Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * mm/mremap.c |
| 3 | * |
| 4 | * (C) Copyright 1996 Linus Torvalds |
| 5 | * |
| 6 | * Address space accounting code <alan@redhat.com> |
| 7 | * (C) Copyright 2002 Red Hat Inc, All Rights Reserved |
| 8 | */ |
| 9 | |
| 10 | #include <linux/mm.h> |
| 11 | #include <linux/hugetlb.h> |
| 12 | #include <linux/slab.h> |
| 13 | #include <linux/shm.h> |
| 14 | #include <linux/mman.h> |
| 15 | #include <linux/swap.h> |
| 16 | #include <linux/fs.h> |
| 17 | #include <linux/highmem.h> |
| 18 | #include <linux/security.h> |
| 19 | #include <linux/syscalls.h> |
| 20 | |
| 21 | #include <asm/uaccess.h> |
| 22 | #include <asm/cacheflush.h> |
| 23 | #include <asm/tlbflush.h> |
| 24 | |
| 25 | static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr) |
| 26 | { |
| 27 | pgd_t *pgd; |
| 28 | pud_t *pud; |
| 29 | pmd_t *pmd; |
| 30 | pte_t *pte = NULL; |
| 31 | |
| 32 | pgd = pgd_offset(mm, addr); |
| 33 | if (pgd_none_or_clear_bad(pgd)) |
| 34 | goto end; |
| 35 | |
| 36 | pud = pud_offset(pgd, addr); |
| 37 | if (pud_none_or_clear_bad(pud)) |
| 38 | goto end; |
| 39 | |
| 40 | pmd = pmd_offset(pud, addr); |
| 41 | if (pmd_none_or_clear_bad(pmd)) |
| 42 | goto end; |
| 43 | |
| 44 | pte = pte_offset_map_nested(pmd, addr); |
| 45 | if (pte_none(*pte)) { |
| 46 | pte_unmap_nested(pte); |
| 47 | pte = NULL; |
| 48 | } |
| 49 | end: |
| 50 | return pte; |
| 51 | } |
| 52 | |
| 53 | static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr) |
| 54 | { |
| 55 | pgd_t *pgd; |
| 56 | pud_t *pud; |
| 57 | pmd_t *pmd; |
| 58 | |
| 59 | pgd = pgd_offset(mm, addr); |
| 60 | if (pgd_none_or_clear_bad(pgd)) |
| 61 | return NULL; |
| 62 | |
| 63 | pud = pud_offset(pgd, addr); |
| 64 | if (pud_none_or_clear_bad(pud)) |
| 65 | return NULL; |
| 66 | |
| 67 | pmd = pmd_offset(pud, addr); |
| 68 | if (pmd_none_or_clear_bad(pmd)) |
| 69 | return NULL; |
| 70 | |
| 71 | return pte_offset_map(pmd, addr); |
| 72 | } |
| 73 | |
| 74 | static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr) |
| 75 | { |
| 76 | pgd_t *pgd; |
| 77 | pud_t *pud; |
| 78 | pmd_t *pmd; |
| 79 | pte_t *pte = NULL; |
| 80 | |
| 81 | pgd = pgd_offset(mm, addr); |
| 82 | |
| 83 | pud = pud_alloc(mm, pgd, addr); |
| 84 | if (!pud) |
| 85 | return NULL; |
| 86 | pmd = pmd_alloc(mm, pud, addr); |
| 87 | if (pmd) |
| 88 | pte = pte_alloc_map(mm, pmd, addr); |
| 89 | return pte; |
| 90 | } |
| 91 | |
| 92 | static int |
| 93 | move_one_page(struct vm_area_struct *vma, unsigned long old_addr, |
| 94 | struct vm_area_struct *new_vma, unsigned long new_addr) |
| 95 | { |
| 96 | struct address_space *mapping = NULL; |
| 97 | struct mm_struct *mm = vma->vm_mm; |
| 98 | int error = 0; |
| 99 | pte_t *src, *dst; |
| 100 | |
| 101 | if (vma->vm_file) { |
| 102 | /* |
| 103 | * Subtle point from Rajesh Venkatasubramanian: before |
| 104 | * moving file-based ptes, we must lock vmtruncate out, |
| 105 | * since it might clean the dst vma before the src vma, |
| 106 | * and we propagate stale pages into the dst afterward. |
