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/security.h>
  19
  20 #include <asm/uaccess.h>
  21 #include <asm/pgalloc.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         pmd_t *pmd;
  29         pte_t *pte = NULL;
  30
  31         pgd = pgd_offset(mm, addr);
  32         if (pgd_none(*pgd))
  33                 goto end;
  34         if (pgd_bad(*pgd)) {
  35                 pgd_ERROR(*pgd);
  36                 pgd_clear(pgd);
  37                 goto end;
  38         }
  39
  40         pmd = pmd_offset(pgd, addr);
  41         if (pmd_none(*pmd))
  42                 goto end;
  43         if (pmd_bad(*pmd)) {
  44                 pmd_ERROR(*pmd);
  45                 pmd_clear(pmd);
  46                 goto end;
  47         }
  48
  49         pte = pte_offset_map_nested(pmd, addr);
  50         if (pte_none(*pte)) {
  51                 pte_unmap_nested(pte);
  52                 pte = NULL;
  53         }
  54 end:
  55         return pte;
  56 }
  57
  58 static inline int page_table_present(struct mm_struct *mm, unsigned long addr)
  59 {
  60         pgd_t *pgd;
  61         pmd_t *pmd;
  62
  63         pgd = pgd_offset(mm, addr);
  64         if (pgd_none(*pgd))
  65                 return 0;
  66         pmd = pmd_offset(pgd, addr);
  67         return pmd_present(*pmd);
  68 }
  69
  70 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  71 {
  72         pmd_t *pmd;
  73         pte_t *pte = NULL;
  74
  75         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  76         if (pmd)
  77                 pte = pte_alloc_map(mm, pmd, addr);
  78         return pte;
  79 }
  80
  81 static int
  82 copy_one_pte(struct vm_area_struct *vma, unsigned long old_addr,
  83              pte_t *src, pte_t *dst)
  84 {
  85         int error = 0;
  86         pte_t pte;
  87         struct page *page = NULL;
  88
  89         if (pte_present(*src))
  90                 page = pte_page(*src);
  91
  92         if (!pte_none(*src)) {
  93                 pte = ptep_clear_flush(vma, old_addr, src);
  94                 if (!dst) {
  95                         /* No dest?  We must put it back. */
  96                         dst = src;
  97                         error++;
  98                 }
  99                 set_pte(dst, pte);
 100         }
 101         return error;
 102 }
 103
 104 static int
 105 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
 106                 unsigned long new_addr)
 107 {
 108         struct mm_struct *mm = vma->vm_mm;
 109         int error = 0;
 110         pte_t *src, *dst;
 111
 112         spin_lock(&mm->page_table_lock);
 113         src = get_one_pte_map_nested(mm, old_addr);
 114         if (src) {
 115                 /*
 116                  * Look to see whether alloc_one_pte_map needs to perform a
 117                  * memory allocation.  If it does then we need to drop the
 118                  * atomic kmap
 119                  */
 120                 if (!page_table_present(mm, new_addr)) {
 121                         pte_unmap_nested(src);
 122                         src = NULL;
 123                 }
 124                 dst = alloc_one_pte_map(mm, new_addr);
 125                 if (src == NULL)
 126                         src = get_one_pte_map_nested(mm, old_addr);
 127                 /*
 128                  * Since alloc_one_pte_map can drop and re-acquire
 129                  * page_table_lock, we should re-check the src entry...
 130                  */
 131                 if (src) {
 132                         error = copy_one_pte(vma, old_addr, src, dst);
 133                         pte_unmap_nested(src);
 134                 }
 135                 pte_unmap(dst);
 136         }
 137         spin_unlock(&mm->page_table_lock);
 138         return error;
 139 }
 140
 141 static int move_page_tables(struct vm_area_struct *vma,
 142         unsigned long new_addr, unsigned long old_addr, unsigned long len)
 143 {
 144         unsigned long offset = len;
 145
 146         flush_cache_range(vma, old_addr, old_addr + len);
 147
 148         /*
 149          * This is not the clever way to do this, but we're taking the
 150          * easy way out on the assumption that most remappings will be
 151          * only a few pages.. This also makes error recovery easier.
