[PATCH] ARM: Fix Xscale copy_page implementation

The ARM copypage changes in 2.6.12-rc4-git1 removed the preempt locking
from the copypage functions which broke the XScale implementation.
This patch fixes the locking on XScale and removes the now unneeded
minicache code.
Signed-off-by: Russell King <rmk@arm.linux.org.uk>
Checked-by: Richard Purdie

arch/arm/mm/Kconfig

@@ -228,7 +228,6 @@ config CPU_SA1100
 	select CPU_CACHE_V4WB
 	select CPU_CACHE_VIVT
 	select CPU_TLB_V4WB
-	select CPU_MINICACHE
 
 # XScale
 config CPU_XSCALE
@@ -239,7 +238,6 @@ config CPU_XSCALE
 	select CPU_ABRT_EV5T
 	select CPU_CACHE_VIVT
 	select CPU_TLB_V4WBI
-	select CPU_MINICACHE
 
 # ARMv6
 config CPU_V6
@@ -345,11 +343,6 @@ config CPU_TLB_V4WBI
 config CPU_TLB_V6
 	bool
 
-config CPU_MINICACHE
-	bool
-	help
-	  Processor has a minicache.
-
 comment "Processor Features"
 
 config ARM_THUMB

arch/arm/mm/Makefile

@@ -31,8 +31,6 @@ obj-$(CONFIG_CPU_COPY_V6)	+= copypage-v6.o mmu.o
 obj-$(CONFIG_CPU_SA1100)	+= copypage-v4mc.o
 obj-$(CONFIG_CPU_XSCALE)	+= copypage-xscale.o
 
-obj-$(CONFIG_CPU_MINICACHE)	+= minicache.o
-
 obj-$(CONFIG_CPU_TLB_V3)	+= tlb-v3.o
 obj-$(CONFIG_CPU_TLB_V4WT)	+= tlb-v4.o
 obj-$(CONFIG_CPU_TLB_V4WB)	+= tlb-v4wb.o

arch/arm/mm/copypage-xscale.S

-/*
- *  linux/arch/arm/lib/copypage-xscale.S
- *
- *  Copyright (C) 2001 Russell King
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 as
- * published by the Free Software Foundation.
- */
-#include <linux/linkage.h>
-#include <linux/init.h>
-#include <asm/constants.h>
-
-/*
- * General note:
- *  We don't really want write-allocate cache behaviour for these functions
- *  since that will just eat through 8K of the cache.
- */
-
-	.text
-	.align	5
-/*
- * XScale optimised copy_user_page
- *  r0 = destination
- *  r1 = source
- *  r2 = virtual user address of ultimate destination page
- *
- * The source page may have some clean entries in the cache already, but we
- * can safely ignore them - break_cow() will flush them out of the cache
- * if we eventually end up using our copied page.
- *
- * What we could do is use the mini-cache to buffer reads from the source
- * page.  We rely on the mini-cache being smaller than one page, so we'll
- * cycle through the complete cache anyway.
- */
-ENTRY(xscale_mc_copy_user_page)
-	stmfd	sp!, {r4, r5, lr}
-	mov	r5, r0
-	mov	r0, r1
-	bl	map_page_minicache
-	mov	r1, r5
-	mov	lr, #PAGE_SZ/64-1
-
-	/*
-	 * Strangely enough, best performance is achieved
-	 * when prefetching destination as well.  (NP)
-	 */
-	pld	[r0, #0]
-	pld	[r0, #32]
-	pld	[r1, #0]
-	pld	[r1, #32]
-
-1:	pld	[r0, #64]
-	pld	[r0, #96]
-	pld	[r1, #64]
-	pld	[r1, #96]
-
-2:	ldrd	r2, [r0], #8
-	ldrd	r4, [r0], #8
-	mov	ip, r1
-	strd	r2, [r1], #8
-	ldrd	r2, [r0], #8
-	strd	r4, [r1], #8
-	ldrd	r4, [r0], #8
-	strd	r2, [r1], #8
-	strd	r4, [r1], #8
-	mcr	p15, 0, ip, c7, c10, 1		@ clean D line
-	ldrd	r2, [r0], #8
-	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line
-	ldrd	r4, [r0], #8
-	mov	ip, r1
-	strd	r2, [r1], #8
-	ldrd	r2, [r0], #8
-	strd	r4, [r1], #8
-	ldrd	r4, [r0], #8
-	strd	r2, [r1], #8
-	strd	r4, [r1], #8
-	mcr	p15, 0, ip, c7, c10, 1		@ clean D line
-	subs	lr, lr, #1
-	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line
-	bgt	1b
-	beq	2b
-
-	ldmfd	sp!, {r4, r5, pc}
-
-	.align	5
-/*
- * XScale optimised clear_user_page
- *  r0 = destination
- *  r1 = virtual user address of ultimate destination page
- */
-ENTRY(xscale_mc_clear_user_page)
-	mov	r1, #PAGE_SZ/32
-	mov	r2, #0
-	mov	r3, #0
-1:	mov	ip, r0
-	strd	r2, [r0], #8
-	strd	r2, [r0], #8
-	strd	r2, [r0], #8
-	strd	r2, [r0], #8
-	mcr	p15, 0, ip, c7, c10, 1		@ clean D line
-	subs	r1, r1, #1
-	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line
-	bne	1b
-	mov	pc, lr
-
-	__INITDATA
-
-	.type	xscale_mc_user_fns, #object
-ENTRY(xscale_mc_user_fns)
-	.long	xscale_mc_clear_user_page
-	.long	xscale_mc_copy_user_page
-	.size	xscale_mc_user_fns, . - xscale_mc_user_fns

