1 /* 2 * linux/arch/arm/lib/uaccess_with_memcpy.c 3 * 4 * Written by: Lennert Buytenhek and Nicolas Pitre 5 * Copyright (C) 2009 Marvell Semiconductor 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License version 2 as 9 * published by the Free Software Foundation. 10 */ 11 12 #include <linux/kernel.h> 13 #include <linux/ctype.h> 14 #include <linux/uaccess.h> 15 #include <linux/rwsem.h> 16 #include <linux/mm.h> 17 #include <linux/sched.h> 18 #include <linux/hardirq.h> /* for in_atomic() */ 19 #include <linux/gfp.h> 20 #include <linux/highmem.h> 21 #include <linux/hugetlb.h> 22 #include <linux/export.h> 23 #include <asm/current.h> 24 #include <asm/page.h> 25 26 static int 27 pin_page_for_write(const void __user *_addr, pte_t **ptep, spinlock_t **ptlp) 28 { 29 unsigned long addr = (unsigned long)_addr; 30 pgd_t *pgd; 31 pmd_t *pmd; 32 pte_t *pte; 33 pud_t *pud; 34 spinlock_t *ptl; 35 36 pgd = pgd_offset(current->mm, addr); 37 if (unlikely(pgd_none(*pgd) || pgd_bad(*pgd))) 38 return 0; 39 40 pud = pud_offset(pgd, addr); 41 if (unlikely(pud_none(*pud) || pud_bad(*pud))) 42 return 0; 43 44 pmd = pmd_offset(pud, addr); 45 if (unlikely(pmd_none(*pmd))) 46 return 0; 47 48 /* 49 * A pmd can be bad if it refers to a HugeTLB or THP page. 50 * 51 * Both THP and HugeTLB pages have the same pmd layout 52 * and should not be manipulated by the pte functions. 53 * 54 * Lock the page table for the destination and check 55 * to see that it's still huge and whether or not we will 56 * need to fault on write. 57 */ 58 if (unlikely(pmd_thp_or_huge(*pmd))) { 59 ptl = ¤t->mm->page_table_lock; 60 spin_lock(ptl); 61 if (unlikely(!pmd_thp_or_huge(*pmd) 62 || pmd_hugewillfault(*pmd))) { 63 spin_unlock(ptl); 64 return 0; 65 } 66 67 *ptep = NULL; 68 *ptlp = ptl; 69 return 1; 70 } 71 72 if (unlikely(pmd_bad(*pmd))) 73 return 0; 74 75 pte = pte_offset_map_lock(current->mm, pmd, addr, &ptl); 76 if (unlikely(!pte_present(*pte) || !pte_young(*pte) || 77 !pte_write(*pte) || !pte_dirty(*pte))) { 78 pte_unmap_unlock(pte, ptl); 79 return 0; 80 } 81 82 *ptep = pte; 83 *ptlp = ptl; 84 85 return 1; 86 } 87 88 static unsigned long noinline 89 __copy_to_user_memcpy(void __user *to, const void *from, unsigned long n) 90 { 91 unsigned long ua_flags; 92 int atomic; 93 94 if (unlikely(segment_eq(get_fs(), KERNEL_DS))) { 95 memcpy((void *)to, from, n); 96 return 0; 97 } 98 99 /* the mmap semaphore is taken only if not in an atomic context */ 100 atomic = faulthandler_disabled(); 101 102 if (!atomic) 103 down_read(¤t->mm->mmap_sem); 104 while (n) { 105 pte_t *pte; 106 spinlock_t *ptl; 107 int tocopy; 108 109 while (!pin_page_for_write(to, &pte, &ptl)) { 110 if (!atomic) 111 up_read(¤t->mm->mmap_sem); 112 if (__put_user(0, (char __user *)to)) 113 goto out; 114 if (!atomic) 115 down_read(¤t->mm->mmap_sem); 116 } 117 118 tocopy = (~(unsigned long)to & ~PAGE_MASK) + 1; 119 if (tocopy > n) 120 tocopy = n; 121 122 ua_flags = uaccess_save_and_enable(); 123 memcpy((void *)to, from, tocopy); 124 uaccess_restore(ua_flags); 125 to += tocopy; 126 from += tocopy; 127 n -= tocopy; 128 129 if (pte) 130 pte_unmap_unlock(pte, ptl); 131 else 132 spin_unlock(ptl); 133 } 134 if (!atomic) 135 up_read(¤t->mm->mmap_sem); 136 137 out: 138 return n; 139 } 140 141 unsigned long 142 arm_copy_to_user(void __user *to, const void *from, unsigned long n) 143 { 144 /* 145 * This test is stubbed out of the main function above to keep 146 * the overhead for small copies low by avoiding a large 147 * register dump on the stack just to reload them right away. 148 * With frame pointer disabled, tail call optimization kicks in 149 * as well making this test almost invisible. 