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