 152          */
 153         while (offset) {
 154                 offset -= PAGE_SIZE;
 155                 if (move_one_page(vma, old_addr + offset, new_addr + offset))
 156                         goto oops_we_failed;
 157         }
 158         return 0;
 159
 160         /*
 161          * Ok, the move failed because we didn't have enough pages for
 162          * the new page table tree. This is unlikely, but we have to
 163          * take the possibility into account. In that case we just move
 164          * all the pages back (this will work, because we still have
 165          * the old page tables)
 166          */
 167 oops_we_failed:
 168         flush_cache_range(vma, new_addr, new_addr + len);
 169         while ((offset += PAGE_SIZE) < len)
 170                 move_one_page(vma, new_addr + offset, old_addr + offset);
 171         zap_page_range(vma, new_addr, len);
 172         return -1;
 173 }
 174
 175 static unsigned long move_vma(struct vm_area_struct *vma,
 176         unsigned long addr, unsigned long old_len, unsigned long new_len,
 177         unsigned long new_addr)
 178 {
 179 #if VMA_MERGING_FIXUP
 180         struct mm_struct *mm = vma->vm_mm;
 181         struct vm_area_struct *new_vma, *next, *prev;
 182 #else
 183         struct vm_area_struct *new_vma;
 184 #endif
 185         int allocated_vma;
 186         int split = 0;
 187
 188         new_vma = NULL;
 189 #if VMA_MERGING_FIXUP
 190         next = find_vma_prev(mm, new_addr, &prev);
 191         if (next) {
 192                 if (prev && prev->vm_end == new_addr &&
 193                     mpol_equal(prev->vm_policy, next->vm_policy) &&
 194                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file &&
 195                                         !(vma->vm_flags & VM_SHARED)) {
 196                         spin_lock(&mm->page_table_lock);
 197                         prev->vm_end = new_addr + new_len;
 198                         spin_unlock(&mm->page_table_lock);
 199                         new_vma = prev;
 200                         if (next != prev->vm_next)
 201                                 BUG();
 202                         if (prev->vm_end == next->vm_start &&
 203                                     mpol_equal(next->vm_policy, prev->vm_policy) &&
 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                                 mpol_free(next->vm_policy);
 213                                 kmem_cache_free(vm_area_cachep, next);
 214                         }
 215                 } else if (next->vm_start == new_addr + new_len &&
 216                                 can_vma_merge(next, vma->vm_flags) &&
 217                                 mpol_equal(next->vm_policy, vma->vm_policy) &&
 218                                 !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 219                         spin_lock(&mm->page_table_lock);
 220                         next->vm_start = new_addr;
 221                         spin_unlock(&mm->page_table_lock);
 222                         new_vma = next;
 223                 }
 224         } else {
 225                 prev = find_vma(mm, new_addr-1);
 226                 if (prev && prev->vm_end == new_addr &&
 227                     mpol_equal(prev->vm_policy, vma->vm_policy) &&
 228                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file &&
 229                                 !(vma->vm_flags & VM_SHARED)) {
 230                         spin_lock(&mm->page_table_lock);
 231                         prev->vm_end = new_addr + new_len;
 232                         spin_unlock(&mm->page_table_lock);
 233                         new_vma = prev;
 234                 }
 235         }
 236 #endif
 237
 238         allocated_vma = 0;
 239         if (!new_vma) {
 240                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
 241                 if (!new_vma)
 242                         goto out;
 243                 allocated_vma = 1;
 244         }
 245
 246         if (!move_page_tables(vma, new_addr, addr, old_len)) {
 247                 unsigned long vm_locked = vma->vm_flags & VM_LOCKED;
 248
 249                 if (allocated_vma) {
 250                         *new_vma = *vma;
 251                         new_vma->vm_policy =mpol_copy(new_vma->vm_policy);
 252                         if (IS_ERR(new_vma->vm_policy))
 253                                 goto out_vma;
 254                         INIT_VMA_SHARED(new_vma);
 255                         new_vma->vm_start = new_addr;
 256                         new_vma->vm_end = new_addr+new_len;
 257                         new_vma->vm_pgoff += (addr-vma->vm_start) >> PAGE_SHIFT;
 258                         if (new_vma->vm_file)
