mm/mmap.c

   1 /*
   2  * mm/mmap.c
   3  *
   4  * Written by obz.
   5  *
   6  * Address space accounting code        <alan@redhat.com>
   7  */
   8
   9 #include <linux/slab.h>
  10 #include <linux/shm.h>
  11 #include <linux/mman.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/swap.h>
  14 #include <linux/syscalls.h>
  15 #include <linux/init.h>
  16 #include <linux/file.h>
  17 #include <linux/fs.h>
  18 #include <linux/personality.h>
  19 #include <linux/security.h>
  20 #include <linux/hugetlb.h>
  21 #include <linux/profile.h>
  22 #include <linux/module.h>
  23 #include <linux/mount.h>
  24 #include <linux/mempolicy.h>
  25 #include <linux/rmap.h>
  26
  27 #include <asm/uaccess.h>
  28 #include <asm/pgalloc.h>
  29 #include <asm/cacheflush.h>
  30 #include <asm/tlb.h>
  31
  32 /*
  33  * WARNING: the debugging will use recursive algorithms so never enable this
  34  * unless you know what you are doing.
  35  */
  36 #undef DEBUG_MM_RB
  37
  38 /* description of effects of mapping type and prot in current implementation.
  39  * this is due to the limited x86 page protection hardware.  The expected
  40  * behavior is in parens:
  41  *
  42  * map_type     prot
  43  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
  44  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
  45  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
  46  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
  47  *
  48  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
  49  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
  50  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
  51  *
  52  */
  53 pgprot_t protection_map[16] = {
  54         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
  55         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
  56 };
  57
  58 int sysctl_overcommit_memory = 0;       /* default is heuristic overcommit */
  59 int sysctl_overcommit_ratio = 50;       /* default is 50% */
  60 int sysctl_max_map_count = DEFAULT_MAX_MAP_COUNT;
  61 atomic_t vm_committed_space = ATOMIC_INIT(0);
  62
  63 EXPORT_SYMBOL(sysctl_overcommit_memory);
  64 EXPORT_SYMBOL(sysctl_overcommit_ratio);
  65 EXPORT_SYMBOL(sysctl_max_map_count);
  66 EXPORT_SYMBOL(vm_committed_space);
  67
  68 /*
  69  * Requires inode->i_mapping->i_mmap_lock
  70  */
  71 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
  72                 struct file *file, struct address_space *mapping)
  73 {
  74         if (vma->vm_flags & VM_DENYWRITE)
  75                 atomic_inc(&file->f_dentry->d_inode->i_writecount);
  76         if (vma->vm_flags & VM_SHARED)
  77                 mapping->i_mmap_writable--;
  78
  79         flush_dcache_mmap_lock(mapping);
  80         if (unlikely(vma->vm_flags & VM_NONLINEAR))
  81                 list_del_init(&vma->shared.vm_set.list);
  82         else
  83                 vma_prio_tree_remove(vma, &mapping->i_mmap);
  84         flush_dcache_mmap_unlock(mapping);
  85 }
  86
  87 /*
  88  * Remove one vm structure and free it.
  89  */
  90 static void remove_vm_struct(struct vm_area_struct *vma)
  91 {
  92         struct file *file = vma->vm_file;
  93
  94         if (file) {
  95                 struct address_space *mapping = file->f_mapping;
  96                 spin_lock(&mapping->i_mmap_lock);
  97                 __remove_shared_vm_struct(vma, file, mapping);
  98                 spin_unlock(&mapping->i_mmap_lock);
  99         }
 100         if (vma->vm_ops && vma->vm_ops->close)
 101                 vma->vm_ops->close(vma);
 102         if (file)
 103                 fput(file);
 104         anon_vma_unlink(vma);
 105         mpol_free(vma_policy(vma));
 106         kmem_cache_free(vm_area_cachep, vma);
 107 }
 108
 109 /*
 110  *  sys_brk() for the most part doesn't need the global kernel
 111  *  lock, except when an application is doing something nasty
 112  *  like trying to un-brk an area that has already been mapped
 113  *  to a regular file.  in this case, the unmapping will need
 114  *  to invoke file system routines that need the global lock.
 115  */
 116 asmlinkage unsigned long sys_brk(unsigned long brk)
 117 {
 118         unsigned long rlim, retval;
 119         unsigned long newbrk, oldbrk;
 120         struct mm_struct *mm = current->mm;
 121
 122         down_write(&mm->mmap_sem);
 123
 124         if (brk < mm->end_code)
 125                 goto out;
 126         newbrk = PAGE_ALIGN(brk);
 127         oldbrk = PAGE_ALIGN(mm->brk);
 128         if (oldbrk == newbrk)
 129                 goto set_brk;
 130
 131         /* Always allow shrinking brk. */
 132         if (brk <= mm->brk) {
 133                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
 134                         goto set_brk;
 135                 goto out;
 136         }
 137
 138         /* Check against rlimit.. */
 139         rlim = current->rlim[RLIMIT_DATA].rlim_cur;
 140         if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
 141                 goto out;
 142
 143         /* Check against existing mmap mappings. */
 144         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
 145                 goto out;
 146
 147         /* Ok, looks good - let it rip. */
 148         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
 149                 goto out;
 150 set_brk:
 151         mm->brk = brk;
 152 out:
 153         retval = mm->brk;
 154         up_write(&mm->mmap_sem);
 155         return retval;
 156 }
 157
 158 #ifdef DEBUG_MM_RB
 159 static int browse_rb(struct rb_root *root)
 160 {
 161         int i = 0, j;
 162         struct rb_node *nd, *pn = NULL;
 163         unsigned long prev = 0, pend = 0;
 164
 165         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
 166                 struct vm_area_struct *vma;
 167                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
 168                 if (vma->vm_start < prev)
 169                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
 170                 if (vma->vm_start < pend)
 171                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
 172                 if (vma->vm_start > vma->vm_end)
 173                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
 174                 i++;
 175                 pn = nd;
 176         }
 177         j = 0;
 178         for (nd = pn; nd; nd = rb_prev(nd)) {
 179                 j++;
 180         }
 181         if (i != j)
 182                 printk("backwards %d, forwards %d\n", j, i), i = 0;
 183         return i;
 184 }
 185
 186 void validate_mm(struct mm_struct *mm)
 187 {
 188         int bug = 0;
 189         int i = 0;
 190         struct vm_area_struct *tmp = mm->mmap;
 191         while (tmp) {
 192                 tmp = tmp->vm_next;
 193                 i++;
 194         }
 195         if (i != mm->map_count)
 196                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
 197         i = browse_rb(&mm->mm_rb);
 198         if (i != mm->map_count)
 199                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
 200         if (bug)
 201                 BUG();
 202 }
 203 #else
 204 #define validate_mm(mm) do { } while (0)
 205 #endif
 206
 207 static struct vm_area_struct *
 208 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
 209                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
 210                 struct rb_node ** rb_parent)
 211 {
 212         struct vm_area_struct * vma;
 213         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
 214
 215         __rb_link = &mm->mm_rb.rb_node;