150 */ 151 if (n < 64) { 152 unsigned long ua_flags = uaccess_save_and_enable(); 153 n = __copy_to_user_std(to, from, n); 154 uaccess_restore(ua_flags); 155 } else { 156 n = __copy_to_user_memcpy(to, from, n); 157 } 158 return n; 159 } 160 EXPORT_SYMBOL(arm_copy_to_user); 161 162 static unsigned long noinline 163 __clear_user_memset(void __user *addr, unsigned long n) 164 { 165 unsigned long ua_flags; 166 167 if (unlikely(segment_eq(get_fs(), KERNEL_DS))) { 168 memset((void *)addr, 0, n); 169 return 0; 170 } 171 172 down_read(¤t->mm->mmap_sem); 173 while (n) { 174 pte_t *pte; 175 spinlock_t *ptl; 176 int tocopy; 177 178 while (!pin_page_for_write(addr, &pte, &ptl)) { 179 up_read(¤t->mm->mmap_sem); 180 if (__put_user(0, (char __user *)addr)) 181 goto out; 182 down_read(¤t->mm->mmap_sem); 183 } 184 185 tocopy = (~(unsigned long)addr & ~PAGE_MASK) + 1; 186 if (tocopy > n) 187 tocopy = n; 188 189 ua_flags = uaccess_save_and_enable(); 190 memset((void *)addr, 0, tocopy); 191 uaccess_restore(ua_flags); 192 addr += tocopy; 193 n -= tocopy; 194 195 if (pte) 196 pte_unmap_unlock(pte, ptl); 197 else 198 spin_unlock(ptl); 199 } 200 up_read(¤t->mm->mmap_sem); 201 202 out: 203 return n; 204 } 205 206 unsigned long arm_clear_user(void __user *addr, unsigned long n) 207 { 208 /* See rational for this in __copy_to_user() above. */ 209 if (n < 64) { 210 unsigned long ua_flags = uaccess_save_and_enable(); 211 n = __clear_user_std(addr, n); 212 uaccess_restore(ua_flags); 213 } else { 214 n = __clear_user_memset(addr, n); 215 } 216 return n; 217 } 218 EXPORT_SYMBOL(arm_clear_user); 219 220 #if 0 221 222 /* 223 * This code is disabled by default, but kept around in case the chosen 224 * thresholds need to be revalidated. Some overhead (small but still) 225 * would be implied by a runtime determined variable threshold, and 226 * so far the measurement on concerned targets didn't show a worthwhile 227 * variation. 228 * 229 * Note that a fairly precise sched_clock() implementation is needed 230 * for results to make some sense. 231 */ 232 233 #include <linux/vmalloc.h> 234 235 static int __init test_size_treshold(void) 236 { 237 struct page *src_page, *dst_page; 238 void *user_ptr, *kernel_ptr; 239 unsigned long long t0, t1, t2; 240 int size, ret; 241 242 ret = -ENOMEM; 243 src_page = alloc_page(GFP_KERNEL); 244 if (!src_page) 245 goto no_src; 246 dst_page = alloc_page(GFP_KERNEL); 247 if (!dst_page) 248 goto no_dst; 249 kernel_ptr = page_address(src_page); 250 user_ptr = vmap(&dst_page, 1, VM_IOREMAP, __pgprot(__P010)); 251 if (!user_ptr) 252 goto no_vmap; 253 254 /* warm up the src page dcache */ 255 ret = __copy_to_user_memcpy(user_ptr, kernel_ptr, PAGE_SIZE); 256 257 for (size = PAGE_SIZE; size >= 4; size /= 2) { 258 t0 = sched_clock(); 259 ret |= __copy_to_user_memcpy(user_ptr, kernel_ptr, size); 260 t1 = sched_clock(); 261 ret |= __copy_to_user_std(user_ptr, kernel_ptr, size); 262 t2 = sched_clock(); 263 printk("copy_to_user: %d %llu %llu\n", size, t1 - t0, t2 - t1); 264 } 265 266 for (size = PAGE_SIZE; size >= 4; size /= 2) { 267 t0 = sched_clock(); 268 ret |= __clear_user_memset(user_ptr, size); 269 t1 = sched_clock(); 270 ret |= __clear_user_std(user_ptr, size); 271 t2 = sched_clock(); 272 printk("clear_user: %d %llu %llu\n", size, t1 - t0, t2 - t1); 273 } 274 275 if (ret) 276 ret = -EFAULT; 277 278 vunmap(user_ptr); 279 no_vmap: 280 put_page(dst_page); 281 no_dst: 282 put_page(src_page); 283 no_src: 284 return ret; 285 } 286 287 subsys_initcall(test_size_treshold); 288 289 #endif 290