mm/mremap.c

   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/rmap-locking.h>
  19 #include <linux/security.h>
  20
  21 #include <asm/uaccess.h>
  22 #include <asm/pgalloc.h>
  23 #include <asm/cacheflush.h>
  24 #include <asm/tlbflush.h>
  25
  26 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
  27 {
  28         pgd_t *pgd;
  29         pmd_t *pmd;
  30         pte_t *pte = NULL;
  31
  32         pgd = pgd_offset(mm, addr);
  33         if (pgd_none(*pgd))
  34                 goto end;
  35         if (pgd_bad(*pgd)) {
  36                 pgd_ERROR(*pgd);
  37                 pgd_clear(pgd);
  38                 goto end;
  39         }
  40
  41         pmd = pmd_offset(pgd, addr);
  42         if (pmd_none(*pmd))
  43                 goto end;
  44         if (pmd_bad(*pmd)) {
  45                 pmd_ERROR(*pmd);
  46                 pmd_clear(pmd);
  47                 goto end;
  48         }
  49
  50         pte = pte_offset_map_nested(pmd, addr);
  51         if (pte_none(*pte)) {
  52                 pte_unmap_nested(pte);
  53                 pte = NULL;
  54         }
  55 end:
  56         return pte;
  57 }
  58
  59 static inline int page_table_present(struct mm_struct *mm, unsigned long addr)
  60 {
  61         pgd_t *pgd;
  62         pmd_t *pmd;
  63
  64         pgd = pgd_offset(mm, addr);
  65         if (pgd_none(*pgd))
  66                 return 0;
  67         pmd = pmd_offset(pgd, addr);
  68         return pmd_present(*pmd);
  69 }
  70
  71 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  72 {
  73         pmd_t *pmd;
  74         pte_t *pte = NULL;
  75
  76         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  77         if (pmd)
  78                 pte = pte_alloc_map(mm, pmd, addr);
  79         return pte;
  80 }
  81
  82 static void
  83 copy_one_pte(struct vm_area_struct *vma, unsigned long old_addr,
  84              pte_t *src, pte_t *dst, struct pte_chain **pte_chainp)
  85 {
  86         pte_t pte = ptep_clear_flush(vma, old_addr, src);
  87         set_pte(dst, pte);
  88
  89         if (pte_present(pte)) {
  90                 unsigned long pfn = pte_pfn(pte);
  91                 if (pfn_valid(pfn)) {
  92                         struct page *page = pfn_to_page(pfn);
  93                         page_remove_rmap(page, src);
  94                         *pte_chainp = page_add_rmap(page, dst, *pte_chainp);
  95                 }
  96         }
  97 }
  98
  99 static int
 100 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
 101                 unsigned long new_addr)
 102 {
 103         struct mm_struct *mm = vma->vm_mm;
 104         int error = 0;
 105         pte_t *src, *dst;
 106         struct pte_chain *pte_chain;
 107
 108         pte_chain = pte_chain_alloc(GFP_KERNEL);
 109         if (!pte_chain) {
 110                 error = -ENOMEM;
 111                 goto out;
 112         }
 113         spin_lock(&mm->page_table_lock);
 114         src = get_one_pte_map_nested(mm, old_addr);
 115         if (src) {
 116                 /*
 117                  * Look to see whether alloc_one_pte_map needs to perform a
 118                  * memory allocation.  If it does then we need to drop the
 119                  * atomic kmap
 120                  */
 121                 if (!page_table_present(mm, new_addr)) {
 122                         pte_unmap_nested(src);
 123                         src = NULL;
 124                 }
 125                 dst = alloc_one_pte_map(mm, new_addr);
 126                 if (src == NULL)
 127                         src = get_one_pte_map_nested(mm, old_addr);
 128                 /*
 129                  * Since alloc_one_pte_map can drop and re-acquire
 130                  * page_table_lock, we should re-check the src entry...
