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 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                 if (dst)
 141                         pte_unmap(dst);
 142         }
 143         spin_unlock(&mm->page_table_lock);
 144         pte_chain_free(pte_chain);
 145 out:
 146         return error;
 147 }
 148
 149 static unsigned long 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;
 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         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 162                 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
 163                         break;
 164         }
 165         return offset;
 166 }
 167
 168 static unsigned long move_vma(struct vm_area_struct *vma,
 169                 unsigned long old_addr, unsigned long old_len,
 170                 unsigned long new_len, unsigned long new_addr)
 171 {
 172         struct mm_struct *mm = vma->vm_mm;
 173         struct address_space *mapping = NULL;
 174         struct vm_area_struct *new_vma;
 175         unsigned long vm_flags = vma->vm_flags;
 176         unsigned long new_pgoff;
 177         unsigned long moved_len;
 178         unsigned long excess = 0;
 179         int split = 0;
 180
 181         /*
 182          * We'd prefer to avoid failure later on in do_munmap:
 183          * which may split one vma into three before unmapping.
 184          */
 185         if (mm->map_count >= sysctl_max_map_count - 3)
 186                 return -ENOMEM;
 187
 188         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 189         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 190         if (!new_vma)
 191                 return -ENOMEM;
 192
 193         if (vma->vm_file) {
 194                 /*
 195                  * Subtle point from Rajesh Venkatasubramanian: before
 196                  * moving file-based ptes, we must lock vmtruncate out,
 197                  * since it might clean the dst vma before the src vma,
 198                  * and we propagate stale pages into the dst afterward.
 199                  */
 200                 mapping = vma->vm_file->f_mapping;
 201                 down(&mapping->i_shared_sem);
 202         }
 203         moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
 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);
 211                 vma = new_vma;
 212                 old_len = new_len;
 213                 old_addr = new_addr;
 214                 new_addr = -ENOMEM;
 215         }
 216         if (mapping)
 217                 up(&mapping->i_shared_sem);
 218
 219         /* Conceal VM_ACCOUNT so old reservation is not undone */
 220         if (vm_flags & VM_ACCOUNT) {
 221                 vma->vm_flags &= ~VM_ACCOUNT;
 222                 excess = vma->vm_end - vma->vm_start - old_len;
 223                 if (old_addr > vma->vm_start &&
 224                     old_addr + old_len < vma->vm_end)
 225                         split = 1;
 226         }
 227
 228         if (do_munmap(mm, old_addr, old_len) < 0) {
 229                 /* OOM: unable to split vma, just get accounts right */
 230                 vm_unacct_memory(excess >> PAGE_SHIFT);
 231                 excess = 0;
 232         }
 233
 234         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 235         if (excess) {
 236                 vma->vm_flags |= VM_ACCOUNT;
 237                 if (split)
 238                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 239         }
 240
 241         mm->total_vm += new_len >> PAGE_SHIFT;
 242         if (vm_flags & VM_LOCKED) {
 243                 mm->locked_vm += new_len >> PAGE_SHIFT;
 244                 if (new_len > old_len)
 245                         make_pages_present(new_addr + old_len,
 246                                            new_addr + new_len);
 247         }
 248
 249         return new_addr;
 250 }
 251
 252 /*
 253  * Expand (or shrink) an existing mapping, potentially moving it at the
 254  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 255  *
 256  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 257  * This option implies MREMAP_MAYMOVE.
 258  */
 259 unsigned long do_mremap(unsigned long addr,
 260         unsigned long old_len, unsigned long new_len,
 261         unsigned long flags, unsigned long new_addr)
 262 {
 263         struct vm_area_struct *vma;
 264         unsigned long ret = -EINVAL;
 265         unsigned long charged = 0;
 266
 267         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 268                 goto out;
 269
 270         if (addr & ~PAGE_MASK)
 271                 goto out;
 272
 273         old_len = PAGE_ALIGN(old_len);
 274         new_len = PAGE_ALIGN(new_len);
 275
 276         /*
 277          * We allow a zero old-len as a special case
 278          * for DOS-emu "duplicate shm area" thing. But
 279          * a zero new-len is nonsensical.