| 107 | */ |
| 108 | mapping = vma->vm_file->f_mapping; |
| 109 | spin_lock(&mapping->i_mmap_lock); |
| 110 | if (new_vma->vm_truncate_count && |
| 111 | new_vma->vm_truncate_count != vma->vm_truncate_count) |
| 112 | new_vma->vm_truncate_count = 0; |
| 113 | } |
| 114 | spin_lock(&mm->page_table_lock); |
| 115 | |
| 116 | src = get_one_pte_map_nested(mm, old_addr); |
| 117 | if (src) { |
| 118 | /* |
| 119 | * Look to see whether alloc_one_pte_map needs to perform a |
| 120 | * memory allocation. If it does then we need to drop the |
| 121 | * atomic kmap |
| 122 | */ |
| 123 | dst = get_one_pte_map(mm, new_addr); |
| 124 | if (unlikely(!dst)) { |
| 125 | pte_unmap_nested(src); |
| 126 | if (mapping) |
| 127 | spin_unlock(&mapping->i_mmap_lock); |
| 128 | dst = alloc_one_pte_map(mm, new_addr); |
| 129 | if (mapping && !spin_trylock(&mapping->i_mmap_lock)) { |
| 130 | spin_unlock(&mm->page_table_lock); |
| 131 | spin_lock(&mapping->i_mmap_lock); |
| 132 | spin_lock(&mm->page_table_lock); |
| 133 | } |
| 134 | src = get_one_pte_map_nested(mm, old_addr); |
| 135 | } |
| 136 | /* |
| 137 | * Since alloc_one_pte_map can drop and re-acquire |
| 138 | * page_table_lock, we should re-check the src entry... |
| 139 | */ |
| 140 | if (src) { |
| 141 | if (dst) { |
| 142 | pte_t pte; |
| 143 | pte = ptep_clear_flush(vma, old_addr, src); |
Nick Piggin | 8b1f312 | 2005-09-27 21:45:18 -0700 | [diff] [blame] | 144 | |
Nick Piggin | 9a61c34 | 2005-09-03 15:54:49 -0700 | [diff] [blame] | 145 | /* ZERO_PAGE can be dependant on virtual addr */ |
Nick Piggin | 8b1f312 | 2005-09-27 21:45:18 -0700 | [diff] [blame] | 146 | pte = move_pte(pte, new_vma->vm_page_prot, |
| 147 | old_addr, new_addr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 148 | set_pte_at(mm, new_addr, dst, pte); |
| 149 | } else |
| 150 | error = -ENOMEM; |
| 151 | pte_unmap_nested(src); |
| 152 | } |
| 153 | if (dst) |
| 154 | pte_unmap(dst); |
| 155 | } |
| 156 | spin_unlock(&mm->page_table_lock); |
| 157 | if (mapping) |
| 158 | spin_unlock(&mapping->i_mmap_lock); |
| 159 | return error; |
| 160 | } |
| 161 | |
| 162 | static unsigned long move_page_tables(struct vm_area_struct *vma, |
| 163 | unsigned long old_addr, struct vm_area_struct *new_vma, |
| 164 | unsigned long new_addr, unsigned long len) |
| 165 | { |
| 166 | unsigned long offset; |
| 167 | |
| 168 | flush_cache_range(vma, old_addr, old_addr + len); |
| 169 | |
| 170 | /* |
| 171 | * This is not the clever way to do this, but we're taking the |
| 172 | * easy way out on the assumption that most remappings will be |
| 173 | * only a few pages.. This also makes error recovery easier. |
| 174 | */ |
| 175 | for (offset = 0; offset < len; offset += PAGE_SIZE) { |
| 176 | if (move_one_page(vma, old_addr + offset, |
| 177 | new_vma, new_addr + offset) < 0) |
| 178 | break; |
| 179 | cond_resched(); |
| 180 | } |
| 181 | return offset; |
| 182 | } |
| 183 | |
| 184 | static unsigned long move_vma(struct vm_area_struct *vma, |
| 185 | unsigned long old_addr, unsigned long old_len, |