 131                  */
 132                 if (src) {
 133                         if (dst)
 134                                 copy_one_pte(vma, old_addr, src,
 135                                                 dst, &pte_chain);
 136                         else
 137                                 error = -ENOMEM;
 138                         pte_unmap_nested(src);
 139                 }
 140                 pte_unmap(dst);
 141         }
 142         spin_unlock(&mm->page_table_lock);
 143         pte_chain_free(pte_chain);
 144 out:
 145         return error;
 146 }
 147
 148 static int move_page_tables(struct vm_area_struct *vma,
 149         unsigned long new_addr, unsigned long old_addr, unsigned long len)
 150 {
 151         unsigned long offset = len;
 152
 153         flush_cache_range(vma, old_addr, old_addr + len);
 154
 155         /*
 156          * This is not the clever way to do this, but we're taking the
 157          * easy way out on the assumption that most remappings will be
 158          * only a few pages.. This also makes error recovery easier.
 159          */
 160         while (offset) {
 161                 offset -= PAGE_SIZE;
 162                 if (move_one_page(vma, old_addr + offset, new_addr + offset))
 163                         goto oops_we_failed;
 164         }
 165         return 0;
 166
 167         /*
 168          * Ok, the move failed because we didn't have enough pages for
 169          * the new page table tree. This is unlikely, but we have to
 170          * take the possibility into account. In that case we just move
 171          * all the pages back (this will work, because we still have
 172          * the old page tables)
 173          */
 174 oops_we_failed:
 175         flush_cache_range(vma, new_addr, new_addr + len);
 176         while ((offset += PAGE_SIZE) < len)
 177                 move_one_page(vma, new_addr + offset, old_addr + offset);
 178         zap_page_range(vma, new_addr, len);
 179         return -1;
 180 }
 181
 182 static unsigned long move_vma(struct vm_area_struct *vma,
 183         unsigned long addr, unsigned long old_len, unsigned long new_len,
 184         unsigned long new_addr)
 185 {
 186         struct mm_struct *mm = vma->vm_mm;
 187         struct vm_area_struct *new_vma, *next, *prev;
 188         int allocated_vma;
 189         int split = 0;
 190
 191         new_vma = NULL;
 192         next = find_vma_prev(mm, new_addr, &prev);
 193         if (next) {
 194                 if (prev && prev->vm_end == new_addr &&
 195                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file &&
 196                                         !(vma->vm_flags & VM_SHARED)) {
 197                         spin_lock(&mm->page_table_lock);
 198                         prev->vm_end = new_addr + new_len;
 199                         spin_unlock(&mm->page_table_lock);
 200                         new_vma = prev;
 201                         if (next != prev->vm_next)
 202                                 BUG();
 203                         if (prev->vm_end == next->vm_start &&
 204                                         can_vma_merge(next, prev->vm_flags)) {
 205                                 spin_lock(&mm->page_table_lock);
 206                                 prev->vm_end = next->vm_end;
 207                                 __vma_unlink(mm, next, prev);
 208                                 spin_unlock(&mm->page_table_lock);
 209                                 if (vma == next)
 210                                         vma = prev;
 211                                 mm->map_count--;
 212                                 kmem_cache_free(vm_area_cachep, next);
 213                         }
 214                 } else if (next->vm_start == new_addr + new_len &&
 215                                 can_vma_merge(next, vma->vm_flags) &&
 216                                 !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 217                         spin_lock(&mm->page_table_lock);
 218                         next->vm_start = new_addr;
 219                         spin_unlock(&mm->page_table_lock);
 220                         new_vma = next;
 221                 }
 222         } else {
 223                 prev = find_vma(mm, new_addr-1);
 224                 if (prev && prev->vm_end == new_addr &&
 225                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file &&
 226                                 !(vma->vm_flags & VM_SHARED)) {
 227                         spin_lock(&mm->page_table_lock);