 259                                 get_file(new_vma->vm_file);
 260                         if (new_vma->vm_ops && new_vma->vm_ops->open)
 261                                 new_vma->vm_ops->open(new_vma);
 262                         insert_vm_struct(current->mm, new_vma);
 263                 }
 264
 265                 /* Conceal VM_ACCOUNT so old reservation is not undone */
 266                 if (vma->vm_flags & VM_ACCOUNT) {
 267                         vma->vm_flags &= ~VM_ACCOUNT;
 268                         if (addr > vma->vm_start) {
 269                                 if (addr + old_len < vma->vm_end)
 270                                         split = 1;
 271                         } else if (addr + old_len == vma->vm_end)
 272                                 vma = NULL;     /* it will be removed */
 273                 } else
 274                         vma = NULL;             /* nothing more to do */
 275
 276                 do_munmap(current->mm, addr, old_len);
 277
 278                 /* Restore VM_ACCOUNT if one or two pieces of vma left */
 279                 if (vma) {
 280                         vma->vm_flags |= VM_ACCOUNT;
 281                         if (split)
 282                                 vma->vm_next->vm_flags |= VM_ACCOUNT;
 283                 }
 284
 285                 current->mm->total_vm += new_len >> PAGE_SHIFT;
 286                 if (vm_locked) {
 287                         current->mm->locked_vm += new_len >> PAGE_SHIFT;
 288                         if (new_len > old_len)
 289                                 make_pages_present(new_addr + old_len,
 290                                                    new_addr + new_len);
 291                 }
 292                 return new_addr;
 293         }
 294  out_vma:
 295         if (allocated_vma)
 296                 kmem_cache_free(vm_area_cachep, new_vma);
 297  out:
 298         return -ENOMEM;
 299 }
 300
 301 /*
 302  * Expand (or shrink) an existing mapping, potentially moving it at the
 303  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 304  *
 305  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 306  * This option implies MREMAP_MAYMOVE.
 307  */
 308 unsigned long do_mremap(unsigned long addr,
 309         unsigned long old_len, unsigned long new_len,
 310         unsigned long flags, unsigned long new_addr)
 311 {
 312         struct vm_area_struct *vma;
 313         struct address_space *mapping = NULL;
 314         struct prio_tree_root *root = NULL;
 315         unsigned long ret = -EINVAL;
 316         unsigned long charged = 0;
 317
 318         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 319                 goto out;
 320
 321         if (addr & ~PAGE_MASK)
 322                 goto out;
 323
 324         old_len = PAGE_ALIGN(old_len);
 325         new_len = PAGE_ALIGN(new_len);
 326
 327         /*
 328          * We allow a zero old-len as a special case
 329          * for DOS-emu "duplicate shm area" thing. But
 330          * a zero new-len is nonsensical.
 331          */
 332         if (!new_len)
 333                 goto out;
 334
 335         /* new_addr is only valid if MREMAP_FIXED is specified */
 336         if (flags & MREMAP_FIXED) {
 337                 if (new_addr & ~PAGE_MASK)
 338                         goto out;
 339                 if (!(flags & MREMAP_MAYMOVE))
 340                         goto out;
 341
 342                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 343                         goto out;
 344
 345                 /* Check if the location we're moving into overlaps the
 346                  * old location at all, and fail if it does.
 347                  */
 348                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 349                         goto out;
 350
 351                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 352                         goto out;
 353
 354                 ret = do_munmap(current->mm, new_addr, new_len);
 355                 if (ret)
 356                         goto out;
 357         }
 358
 359         /*
 360          * Always allow a shrinking remap: that just unmaps
 361          * the unnecessary pages..
 362          * do_munmap does all the needed commit accounting
 363          */
 364         if (old_len >= new_len) {
 365                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 366                 if (ret && old_len != new_len)
 367                         goto out;
 368                 ret = addr;
 369                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 370                         goto out;
 371                 old_len = new_len;
 372         }
 373
 374         /*
 375          * Ok, we need to grow..  or relocate.