 216         rb_prev = __rb_parent = NULL;
 217         vma = NULL;
 218
 219         while (*__rb_link) {
 220                 struct vm_area_struct *vma_tmp;
 221
 222                 __rb_parent = *__rb_link;
 223                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
 224
 225                 if (vma_tmp->vm_end > addr) {
 226                         vma = vma_tmp;
 227                         if (vma_tmp->vm_start <= addr)
 228                                 return vma;
 229                         __rb_link = &__rb_parent->rb_left;
 230                 } else {
 231                         rb_prev = __rb_parent;
 232                         __rb_link = &__rb_parent->rb_right;
 233                 }
 234         }
 235
 236         *pprev = NULL;
 237         if (rb_prev)
 238                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 239         *rb_link = __rb_link;
 240         *rb_parent = __rb_parent;
 241         return vma;
 242 }
 243
 244 static inline void
 245 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
 246                 struct vm_area_struct *prev, struct rb_node *rb_parent)
 247 {
 248         if (vma->vm_flags & VM_EXEC)
 249                 arch_add_exec_range(mm, vma->vm_end);
 250         if (prev) {
 251                 vma->vm_next = prev->vm_next;
 252                 prev->vm_next = vma;
 253         } else {
 254                 mm->mmap = vma;
 255                 if (rb_parent)
 256                         vma->vm_next = rb_entry(rb_parent,
 257                                         struct vm_area_struct, vm_rb);
 258                 else
 259                         vma->vm_next = NULL;
 260         }
 261 }
 262
 263 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
 264                 struct rb_node **rb_link, struct rb_node *rb_parent)
 265 {
 266         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
 267         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
 268 }
 269
 270 static inline void __vma_link_file(struct vm_area_struct *vma)
 271 {
 272         struct file * file;
 273
 274         file = vma->vm_file;
 275         if (file) {
 276                 struct address_space *mapping = file->f_mapping;
 277
 278                 if (vma->vm_flags & VM_DENYWRITE)
 279                         atomic_dec(&file->f_dentry->d_inode->i_writecount);
 280                 if (vma->vm_flags & VM_SHARED)
 281                         mapping->i_mmap_writable++;
 282
 283                 flush_dcache_mmap_lock(mapping);
 284                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
 285                         list_add_tail(&vma->shared.vm_set.list,
 286                                         &mapping->i_mmap_nonlinear);
 287                 else
 288                         vma_prio_tree_insert(vma, &mapping->i_mmap);
 289                 flush_dcache_mmap_unlock(mapping);
 290         }
 291 }
 292
 293 static void
 294 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 295         struct vm_area_struct *prev, struct rb_node **rb_link,
 296         struct rb_node *rb_parent)
 297 {
 298         vma_prio_tree_init(vma);
 299         __vma_link_list(mm, vma, prev, rb_parent);
 300         __vma_link_rb(mm, vma, rb_link, rb_parent);
 301         __vma_link_file(vma);
 302         __anon_vma_link(vma);
 303 }
 304
 305 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
 306                         struct vm_area_struct *prev, struct rb_node **rb_link,
 307                         struct rb_node *rb_parent)
 308 {
 309         struct address_space *mapping = NULL;
 310
 311         if (vma->vm_file)
 312                 mapping = vma->vm_file->f_mapping;
 313
 314         if (mapping)
 315                 spin_lock(&mapping->i_mmap_lock);
 316         anon_vma_lock(vma);
 317         __vma_link(mm, vma, prev, rb_link, rb_parent);
 318         anon_vma_unlock(vma);
 319         if (mapping)
 320                 spin_unlock(&mapping->i_mmap_lock);
 321
 322         mark_mm_hugetlb(mm, vma);
 323         mm->map_count++;
 324         validate_mm(mm);
 325 }
 326
 327 /*
 328  * Insert vm structure into process list sorted by address and into the
 329  * inode's i_mmap tree. The caller should hold mm->mmap_sem and
 330  * ->f_mappping->i_mmap_lock if vm_file is non-NULL.
 331  */
 332 static void
 333 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
 334 {
 335         struct vm_area_struct * __vma, * prev;
 336         struct rb_node ** rb_link, * rb_parent;
 337
 338         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
 339         if (__vma && __vma->vm_start < vma->vm_end)
 340                 BUG();
 341         __vma_link(mm, vma, prev, rb_link, rb_parent);
 342         mark_mm_hugetlb(mm, vma);
 343         mm->map_count++;
 344         validate_mm(mm);
 345 }
 346
 347 static inline void __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
 348                   struct vm_area_struct *prev)
 349 {
 350         prev->vm_next = vma->vm_next;
 351         rb_erase(&vma->vm_rb, &mm->mm_rb);
 352         if (mm->mmap_cache == vma)
 353                 mm->mmap_cache = prev;
 354         if (vma->vm_flags & VM_EXEC)
 355                 arch_remove_exec_range(mm, vma->vm_end);
 356 }
 357
 358 /*
 359  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
 360  * is already present in an i_mmap tree without adjusting the tree.
 361  * The following helper function should be used when such adjustments
 362  * are necessary.  The "insert" vma (if any) is to be inserted
 363  * before we drop the necessary locks.
 364  */
 365 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
 366         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
 367 {
 368         struct mm_struct *mm = vma->vm_mm;
 369         struct vm_area_struct *next = vma->vm_next;
 370         struct address_space *mapping = NULL;
 371         struct prio_tree_root *root = NULL;
 372         struct file *file = vma->vm_file;
 373         struct anon_vma *anon_vma = NULL;
 374         long adjust_next = 0;
 375         int remove_next = 0;
 376
 377         if (next && !insert) {
 378                 if (end >= next->vm_end) {
 379 again:                  remove_next = 1 + (end > next->vm_end);
 380                         end = next->vm_end;
 381                         anon_vma = next->anon_vma;
 382                 } else if (end < vma->vm_end || end > next->vm_start) {
 383                         /*
 384                          * vma shrinks, and !insert tells it's not
 385                          * split_vma inserting another: so it must
 386                          * be mprotect shifting the boundary down.
 387                          *   Or:
 388                          * vma expands, overlapping part of the next:
 389                          * must be mprotect shifting the boundary up.
 390                          */
 391                         BUG_ON(vma->vm_end != next->vm_start);
 392                         adjust_next = end - next->vm_start;
 393                         anon_vma = next->anon_vma;
 394                 }
 395         }
 396
 397         if (file) {
 398                 mapping = file->f_mapping;
 399                 if (!(vma->vm_flags & VM_NONLINEAR))
 400                         root = &mapping->i_mmap;
 401                 spin_lock(&mapping->i_mmap_lock);
 402         }
 403
 404         /*
 405          * When changing only vma->vm_end, we don't really need
 406          * anon_vma lock: but is that case worth optimizing out?