 228                         prev->vm_end = new_addr + new_len;
 229                         spin_unlock(&mm->page_table_lock);
 230                         new_vma = prev;
 231                 }
 232         }
 233
 234         allocated_vma = 0;
 235         if (!new_vma) {
 236                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
 237                 if (!new_vma)
 238                         goto out;
 239                 allocated_vma = 1;
 240         }
 241
 242         if (!move_page_tables(vma, new_addr, addr, old_len)) {
 243                 unsigned long vm_locked = vma->vm_flags & VM_LOCKED;
 244
 245                 if (allocated_vma) {
 246                         *new_vma = *vma;
 247                         INIT_LIST_HEAD(&new_vma->shared);
 248                         new_vma->vm_start = new_addr;
 249                         new_vma->vm_end = new_addr+new_len;
 250                         new_vma->vm_pgoff += (addr-vma->vm_start) >> PAGE_SHIFT;
 251                         if (new_vma->vm_file)
 252                                 get_file(new_vma->vm_file);
 253                         if (new_vma->vm_ops && new_vma->vm_ops->open)
 254                                 new_vma->vm_ops->open(new_vma);
 255                         insert_vm_struct(current->mm, new_vma);
 256                 }
 257
 258                 /* Conceal VM_ACCOUNT so old reservation is not undone */
 259                 if (vma->vm_flags & VM_ACCOUNT) {
 260                         vma->vm_flags &= ~VM_ACCOUNT;
 261                         if (addr > vma->vm_start) {
 262                                 if (addr + old_len < vma->vm_end)
 263                                         split = 1;
 264                         } else if (addr + old_len == vma->vm_end)
 265                                 vma = NULL;     /* it will be removed */
 266                 } else
 267                         vma = NULL;             /* nothing more to do */
 268
 269                 do_munmap(current->mm, addr, old_len);
 270
 271                 /* Restore VM_ACCOUNT if one or two pieces of vma left */
 272                 if (vma) {
 273                         vma->vm_flags |= VM_ACCOUNT;
 274                         if (split)
 275                                 vma->vm_next->vm_flags |= VM_ACCOUNT;
 276                 }
 277
 278                 current->mm->total_vm += new_len >> PAGE_SHIFT;
 279                 if (vm_locked) {
 280                         current->mm->locked_vm += new_len >> PAGE_SHIFT;
 281                         if (new_len > old_len)
 282                                 make_pages_present(new_addr + old_len,
 283                                                    new_addr + new_len);
 284                 }
 285                 return new_addr;
 286         }
 287         if (allocated_vma)
 288                 kmem_cache_free(vm_area_cachep, new_vma);
 289  out:
 290         return -ENOMEM;
 291 }
 292
 293 /*
 294  * Expand (or shrink) an existing mapping, potentially moving it at the
 295  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 296  *
 297  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 298  * This option implies MREMAP_MAYMOVE.
 299  */
 300 unsigned long do_mremap(unsigned long addr,
 301         unsigned long old_len, unsigned long new_len,
 302         unsigned long flags, unsigned long new_addr)
 303 {
 304         struct vm_area_struct *vma;
 305         unsigned long ret = -EINVAL;
 306         unsigned long charged = 0;
 307
 308         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 309                 goto out;
 310
 311         if (addr & ~PAGE_MASK)
 312                 goto out;
 313
 314         old_len = PAGE_ALIGN(old_len);
 315         new_len = PAGE_ALIGN(new_len);
 316
 317         /*
 318          * We allow a zero old-len as a special case
 319          * for DOS-emu "duplicate shm area" thing. But
 320          * a zero new-len is nonsensical.
 321          */
 322         if (!new_len)
 323                 goto out;
 324
 325         /* new_addr is only valid if MREMAP_FIXED is specified */
 326         if (flags & MREMAP_FIXED) {
 327                 if (new_addr & ~PAGE_MASK)
 328                         goto out;
 329                 if (!(flags & MREMAP_MAYMOVE))
 330                         goto out;
 331
 332                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 333                         goto out;
 334
 335                 /* Check if the location we're moving into overlaps the
 336                  * old location at all, and fail if it does.