| 186 | unsigned long new_len, unsigned long new_addr) |
| 187 | { |
| 188 | struct mm_struct *mm = vma->vm_mm; |
| 189 | struct vm_area_struct *new_vma; |
| 190 | unsigned long vm_flags = vma->vm_flags; |
| 191 | unsigned long new_pgoff; |
| 192 | unsigned long moved_len; |
| 193 | unsigned long excess = 0; |
| 194 | int split = 0; |
| 195 | |
| 196 | /* |
| 197 | * We'd prefer to avoid failure later on in do_munmap: |
| 198 | * which may split one vma into three before unmapping. |
| 199 | */ |
| 200 | if (mm->map_count >= sysctl_max_map_count - 3) |
| 201 | return -ENOMEM; |
| 202 | |
| 203 | new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT); |
| 204 | new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff); |
| 205 | if (!new_vma) |
| 206 | return -ENOMEM; |
| 207 | |
| 208 | moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len); |
| 209 | if (moved_len < old_len) { |
| 210 | /* |
| 211 | * On error, move entries back from new area to old, |
| 212 | * which will succeed since page tables still there, |
| 213 | * and then proceed to unmap new area instead of old. |
| 214 | */ |
| 215 | move_page_tables(new_vma, new_addr, vma, old_addr, moved_len); |
| 216 | vma = new_vma; |
| 217 | old_len = new_len; |
| 218 | old_addr = new_addr; |
| 219 | new_addr = -ENOMEM; |
| 220 | } |
| 221 | |
| 222 | /* Conceal VM_ACCOUNT so old reservation is not undone */ |
| 223 | if (vm_flags & VM_ACCOUNT) { |
| 224 | vma->vm_flags &= ~VM_ACCOUNT; |
| 225 | excess = vma->vm_end - vma->vm_start - old_len; |
| 226 | if (old_addr > vma->vm_start && |
| 227 | old_addr + old_len < vma->vm_end) |
| 228 | split = 1; |
| 229 | } |
| 230 | |
Kirill Korotaev | 7179906 | 2005-05-16 21:53:18 -0700 | [diff] [blame] | 231 | /* |
| 232 | * if we failed to move page tables we still do total_vm increment |
| 233 | * since do_munmap() will decrement it by old_len == new_len |
| 234 | */ |
| 235 | mm->total_vm += new_len >> PAGE_SHIFT; |
Hugh Dickins | ab50b8e | 2005-10-29 18:15:56 -0700 | [diff] [blame] | 236 | vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT); |
Kirill Korotaev | 7179906 | 2005-05-16 21:53:18 -0700 | [diff] [blame] | 237 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 238 | if (do_munmap(mm, old_addr, old_len) < 0) { |
| 239 | /* OOM: unable to split vma, just get accounts right */ |
| 240 | vm_unacct_memory(excess >> PAGE_SHIFT); |
| 241 | excess = 0; |
| 242 | } |
| 243 | |
| 244 | /* Restore VM_ACCOUNT if one or two pieces of vma left */ |
| 245 | if (excess) { |
| 246 | vma->vm_flags |= VM_ACCOUNT; |
| 247 | if (split) |
| 248 | vma->vm_next->vm_flags |= VM_ACCOUNT; |
| 249 | } |
| 250 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 251 | if (vm_flags & VM_LOCKED) { |
| 252 | mm->locked_vm += new_len >> PAGE_SHIFT; |
| 253 | if (new_len > old_len) |
| 254 | make_pages_present(new_addr + old_len, |
| 255 | new_addr + new_len); |
| 256 | } |
| 257 | |
| 258 | return new_addr; |
| 259 | } |
| 260 | |
| 261 | /* |
| 262 | * Expand (or shrink) an existing mapping, potentially moving it at the |
