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 #ifdef CONFIG_HIGHPTE   /* Save a few cycles on the sane machines */
  60 static inline int page_table_present(struct mm_struct *mm, unsigned long addr)
  61 {
  62         pgd_t *pgd;
  63         pmd_t *pmd;
  64
  65         pgd = pgd_offset(mm, addr);
  66         if (pgd_none(*pgd))
  67                 return 0;
  68         pmd = pmd_offset(pgd, addr);
  69         return pmd_present(*pmd);
  70 }
  71 #else
  72 #define page_table_present(mm, addr)    (1)
  73 #endif
  74
  75 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
  76 {
  77         pmd_t *pmd;
  78         pte_t *pte = NULL;
  79
  80         pmd = pmd_alloc(mm, pgd_offset(mm, addr), addr);
  81         if (pmd)
  82                 pte = pte_alloc_map(mm, pmd, addr);
  83         return pte;
  84 }
  85
  86 static int
  87 copy_one_pte(struct mm_struct *mm, pte_t *src, pte_t *dst,
  88                 struct pte_chain **pte_chainp)
  89 {
  90         int error = 0;
  91         pte_t pte;
  92         struct page *page = NULL;
  93
  94         if (pte_present(*src))
  95                 page = pte_page(*src);
  96
  97         if (!pte_none(*src)) {
  98                 if (page)
  99                         page_remove_rmap(page, src);
 100                 pte = ptep_get_and_clear(src);
 101                 if (!dst) {
 102                         /* No dest?  We must put it back. */
 103                         dst = src;
 104                         error++;
 105                 }
 106                 set_pte(dst, pte);
 107                 if (page)
 108                         *pte_chainp = page_add_rmap(page, dst, *pte_chainp);
 109         }
 110         return error;
 111 }
 112
 113 static int
 114 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
 115                 unsigned long new_addr)
 116 {
 117         struct mm_struct *mm = vma->vm_mm;
 118         int error = 0;
 119         pte_t *src, *dst;
 120         struct pte_chain *pte_chain;
 121
 122         pte_chain = pte_chain_alloc(GFP_KERNEL);
 123         if (!pte_chain) {
 124                 error = -ENOMEM;
 125                 goto out;
 126         }
 127         spin_lock(&mm->page_table_lock);
 128         src = get_one_pte_map_nested(mm, old_addr);
 129         if (src) {
 130                 /*
 131                  * Look to see whether alloc_one_pte_map needs to perform a
 132                  * memory allocation.  If it does then we need to drop the
 133                  * atomic kmap
 134                  */
 135                 if (!page_table_present(mm, new_addr)) {
 136                         pte_unmap_nested(src);
 137                         src = NULL;
 138                 }
 139                 dst = alloc_one_pte_map(mm, new_addr);
 140                 if (src == NULL)
 141                         src = get_one_pte_map_nested(mm, old_addr);
 142                 error = copy_one_pte(mm, src, dst, &pte_chain);
 143                 pte_unmap_nested(src);
 144                 pte_unmap(dst);
 145         }
 146         flush_tlb_page(vma, old_addr);
 147         spin_unlock(&mm->page_table_lock);
 148         pte_chain_free(pte_chain);
 149 out:
 150         return error;
 151 }
 152
 153 static int move_page_tables(struct vm_area_struct *vma,
 154         unsigned long new_addr, unsigned long old_addr, unsigned long len)
 155 {
 156         unsigned long offset = len;
 157
 158         flush_cache_range(vma, old_addr, old_addr + len);
 159
 160         /*
 161          * This is not the clever way to do this, but we're taking the
 162          * easy way out on the assumption that most remappings will be
 163          * only a few pages.. This also makes error recovery easier.
