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.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 inline int
  83 can_move_one_pte(pte_t *src, unsigned long new_addr)
  84 {
  85         int move = 1;
  86         if (pte_present(*src)) {
  87                 unsigned long pfn = pte_pfn(*src);
  88                 if (pfn_valid(pfn)) {
  89                         struct page *page = pfn_to_page(pfn);
  90                         if (PageAnon(page))
  91                                 move = mremap_move_anon_rmap(page, new_addr);
  92                 }
  93         }
  94         return move;
  95 }
  96
  97 static int
  98 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
  99                 unsigned long new_addr)
 100 {
 101         struct mm_struct *mm = vma->vm_mm;
 102         int error = 0;
 103         pte_t *src, *dst;
 104
 105         spin_lock(&mm->page_table_lock);
 106         src = get_one_pte_map_nested(mm, old_addr);
 107         if (src) {
 108                 /*
 109                  * Look to see whether alloc_one_pte_map needs to perform a
 110                  * memory allocation.  If it does then we need to drop the
 111                  * atomic kmap
 112                  */
 113                 if (!page_table_present(mm, new_addr)) {
 114                         pte_unmap_nested(src);
 115                         src = NULL;
 116                 }
 117                 dst = alloc_one_pte_map(mm, new_addr);
 118                 if (src == NULL)
 119                         src = get_one_pte_map_nested(mm, old_addr);
 120                 /*
 121                  * Since alloc_one_pte_map can drop and re-acquire
 122                  * page_table_lock, we should re-check the src entry...
 123                  */
 124                 if (src) {
 125                         if (!dst)
 126                                 error = -ENOMEM;
 127                         else if (!can_move_one_pte(src, new_addr))
 128                                 error = -EAGAIN;
 129                         else {
 130                                 pte_t pte;
 131                                 pte = ptep_clear_flush(vma, old_addr, src);
 132                                 set_pte(dst, pte);
 133                         }
 134                         pte_unmap_nested(src);
 135                 }
 136                 if (dst)
 137                         pte_unmap(dst);
 138         }
 139         spin_unlock(&mm->page_table_lock);
 140         return error;
 141 }
 142
 143 static unsigned long move_page_tables(struct vm_area_struct *vma,
 144                 unsigned long new_addr, unsigned long old_addr,
 145                 unsigned long len, int *cows)
 146 {
 147         unsigned long offset;
 148
 149         flush_cache_range(vma, old_addr, old_addr + len);
 150
 151         /*
 152          * This is not the clever way to do this, but we're taking the
 153          * easy way out on the assumption that most remappings will be
 154          * only a few pages.. This also makes error recovery easier.
 155          */
 156         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 157                 int ret = move_one_page(vma, old_addr+offset, new_addr+offset);
 158                 /*
 159                  * The anonmm objrmap can only track anon page movements
 160                  * if the page is exclusive to one mm.  In the rare case
 161                  * when mremap move is applied to a shared page, break
 162                  * COW (take a copy of the page) to make it exclusive.
 163                  * If shared while on swap, page will be copied when
 164                  * brought back in (if it's still shared by then).
 165                  */
 166                 if (ret == -EAGAIN) {
 167                         ret = make_page_exclusive(vma, old_addr+offset);
 168                         offset -= PAGE_SIZE;
 169                         (*cows)++;
 170                 }
 171                 if (ret)
 172                         break;
 173                 cond_resched();
 174         }
 175         return offset;
 176 }
 177
 178 static unsigned long move_vma(struct vm_area_struct *vma,
 179                 unsigned long old_addr, unsigned long old_len,
 180                 unsigned long new_len, unsigned long new_addr)
 181 {
 182         struct mm_struct *mm = vma->vm_mm;
 183         struct vm_area_struct *new_vma;
 184         unsigned long vm_flags = vma->vm_flags;
 185         unsigned long new_pgoff;
 186         unsigned long moved_len;
 187         unsigned long excess = 0;
 188         int split = 0;
 189         int cows = 0;
 190
 191         /*
 192          * We'd prefer to avoid failure later on in do_munmap:
 193          * which may split one vma into three before unmapping.
