy_page(page_to_phys(page), addr);
-	mb();
+	/*
+	 * Discard any existing kernel-originated lines of the wrong
+	 * colour (as above)
+	 */
+	sh64_dcache_purge_coloured_phy_page(__pa(to), (unsigned long)to);
 
-	if (vma->vm_flags & VM_EXEC) {
-		sh64_icache_inv_user_small_range(vma->vm_mm, addr, len);
-	}
-}
+	coloured_to = (void *)sh64_make_unique_eaddr(address, __pa(to));
+	clear_page(coloured_to);
 
-/*##########################################################################
-			ARCH/SH64 PRIVATE CALLABLE API.
-  ##########################################################################*/
+	sh64_teardown_dtlb_cache_slot();
+}
 
-void flush_cache_sigtramp(unsigned long vaddr)
+/*
+ * 'from' and 'to' are kernel virtual addresses (within the superpage
+ * mapping of the physical RAM).  'address' is the user virtual address
+ * where the copy 'to' will be mapped after.  This allows a custom
+ * mapping to be used to ensure that the new copy is placed in the
+ * right cache sets for the user to see it without having to bounce it
+ * out via memory.  Note however : the call to flush_page_to_ram in
+ * (generic)/mm/memory.c:(break_cow) undoes all this good work in that one
+ * very important case!
+ *
+ * TBD : can we guarantee that on every call, any cache entries for
+ * 'from' are in the same colour sets as 'address' also?  i.e. is this
+ * always used just to deal with COW?  (I suspect not).
+ *
+ * There are two possibilities here for when the page 'from' was last accessed:
+ * - by the kernel : this is OK, no purge required.
+ * - by the/a user (e.g. for break_COW) : need to purge.
+ *
+ * If the potential user mapping at 'address' is the same colour as
+ * 'from' there is no need to purge any cache lines from the 'from'
+ * page mapped into cache sets of colour 'address'.  (The copy will be
+ * accessing the page through 'from').
+ */
+void copy_user_page(void *to, void *from, unsigned long address,
+		    struct page *page)
 {
-	unsigned long end = vaddr + L1_CACHE_BYTES;
-
-	/* For the address range [start,end), write back the data from the
-	   D-cache and invalidate the corresponding region of the I-cache for
-	   the current process.  Used to flush signal trampolines on the stack
-	   to make them executable. */
+	if (((address ^ (unsigned long) from) & CACHE_OC_SYN_MASK) != 0)
+		sh64_dcache_purge_coloured_phy_page(__pa(from), address);
 
-	sh64_dcache_wback_current_user_range(vaddr, end);
-	wmb();
-	sh64_icache_inv_current_user_range(vaddr, end);
+	if (((address ^ (unsigned long) to) & CACHE_OC_SYN_MASK) == 0)
+		copy_page(to, from);
+	else
+		sh64_copy_user_page_coloured(to, from, address);
 }
 
+/*
+ * 'to' is a kernel virtual address (within the superpage mapping of the
+ * physical RAM).  'address' is the user virtual address where the 'to'
+ * page will be mapped after.  This allows a custom mapping to be used to
+ * ensure that the new copy is placed in the right cache sets for the
+ * user to see it without having to bounce it out via memory.
+ */
+void clear_user_page(void *to, unsigned long address, struct page *page)
+{
+	if (((address ^ (unsigned long) to) & CACHE_OC_SYN_MASK) == 0)
+		clear_page(to);
+	else
+		sh64_clear_user_page_coloured(to, address);
+}
diff --git a/include/asm-sh/cpu-sh5/cacheflush.h b/include/asm-sh/cpu-sh5/cacheflush.h
index f935acb..5a11f0b 100644
--- a/include/asm-sh/cpu-sh5/cacheflush.h
+++ b/include/asm-sh/cpu-sh5/cacheflush.h
@@ -3,8 +3,6 @@
 
 #ifndef __ASSEMBLY__
 
-#include <asm/page.h>
-
 struct vm_area_struct;
 struct page;
 struct mm_struct;
@@ -27,7 +25,7 @@ extern void flush_icache_user_range(struct vm_area_struct *vma,
 #define flush_dcache_mmap_unlock(mapping)	do { } while (0)
 
 #define flush_icache_page(vma, page)	do { } while (0)
-#define p3_cache_init()			do { } while (0)
+void p3_cache_init(void);
 
 #endif /* __ASSEMBLY__ */
 
diff --git a/include/asm-sh/mmu_context_64.h b/include/asm-sh/mmu_context_64.h
index 020be74..9649f1c 100644
--- a/include/asm-sh/mmu_context_64.h
+++ b/include/asm-sh/mmu_context_64.h
@@ -66,6 +66,9 @@ static inline void set_asid(unsigned long asid)
 		      : "=r" (sr), "=r" (pc) : "0" (sr));
 }
 
+/* arch/sh/kernel/cpu/sh5/entry.S */
+extern unsigned long switch_and_save_asid(unsigned long new_asid);
+
 /* No spare register to twiddle, so use a software cache */
 extern pgd_t *mmu_pdtp_cache;
 
diff --git a/include/asm-sh/page.h b/include/asm-sh/page.h
index 134562d..304c30b 100644
--- a/include/asm-sh/page.h
+++ b/include/asm-sh/page.h
@@ -55,11 +55,14 @@ extern void clear_page(void *to);
 extern void copy_page(void *to, void *from);
 
 #if !defined(CONFIG_CACHE_OFF) && defined(CONFIG_MMU) && \
-	(defined(CONFIG_CPU_SH4) || defined(CONFIG_SH7705_CACHE_32KB))
+	(defined(CONFIG_CPU_SH5) || defined(CONFIG_CPU_SH4) || \
+	 defined(CONFIG_SH7705_CACHE_32KB))
 struct page;
 struct vm_area_struct;
 extern void clear_user_page(void *to, unsigned long address, struct page *page);
-#ifdef CONFIG_CPU_SH4
+extern void copy_user_page(void *to, void *from, unsigned long address,
+			   struct page *page);
+#if defined(CONFIG_CPU_SH4)
 extern void copy_user_highpage(struct page *to, struct page *from,
 			       unsigned long vaddr, struct vm_area_struct *vma);
 #define __HAVE_ARCH_COPY_USER_HIGHPAGE