1 /* 2 * linux/arch/arm/lib/copypage-xscale.S 3 * 4 * Copyright (C) 1995-2005 Russell King 5 * 6 * This program is free software; you can redistribute it and/or modify 7 * it under the terms of the GNU General Public License version 2 as 8 * published by the Free Software Foundation. 9 * 10 * This handles the mini data cache, as found on SA11x0 and XScale 11 * processors. When we copy a user page page, we map it in such a way 12 * that accesses to this page will not touch the main data cache, but 13 * will be cached in the mini data cache. This prevents us thrashing 14 * the main data cache on page faults. 15 */ 16 #include <linux/init.h> 17 #include <linux/mm.h> 18 #include <linux/highmem.h> 19 20 #include <asm/pgtable.h> 21 #include <asm/tlbflush.h> 22 #include <asm/cacheflush.h> 23 24 #include "mm.h" 25 26 #define minicache_pgprot __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | \ 27 L_PTE_MT_MINICACHE) 28 29 static DEFINE_RAW_SPINLOCK(minicache_lock); 30 31 /* 32 * XScale mini-dcache optimised copy_user_highpage 33 * 34 * We flush the destination cache lines just before we write the data into the 35 * corresponding address. Since the Dcache is read-allocate, this removes the 36 * Dcache aliasing issue. The writes will be forwarded to the write buffer, 37 * and merged as appropriate. 38 */ 39 static void mc_copy_user_page(void *from, void *to) 40 { 41 int tmp; 42 43 /* 44 * Strangely enough, best performance is achieved 45 * when prefetching destination as well. (NP) 46 */ 47 asm volatile ("\ 48 pld [%0, #0] \n\ 49 pld [%0, #32] \n\ 50 pld [%1, #0] \n\ 51 pld [%1, #32] \n\ 52 1: pld [%0, #64] \n\ 53 pld [%0, #96] \n\ 54 pld [%1, #64] \n\ 55 pld [%1, #96] \n\ 56 2: ldrd r2, [%0], #8 \n\ 57 ldrd r4, [%0], #8 \n\ 58 mov ip, %1 \n\ 59 strd r2, [%1], #8 \n\ 60 ldrd r2, [%0], #8 \n\ 61 strd r4, [%1], #8 \n\ 62 ldrd r4, [%0], #8 \n\ 63 strd r2, [%1], #8 \n\ 64 strd r4, [%1], #8 \n\ 65 mcr p15, 0, ip, c7, c10, 1 @ clean D line\n\ 66 ldrd r2, [%0], #8 \n\ 67 mcr p15, 0, ip, c7, c6, 1 @ invalidate D line\n\ 68 ldrd r4, [%0], #8 \n\ 69 mov ip, %1 \n\ 70 strd r2, [%1], #8 \n\ 71 ldrd r2, [%0], #8 \n\ 72 strd r4, [%1], #8 \n\ 73 ldrd r4, [%0], #8 \n\ 74 strd r2, [%1], #8 \n\ 75 strd r4, [%1], #8 \n\ 76 mcr p15, 0, ip, c7, c10, 1 @ clean D line\n\ 77 subs %2, %2, #1 \n\ 78 mcr p15, 0, ip, c7, c6, 1 @ invalidate D line\n\ 79 bgt 1b \n\ 80 beq 2b " 81 : "+&r" (from), "+&r" (to), "=&r" (tmp) 82 : "2" (PAGE_SIZE / 64 - 1) 83 : "r2", "r3", "r4", "r5", "ip"); 84 } 85 86 void xscale_mc_copy_user_highpage(struct page *to, struct page *from, 87 unsigned long vaddr, struct vm_area_struct *vma) 88 { 89 void *kto = kmap_atomic(to); 90 91 if (!test_and_set_bit(PG_dcache_clean, &from->flags)) 92 __flush_dcache_page(page_mapping_file(from), from); 93 94 raw_spin_lock(&minicache_lock); 95 96 set_top_pte(COPYPAGE_MINICACHE, mk_pte(from, minicache_pgprot)); 97 98 mc_copy_user_page((void *)COPYPAGE_MINICACHE, kto); 99 100 raw_spin_unlock(&minicache_lock); 101 102 kunmap_atomic(kto); 103 } 104 105 /* 106 * XScale optimised clear_user_page 107 */ 108 void 109 xscale_mc_clear_user_highpage(struct page *page, unsigned long vaddr) 110 { 111 void *ptr, *kaddr = kmap_atomic(page); 112 asm volatile( 113 "mov r1, %2 \n\ 114 mov r2, #0 \n\ 115 mov r3, #0 \n\ 116 1: mov ip, %0 \n\ 117 strd r2, [%0], #8 \n\ 118 strd r2, [%0], #8 \n\ 119 strd r2, [%0], #8 \n\ 120 strd r2, [%0], #8 \n\ 121 mcr p15, 0, ip, c7, c10, 1 @ clean D line\n\ 122 subs r1, r1, #1 \n\ 123 mcr p15, 0, ip, c7, c6, 1 @ invalidate D line\n\ 124 bne 1b" 125 : "=r" (ptr) 126 : "0" (kaddr), "I" (PAGE_SIZE / 32) 127 : "r1", "r2", "r3", "ip"); 128 kunmap_atomic(kaddr); 129 } 130 131 struct cpu_user_fns xscale_mc_user_fns __initdata = { 132 .cpu_clear_user_highpage = xscale_mc_clear_user_highpage, 133 .cpu_copy_user_highpage = xscale_mc_copy_user_highpage, 134 }; 135