 407          */
 408         if (vma->anon_vma)
 409                 anon_vma = vma->anon_vma;
 410         if (anon_vma)
 411                 spin_lock(&anon_vma->lock);
 412
 413         if (root) {
 414                 flush_dcache_mmap_lock(mapping);
 415                 vma_prio_tree_remove(vma, root);
 416                 if (adjust_next)
 417                         vma_prio_tree_remove(next, root);
 418         }
 419
 420         vma->vm_start = start;
 421         vma->vm_end = end;
 422         vma->vm_pgoff = pgoff;
 423         if (adjust_next) {
 424                 next->vm_start += adjust_next;
 425                 next->vm_pgoff += adjust_next >> PAGE_SHIFT;
 426         }
 427
 428         if (root) {
 429                 if (adjust_next) {
 430                         vma_prio_tree_init(next);
 431                         vma_prio_tree_insert(next, root);
 432                 }
 433                 vma_prio_tree_init(vma);
 434                 vma_prio_tree_insert(vma, root);
 435                 flush_dcache_mmap_unlock(mapping);
 436         }
 437
 438         if (remove_next) {
 439                 /*
 440                  * vma_merge has merged next into vma, and needs
 441                  * us to remove next before dropping the locks.
 442                  */
 443                 __vma_unlink(mm, next, vma);
 444                 if (file)
 445                         __remove_shared_vm_struct(next, file, mapping);
 446                 if (next->anon_vma)
 447                         __anon_vma_merge(vma, next);
 448         } else if (insert) {
 449                 /*
 450                  * split_vma has split insert from vma, and needs
 451                  * us to insert it before dropping the locks
 452                  * (it may either follow vma or precede it).
 453                  */
 454                 __insert_vm_struct(mm, insert);
 455         }
 456
 457         if (anon_vma)
 458                 spin_unlock(&anon_vma->lock);
 459         if (mapping)
 460                 spin_unlock(&mapping->i_mmap_lock);
 461
 462         if (remove_next) {
 463                 if (file)
 464                         fput(file);
 465                 mm->map_count--;
 466                 mpol_free(vma_policy(next));
 467                 kmem_cache_free(vm_area_cachep, next);
 468                 /*
 469                  * In mprotect's case 6 (see comments on vma_merge),
 470                  * we must remove another next too. It would clutter
 471                  * up the code too much to do both in one go.
 472                  */
 473                 if (remove_next == 2) {
 474                         next = vma->vm_next;
 475                         goto again;
 476                 }
 477         }
 478
 479         validate_mm(mm);
 480 }
 481
 482 /*
 483  * If the vma has a ->close operation then the driver probably needs to release
 484  * per-vma resources, so we don't attempt to merge those.
 485  */
 486 #define VM_SPECIAL (VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_RESERVED)
 487
 488 static inline int is_mergeable_vma(struct vm_area_struct *vma,
 489                         struct file *file, unsigned long vm_flags)
 490 {
 491         if (vma->vm_flags != vm_flags)
 492                 return 0;
 493         if (vma->vm_file != file)
 494                 return 0;
 495         if (vma->vm_ops && vma->vm_ops->close)
 496                 return 0;
 497         return 1;
 498 }
 499
 500 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
 501                                         struct anon_vma *anon_vma2)
 502 {
 503         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
 504 }
 505
 506 /*
 507  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 508  * in front of (at a lower virtual address and file offset than) the vma.
 509  *
 510  * We cannot merge two vmas if they have differently assigned (non-NULL)
 511  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 512  *
 513  * We don't check here for the merged mmap wrapping around the end of pagecache
 514  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
 515  * wrap, nor mmaps which cover the final page at index -1UL.
 516  */
 517 static int
 518 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
 519         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
 520 {
 521         if (is_mergeable_vma(vma, file, vm_flags) &&
 522             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
 523                 if (vma->vm_pgoff == vm_pgoff)
 524                         return 1;
 525         }
 526         return 0;
 527 }
 528
 529 /*
 530  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
 531  * beyond (at a higher virtual address and file offset than) the vma.
 532  *
 533  * We cannot merge two vmas if they have differently assigned (non-NULL)
 534  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
 535  */
 536 static int
 537 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
 538         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
 539 {
 540         if (is_mergeable_vma(vma, file, vm_flags) &&
 541             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
 542                 pgoff_t vm_pglen;
 543                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 544                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
 545                         return 1;
 546         }
 547         return 0;
 548 }
 549
 550 /*
 551  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
 552  * whether that can be merged with its predecessor or its successor.
 553  * Or both (it neatly fills a hole).
 554  *
 555  * In most cases - when called for mmap, brk or mremap - [addr,end) is
 556  * certain not to be mapped by the time vma_merge is called; but when
 557  * called for mprotect, it is certain to be already mapped (either at
 558  * an offset within prev, or at the start of next), and the flags of
 559  * this area are about to be changed to vm_flags - and the no-change
 560  * case has already been eliminated.
 561  *
 562  * The following mprotect cases have to be considered, where AAAA is
 563  * the area passed down from mprotect_fixup, never extending beyond one
 564  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
 565  *
 566  *     AAAA             AAAA                AAAA          AAAA
 567  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
 568  *    cannot merge    might become    might become    might become
 569  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
 570  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
 571  *    mremap move:                                    PPPPNNNNNNNN 8
 572  *        AAAA
 573  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
 574  *    might become    case 1 below    case 2 below    case 3 below
 575  *
 576  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
 577  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
 578  */
 579 struct vm_area_struct *vma_merge(struct mm_struct *mm,
 580                         struct vm_area_struct *prev, unsigned long addr,
 581                         unsigned long end, unsigned long vm_flags,
 582                         struct anon_vma *anon_vma, struct file *file,
 583                         pgoff_t pgoff, struct mempolicy *policy)
 584 {
 585         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
 586         struct vm_area_struct *area, *next;
 587
 588         /*
 589          * We later require that vma->vm_flags == vm_flags,
 590          * so this tests vma->vm_flags & VM_SPECIAL, too.
 591          */
 592         if (vm_flags & VM_SPECIAL)
 593                 return NULL;
 594
 595         if (prev)
 596                 next = prev->vm_next;
 597         else
 598                 next = mm->mmap;
 599         area = next;
 600         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
 601                 next = next->vm_next;
 602
 603         /*
 604          * Can it merge with the predecessor?
 605          */
 606         if (prev && prev->vm_end == addr &&
 607                         mpol_equal(vma_policy(prev), policy) &&
 608                         can_vma_merge_after(prev, vm_flags,
 609                                                 anon_vma, file, pgoff)) {
 610                 /*
 611                  * OK, it can.  Can we now merge in the successor as well?
 612                  */
 613                 if (next && end == next->vm_start &&
 614                                 mpol_equal(policy, vma_policy(next)) &&
 615                                 can_vma_merge_before(next, vm_flags,
 616                                         anon_vma, file, pgoff+pglen) &&
 617                                 is_mergeable_anon_vma(prev->anon_vma,
 618                                                       next->anon_vma)) {
 619                                                         /* cases 1, 6 */
 620                         vma_adjust(prev, prev->vm_start,
 621                                 next->vm_end, prev->vm_pgoff, NULL);
 622                 } else                                  /* cases 2, 5, 7 */
 623                         vma_adjust(prev, prev->vm_start,
 624                                 end, prev->vm_pgoff, NULL);
 625                 if (prev->vm_flags & VM_EXEC)
 626                         arch_add_exec_range(mm, prev->vm_end);
 627                 return prev;
 628         }
 629
 630         /*
 631          * Can this new request be merged in front of next?