 164          */
 165         while (offset) {
 166                 offset -= PAGE_SIZE;
 167                 if (move_one_page(vma, old_addr + offset, new_addr + offset))
 168                         goto oops_we_failed;
 169         }
 170         return 0;
 171
 172         /*
 173          * Ok, the move failed because we didn't have enough pages for
 174          * the new page table tree. This is unlikely, but we have to
 175          * take the possibility into account. In that case we just move
 176          * all the pages back (this will work, because we still have
 177          * the old page tables)
 178          */
 179 oops_we_failed:
 180         flush_cache_range(vma, new_addr, new_addr + len);
 181         while ((offset += PAGE_SIZE) < len)
 182                 move_one_page(vma, new_addr + offset, old_addr + offset);
 183         zap_page_range(vma, new_addr, len);
 184         return -1;
 185 }
 186
 187 static unsigned long move_vma(struct vm_area_struct *vma,
 188         unsigned long addr, unsigned long old_len, unsigned long new_len,
 189         unsigned long new_addr)
 190 {
 191         struct mm_struct *mm = vma->vm_mm;
 192         struct vm_area_struct *new_vma, *next, *prev;
 193         int allocated_vma;
 194         int split = 0;
 195
 196         new_vma = NULL;
 197         next = find_vma_prev(mm, new_addr, &prev);
 198         if (next) {
 199                 if (prev && prev->vm_end == new_addr &&
 200                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file &&
 201                                         !(vma->vm_flags & VM_SHARED)) {
 202                         spin_lock(&mm->page_table_lock);
 203                         prev->vm_end = new_addr + new_len;
 204                         spin_unlock(&mm->page_table_lock);
 205                         new_vma = prev;
 206                         if (next != prev->vm_next)
 207                                 BUG();
 208                         if (prev->vm_end == next->vm_start &&
 209                                         can_vma_merge(next, prev->vm_flags)) {
 210                                 spin_lock(&mm->page_table_lock);
 211                                 prev->vm_end = next->vm_end;
 212                                 __vma_unlink(mm, next, prev);
 213                                 spin_unlock(&mm->page_table_lock);
 214                                 if (vma == next)
 215                                         vma = prev;
 216                                 mm->map_count--;
 217                                 kmem_cache_free(vm_area_cachep, next);
 218                         }
 219                 } else if (next->vm_start == new_addr + new_len &&
 220                                 can_vma_merge(next, vma->vm_flags) &&
 221                                 !vma->vm_file && !(vma->vm_flags & VM_SHARED)) {
 222                         spin_lock(&mm->page_table_lock);
 223                         next->vm_start = new_addr;
 224                         spin_unlock(&mm->page_table_lock);
 225                         new_vma = next;
 226                 }
 227         } else {
 228                 prev = find_vma(mm, new_addr-1);
 229                 if (prev && prev->vm_end == new_addr &&
 230                     can_vma_merge(prev, vma->vm_flags) && !vma->vm_file &&
 231                                 !(vma->vm_flags & VM_SHARED)) {
 232                         spin_lock(&mm->page_table_lock);
 233                         prev->vm_end = new_addr + new_len;
 234                         spin_unlock(&mm->page_table_lock);
 235                         new_vma = prev;
 236                 }
 237         }
 238
 239         allocated_vma = 0;
 240         if (!new_vma) {
 241                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
 242                 if (!new_vma)
 243                         goto out;
 244                 allocated_vma = 1;
 245         }
 246
 247         if (!move_page_tables(vma, new_addr, addr, old_len)) {
 248                 unsigned long vm_locked = vma->vm_flags & VM_LOCKED;
 249
 250                 if (allocated_vma) {
 251                         *new_vma = *vma;
 252                         INIT_LIST_HEAD(&new_vma->shared);
 253                         new_vma->vm_start = new_addr;
 254                         new_vma->vm_end = new_addr+new_len;
 255                         new_vma->vm_pgoff += (addr-vma->vm_start) >> PAGE_SHIFT;
 256                         if (new_vma->vm_file)
 257                                 get_file(new_vma->vm_file);
 258                         if (new_vma->vm_ops && new_vma->vm_ops->open)
 259                                 new_vma->vm_ops->open(new_vma);
 260                         insert_vm_struct(current->mm, new_vma);
 261                 }
 262
 263                 /* Conceal VM_ACCOUNT so old reservation is not undone */
 264                 if (vma->vm_flags & VM_ACCOUNT) {
 265                         vma->vm_flags &= ~VM_ACCOUNT;
 266                         if (addr > vma->vm_start) {
 267                                 if (addr + old_len < vma->vm_end)
 268                                         split = 1;
 269                         } else if (addr + old_len == vma->vm_end)
 270                                 vma = NULL;     /* it will be removed */
 271                 } else
 272                         vma = NULL;             /* nothing more to do */
 273
 274                 do_munmap(current->mm, addr, old_len);
 275
 276                 /* Restore VM_ACCOUNT if one or two pieces of vma left */
 277                 if (vma) {
 278                         vma->vm_flags |= VM_ACCOUNT;
 279                         if (split)
 280                                 vma->vm_next->vm_flags |= VM_ACCOUNT;
 281                 }
 282
 283                 current->mm->total_vm += new_len >> PAGE_SHIFT;
 284                 if (vm_locked) {
 285                         current->mm->locked_vm += new_len >> PAGE_SHIFT;
 286                         if (new_len > old_len)
 287                                 make_pages_present(new_addr + old_len,
 288                                                    new_addr + new_len);
 289                 }
 290                 return new_addr;
 291         }
 292         if (allocated_vma)
 293                 kmem_cache_free(vm_area_cachep, new_vma);
 294  out:
 295         return -ENOMEM;
 296 }
 297
 298 /*
 299  * Expand (or shrink) an existing mapping, potentially moving it at the
 300  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
 301  *
 302  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
 303  * This option implies MREMAP_MAYMOVE.