| 263 | * same time (controlled by the MREMAP_MAYMOVE flag and available VM space) |
| 264 | * |
| 265 | * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise |
| 266 | * This option implies MREMAP_MAYMOVE. |
| 267 | */ |
| 268 | unsigned long do_mremap(unsigned long addr, |
| 269 | unsigned long old_len, unsigned long new_len, |
| 270 | unsigned long flags, unsigned long new_addr) |
| 271 | { |
| 272 | struct vm_area_struct *vma; |
| 273 | unsigned long ret = -EINVAL; |
| 274 | unsigned long charged = 0; |
| 275 | |
| 276 | if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE)) |
| 277 | goto out; |
| 278 | |
| 279 | if (addr & ~PAGE_MASK) |
| 280 | goto out; |
| 281 | |
| 282 | old_len = PAGE_ALIGN(old_len); |
| 283 | new_len = PAGE_ALIGN(new_len); |
| 284 | |
| 285 | /* |
| 286 | * We allow a zero old-len as a special case |
| 287 | * for DOS-emu "duplicate shm area" thing. But |
| 288 | * a zero new-len is nonsensical. |
| 289 | */ |
| 290 | if (!new_len) |
| 291 | goto out; |
| 292 | |
| 293 | /* new_addr is only valid if MREMAP_FIXED is specified */ |
| 294 | if (flags & MREMAP_FIXED) { |
| 295 | if (new_addr & ~PAGE_MASK) |
| 296 | goto out; |
| 297 | if (!(flags & MREMAP_MAYMOVE)) |
| 298 | goto out; |
| 299 | |
| 300 | if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len) |
| 301 | goto out; |
| 302 | |
| 303 | /* Check if the location we're moving into overlaps the |
| 304 | * old location at all, and fail if it does. |
| 305 | */ |
| 306 | if ((new_addr <= addr) && (new_addr+new_len) > addr) |
| 307 | goto out; |
| 308 | |
| 309 | if ((addr <= new_addr) && (addr+old_len) > new_addr) |
| 310 | goto out; |
| 311 | |
| 312 | ret = do_munmap(current->mm, new_addr, new_len); |
| 313 | if (ret) |
| 314 | goto out; |
| 315 | } |
| 316 | |
| 317 | /* |
| 318 | * Always allow a shrinking remap: that just unmaps |
| 319 | * the unnecessary pages.. |
| 320 | * do_munmap does all the needed commit accounting |
| 321 | */ |
| 322 | if (old_len >= new_len) { |
| 323 | ret = do_munmap(current->mm, addr+new_len, old_len - new_len); |
| 324 | if (ret && old_len != new_len) |
| 325 | goto out; |
| 326 | ret = addr; |
| 327 | if (!(flags & MREMAP_FIXED) || (new_addr == addr)) |
| 328 | goto out; |
| 329 | old_len = new_len; |
| 330 | } |
| 331 | |
| 332 | /* |
| 333 | * Ok, we need to grow.. or relocate. |
| 334 | */ |
| 335 | ret = -EFAULT; |
| 336 | vma = find_vma(current->mm, addr); |
| 337 | if (!vma || vma->vm_start > addr) |
| 338 | goto out; |
| 339 | if (is_vm_hugetlb_page(vma)) { |
| 340 | ret = -EINVAL; |
| 341 | goto out; |
| 342 | } |
| 343 | /* We can't remap across vm area boundaries */ |
| 344 | if (old_len > vma->vm_end - addr) |
| 345 | goto out; |
| 346 | if (vma->vm_flags & VM_DONTEXPAND) { |
| 347 | if (new_len > old_len) |
| 348 | goto out; |
| 349 | } |
| 350 | if (vma->vm_flags & VM_LOCKED) { |
| 351 | unsigned long locked, lock_limit; |
| 352 | locked = current->mm->locked_vm << PAGE_SHIFT; |
| 353 | lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur; |
| 354 | locked += new_len - old_len; |