 632          */
 633         if (next && end == next->vm_start &&
 634                         mpol_equal(policy, vma_policy(next)) &&
 635                         can_vma_merge_before(next, vm_flags,
 636                                         anon_vma, file, pgoff+pglen)) {
 637                 if (prev && addr < prev->vm_end)        /* case 4 */
 638                         vma_adjust(prev, prev->vm_start,
 639                                 addr, prev->vm_pgoff, NULL);
 640                 else                                    /* cases 3, 8 */
 641                         vma_adjust(area, addr, next->vm_end,
 642                                 next->vm_pgoff - pglen, NULL);
 643                 return area;
 644         }
 645
 646         return NULL;
 647 }
 648
 649 /*
 650  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
 651  * neighbouring vmas for a suitable anon_vma, before it goes off
 652  * to allocate a new anon_vma.  It checks because a repetitive
 653  * sequence of mprotects and faults may otherwise lead to distinct
 654  * anon_vmas being allocated, preventing vma merge in subsequent
 655  * mprotect.
 656  */
 657 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
 658 {
 659         struct vm_area_struct *near;
 660         unsigned long vm_flags;
 661
 662         near = vma->vm_next;
 663         if (!near)
 664                 goto try_prev;
 665
 666         /*
 667          * Since only mprotect tries to remerge vmas, match flags
 668          * which might be mprotected into each other later on.
 669          * Neither mlock nor madvise tries to remerge at present,
 670          * so leave their flags as obstructing a merge.
 671          */
 672         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
 673         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
 674
 675         if (near->anon_vma && vma->vm_end == near->vm_start &&
 676                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
 677                         can_vma_merge_before(near, vm_flags,
 678                                 NULL, vma->vm_file, vma->vm_pgoff +
 679                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
 680                 return near->anon_vma;
 681 try_prev:
 682         /*
 683          * It is potentially slow to have to call find_vma_prev here.
 684          * But it's only on the first write fault on the vma, not
 685          * every time, and we could devise a way to avoid it later
 686          * (e.g. stash info in next's anon_vma_node when assigning
 687          * an anon_vma, or when trying vma_merge).  Another time.
 688          */
 689         if (find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma)
 690                 BUG();
 691         if (!near)
 692                 goto none;
 693
 694         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
 695         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
 696
 697         if (near->anon_vma && near->vm_end == vma->vm_start &&
 698                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
 699                         can_vma_merge_after(near, vm_flags,
 700                                 NULL, vma->vm_file, vma->vm_pgoff))
 701                 return near->anon_vma;
 702 none:
 703         /*
 704          * There's no absolute need to look only at touching neighbours:
 705          * we could search further afield for "compatible" anon_vmas.
 706          * But it would probably just be a waste of time searching,
 707          * or lead to too many vmas hanging off the same anon_vma.
 708          * We're trying to allow mprotect remerging later on,
 709          * not trying to minimize memory used for anon_vmas.
 710          */
 711         return NULL;
 712 }
 713
 714 /*
 715  * The caller must hold down_write(current->mm->mmap_sem).
 716  */
 717
 718 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
 719                         unsigned long len, unsigned long prot,
 720                         unsigned long flags, unsigned long pgoff)
 721 {
 722         struct mm_struct * mm = current->mm;
 723         struct vm_area_struct * vma, * prev;
 724         struct inode *inode;
 725         unsigned int vm_flags;
 726         int correct_wcount = 0;
 727         int error;
 728         struct rb_node ** rb_link, * rb_parent;
 729         int accountable = 1;
 730         unsigned long charged = 0;
 731
 732         if (file) {
 733                 if (is_file_hugepages(file))
 734                         accountable = 0;
 735
 736                 if (!file->f_op || !file->f_op->mmap)
 737                         return -ENODEV;
 738
 739                 if ((prot & PROT_EXEC) &&
 740                     (file->f_vfsmnt->mnt_flags & MNT_NOEXEC))
 741                         return -EPERM;
 742         }
 743
 744         if (!len)
 745                 return addr;
 746
 747         /* Careful about overflows.. */
 748         len = PAGE_ALIGN(len);
 749         if (!len || len > TASK_SIZE)
 750                 return -EINVAL;
 751
 752         /* offset overflow? */
 753         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
 754                 return -EINVAL;
 755
 756         /* Too many mappings? */
 757         if (mm->map_count > sysctl_max_map_count)
 758                 return -ENOMEM;
 759
 760         /* Obtain the address to map to. we verify (or select) it and ensure
 761          * that it represents a valid section of the address space.
 762          */
 763         addr = get_unmapped_area(file, addr, len, pgoff, flags, prot & PROT_EXEC);
 764         if (addr & ~PAGE_MASK)
 765                 return addr;
 766
 767         /* Do simple checking here so the lower-level routines won't have
 768          * to. we assume access permissions have been handled by the open
 769          * of the memory object, so we don't do any here.
 770          */
 771         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
 772                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
 773
 774         if (flags & MAP_LOCKED) {
 775                 if (!capable(CAP_IPC_LOCK))
 776                         return -EPERM;
 777                 vm_flags |= VM_LOCKED;
 778         }
 779         /* mlock MCL_FUTURE? */
 780         if (vm_flags & VM_LOCKED) {
 781                 unsigned long locked = mm->locked_vm << PAGE_SHIFT;
 782                 locked += len;
 783                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
 784                         return -EAGAIN;
 785         }
 786
 787         inode = file ? file->f_dentry->d_inode : NULL;
 788
 789         if (file) {
 790                 switch (flags & MAP_TYPE) {
 791                 case MAP_SHARED:
 792                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
 793                                 return -EACCES;
 794
 795                         /*
 796                          * Make sure we don't allow writing to an append-only
 797                          * file..
 798                          */
 799                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
 800                                 return -EACCES;
 801
 802                         /*
 803                          * Make sure there are no mandatory locks on the file.
 804                          */
 805                         if (locks_verify_locked(inode))
 806                                 return -EAGAIN;
 807
 808                         vm_flags |= VM_SHARED | VM_MAYSHARE;
 809                         if (!(file->f_mode & FMODE_WRITE))
 810                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
 811
 812                         /* fall through */
 813                 case MAP_PRIVATE:
 814                         if (!(file->f_mode & FMODE_READ))
 815                                 return -EACCES;
 816                         break;
 817
 818                 default:
 819                         return -EINVAL;
 820                 }
 821         } else {
 822                 switch (flags & MAP_TYPE) {
 823                 case MAP_SHARED:
 824                         vm_flags |= VM_SHARED | VM_MAYSHARE;
 825                         break;
 826                 case MAP_PRIVATE:
 827                         /*
 828                          * Set pgoff according to addr for anon_vma.