 280          */
 281         if (!new_len)
 282                 goto out;
 283
 284         /* new_addr is only valid if MREMAP_FIXED is specified */
 285         if (flags & MREMAP_FIXED) {
 286                 if (new_addr & ~PAGE_MASK)
 287                         goto out;
 288                 if (!(flags & MREMAP_MAYMOVE))
 289                         goto out;
 290
 291                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 292                         goto out;
 293
 294                 /* Check if the location we're moving into overlaps the
 295                  * old location at all, and fail if it does.
 296                  */
 297                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 298                         goto out;
 299
 300                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 301                         goto out;
 302
 303                 ret = do_munmap(current->mm, new_addr, new_len);
 304                 if (ret)
 305                         goto out;
 306         }
 307
 308         /*
 309          * Always allow a shrinking remap: that just unmaps
 310          * the unnecessary pages..
 311          * do_munmap does all the needed commit accounting
 312          */
 313         if (old_len >= new_len) {
 314                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
 315                 if (ret && old_len != new_len)
 316                         goto out;
 317                 ret = addr;
 318                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 319                         goto out;
 320                 old_len = new_len;
 321         }
 322
 323         /*
 324          * Ok, we need to grow..  or relocate.
 325          */
 326         ret = -EFAULT;
 327         vma = find_vma(current->mm, addr);
 328         if (!vma || vma->vm_start > addr)
 329                 goto out;
 330         if (is_vm_hugetlb_page(vma)) {
 331                 ret = -EINVAL;
 332                 goto out;
 333         }
 334         /* We can't remap across vm area boundaries */
 335         if (old_len > vma->vm_end - addr)
 336                 goto out;
 337         if (vma->vm_flags & VM_DONTEXPAND) {
 338                 if (new_len > old_len)
 339                         goto out;
 340         }
 341         if (vma->vm_flags & VM_LOCKED) {
 342                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 343                 locked += new_len - old_len;
 344                 ret = -EAGAIN;
 345                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 346                         goto out;
 347         }
 348         ret = -ENOMEM;
 349         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 350             > current->rlim[RLIMIT_AS].rlim_cur)
 351                 goto out;
 352
 353         if (vma->vm_flags & VM_ACCOUNT) {
 354                 charged = (new_len - old_len) >> PAGE_SHIFT;
 355                 if (security_vm_enough_memory(charged))
 356                         goto out_nc;
 357         }
 358
 359         /* old_len exactly to the end of the area..
 360          * And we're not relocating the area.
 361          */
 362         if (old_len == vma->vm_end - addr &&
 363             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 364             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 365                 unsigned long max_addr = TASK_SIZE;
 366                 if (vma->vm_next)
 367                         max_addr = vma->vm_next->vm_start;
 368                 /* can we just expand the current mapping? */
 369                 if (max_addr - addr >= new_len) {
 370                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 371                         spin_lock(&vma->vm_mm->page_table_lock);
 372                         vma->vm_end = addr + new_len;
 373                         spin_unlock(&vma->vm_mm->page_table_lock);
 374                         current->mm->total_vm += pages;
 375                         if (vma->vm_flags & VM_LOCKED) {
 376                                 current->mm->locked_vm += pages;
 377                                 make_pages_present(addr + old_len,
 378                                                    addr + new_len);
 379                         }
 380                         ret = addr;
 381                         goto out;
 382                 }
 383         }
 384
 385         /*
 386          * We weren't able to just expand or shrink the area,
 387          * we need to create a new one and move it..
 388          */
 389         ret = -ENOMEM;
 390         if (flags & MREMAP_MAYMOVE) {
 391                 if (!(flags & MREMAP_FIXED)) {
 392                         unsigned long map_flags = 0;
 393                         if (vma->vm_flags & VM_MAYSHARE)
 394                                 map_flags |= MAP_SHARED;
 395
 396                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 397                                                 vma->vm_pgoff, map_flags);
 398                         ret = new_addr;
 399                         if (new_addr & ~PAGE_MASK)
 400                                 goto out;
 401                 }
 402                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 403         }
 404 out:
 405         if (ret & ~PAGE_MASK)
 406                 vm_unacct_memory(charged);
 407 out_nc:
 408         return ret;
 409 }
 410
 411 asmlinkage unsigned long sys_mremap(unsigned long addr,
 412         unsigned long old_len, unsigned long new_len,
 413         unsigned long flags, unsigned long new_addr)
 414 {
 415         unsigned long ret;
 416
 417         down_write(&current->mm->mmap_sem);
 418         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 419         up_write(&current->mm->mmap_sem);
 420         return ret;
 421 }