arch/arm/mm/copypage-xscale.c

+/*
+ *  linux/arch/arm/lib/copypage-xscale.S
+ *
+ *  Copyright (C) 1995-2005 Russell King
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ *
+ * This handles the mini data cache, as found on SA11x0 and XScale
+ * processors.  When we copy a user page page, we map it in such a way
+ * that accesses to this page will not touch the main data cache, but
+ * will be cached in the mini data cache.  This prevents us thrashing
+ * the main data cache on page faults.
+ */
+#include <linux/init.h>
+#include <linux/mm.h>
+
+#include <asm/page.h>
+#include <asm/pgtable.h>
+#include <asm/tlbflush.h>
+
+/*
+ * 0xffff8000 to 0xffffffff is reserved for any ARM architecture
+ * specific hacks for copying pages efficiently.
+ */
+#define COPYPAGE_MINICACHE	0xffff8000
+
+#define minicache_pgprot __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | \
+				  L_PTE_CACHEABLE)
+
+#define TOP_PTE(x)	pte_offset_kernel(top_pmd, x)
+
+static DEFINE_SPINLOCK(minicache_lock);
+
+/*
+ * XScale mini-dcache optimised copy_user_page
+ *
+ * We flush the destination cache lines just before we write the data into the
+ * corresponding address.  Since the Dcache is read-allocate, this removes the
+ * Dcache aliasing issue.  The writes will be forwarded to the write buffer,
+ * and merged as appropriate.
+ */
+static void __attribute__((naked))
+mc_copy_user_page(void *from, void *to)
+{
+	/*
+	 * Strangely enough, best performance is achieved
+	 * when prefetching destination as well.  (NP)
+	 */
+	asm volatile(
+	"stmfd	sp!, {r4, r5, lr}		\n\
+	mov	lr, %2				\n\
+	pld	[r0, #0]			\n\
+	pld	[r0, #32]			\n\
+	pld	[r1, #0]			\n\
+	pld	[r1, #32]			\n\
+1:	pld	[r0, #64]			\n\
+	pld	[r0, #96]			\n\
+	pld	[r1, #64]			\n\
+	pld	[r1, #96]			\n\
+2:	ldrd	r2, [r0], #8			\n\
+	ldrd	r4, [r0], #8			\n\
+	mov	ip, r1				\n\
+	strd	r2, [r1], #8			\n\
+	ldrd	r2, [r0], #8			\n\
+	strd	r4, [r1], #8			\n\
+	ldrd	r4, [r0], #8			\n\
+	strd	r2, [r1], #8			\n\
+	strd	r4, [r1], #8			\n\
+	mcr	p15, 0, ip, c7, c10, 1		@ clean D line\n\
+	ldrd	r2, [r0], #8			\n\
+	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line\n\
+	ldrd	r4, [r0], #8			\n\
+	mov	ip, r1				\n\
+	strd	r2, [r1], #8			\n\
+	ldrd	r2, [r0], #8			\n\
+	strd	r4, [r1], #8			\n\
+	ldrd	r4, [r0], #8			\n\
+	strd	r2, [r1], #8			\n\
+	strd	r4, [r1], #8			\n\
+	mcr	p15, 0, ip, c7, c10, 1		@ clean D line\n\
+	subs	lr, lr, #1			\n\
+	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line\n\
+	bgt	1b				\n\
+	beq	2b				\n\
+	ldmfd	sp!, {r4, r5, pc}		"
+	:
+	: "r" (from), "r" (to), "I" (PAGE_SIZE / 64 - 1));
+}
+
+void xscale_mc_copy_user_page(void *kto, const void *kfrom, unsigned long vaddr)
+{
+	spin_lock(&minicache_lock);
+
+	set_pte(TOP_PTE(COPYPAGE_MINICACHE), pfn_pte(__pa(kfrom) >> PAGE_SHIFT, minicache_pgprot));
+	flush_tlb_kernel_page(COPYPAGE_MINICACHE);
+
+	mc_copy_user_page((void *)COPYPAGE_MINICACHE, kto);
+
+	spin_unlock(&minicache_lock);
+}
+
+/*
+ * XScale optimised clear_user_page
+ */
+void __attribute__((naked))
+xscale_mc_clear_user_page(void *kaddr, unsigned long vaddr)
+{
+	asm volatile(
+	"mov	r1, %0				\n\
+	mov	r2, #0				\n\
+	mov	r3, #0				\n\
+1:	mov	ip, r0				\n\
+	strd	r2, [r0], #8			\n\
+	strd	r2, [r0], #8			\n\
+	strd	r2, [r0], #8			\n\
+	strd	r2, [r0], #8			\n\
+	mcr	p15, 0, ip, c7, c10, 1		@ clean D line\n\
+	subs	r1, r1, #1			\n\
+	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line\n\
+	bne	1b				\n\
+	mov	pc, lr"
+	:
+	: "I" (PAGE_SIZE / 32));
+}
+
+struct cpu_user_fns xscale_mc_user_fns __initdata = {
+	.cpu_clear_user_page	= xscale_mc_clear_user_page, 
+	.cpu_copy_user_page	= xscale_mc_copy_user_page,
+};