 829                          */
 830                         pgoff = addr >> PAGE_SHIFT;
 831                         break;
 832                 default:
 833                         return -EINVAL;
 834                 }
 835         }
 836
 837         error = security_file_mmap(file, prot, flags);
 838         if (error)
 839                 return error;
 840
 841         /* Clear old maps */
 842         error = -ENOMEM;
 843 munmap_back:
 844         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
 845         if (vma && vma->vm_start < addr + len) {
 846                 if (do_munmap(mm, addr, len))
 847                         return -ENOMEM;
 848                 goto munmap_back;
 849         }
 850
 851         /* Check against address space limit. */
 852         if ((mm->total_vm << PAGE_SHIFT) + len
 853             > current->rlim[RLIMIT_AS].rlim_cur)
 854                 return -ENOMEM;
 855
 856         if (accountable && (!(flags & MAP_NORESERVE) ||
 857                         sysctl_overcommit_memory > 1)) {
 858                 if (vm_flags & VM_SHARED) {
 859                         /* Check memory availability in shmem_file_setup? */
 860                         vm_flags |= VM_ACCOUNT;
 861                 } else if (vm_flags & VM_WRITE) {
 862                         /*
 863                          * Private writable mapping: check memory availability
 864                          */
 865                         charged = len >> PAGE_SHIFT;
 866                         if (security_vm_enough_memory(charged))
 867                                 return -ENOMEM;
 868                         vm_flags |= VM_ACCOUNT;
 869                 }
 870         }
 871
 872         /*
 873          * Can we just expand an old private anonymous mapping?
 874          * The VM_SHARED test is necessary because shmem_zero_setup
 875          * will create the file object for a shared anonymous map below.
 876          */
 877         if (!file && !(vm_flags & VM_SHARED) &&
 878             vma_merge(mm, prev, addr, addr + len, vm_flags,
 879                                         NULL, NULL, pgoff, NULL))
 880                 goto out;
 881
 882         /*
 883          * Determine the object being mapped and call the appropriate
 884          * specific mapper. the address has already been validated, but
 885          * not unmapped, but the maps are removed from the list.
 886          */
 887         vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
 888         if (!vma) {
 889                 error = -ENOMEM;
 890                 goto unacct_error;
 891         }
 892         memset(vma, 0, sizeof(*vma));
 893
 894         vma->vm_mm = mm;
 895         vma->vm_start = addr;
 896         vma->vm_end = addr + len;
 897         vma->vm_flags = vm_flags;
 898         vma->vm_page_prot = protection_map[vm_flags & 0x0f];
 899         vma->vm_pgoff = pgoff;
 900
 901         if (file) {
 902                 error = -EINVAL;
 903                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
 904                         goto free_vma;
 905                 if (vm_flags & VM_DENYWRITE) {
 906                         error = deny_write_access(file);
 907                         if (error)
 908                                 goto free_vma;
 909                         correct_wcount = 1;
 910                 }
 911                 vma->vm_file = file;
 912                 get_file(file);
 913                 error = file->f_op->mmap(file, vma);
 914                 if (error)
 915                         goto unmap_and_free_vma;
 916         } else if (vm_flags & VM_SHARED) {
 917                 error = shmem_zero_setup(vma);
 918                 if (error)
 919                         goto free_vma;
 920         }
 921
 922         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
 923          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
 924          * that memory reservation must be checked; but that reservation
 925          * belongs to shared memory object, not to vma: so now clear it.
 926          */
 927         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
 928                 vma->vm_flags &= ~VM_ACCOUNT;
 929
 930         /* Can addr have changed??
 931          *
 932          * Answer: Yes, several device drivers can do it in their
 933          *         f_op->mmap method. -DaveM
 934          */
 935         addr = vma->vm_start;
 936
 937         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
 938                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
 939                 vma_link(mm, vma, prev, rb_link, rb_parent);
 940                 if (correct_wcount)
 941                         atomic_inc(&inode->i_writecount);
 942         } else {
 943                 if (file) {
 944                         if (correct_wcount)
 945                                 atomic_inc(&inode->i_writecount);
 946                         fput(file);
 947                 }
 948                 mpol_free(vma_policy(vma));
 949                 kmem_cache_free(vm_area_cachep, vma);
 950         }
 951 out:
 952         mm->total_vm += len >> PAGE_SHIFT;
 953         if (vm_flags & VM_LOCKED) {
 954                 mm->locked_vm += len >> PAGE_SHIFT;
 955                 make_pages_present(addr, addr + len);
 956         }
 957         if (flags & MAP_POPULATE) {
 958                 up_write(&mm->mmap_sem);
 959                 sys_remap_file_pages(addr, len, 0,
 960                                         pgoff, flags & MAP_NONBLOCK);
 961                 down_write(&mm->mmap_sem);
 962         }
 963         return addr;
 964
 965 unmap_and_free_vma:
 966         if (correct_wcount)
 967                 atomic_inc(&inode->i_writecount);
 968         vma->vm_file = NULL;
 969         fput(file);
 970
 971         /* Undo any partial mapping done by a device driver. */
 972         zap_page_range(vma, vma->vm_start, vma->vm_end - vma->vm_start, NULL);
 973 free_vma:
 974         kmem_cache_free(vm_area_cachep, vma);
 975 unacct_error:
 976         if (charged)
 977                 vm_unacct_memory(charged);
 978         return error;
 979 }
 980
 981 EXPORT_SYMBOL(do_mmap_pgoff);
 982
 983 /* Get an address range which is currently unmapped.
 984  * For shmat() with addr=0.
 985  *
 986  * Ugly calling convention alert:
 987  * Return value with the low bits set means error value,
 988  * ie
 989  *      if (ret & ~PAGE_MASK)
 990  *              error = ret;
 991  *
 992  * This function "knows" that -ENOMEM has the bits set.
 993  */
 994 #ifndef HAVE_ARCH_UNMAPPED_AREA
 995 static inline unsigned long
 996 arch_get_unmapped_area(struct file *filp, unsigned long addr,
 997                 unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
 998 {
 999         struct mm_struct *mm = current->mm;
1000         struct vm_area_struct *vma;
1001         unsigned long start_addr;
1002
1003         if (len > TASK_SIZE)
1004                 return -ENOMEM;
1005
1006         if (addr) {
1007                 addr = PAGE_ALIGN(addr);
1008                 vma = find_vma(mm, addr);
1009                 if (TASK_SIZE - len >= addr &&
1010                     (!vma || addr + len <= vma->vm_start))
1011                         return addr;
1012         }
1013         start_addr = addr = mm->free_area_cache;
1014
1015 full_search:
1016         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1017                 /* At this point:  (!vma || addr < vma->vm_end). */
1018                 if (TASK_SIZE - len < addr) {
1019                         /*
1020                          * Start a new search - just in case we missed
1021                          * some holes.
1022                          */
1023                         if (start_addr != TASK_UNMAPPED_BASE) {
1024                                 start_addr = addr = TASK_UNMAPPED_BASE;
1025                                 goto full_search;
1026                         }
1027                         return -ENOMEM;
1028                 }
1029                 if (!vma || addr + len <= vma->vm_start) {
1030                         /*
1031                          * Remember the place where we stopped the search:
1032                          */
1033                         mm->free_area_cache = addr + len;
1034                         return addr;
1035                 }
1036                 addr = vma->vm_end;
1037         }
1038 }
1039 #else
1040 extern unsigned long
1041 arch_get_unmapped_area(struct file *, unsigned long, unsigned long,
1042                         unsigned long, unsigned long, unsigned long);
1043 #endif
1044
1045 unsigned long
1046 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1047                 unsigned long pgoff, unsigned long flags, unsigned long exec)
1048 {
1049         if (flags & MAP_FIXED) {
1050                 unsigned long ret;
1051
1052                 if (addr > TASK_SIZE - len)
1053                         return -ENOMEM;
1054                 if (addr & ~PAGE_MASK)
1055                         return -EINVAL;
1056                 if (file && is_file_hugepages(file))  {
1057                         /*
1058                          * Check if the given range is hugepage aligned, and
1059                          * can be made suitable for hugepages.