 194          */
 195         if (mm->map_count >= sysctl_max_map_count - 3)
 196                 return -ENOMEM;
 197
 198         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 199         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 200         if (!new_vma)
 201                 return -ENOMEM;
 202
 203         moved_len = move_page_tables(vma, new_addr, old_addr, old_len, &cows);
 204         if (moved_len < old_len) {
 205                 /*
 206                  * On error, move entries back from new area to old,
 207                  * which will succeed since page tables still there,
 208                  * and then proceed to unmap new area instead of old.
 209                  */
 210                 move_page_tables(new_vma, old_addr, new_addr, moved_len, &cows);
 211                 vma = new_vma;
 212                 old_len = new_len;
 213                 old_addr = new_addr;
 214                 new_addr = -ENOMEM;
 215         }
 216         if (cows)       /* Downgrade or remove this message later */
 217                 printk(KERN_WARNING "%s: mremap moved %d cows\n",
 218                                                         current->comm, cows);
 219
 220         /* Conceal VM_ACCOUNT so old reservation is not undone */
 221         if (vm_flags & VM_ACCOUNT) {
 222                 vma->vm_flags &= ~VM_ACCOUNT;
 223                 excess = vma->vm_end - vma->vm_start - old_len;
 224                 if (old_addr > vma->vm_start &&
 225                     old_addr + old_len < vma->vm_end)
 226                         split = 1;
 227         }
 228
 229         if (do_munmap(mm, old_addr, old_len) < 0) {
 230                 /* OOM: unable to split vma, just get accounts right */
 231                 vm_unacct_memory(excess >> PAGE_SHIFT);
 232                 excess = 0;
 233         }
 234
 235         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 236         if (excess) {
 237                 vma->vm_flags |= VM_ACCOUNT;
 238                 if (split)
 239                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 240         }
 241
 242         mm->total_vm += new_len >> PAGE_SHIFT;
 243         if (vm_flags & VM_LOCKED) {
 244                 mm->locked_vm += new_len >> PAGE_SHIFT;
 245                 if (new_len > old_len)
 246                         make_pages_present(new_addr + old_len,
 247                                            new_addr + new_len);
 248         }
 249
 250         return new_addr;
 251 }
 252
 253 /*
 254  * Expand (or shrink) an existing mapping, potentially moving it at the
 255  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 256  *
 257  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 258  * This option implies MREMAP_MAYMOVE.
 259  */
 260 unsigned long do_mremap(unsigned long addr,
 261         unsigned long old_len, unsigned long new_len,
 262         unsigned long flags, unsigned long new_addr)
 263 {
 264         struct vm_area_struct *vma;
 265         unsigned long ret = -EINVAL;
 266         unsigned long charged = 0;
 267
 268         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 269                 goto out;
 270
 271         if (addr & ~PAGE_MASK)
 272                 goto out;
 273
 274         old_len = PAGE_ALIGN(old_len);
 275         new_len = PAGE_ALIGN(new_len);
 276
 277         /*
 278          * We allow a zero old-len as a special case
 279          * for DOS-emu "duplicate shm area" thing. But
 280          * a zero new-len is nonsensical.
 281          */
 282         if (!new_len)
 283                 goto out;
 284
 285         /* new_addr is only valid if MREMAP_FIXED is specified */
 286         if (flags & MREMAP_FIXED) {
 287                 if (new_addr & ~PAGE_MASK)
 288                         goto out;
 289                 if (!(flags & MREMAP_MAYMOVE))
 290                         goto out;
 291
 292                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 293                         goto out;
 294
 295                 /* Check if the location we're moving into overlaps the
 296                  * old location at all, and fail if it does.