 304  */
 305 unsigned long do_mremap(unsigned long addr,
 306         unsigned long old_len, unsigned long new_len,
 307         unsigned long flags, unsigned long new_addr)
 308 {
 309         struct vm_area_struct *vma;
 310         unsigned long ret = -EINVAL;
 311         unsigned long charged = 0;
 312
 313         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
 314                 goto out;
 315
 316         if (addr & ~PAGE_MASK)
 317                 goto out;
 318
 319         old_len = PAGE_ALIGN(old_len);
 320         new_len = PAGE_ALIGN(new_len);
 321
 322         /* new_addr is only valid if MREMAP_FIXED is specified */
 323         if (flags & MREMAP_FIXED) {
 324                 if (new_addr & ~PAGE_MASK)
 325                         goto out;
 326                 if (!(flags & MREMAP_MAYMOVE))
 327                         goto out;
 328
 329                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
 330                         goto out;
 331
 332                 /* Check if the location we're moving into overlaps the
 333                  * old location at all, and fail if it does.
 334                  */
 335                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
 336                         goto out;
 337
 338                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
 339                         goto out;
 340
 341                 do_munmap(current->mm, new_addr, new_len);
 342         }
 343
 344         /*
 345          * Always allow a shrinking remap: that just unmaps
 346          * the unnecessary pages..
 347          * do_munmap does all the needed commit accounting
 348          */
 349         ret = addr;
 350         if (old_len >= new_len) {
 351                 do_munmap(current->mm, addr+new_len, old_len - new_len);
 352                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
 353                         goto out;
 354                 old_len = new_len;
 355         }
 356
 357         /*
 358          * Ok, we need to grow..  or relocate.
 359          */
 360         ret = -EFAULT;
 361         vma = find_vma(current->mm, addr);
 362         if (!vma || vma->vm_start > addr)
 363                 goto out;
 364         if (is_vm_hugetlb_page(vma)) {
 365                 ret = -EINVAL;
 366                 goto out;
 367         }
 368         /* We can't remap across vm area boundaries */
 369         if (old_len > vma->vm_end - addr)
 370                 goto out;
 371         if (vma->vm_flags & VM_DONTEXPAND) {
 372                 if (new_len > old_len)
 373                         goto out;
 374         }
 375         if (vma->vm_flags & VM_LOCKED) {
 376                 unsigned long locked = current->mm->locked_vm << PAGE_SHIFT;
 377                 locked += new_len - old_len;
 378                 ret = -EAGAIN;
 379                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 380                         goto out;
 381         }
 382         ret = -ENOMEM;
 383         if ((current->mm->total_vm << PAGE_SHIFT) + (new_len - old_len)
 384             > current->rlim[RLIMIT_AS].rlim_cur)
 385                 goto out;
 386
 387         if (vma->vm_flags & VM_ACCOUNT) {
 388                 charged = (new_len - old_len) >> PAGE_SHIFT;
 389                 if (security_vm_enough_memory(charged))
 390                         goto out_nc;
 391         }
 392
 393         /* old_len exactly to the end of the area..
 394          * And we're not relocating the area.
 395          */
 396         if (old_len == vma->vm_end - addr &&
 397             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
 398             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
 399                 unsigned long max_addr = TASK_SIZE;
 400                 if (vma->vm_next)
 401                         max_addr = vma->vm_next->vm_start;
 402                 /* can we just expand the current mapping? */
 403                 if (max_addr - addr >= new_len) {
 404                         int pages = (new_len - old_len) >> PAGE_SHIFT;
 405                         spin_lock(&vma->vm_mm->page_table_lock);
 406                         vma->vm_end = addr + new_len;
 407                         spin_unlock(&vma->vm_mm->page_table_lock);
 408                         current->mm->total_vm += pages;
 409                         if (vma->vm_flags & VM_LOCKED) {
 410                                 current->mm->locked_vm += pages;
 411                                 make_pages_present(addr + old_len,
 412                                                    addr + new_len);
 413                         }
 414                         ret = addr;
 415                         goto out;
 416                 }
 417         }
 418
 419         /*
 420          * We weren't able to just expand or shrink the area,
 421          * we need to create a new one and move it..
 422          */
 423         ret = -ENOMEM;
 424         if (flags & MREMAP_MAYMOVE) {
 425                 if (!(flags & MREMAP_FIXED)) {
 426                         unsigned long map_flags = 0;
 427                         if (vma->vm_flags & VM_SHARED)
 428                                 map_flags |= MAP_SHARED;
 429
 430                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
 431                                                 vma->vm_pgoff, map_flags);
 432                         ret = new_addr;
 433                         if (new_addr & ~PAGE_MASK)
 434                                 goto out;
 435                 }
 436                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
 437         }
 438 out:
 439         if (ret & ~PAGE_MASK)
 440                 vm_unacct_memory(charged);
 441 out_nc:
 442         return ret;
 443 }
 444
 445 asmlinkage unsigned long sys_mremap(unsigned long addr,
 446         unsigned long old_len, unsigned long new_len,
 447         unsigned long flags, unsigned long new_addr)
 448 {
 449         unsigned long ret;
 450
 451         down_write(&current->mm->mmap_sem);
 452         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
 453         up_write(&current->mm->mmap_sem);
 454         return ret;
 455 }