1060                          */
1061                         ret = prepare_hugepage_range(addr, len);
1062                 } else {
1063                         /*
1064                          * Ensure that a normal request is not falling in a
1065                          * reserved hugepage range.  For some archs like IA-64,
1066                          * there is a separate region for hugepages.
1067                          */
1068                         ret = is_hugepage_only_range(addr, len);
1069                 }
1070                 if (ret)
1071                         return -EINVAL;
1072                 return addr;
1073         }
1074
1075         if (file && file->f_op && file->f_op->get_unmapped_area)
1076                 return file->f_op->get_unmapped_area(file, addr, len,
1077                                                 pgoff, flags);
1078
1079         return arch_get_unmapped_area(file, addr, len, pgoff, flags, exec);
1080 }
1081
1082 EXPORT_SYMBOL(get_unmapped_area);
1083
1084 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1085 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1086 {
1087         struct vm_area_struct *vma = NULL;
1088
1089         if (mm) {
1090                 /* Check the cache first. */
1091                 /* (Cache hit rate is typically around 35%.) */
1092                 vma = mm->mmap_cache;
1093                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1094                         struct rb_node * rb_node;
1095
1096                         rb_node = mm->mm_rb.rb_node;
1097                         vma = NULL;
1098
1099                         while (rb_node) {
1100                                 struct vm_area_struct * vma_tmp;
1101
1102                                 vma_tmp = rb_entry(rb_node,
1103                                                 struct vm_area_struct, vm_rb);
1104
1105                                 if (vma_tmp->vm_end > addr) {
1106                                         vma = vma_tmp;
1107                                         if (vma_tmp->vm_start <= addr)
1108                                                 break;
1109                                         rb_node = rb_node->rb_left;
1110                                 } else
1111                                         rb_node = rb_node->rb_right;
1112                         }
1113                         if (vma)
1114                                 mm->mmap_cache = vma;
1115                 }
1116         }
1117         return vma;
1118 }
1119
1120 EXPORT_SYMBOL(find_vma);
1121
1122 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1123 struct vm_area_struct *
1124 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1125                         struct vm_area_struct **pprev)
1126 {
1127         struct vm_area_struct *vma = NULL, *prev = NULL;
1128         struct rb_node * rb_node;
1129         if (!mm)
1130                 goto out;
1131
1132         /* Guard against addr being lower than the first VMA */
1133         vma = mm->mmap;
1134
1135         /* Go through the RB tree quickly. */
1136         rb_node = mm->mm_rb.rb_node;
1137
1138         while (rb_node) {
1139                 struct vm_area_struct *vma_tmp;
1140                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1141
1142                 if (addr < vma_tmp->vm_end) {
1143                         rb_node = rb_node->rb_left;
1144                 } else {
1145                         prev = vma_tmp;
1146                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1147                                 break;
1148                         rb_node = rb_node->rb_right;
1149                 }
1150         }
1151
1152 out:
1153         *pprev = prev;
1154         return prev ? prev->vm_next : vma;
1155 }
1156
1157
1158 static int over_stack_limit(unsigned long sz)
1159 {
1160         if (sz < EXEC_STACK_BIAS)
1161                 return 0;
1162         return (sz - EXEC_STACK_BIAS) > current->rlim[RLIMIT_STACK].rlim_cur;
1163 }
1164
1165 #ifdef CONFIG_STACK_GROWSUP
1166 /*
1167  * vma is the first one with address > vma->vm_end.  Have to extend vma.
1168  */
1169 int expand_stack(struct vm_area_struct * vma, unsigned long address)
1170 {
1171         unsigned long grow;
1172
1173         if (!(vma->vm_flags & VM_GROWSUP))
1174                 return -EFAULT;
1175
1176         /*
1177          * We must make sure the anon_vma is allocated
1178          * so that the anon_vma locking is not a noop.
1179          */
1180         if (unlikely(anon_vma_prepare(vma)))
1181                 return -ENOMEM;
1182         anon_vma_lock(vma);
1183
1184         /*
1185          * vma->vm_start/vm_end cannot change under us because the caller
1186          * is required to hold the mmap_sem in read mode.  We need the
1187          * anon_vma lock to serialize against concurrent expand_stacks.
1188          */
1189         address += 4 + PAGE_SIZE - 1;
1190         address &= PAGE_MASK;
1191         grow = (address - vma->vm_end) >> PAGE_SHIFT;
1192
1193         /* Overcommit.. */
1194         if (security_vm_enough_memory(grow)) {
1195                 anon_vma_unlock(vma);
1196                 return -ENOMEM;
1197         }
1198
1199         if (over_stack_limit(address - vma->vm_start) ||
1200                         ((vma->vm_mm->total_vm + grow) << PAGE_SHIFT) >
1201                         current->rlim[RLIMIT_AS].rlim_cur) {
1202                 anon_vma_unlock(vma);
1203                 vm_unacct_memory(grow);
1204                 return -ENOMEM;
1205         }
1206         vma->vm_end = address;
1207         vma->vm_mm->total_vm += grow;
1208         if (vma->vm_flags & VM_LOCKED)
1209                 vma->vm_mm->locked_vm += grow;
1210         anon_vma_unlock(vma);
1211         return 0;
1212 }
1213
1214 struct vm_area_struct *
1215 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1216 {
1217         struct vm_area_struct *vma, *prev;
1218
1219         addr &= PAGE_MASK;
1220         vma = find_vma_prev(mm, addr, &prev);
1221         if (vma && (vma->vm_start <= addr))
1222                 return vma;
1223         if (!prev || expand_stack(prev, addr))
1224                 return NULL;
1225         if (prev->vm_flags & VM_LOCKED) {
1226                 make_pages_present(addr, prev->vm_end);
1227         }
1228         return prev;
1229 }
1230 #else
1231 /*
1232  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1233  */
1234 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1235 {
1236         unsigned long grow;
1237
1238         /*
1239          * We must make sure the anon_vma is allocated
1240          * so that the anon_vma locking is not a noop.
1241          */
1242         if (unlikely(anon_vma_prepare(vma)))
1243                 return -ENOMEM;
1244         anon_vma_lock(vma);
1245
1246         /*
1247          * vma->vm_start/vm_end cannot change under us because the caller
1248          * is required to hold the mmap_sem in read mode.  We need the
1249          * anon_vma lock to serialize against concurrent expand_stacks.