| 355 | ret = -EAGAIN; |
| 356 | if (locked > lock_limit && !capable(CAP_IPC_LOCK)) |
| 357 | goto out; |
| 358 | } |
akpm@osdl.org | 119f657 | 2005-05-01 08:58:35 -0700 | [diff] [blame] | 359 | if (!may_expand_vm(current->mm, (new_len - old_len) >> PAGE_SHIFT)) { |
| 360 | ret = -ENOMEM; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 361 | goto out; |
akpm@osdl.org | 119f657 | 2005-05-01 08:58:35 -0700 | [diff] [blame] | 362 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 363 | |
| 364 | if (vma->vm_flags & VM_ACCOUNT) { |
| 365 | charged = (new_len - old_len) >> PAGE_SHIFT; |
| 366 | if (security_vm_enough_memory(charged)) |
| 367 | goto out_nc; |
| 368 | } |
| 369 | |
| 370 | /* old_len exactly to the end of the area.. |
| 371 | * And we're not relocating the area. |
| 372 | */ |
| 373 | if (old_len == vma->vm_end - addr && |
| 374 | !((flags & MREMAP_FIXED) && (addr != new_addr)) && |
| 375 | (old_len != new_len || !(flags & MREMAP_MAYMOVE))) { |
| 376 | unsigned long max_addr = TASK_SIZE; |
| 377 | if (vma->vm_next) |
| 378 | max_addr = vma->vm_next->vm_start; |
| 379 | /* can we just expand the current mapping? */ |
| 380 | if (max_addr - addr >= new_len) { |
| 381 | int pages = (new_len - old_len) >> PAGE_SHIFT; |
| 382 | |
| 383 | vma_adjust(vma, vma->vm_start, |
| 384 | addr + new_len, vma->vm_pgoff, NULL); |
| 385 | |
| 386 | current->mm->total_vm += pages; |
Hugh Dickins | ab50b8e | 2005-10-29 18:15:56 -0700 | [diff] [blame] | 387 | vm_stat_account(vma->vm_mm, vma->vm_flags, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 388 | vma->vm_file, pages); |
| 389 | if (vma->vm_flags & VM_LOCKED) { |
| 390 | current->mm->locked_vm += pages; |
| 391 | make_pages_present(addr + old_len, |
| 392 | addr + new_len); |
| 393 | } |
| 394 | ret = addr; |
| 395 | goto out; |
| 396 | } |
| 397 | } |
| 398 | |
| 399 | /* |
| 400 | * We weren't able to just expand or shrink the area, |
| 401 | * we need to create a new one and move it.. |
| 402 | */ |
| 403 | ret = -ENOMEM; |
| 404 | if (flags & MREMAP_MAYMOVE) { |
| 405 | if (!(flags & MREMAP_FIXED)) { |
| 406 | unsigned long map_flags = 0; |
| 407 | if (vma->vm_flags & VM_MAYSHARE) |
| 408 | map_flags |= MAP_SHARED; |
| 409 | |
| 410 | new_addr = get_unmapped_area(vma->vm_file, 0, new_len, |
| 411 | vma->vm_pgoff, map_flags); |
| 412 | ret = new_addr; |
| 413 | if (new_addr & ~PAGE_MASK) |
| 414 | goto out; |
| 415 | } |
| 416 | ret = move_vma(vma, addr, old_len, new_len, new_addr); |
| 417 | } |
| 418 | out: |
| 419 | if (ret & ~PAGE_MASK) |
| 420 | vm_unacct_memory(charged); |
| 421 | out_nc: |
| 422 | return ret; |
| 423 | } |
| 424 | |
| 425 | asmlinkage unsigned long sys_mremap(unsigned long addr, |
| 426 | unsigned long old_len, unsigned long new_len, |
| 427 | unsigned long flags, unsigned long new_addr) |
| 428 | { |
| 429 | unsigned long ret; |
| 430 | |
| 431 | down_write(¤t->mm->mmap_sem); |
| 432 | ret = do_mremap(addr, old_len, new_len, flags, new_addr); |
| 433 | up_write(¤t->mm->mmap_sem); |
| 434 | return ret; |
| 435 | } |