 337                  */
 338                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 339                         goto out;
 340
 341                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 342                         goto out;
 343
 344                 ret = do_munmap(current->mm, new_addr, new_len);
 345                 if (ret)
 346                         goto out;
 347         }
 348
 349         /*
 350          * Always allow a shrinking remap: that just unmaps
 351          * the unnecessary pages..
 352          * do_munmap does all the needed commit accounting
 353          */
 354         if (old_len >= new_len) {
 355                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 356                 if (ret && old_len != new_len)
 357                         goto out;
 358                 ret = addr;
 359                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 360                         goto out;
 361                 old_len = new_len;
 362         }
 363
 364         /*
 365          * Ok, we need to grow..  or relocate.
 366          */
 367         ret = -EFAULT;
 368         vma = find_vma(current->mm, addr);
 369         if (!vma || vma->vm_start > addr)
 370                 goto out;
 371         if (is_vm_hugetlb_page(vma)) {
 372                 ret = -EINVAL;
 373                 goto out;
 374         }
 375         /* We can't remap across vm area boundaries */
 376         if (old_len > vma->vm_end - addr)
 377                 goto out;
 378         if (vma->vm_flags & VM_DONTEXPAND) {
 379                 if (new_len > old_len)
 380                         goto out;
 381         }
 382         if (vma->vm_flags & VM_LOCKED) {
 383                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 384                 locked += new_len - old_len;
 385                 ret = -EAGAIN;
 386                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 387                         goto out;
 388         }
 389         ret = -ENOMEM;
 390         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 391             > current->rlim[RLIMIT_AS].rlim_cur)
 392                 goto out;
 393
 394         if (vma->vm_flags & VM_ACCOUNT) {
 395                 charged = (new_len - old_len) >> PAGE_SHIFT;
 396                 if (security_vm_enough_memory(charged))
 397                         goto out_nc;
 398         }
 399
 400         /* old_len exactly to the end of the area..
 401          * And we're not relocating the area.
 402          */
 403         if (old_len == vma->vm_end - addr &&
 404             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 405             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 406                 unsigned long max_addr = TASK_SIZE;
 407                 if (vma->vm_next)
 408                         max_addr = vma->vm_next->vm_start;
 409                 /* can we just expand the current mapping? */
 410                 if (max_addr - addr >= new_len) {
 411                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 412                         spin_lock(&vma->vm_mm->page_table_lock);
 413                         vma->vm_end = addr + new_len;
 414                         spin_unlock(&vma->vm_mm->page_table_lock);
 415                         current->mm->total_vm += pages;
 416                         if (vma->vm_flags & VM_LOCKED) {
 417                                 current->mm->locked_vm += pages;
 418                                 make_pages_present(addr + old_len,
 419                                                    addr + new_len);
 420                         }
 421                         ret = addr;
 422                         goto out;
 423                 }
 424         }
 425
 426         /*
 427          * We weren't able to just expand or shrink the area,
 428          * we need to create a new one and move it..
 429          */
 430         ret = -ENOMEM;
 431         if (flags & MREMAP_MAYMOVE) {
 432                 if (!(flags & MREMAP_FIXED)) {
 433                         unsigned long map_flags = 0;
 434                         if (vma->vm_flags & VM_MAYSHARE)
 435                                 map_flags |= MAP_SHARED;
 436
 437                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 438                                                 vma->vm_pgoff, map_flags);
 439                         ret = new_addr;
 440                         if (new_addr & ~PAGE_MASK)
 441                                 goto out;
 442                 }
 443                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 444         }
 445 out:
 446         if (ret & ~PAGE_MASK)
 447                 vm_unacct_memory(charged);
 448 out_nc:
 449         return ret;
 450 }
 451
 452 asmlinkage unsigned long sys_mremap(unsigned long addr,
 453         unsigned long old_len, unsigned long new_len,
 454         unsigned long flags, unsigned long new_addr)
 455 {
 456         unsigned long ret;
 457
 458         down_write(&current->mm->mmap_sem);
 459         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 460         up_write(&current->mm->mmap_sem);
 461         return ret;
 462 }