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;
 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         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 161                 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
 162                         break;
 163         }
 164         return offset;
 165 }
 166
 167 static unsigned long move_vma(struct vm_area_struct *vma,
 168                 unsigned long old_addr, unsigned long old_len,
 169                 unsigned long new_len, unsigned long new_addr)
 170 {
 171         struct mm_struct *mm = vma->vm_mm;
 172         struct vm_area_struct *new_vma;
 173         unsigned long vm_flags = vma->vm_flags;
 174         unsigned long new_pgoff;
 175         unsigned long moved_len;
 176         unsigned long excess = 0;
 177         int split = 0;
 178
 179         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 180         new_vma = copy_vma(vma, new_addr, new_len, new_pgoff);
 181         if (!new_vma)
 182                 return -ENOMEM;
 183
 184         moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
 185         if (moved_len < old_len) {
 186                 /*
 187                  * On error, move entries back from new area to old,
 188                  * which will succeed since page tables still there,
 189                  * and then proceed to unmap new area instead of old.
 190                  *
 191                  * Subtle point from Rajesh Venkatasubramanian: before
 192                  * moving file-based ptes, move new_vma before old vma
 193                  * in the i_mmap or i_mmap_shared list, so when racing
 194                  * against vmtruncate we cannot propagate pages to be
 195                  * truncated back from new_vma into just cleaned old.
 196                  */
 197                 vma_relink_file(vma, new_vma);
 198                 move_page_tables(new_vma, old_addr, new_addr, moved_len);
 199                 vma = new_vma;
 200                 old_len = new_len;
 201                 old_addr = new_addr;
 202                 new_addr = -ENOMEM;
 203         }
 204
 205         /* Conceal VM_ACCOUNT so old reservation is not undone */
 206         if (vm_flags & VM_ACCOUNT) {
 207                 vma->vm_flags &= ~VM_ACCOUNT;
 208                 excess = vma->vm_end - vma->vm_start - old_len;
 209                 if (old_addr > vma->vm_start &&
 210                     old_addr + old_len < vma->vm_end)
 211                         split = 1;
 212         }
 213
 214         if (do_munmap(mm, old_addr, old_len) < 0) {
 215                 /* OOM: unable to split vma, just get accounts right */
 216                 vm_unacct_memory(excess >> PAGE_SHIFT);
 217                 excess = 0;
 218         }
 219
 220         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 221         if (excess) {
 222                 vma->vm_flags |= VM_ACCOUNT;
 223                 if (split)
 224                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 225         }
 226
 227         mm->total_vm += new_len >> PAGE_SHIFT;
 228         if (vm_flags & VM_LOCKED) {
 229                 mm->locked_vm += new_len >> PAGE_SHIFT;
 230                 if (new_len > old_len)
 231                         make_pages_present(new_addr + old_len,
 232                                            new_addr + new_len);
 233         }
 234
 235         return new_addr;
 236 }
 237
 238 /*
 239  * Expand (or shrink) an existing mapping, potentially moving it at the
 240  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 241  *
 242  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 243  * This option implies MREMAP_MAYMOVE.
 244  */
 245 unsigned long do_mremap(unsigned long addr,
 246         unsigned long old_len, unsigned long new_len,
 247         unsigned long flags, unsigned long new_addr)
 248 {
 249         struct vm_area_struct *vma;
 250         unsigned long ret = -EINVAL;
 251         unsigned long charged = 0;
 252
 253         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 254                 goto out;
 255
 256         if (addr & ~PAGE_MASK)
 257                 goto out;
 258
 259         old_len = PAGE_ALIGN(old_len);
 260         new_len = PAGE_ALIGN(new_len);
 261
 262         /*
 263          * We allow a zero old-len as a special case
 264          * for DOS-emu "duplicate shm area" thing. But
 265          * a zero new-len is nonsensical.
 266          */
 267         if (!new_len)
 268                 goto out;
 269
 270         /* new_addr is only valid if MREMAP_FIXED is specified */
 271         if (flags & MREMAP_FIXED) {
 272                 if (new_addr & ~PAGE_MASK)
 273                         goto out;
 274                 if (!(flags & MREMAP_MAYMOVE))
 275                         goto out;
 276
 277                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 278                         goto out;
 279
 280                 /* Check if the location we're moving into overlaps the
 281                  * old location at all, and fail if it does.