 376          */
 377         ret = -EFAULT;
 378         vma = find_vma(current->mm, addr);
 379         if (!vma || vma->vm_start > addr)
 380                 goto out;
 381         if (is_vm_hugetlb_page(vma)) {
 382                 ret = -EINVAL;
 383                 goto out;
 384         }
 385         /* We can't remap across vm area boundaries */
 386         if (old_len > vma->vm_end - addr)
 387                 goto out;
 388         if (vma->vm_flags & VM_DONTEXPAND) {
 389                 if (new_len > old_len)
 390                         goto out;
 391         }
 392         if (vma->vm_flags & VM_LOCKED) {
 393                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 394                 locked += new_len - old_len;
 395                 ret = -EAGAIN;
 396                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 397                         goto out;
 398         }
 399         ret = -ENOMEM;
 400         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 401             > current->rlim[RLIMIT_AS].rlim_cur)
 402                 goto out;
 403
 404         if (vma->vm_flags & VM_ACCOUNT) {
 405                 charged = (new_len - old_len) >> PAGE_SHIFT;
 406                 if (security_vm_enough_memory(charged))
 407                         goto out_nc;
 408         }
 409
 410         /* old_len exactly to the end of the area..
 411          * And we're not relocating the area.
 412          */
 413         if (old_len == vma->vm_end - addr &&
 414             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 415             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 416                 unsigned long max_addr = TASK_SIZE;
 417                 if (vma->vm_next)
 418                         max_addr = vma->vm_next->vm_start;
 419                 /* can we just expand the current mapping? */
 420                 if (max_addr - addr >= new_len) {
 421                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 422
 423                         if (vma->vm_file) {
 424                                 mapping = vma->vm_file->f_mapping;
 425                                 if (vma->vm_flags & VM_SHARED) {
 426                                         if (likely(!(vma->vm_flags & VM_NONLINEAR)))
 427                                                 root = &mapping->i_mmap_shared;
 428                                 }
 429                                 else
 430                                         root = &mapping->i_mmap;
 431                                 down(&mapping->i_shared_sem);
 432                         }
 433
 434                         spin_lock(&vma->vm_mm->page_table_lock);
 435                         __vma_modify(root, vma, vma->vm_start,
 436                                         addr + new_len, vma->vm_pgoff);
 437                         spin_unlock(&vma->vm_mm->page_table_lock);
 438
 439                         if(mapping)
 440                                 up(&mapping->i_shared_sem);
 441
 442                         current->mm->total_vm += pages;
 443                         if (vma->vm_flags & VM_LOCKED) {
 444                                 current->mm->locked_vm += pages;
 445                                 make_pages_present(addr + old_len,
 446                                                    addr + new_len);
 447                         }
 448                         ret = addr;
 449                         goto out;
 450                 }
 451         }
 452
 453         /*
 454          * We weren't able to just expand or shrink the area,
 455          * we need to create a new one and move it..
 456          */
 457         ret = -ENOMEM;
 458         if (flags & MREMAP_MAYMOVE) {
 459                 if (!(flags & MREMAP_FIXED)) {
 460                         unsigned long map_flags = 0;
 461                         if (vma->vm_flags & VM_MAYSHARE)
 462                                 map_flags |= MAP_SHARED;
 463
 464                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 465                                                 vma->vm_pgoff, map_flags);
 466                         ret = new_addr;
 467                         if (new_addr & ~PAGE_MASK)
 468                                 goto out;
 469                 }
 470                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 471         }
 472 out:
 473         if (ret & ~PAGE_MASK)
 474                 vm_unacct_memory(charged);
 475 out_nc:
 476         return ret;
 477 }
 478
 479 asmlinkage unsigned long sys_mremap(unsigned long addr,
 480         unsigned long old_len, unsigned long new_len,
 481         unsigned long flags, unsigned long new_addr)
 482 {
 483         unsigned long ret;
 484
 485         down_write(&current->mm->mmap_sem);
 486         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 487         up_write(&current->mm->mmap_sem);
 488         return ret;
 489 }