xref: /linux/arch/arm/mm/copypage-xscale.c (revision 8083f3d78825c0ea1948339613914b46105bfd0b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/arch/arm/lib/copypage-xscale.S
4  *
5  *  Copyright (C) 1995-2005 Russell King
6  *
7  * This handles the mini data cache, as found on SA11x0 and XScale
8  * processors.  When we copy a user page page, we map it in such a way
9  * that accesses to this page will not touch the main data cache, but
10  * will be cached in the mini data cache.  This prevents us thrashing
11  * the main data cache on page faults.
12  */
13 #include <linux/init.h>
14 #include <linux/mm.h>
15 #include <linux/highmem.h>
16 
17 #include <asm/pgtable.h>
18 #include <asm/tlbflush.h>
19 #include <asm/cacheflush.h>
20 
21 #include "mm.h"
22 
23 #define minicache_pgprot __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | \
24 				  L_PTE_MT_MINICACHE)
25 
26 static DEFINE_RAW_SPINLOCK(minicache_lock);
27 
28 /*
29  * XScale mini-dcache optimised copy_user_highpage
30  *
31  * We flush the destination cache lines just before we write the data into the
32  * corresponding address.  Since the Dcache is read-allocate, this removes the
33  * Dcache aliasing issue.  The writes will be forwarded to the write buffer,
34  * and merged as appropriate.
35  */
36 static void mc_copy_user_page(void *from, void *to)
37 {
38 	int tmp;
39 
40 	/*
41 	 * Strangely enough, best performance is achieved
42 	 * when prefetching destination as well.  (NP)
43 	 */
44 	asm volatile ("\
45 	pld	[%0, #0]			\n\
46 	pld	[%0, #32]			\n\
47 	pld	[%1, #0]			\n\
48 	pld	[%1, #32]			\n\
49 1:	pld	[%0, #64]			\n\
50 	pld	[%0, #96]			\n\
51 	pld	[%1, #64]			\n\
52 	pld	[%1, #96]			\n\
53 2:	ldrd	r2, r3, [%0], #8		\n\
54 	ldrd	r4, r5, [%0], #8		\n\
55 	mov	ip, %1				\n\
56 	strd	r2, r3, [%1], #8		\n\
57 	ldrd	r2, r3, [%0], #8		\n\
58 	strd	r4, r5, [%1], #8		\n\
59 	ldrd	r4, r5, [%0], #8		\n\
60 	strd	r2, r3, [%1], #8		\n\
61 	strd	r4, r5, [%1], #8		\n\
62 	mcr	p15, 0, ip, c7, c10, 1		@ clean D line\n\
63 	ldrd	r2, r3, [%0], #8		\n\
64 	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line\n\
65 	ldrd	r4, r5, [%0], #8		\n\
66 	mov	ip, %1				\n\
67 	strd	r2, r3, [%1], #8		\n\
68 	ldrd	r2, r3, [%0], #8		\n\
69 	strd	r4, r5, [%1], #8		\n\
70 	ldrd	r4, r5, [%0], #8		\n\
71 	strd	r2, r3, [%1], #8		\n\
72 	strd	r4, r5, [%1], #8		\n\
73 	mcr	p15, 0, ip, c7, c10, 1		@ clean D line\n\
74 	subs	%2, %2, #1			\n\
75 	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line\n\
76 	bgt	1b				\n\
77 	beq	2b				"
78 	: "+&r" (from), "+&r" (to), "=&r" (tmp)
79 	: "2" (PAGE_SIZE / 64 - 1)
80 	: "r2", "r3", "r4", "r5", "ip");
81 }
82 
83 void xscale_mc_copy_user_highpage(struct page *to, struct page *from,
84 	unsigned long vaddr, struct vm_area_struct *vma)
85 {
86 	void *kto = kmap_atomic(to);
87 
88 	if (!test_and_set_bit(PG_dcache_clean, &from->flags))
89 		__flush_dcache_page(page_mapping_file(from), from);
90 
91 	raw_spin_lock(&minicache_lock);
92 
93 	set_top_pte(COPYPAGE_MINICACHE, mk_pte(from, minicache_pgprot));
94 
95 	mc_copy_user_page((void *)COPYPAGE_MINICACHE, kto);
96 
97 	raw_spin_unlock(&minicache_lock);
98 
99 	kunmap_atomic(kto);
100 }
101 
102 /*
103  * XScale optimised clear_user_page
104  */
105 void
106 xscale_mc_clear_user_highpage(struct page *page, unsigned long vaddr)
107 {
108 	void *ptr, *kaddr = kmap_atomic(page);
109 	asm volatile(
110 	"mov	r1, %2				\n\
111 	mov	r2, #0				\n\
112 	mov	r3, #0				\n\
113 1:	mov	ip, %0				\n\
114 	strd	r2, r3, [%0], #8		\n\
115 	strd	r2, r3, [%0], #8		\n\
116 	strd	r2, r3, [%0], #8		\n\
117 	strd	r2, r3, [%0], #8		\n\
118 	mcr	p15, 0, ip, c7, c10, 1		@ clean D line\n\
119 	subs	r1, r1, #1			\n\
120 	mcr	p15, 0, ip, c7, c6, 1		@ invalidate D line\n\
121 	bne	1b"
122 	: "=r" (ptr)
123 	: "0" (kaddr), "I" (PAGE_SIZE / 32)
124 	: "r1", "r2", "r3", "ip");
125 	kunmap_atomic(kaddr);
126 }
127 
128 struct cpu_user_fns xscale_mc_user_fns __initdata = {
129 	.cpu_clear_user_highpage = xscale_mc_clear_user_highpage,
130 	.cpu_copy_user_highpage	= xscale_mc_copy_user_highpage,
131 };
132