 297                  */
 298                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 299                         goto out;
 300
 301                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 302                         goto out;
 303
 304                 ret = do_munmap(current->mm, new_addr, new_len);
 305                 if (ret)
 306                         goto out;
 307         }
 308
 309         /*
 310          * Always allow a shrinking remap: that just unmaps
 311          * the unnecessary pages..
 312          * do_munmap does all the needed commit accounting
 313          */
 314         if (old_len >= new_len) {
 315                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 316                 if (ret && old_len != new_len)
 317                         goto out;
 318                 ret = addr;
 319                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 320                         goto out;
 321                 old_len = new_len;
 322         }
 323
 324         /*
 325          * Ok, we need to grow..  or relocate.
 326          */
 327         ret = -EFAULT;
 328         vma = find_vma(current->mm, addr);
 329         if (!vma || vma->vm_start > addr)
 330                 goto out;
 331         if (is_vm_hugetlb_page(vma)) {
 332                 ret = -EINVAL;
 333                 goto out;
 334         }
 335         /* We can't remap across vm area boundaries */
 336         if (old_len > vma->vm_end - addr)
 337                 goto out;
 338         if (vma->vm_flags & VM_DONTEXPAND) {
 339                 if (new_len > old_len)
 340                         goto out;
 341         }
 342         if (vma->vm_flags & VM_LOCKED) {
 343                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 344                 locked += new_len - old_len;
 345                 ret = -EAGAIN;
 346                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 347                         goto out;
 348         }
 349         ret = -ENOMEM;
 350         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 351             > current->rlim[RLIMIT_AS].rlim_cur)
 352                 goto out;
 353
 354         if (vma->vm_flags & VM_ACCOUNT) {
 355                 charged = (new_len - old_len) >> PAGE_SHIFT;
 356                 if (security_vm_enough_memory(charged))
 357                         goto out_nc;
 358         }
 359
 360         /* old_len exactly to the end of the area..
 361          * And we're not relocating the area.
 362          */
 363         if (old_len == vma->vm_end - addr &&
 364             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 365             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 366                 unsigned long max_addr = TASK_SIZE;
 367                 if (vma->vm_next)
 368                         max_addr = vma->vm_next->vm_start;
 369                 /* can we just expand the current mapping? */
 370                 if (max_addr - addr >= new_len) {
 371                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 372                         spin_lock(&vma->vm_mm->page_table_lock);
 373                         vma->vm_end = addr + new_len;
 374                         spin_unlock(&vma->vm_mm->page_table_lock);
 375                         current->mm->total_vm += pages;
 376                         if (vma->vm_flags & VM_LOCKED) {
 377                                 current->mm->locked_vm += pages;
 378                                 make_pages_present(addr + old_len,
 379                                                    addr + new_len);
 380                         }
 381                         ret = addr;
 382                         goto out;
 383                 }
 384         }
 385
 386         /*
 387          * We weren't able to just expand or shrink the area,
 388          * we need to create a new one and move it..
 389          */
 390         ret = -ENOMEM;
 391         if (flags & MREMAP_MAYMOVE) {
 392                 if (!(flags & MREMAP_FIXED)) {
 393                         unsigned long map_flags = 0;
 394                         if (vma->vm_flags & VM_MAYSHARE)
 395                                 map_flags |= MAP_SHARED;
 396
 397                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 398                                                 vma->vm_pgoff, map_flags);
 399                         ret = new_addr;
 400                         if (new_addr & ~PAGE_MASK)
 401                                 goto out;
 402                 }
 403                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 404         }
 405 out:
 406         if (ret & ~PAGE_MASK)
 407                 vm_unacct_memory(charged);
 408 out_nc:
 409         return ret;
 410 }
 411
 412 asmlinkage unsigned long sys_mremap(unsigned long addr,
 413         unsigned long old_len, unsigned long new_len,
 414         unsigned long flags, unsigned long new_addr)
 415 {
 416         unsigned long ret;
 417
 418         down_write(&current->mm->mmap_sem);
 419         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 420         up_write(&current->mm->mmap_sem);
 421         return ret;
 422 }