 282                  */
 283                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 284                         goto out;
 285
 286                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 287                         goto out;
 288
 289                 ret = do_munmap(current->mm, new_addr, new_len);
 290                 if (ret)
 291                         goto out;
 292         }
 293
 294         /*
 295          * Always allow a shrinking remap: that just unmaps
 296          * the unnecessary pages..
 297          * do_munmap does all the needed commit accounting
 298          */
 299         if (old_len >= new_len) {
 300                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 301                 if (ret && old_len != new_len)
 302                         goto out;
 303                 ret = addr;
 304                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 305                         goto out;
 306                 old_len = new_len;
 307         }
 308
 309         /*
 310          * Ok, we need to grow..  or relocate.
 311          */
 312         ret = -EFAULT;
 313         vma = find_vma(current->mm, addr);
 314         if (!vma || vma->vm_start > addr)
 315                 goto out;
 316         if (is_vm_hugetlb_page(vma)) {
 317                 ret = -EINVAL;
 318                 goto out;
 319         }
 320         /* We can't remap across vm area boundaries */
 321         if (old_len > vma->vm_end - addr)
 322                 goto out;
 323         if (vma->vm_flags & VM_DONTEXPAND) {
 324                 if (new_len > old_len)
 325                         goto out;
 326         }
 327         if (vma->vm_flags & VM_LOCKED) {
 328                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 329                 locked += new_len - old_len;
 330                 ret = -EAGAIN;
 331                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 332                         goto out;
 333         }
 334         ret = -ENOMEM;
 335         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 336             > current->rlim[RLIMIT_AS].rlim_cur)
 337                 goto out;
 338
 339         if (vma->vm_flags & VM_ACCOUNT) {
 340                 charged = (new_len - old_len) >> PAGE_SHIFT;
 341                 if (security_vm_enough_memory(charged))
 342                         goto out_nc;
 343         }
 344
 345         /* old_len exactly to the end of the area..
 346          * And we're not relocating the area.
 347          */
 348         if (old_len == vma->vm_end - addr &&
 349             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 350             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 351                 unsigned long max_addr = TASK_SIZE;
 352                 if (vma->vm_next)
 353                         max_addr = vma->vm_next->vm_start;
 354                 /* can we just expand the current mapping? */
 355                 if (max_addr - addr >= new_len) {
 356                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 357                         spin_lock(&vma->vm_mm->page_table_lock);
 358                         vma->vm_end = addr + new_len;
 359                         spin_unlock(&vma->vm_mm->page_table_lock);
 360                         current->mm->total_vm += pages;
 361                         if (vma->vm_flags & VM_LOCKED) {
 362                                 current->mm->locked_vm += pages;
 363                                 make_pages_present(addr + old_len,
 364                                                    addr + new_len);
 365                         }
 366                         ret = addr;
 367                         goto out;
 368                 }
 369         }
 370
 371         /*
 372          * We weren't able to just expand or shrink the area,
 373          * we need to create a new one and move it..
 374          */
 375         ret = -ENOMEM;
 376         if (flags & MREMAP_MAYMOVE) {
 377                 if (!(flags & MREMAP_FIXED)) {
 378                         unsigned long map_flags = 0;
 379                         if (vma->vm_flags & VM_MAYSHARE)
 380                                 map_flags |= MAP_SHARED;
 381
 382                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 383                                                 vma->vm_pgoff, map_flags);
 384                         ret = new_addr;
 385                         if (new_addr & ~PAGE_MASK)
 386                                 goto out;
 387                 }
 388                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 389         }
 390 out:
 391         if (ret & ~PAGE_MASK)
 392                 vm_unacct_memory(charged);
 393 out_nc:
 394         return ret;
 395 }
 396
 397 asmlinkage unsigned long sys_mremap(unsigned long addr,
 398         unsigned long old_len, unsigned long new_len,
 399         unsigned long flags, unsigned long new_addr)
 400 {
 401         unsigned long ret;
 402
 403         down_write(&current->mm->mmap_sem);
 404         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 405         up_write(&current->mm->mmap_sem);
 406         return ret;
 407 }