arch/arm/mm/minicache.c

-/*
- *  linux/arch/arm/mm/minicache.c
- *
- *  Copyright (C) 2001 Russell King
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 as
- * published by the Free Software Foundation.
- *
- * This handles the mini data cache, as found on SA11x0 and XScale
- * processors.  When we copy a user page page, we map it in such a way
- * that accesses to this page will not touch the main data cache, but
- * will be cached in the mini data cache.  This prevents us thrashing
- * the main data cache on page faults.
- */
-#include <linux/init.h>
-#include <linux/mm.h>
-
-#include <asm/page.h>
-#include <asm/pgtable.h>
-#include <asm/tlbflush.h>
-
-/*
- * 0xffff8000 to 0xffffffff is reserved for any ARM architecture
- * specific hacks for copying pages efficiently.
- */
-#define minicache_address (0xffff8000)
-#define minicache_pgprot __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | \
-				  L_PTE_CACHEABLE)
-
-static pte_t *minicache_pte;
-
-/*
- * Note that this is intended to be called only from the copy_user_page
- * asm code; anything else will require special locking to prevent the
- * mini-cache space being re-used.  (Note: probably preempt unsafe).
- *
- * We rely on the fact that the minicache is 2K, and we'll be pushing
- * 4K of data through it, so we don't actually have to specifically
- * flush the minicache when we change the mapping.
- *
- * Note also: assert(PAGE_OFFSET <= virt < high_memory).
- * Unsafe: preempt, kmap.
- */
-unsigned long map_page_minicache(unsigned long virt)
-{
-	set_pte(minicache_pte, pfn_pte(__pa(virt) >> PAGE_SHIFT, minicache_pgprot));
-	flush_tlb_kernel_page(minicache_address);
-
-	return minicache_address;
-}
-
-static int __init minicache_init(void)
-{
-	pgd_t *pgd;
-	pmd_t *pmd;
-
-	spin_lock(&init_mm.page_table_lock);
-
-	pgd = pgd_offset_k(minicache_address);
-	pmd = pmd_alloc(&init_mm, pgd, minicache_address);
-	if (!pmd)
-		BUG();
-	minicache_pte = pte_alloc_kernel(&init_mm, pmd, minicache_address);
-	if (!minicache_pte)
-		BUG();
-
-	spin_unlock(&init_mm.page_table_lock);
-
-	return 0;
-}
-
-core_initcall(minicache_init);