arch/arm/mm/copypage-xscale.c

   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 /*
  27  * 0xffff8000 to 0xffffffff is reserved for any ARM architecture
  28  * specific hacks for copying pages efficiently.
  29  */
  30 #define COPYPAGE_MINICACHE      0xffff8000
  31
  32 #define minicache_pgprot __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | \
  33                                   L_PTE_MT_MINICACHE)
  34
  35 static DEFINE_SPINLOCK(minicache_lock);
  36
  37 /*
  38  * XScale mini-dcache optimised copy_user_highpage
  39  *
  40  * We flush the destination cache lines just before we write the data into the
  41  * corresponding address.  Since the Dcache is read-allocate, this removes the
  42  * Dcache aliasing issue.  The writes will be forwarded to the write buffer,
  43  * and merged as appropriate.
  44  */
  45 static void __naked
  46 mc_copy_user_page(void *from, void *to)
  47 {
  48         /*
  49          * Strangely enough, best performance is achieved
  50          * when prefetching destination as well.  (NP)
  51          */
  52         asm volatile(
  53         "stmfd  sp!, {r4, r5, lr}               \n\
  54         mov     lr, %2                          \n\
  55         pld     [r0, #0]                        \n\
  56         pld     [r0, #32]                       \n\
  57         pld     [r1, #0]                        \n\
  58         pld     [r1, #32]                       \n\
  59 1:      pld     [r0, #64]                       \n\
  60         pld     [r0, #96]                       \n\
  61         pld     [r1, #64]                       \n\
  62         pld     [r1, #96]                       \n\
  63 2:      ldrd    r2, [r0], #8                    \n\
  64         ldrd    r4, [r0], #8                    \n\
  65         mov     ip, r1                          \n\
  66         strd    r2, [r1], #8                    \n\
  67         ldrd    r2, [r0], #8                    \n\
  68         strd    r4, [r1], #8                    \n\
  69         ldrd    r4, [r0], #8                    \n\
  70         strd    r2, [r1], #8                    \n\
  71         strd    r4, [r1], #8                    \n\
  72         mcr     p15, 0, ip, c7, c10, 1          @ clean D line\n\
  73         ldrd    r2, [r0], #8                    \n\
  74         mcr     p15, 0, ip, c7, c6, 1           @ invalidate D line\n\
  75         ldrd    r4, [r0], #8                    \n\
  76         mov     ip, r1                          \n\
  77         strd    r2, [r1], #8                    \n\
  78         ldrd    r2, [r0], #8                    \n\
  79         strd    r4, [r1], #8                    \n\
  80         ldrd    r4, [r0], #8                    \n\
  81         strd    r2, [r1], #8                    \n\
  82         strd    r4, [r1], #8                    \n\
  83         mcr     p15, 0, ip, c7, c10, 1          @ clean D line\n\
  84         subs    lr, lr, #1                      \n\
  85         mcr     p15, 0, ip, c7, c6, 1           @ invalidate D line\n\
  86         bgt     1b                              \n\
  87         beq     2b                              \n\
  88         ldmfd   sp!, {r4, r5, pc}               "
  89         :
  90         : "r" (from), "r" (to), "I" (PAGE_SIZE / 64 - 1));
  91 }
  92
  93 void xscale_mc_copy_user_highpage(struct page *to, struct page *from,
  94         unsigned long vaddr)
  95 {
  96         void *kto = kmap_atomic(to, KM_USER1);
  97
  98         if (test_and_clear_bit(PG_dcache_dirty, &from->flags))
  99                 __flush_dcache_page(page_mapping(from), from);
 100
 101         spin_lock(&minicache_lock);
 102
 103         set_pte_ext(TOP_PTE(COPYPAGE_MINICACHE), pfn_pte(page_to_pfn(from), minicache_pgprot), 0);
 104         flush_tlb_kernel_page(COPYPAGE_MINICACHE);
 105
 106         mc_copy_user_page((void *)COPYPAGE_MINICACHE, kto);
 107
 108         spin_unlock(&minicache_lock);
 109
 110         kunmap_atomic(kto, KM_USER1);
 111 }
 112
 113 /*
 114  * XScale optimised clear_user_page
 115  */
 116 void
 117 xscale_mc_clear_user_highpage(struct page *page, unsigned long vaddr)
 118 {
 119         void *ptr, *kaddr = kmap_atomic(page, KM_USER0);
 120         asm volatile(
 121         "mov    r1, %2                          \n\
 122         mov     r2, #0                          \n\
 123         mov     r3, #0                          \n\
 124 1:      mov     ip, %0                          \n\
 125         strd    r2, [%0], #8                    \n\
 126         strd    r2, [%0], #8                    \n\
 127         strd    r2, [%0], #8                    \n\
 128         strd    r2, [%0], #8                    \n\
 129         mcr     p15, 0, ip, c7, c10, 1          @ clean D line\n\
 130         subs    r1, r1, #1                      \n\
 131         mcr     p15, 0, ip, c7, c6, 1           @ invalidate D line\n\
 132         bne     1b"
 133         : "=r" (ptr)
 134         : "0" (kaddr), "I" (PAGE_SIZE / 32)
 135         : "r1", "r2", "r3", "ip");
 136         kunmap_atomic(kaddr, KM_USER0);
 137 }
 138
 139 struct cpu_user_fns xscale_mc_user_fns __initdata = {
 140         .cpu_clear_user_highpage = xscale_mc_clear_user_highpage,
 141         .cpu_copy_user_highpage = xscale_mc_copy_user_highpage,
 142 };