1250          */
1251         address &= PAGE_MASK;
1252         grow = (vma->vm_start - address) >> PAGE_SHIFT;
1253
1254         /* Overcommit.. */
1255         if (security_vm_enough_memory(grow)) {
1256                 anon_vma_unlock(vma);
1257                 return -ENOMEM;
1258         }
1259
1260         if (over_stack_limit(vma->vm_end - address) ||
1261                         ((vma->vm_mm->total_vm + grow) << PAGE_SHIFT) >
1262                         current->rlim[RLIMIT_AS].rlim_cur) {
1263                 anon_vma_unlock(vma);
1264                 vm_unacct_memory(grow);
1265                 return -ENOMEM;
1266         }
1267         vma->vm_start = address;
1268         vma->vm_pgoff -= grow;
1269         vma->vm_mm->total_vm += grow;
1270         if (vma->vm_flags & VM_LOCKED)
1271                 vma->vm_mm->locked_vm += grow;
1272         anon_vma_unlock(vma);
1273         return 0;
1274 }
1275
1276 struct vm_area_struct *
1277 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1278 {
1279         struct vm_area_struct * vma;
1280         unsigned long start;
1281
1282         addr &= PAGE_MASK;
1283         vma = find_vma(mm,addr);
1284         if (!vma)
1285                 return NULL;
1286         if (vma->vm_start <= addr)
1287                 return vma;
1288         if (!(vma->vm_flags & VM_GROWSDOWN))
1289                 return NULL;
1290         start = vma->vm_start;
1291         if (expand_stack(vma, addr))
1292                 return NULL;
1293         if (vma->vm_flags & VM_LOCKED) {
1294                 make_pages_present(addr, start);
1295         }
1296         return vma;
1297 }
1298 #endif
1299
1300 /*
1301  * Try to free as many page directory entries as we can,
1302  * without having to work very hard at actually scanning
1303  * the page tables themselves.
1304  *
1305  * Right now we try to free page tables if we have a nice
1306  * PGDIR-aligned area that got free'd up. We could be more
1307  * granular if we want to, but this is fast and simple,
1308  * and covers the bad cases.
1309  *
1310  * "prev", if it exists, points to a vma before the one
1311  * we just free'd - but there's no telling how much before.
1312  */
1313 static void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev,
1314         unsigned long start, unsigned long end)
1315 {
1316         unsigned long first = start & PGDIR_MASK;
1317         unsigned long last = end + PGDIR_SIZE - 1;
1318         unsigned long start_index, end_index;
1319         struct mm_struct *mm = tlb->mm;
1320
1321         if (!prev) {
1322                 prev = mm->mmap;
1323                 if (!prev)
1324                         goto no_mmaps;
1325                 if (prev->vm_end > start) {
1326                         if (last > prev->vm_start)
1327                                 last = prev->vm_start;
1328                         goto no_mmaps;
1329                 }
1330         }
1331         for (;;) {
1332                 struct vm_area_struct *next = prev->vm_next;
1333
1334                 if (next) {
1335                         if (next->vm_start < start) {
1336                                 prev = next;
1337                                 continue;
1338                         }
1339                         if (last > next->vm_start)
1340                                 last = next->vm_start;
1341                 }
1342                 if (prev->vm_end > first)
1343                         first = prev->vm_end + PGDIR_SIZE - 1;
1344                 break;
1345         }
1346 no_mmaps:
1347         if (last < first)       /* for arches with discontiguous pgd indices */
1348                 return;
1349         /*
1350          * If the PGD bits are not consecutive in the virtual address, the
1351          * old method of shifting the VA >> by PGDIR_SHIFT doesn't work.
1352          */
1353         start_index = pgd_index(first);
1354         if (start_index < FIRST_USER_PGD_NR)
1355                 start_index = FIRST_USER_PGD_NR;
1356         end_index = pgd_index(last);
1357         if (end_index > start_index) {
1358                 clear_page_tables(tlb, start_index, end_index - start_index);
1359                 flush_tlb_pgtables(mm, first & PGDIR_MASK, last & PGDIR_MASK);
1360         }
1361 }
1362
1363 /* Normal function to fix up a mapping
1364  * This function is the default for when an area has no specific
1365  * function.  This may be used as part of a more specific routine.
1366  *
1367  * By the time this function is called, the area struct has been
1368  * removed from the process mapping list.
1369  */
1370 static void unmap_vma(struct mm_struct *mm, struct vm_area_struct *area)
1371 {
1372         size_t len = area->vm_end - area->vm_start;
1373
1374         area->vm_mm->total_vm -= len >> PAGE_SHIFT;
1375         if (area->vm_flags & VM_LOCKED)
1376                 area->vm_mm->locked_vm -= len >> PAGE_SHIFT;
1377         /*
1378          * Is this a new hole at the lowest possible address?
1379          */
1380         if (area->vm_start >= TASK_UNMAPPED_BASE &&
1381                                 area->vm_start < area->vm_mm->free_area_cache)
1382               area->vm_mm->free_area_cache = area->vm_start;
1383         /*
1384          * Is this a new hole at the highest possible address?
1385          */
1386         if (area->vm_start > area->vm_mm->non_executable_cache)
1387                 area->vm_mm->non_executable_cache = area->vm_start;
1388
1389         remove_vm_struct(area);
1390         if (unlikely(area->vm_flags & VM_EXEC))
1391                 arch_remove_exec_range(mm, area->vm_end);
1392 }
1393
1394 /*
1395  * Update the VMA and inode share lists.
1396  *
1397  * Ok - we have the memory areas we should free on the 'free' list,
1398  * so release them, and do the vma updates.
1399  */
1400 static void unmap_vma_list(struct mm_struct *mm,
1401         struct vm_area_struct *mpnt)
1402 {
1403         do {
1404                 struct vm_area_struct *next = mpnt->vm_next;
1405                 unmap_vma(mm, mpnt);
1406                 mpnt = next;
1407         } while (mpnt != NULL);
1408         validate_mm(mm);
1409 }
1410
1411 /*
1412  * Get rid of page table information in the indicated region.
1413  *
1414  * Called with the page table lock held.
1415  */
1416 static void unmap_region(struct mm_struct *mm,
1417         struct vm_area_struct *vma,
1418         struct vm_area_struct *prev,
1419         unsigned long start,
1420         unsigned long end)
1421 {
1422         struct mmu_gather *tlb;
1423         unsigned long nr_accounted = 0;
1424
1425         lru_add_drain();
1426         tlb = tlb_gather_mmu(mm, 0);
1427         unmap_vmas(&tlb, mm, vma, start, end, &nr_accounted, NULL);
1428         vm_unacct_memory(nr_accounted);
1429
1430         if (is_hugepage_only_range(start, end - start))
1431                 hugetlb_free_pgtables(tlb, prev, start, end);
1432         else
1433                 free_pgtables(tlb, prev, start, end);
1434         tlb_finish_mmu(tlb, start, end);
1435 }
1436
1437 /*
1438  * Create a list of vma's touched by the unmap, removing them from the mm's
1439  * vma list as we go..
