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/cacheflush.h>
  22 #include <asm/tlbflush.h>
  23
  24 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
  25 {
  26         pgd_t *pgd;
  27         pmd_t *pmd;
  28         pte_t *pte = NULL;
  29
  30         pgd = pgd_offset(mm, addr);
  31         if (pgd_none(*pgd))
  32                 goto end;
  33         if (pgd_bad(*pgd)) {
  34                 pgd_ERROR(*pgd);
  35                 pgd_clear(pgd);
  36                 goto end;
  37         }
  38
  39         pmd = pmd_offset(pgd, addr);
  40         if (pmd_none(*pmd))
  41                 goto end;
  42         if (pmd_bad(*pmd)) {
  43                 pmd_ERROR(*pmd);
  44                 pmd_clear(pmd);
  45                 goto end;
  46         }
  47
  48         pte = pte_offset_map_nested(pmd, addr);
  49         if (pte_none(*pte)) {
  50                 pte_unmap_nested(pte);
  51                 pte = NULL;
  52         }
  53 end:
  54         return pte;
  55 }
  56
  57 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
  58 {
  59         pgd_t *pgd;
  60         pmd_t *pmd;
  61
  62         pgd = pgd_offset(mm, addr);
  63         if (pgd_none(*pgd))
  64                 return NULL;
  65         pmd = pmd_offset(pgd, addr);
  66         if (!pmd_present(*pmd))
  67                 return NULL;
  68         return pte_offset_map(pmd, addr);
  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 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
  84                 unsigned long new_addr)
  85 {
  86         struct address_space *mapping = NULL;
  87         struct mm_struct *mm = vma->vm_mm;
  88         int error = 0;
  89         pte_t *src, *dst;
  90
  91         if (vma->vm_file) {
  92                 /*
  93                  * Subtle point from Rajesh Venkatasubramanian: before
  94                  * moving file-based ptes, we must lock vmtruncate out,
  95                  * since it might clean the dst vma before the src vma,
  96                  * and we propagate stale pages into the dst afterward.
  97                  */
  98                 mapping = vma->vm_file->f_mapping;
  99                 spin_lock(&mapping->i_mmap_lock);
 100         }
 101         spin_lock(&mm->page_table_lock);
 102
 103         src = get_one_pte_map_nested(mm, old_addr);
 104         if (src) {
 105                 /*
 106                  * Look to see whether alloc_one_pte_map needs to perform a
 107                  * memory allocation.  If it does then we need to drop the
 108                  * atomic kmap
 109                  */
 110                 dst = get_one_pte_map(mm, new_addr);
 111                 if (unlikely(!dst)) {
 112                         pte_unmap_nested(src);
 113                         if (mapping)
 114                                 spin_unlock(&mapping->i_mmap_lock);
 115                         dst = alloc_one_pte_map(mm, new_addr);
 116                         if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
 117                                 spin_unlock(&mm->page_table_lock);
 118                                 spin_lock(&mapping->i_mmap_lock);
 119                                 spin_lock(&mm->page_table_lock);
 120                         }
 121                         src = get_one_pte_map_nested(mm, old_addr);
 122                 }
 123                 /*
 124                  * Since alloc_one_pte_map can drop and re-acquire
 125                  * page_table_lock, we should re-check the src entry...
 126                  */
 127                 if (src) {
 128                         if (dst) {
 129                                 pte_t pte;
 130                                 pte = ptep_clear_flush(vma, old_addr, src);
 131                                 set_pte(dst, pte);
 132                         } else
 133                                 error = -ENOMEM;
 134                         pte_unmap_nested(src);
 135                 }
 136                 if (dst)
 137                         pte_unmap(dst);
 138         }
 139         spin_unlock(&mm->page_table_lock);
 140         if (mapping)
 141                 spin_unlock(&mapping->i_mmap_lock);
 142         return error;
 143 }
 144
 145 static unsigned long move_page_tables(struct vm_area_struct *vma,
 146                 unsigned long new_addr, unsigned long old_addr,
 147                 unsigned long len)
 148 {
 149         unsigned long offset;
 150
 151         flush_cache_range(vma, old_addr, old_addr + len);
 152
 153         /*
 154          * This is not the clever way to do this, but we're taking the
 155          * easy way out on the assumption that most remappings will be
 156          * only a few pages.. This also makes error recovery easier.
 157          */
 158         for (offset = 0; offset < len; offset += PAGE_SIZE) {
 159                 if (move_one_page(vma, old_addr+offset, new_addr+offset) < 0)
 160                         break;
 161                 cond_resched();
 162         }
 163         return offset;
 164 }
 165
 166 static unsigned long move_vma(struct vm_area_struct *vma,
 167                 unsigned long old_addr, unsigned long old_len,
 168                 unsigned long new_len, unsigned long new_addr)
 169 {
 170         struct mm_struct *mm = vma->vm_mm;
 171         struct vm_area_struct *new_vma;
 172         unsigned long vm_flags = vma->vm_flags;
 173         unsigned long new_pgoff;
 174         unsigned long moved_len;
 175         unsigned long excess = 0;
 176         int split = 0;
 177
 178         /*
 179          * We'd prefer to avoid failure later on in do_munmap:
 180          * which may split one vma into three before unmapping.
 181          */
 182         if (mm->map_count >= sysctl_max_map_count - 3)
 183                 return -ENOMEM;
 184
 185         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
 186         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
 187         if (!new_vma)
 188                 return -ENOMEM;
 189
 190         moved_len = move_page_tables(vma, new_addr, old_addr, old_len);
 191         if (moved_len < old_len) {
 192                 /*
 193                  * On error, move entries back from new area to old,
 194                  * which will succeed since page tables still there,
 195                  * and then proceed to unmap new area instead of old.
 196                  */
 197                 move_page_tables(new_vma, old_addr, new_addr, moved_len);
 198                 vma = new_vma;
 199                 old_len = new_len;
 200                 old_addr = new_addr;
 201                 new_addr = -ENOMEM;
 202         }
 203
 204         /* Conceal VM_ACCOUNT so old reservation is not undone */
 205         if (vm_flags & VM_ACCOUNT) {
 206                 vma->vm_flags &= ~VM_ACCOUNT;
 207                 excess = vma->vm_end - vma->vm_start - old_len;
 208                 if (old_addr > vma->vm_start &&
 209                     old_addr + old_len < vma->vm_end)
 210                         split = 1;
 211         }
 212
 213         if (do_munmap(mm, old_addr, old_len) < 0) {
 214                 /* OOM: unable to split vma, just get accounts right */
 215                 vm_unacct_memory(excess >> PAGE_SHIFT);
 216                 excess = 0;
 217         }
 218
 219         /* Restore VM_ACCOUNT if one or two pieces of vma left */
 220         if (excess) {
 221                 vma->vm_flags |= VM_ACCOUNT;
 222                 if (split)
 223                         vma->vm_next->vm_flags |= VM_ACCOUNT;
 224         }
 225
 226         mm->total_vm += new_len >> PAGE_SHIFT;
 227         __vm_stat_account(mm, vma->vm_flags, vma->vm_file, 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, lock_limit;
 329                 locked = current->mm->locked_vm << PAGE_SHIFT;
 330                 lock_limit = current->rlim[RLIMIT_MEMLOCK].rlim_cur;
 331                 locked += new_len - old_len;
 332                 ret = -EAGAIN;
 333                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
 334                         goto out;
 335         }
 336         ret = -ENOMEM;
 337         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 338             > current->rlim[RLIMIT_AS].rlim_cur)
 339                 goto out;
 340
 341         if (vma->vm_flags & VM_ACCOUNT) {
 342                 charged = (new_len - old_len) >> PAGE_SHIFT;
 343                 if (security_vm_enough_memory(charged))
 344                         goto out_nc;
 345         }
 346
 347         /* old_len exactly to the end of the area..
 348          * And we're not relocating the area.
 349          */
 350         if (old_len == vma->vm_end - addr &&
 351             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 352             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 353                 unsigned long max_addr = TASK_SIZE;
 354                 if (vma->vm_next)
 355                         max_addr = vma->vm_next->vm_start;
 356                 /* can we just expand the current mapping? */
 357                 if (max_addr - addr >= new_len) {
 358                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 359
 360                         vma_adjust(vma, vma->vm_start,
 361                                 addr + new_len, vma->vm_pgoff, NULL);
 362
 363                         current->mm->total_vm += pages;
 364                         __vm_stat_account(vma->vm_mm, vma->vm_flags,
 365                                                         vma->vm_file, pages);
 366                         if (vma->vm_flags & VM_LOCKED) {
 367                                 current->mm->locked_vm += pages;
 368                                 make_pages_present(addr + old_len,
 369                                                    addr + new_len);
 370                         }
 371                         ret = addr;
 372                         goto out;
 373                 }
 374         }
 375
 376         /*
 377          * We weren't able to just expand or shrink the area,
 378          * we need to create a new one and move it..
 379          */
 380         ret = -ENOMEM;
 381         if (flags & MREMAP_MAYMOVE) {
 382                 if (!(flags & MREMAP_FIXED)) {
 383                         unsigned long map_flags = 0;
 384                         if (vma->vm_flags & VM_MAYSHARE)
 385                                 map_flags |= MAP_SHARED;
 386
 387                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 388                                                 vma->vm_pgoff, map_flags);
 389                         ret = new_addr;
 390                         if (new_addr & ~PAGE_MASK)
 391                                 goto out;
 392                 }
 393                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 394         }
 395 out:
 396         if (ret & ~PAGE_MASK)
 397                 vm_unacct_memory(charged);
 398 out_nc:
 399         return ret;
 400 }
 401
 402 asmlinkage unsigned long sys_mremap(unsigned long addr,
 403         unsigned long old_len, unsigned long new_len,
 404         unsigned long flags, unsigned long new_addr)
 405 {
 406         unsigned long ret;
 407
 408         down_write(&current->mm->mmap_sem);
 409         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 410         up_write(&current->mm->mmap_sem);
 411         return ret;
 412 }