1440  */
1441 static void
1442 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1443         struct vm_area_struct *prev, unsigned long end)
1444 {
1445         struct vm_area_struct **insertion_point;
1446         struct vm_area_struct *tail_vma = NULL;
1447
1448         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1449         do {
1450                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1451                 mm->map_count--;
1452                 tail_vma = vma;
1453                 vma = vma->vm_next;
1454         } while (vma && vma->vm_start < end);
1455         *insertion_point = vma;
1456         tail_vma->vm_next = NULL;
1457         mm->mmap_cache = NULL;          /* Kill the cache. */
1458 }
1459
1460 /*
1461  * Split a vma into two pieces at address 'addr', a new vma is allocated
1462  * either for the first part or the the tail.
1463  */
1464 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1465               unsigned long addr, int new_below)
1466 {
1467         struct mempolicy *pol;
1468         struct vm_area_struct *new;
1469
1470         if (mm->map_count >= sysctl_max_map_count)
1471                 return -ENOMEM;
1472
1473         new = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1474         if (!new)
1475                 return -ENOMEM;
1476
1477         /* most fields are the same, copy all, and then fixup */
1478         *new = *vma;
1479
1480         if (new_below) {
1481                 if (vma->vm_flags & VM_EXEC)
1482                         arch_remove_exec_range(mm, new->vm_end);
1483                 new->vm_end = addr;
1484         }
1485         else {
1486                 new->vm_start = addr;
1487                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1488         }
1489
1490         pol = mpol_copy(vma_policy(vma));
1491         if (IS_ERR(pol)) {
1492                 kmem_cache_free(vm_area_cachep, new);
1493                 return PTR_ERR(pol);
1494         }
1495         vma_set_policy(new, pol);
1496
1497         if (new->vm_file)
1498                 get_file(new->vm_file);
1499
1500         if (new->vm_ops && new->vm_ops->open)
1501                 new->vm_ops->open(new);
1502
1503         if (new_below)
1504                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1505                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1506         else
1507                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1508
1509         return 0;
1510 }
1511
1512 /* Munmap is split into 2 main parts -- this part which finds
1513  * what needs doing, and the areas themselves, which do the
1514  * work.  This now handles partial unmappings.
1515  * Jeremy Fitzhardinge <jeremy@goop.org>
1516  */
1517 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1518 {
1519         unsigned long end;
1520         struct vm_area_struct *mpnt, *prev, *last;
1521
1522         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1523                 return -EINVAL;
1524
1525         if ((len = PAGE_ALIGN(len)) == 0)
1526                 return -EINVAL;
1527
1528         /* Find the first overlapping VMA */
1529         mpnt = find_vma_prev(mm, start, &prev);
1530         if (!mpnt)
1531                 return 0;
1532         /* we have  start < mpnt->vm_end  */
1533
1534         if (is_vm_hugetlb_page(mpnt)) {
1535                 int ret = is_aligned_hugepage_range(start, len);
1536
1537                 if (ret)
1538                         return ret;
1539         }
1540
1541         /* if it doesn't overlap, we have nothing.. */
1542         end = start + len;
1543         if (mpnt->vm_start >= end)
1544                 return 0;
1545
1546         /* Something will probably happen, so notify. */
1547         if (mpnt->vm_file && (mpnt->vm_flags & VM_EXEC))
1548                 profile_exec_unmap(mm);
1549
1550         /*
1551          * If we need to split any vma, do it now to save pain later.
1552          *
1553          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1554          * unmapped vm_area_struct will remain in use: so lower split_vma
1555          * places tmp vma above, and higher split_vma places tmp vma below.
1556          */
1557         if (start > mpnt->vm_start) {
1558                 if (split_vma(mm, mpnt, start, 0))
1559                         return -ENOMEM;
1560                 prev = mpnt;
1561         }
1562
1563         /* Does it split the last one? */
1564         last = find_vma(mm, end);
1565         if (last && end > last->vm_start) {
1566                 if (split_vma(mm, last, end, 1))
1567                         return -ENOMEM;
1568         }
1569         mpnt = prev? prev->vm_next: mm->mmap;
1570
1571         /*
1572          * Remove the vma's, and unmap the actual pages
1573          */
1574         detach_vmas_to_be_unmapped(mm, mpnt, prev, end);
1575         spin_lock(&mm->page_table_lock);
1576         unmap_region(mm, mpnt, prev, start, end);
1577         spin_unlock(&mm->page_table_lock);
1578
1579         /* Fix up all other VM information */
1580         unmap_vma_list(mm, mpnt);
1581
1582         return 0;
1583 }
1584
1585 EXPORT_SYMBOL(do_munmap);
1586
1587 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1588 {
1589         int ret;
1590         struct mm_struct *mm = current->mm;
1591
1592         down_write(&mm->mmap_sem);
1593         ret = do_munmap(mm, addr, len);
1594         up_write(&mm->mmap_sem);
1595         return ret;
1596 }
1597
1598 /*
1599  *  this is really a simplified "do_mmap".  it only handles
1600  *  anonymous maps.  eventually we may be able to do some
1601  *  brk-specific accounting here.
1602  */
1603 unsigned long do_brk(unsigned long addr, unsigned long len)
1604 {
1605         struct mm_struct * mm = current->mm;
1606         struct vm_area_struct * vma, * prev;
1607         unsigned long flags;
1608         struct rb_node ** rb_link, * rb_parent;
1609         pgoff_t pgoff = addr >> PAGE_SHIFT;
1610
1611         len = PAGE_ALIGN(len);
1612         if (!len)
1613                 return addr;
1614
1615         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1616                 return -EINVAL;
1617
1618         /*
1619          * mlock MCL_FUTURE?
1620          */
1621         if (mm->def_flags & VM_LOCKED) {
1622                 unsigned long locked = mm->locked_vm << PAGE_SHIFT;
1623                 locked += len;
1624                 if (locked > current->rlim[RLIMIT_MEMLOCK].rlim_cur)
1625                         return -EAGAIN;
1626         }
1627
1628         /*
1629          * Clear old maps.  this also does some error checking for us
1630          */
1631  munmap_back:
1632         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1633         if (vma && vma->vm_start < addr + len) {
1634                 if (do_munmap(mm, addr, len))
1635                         return -ENOMEM;
1636                 goto munmap_back;
1637         }
1638
1639         /* Check against address space limits *after* clearing old maps... */
1640         if ((mm->total_vm << PAGE_SHIFT) + len
1641             > current->rlim[RLIMIT_AS].rlim_cur)
1642                 return -ENOMEM;
1643
1644         if (mm->map_count > sysctl_max_map_count)
1645                 return -ENOMEM;
1646
1647         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1648                 return -ENOMEM;
1649
1650         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1651
1652         /* Can we just expand an old private anonymous mapping? */
1653         if (vma_merge(mm, prev, addr, addr + len, flags,
1654                                         NULL, NULL, pgoff, NULL))
1655                 goto out;
1656
1657         /*
1658          * create a vma struct for an anonymous mapping
1659          */
1660         vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1661         if (!vma) {
1662                 vm_unacct_memory(len >> PAGE_SHIFT);
1663                 return -ENOMEM;
1664         }
1665         memset(vma, 0, sizeof(*vma));
1666
1667         vma->vm_mm = mm;
1668         vma->vm_start = addr;
1669         vma->vm_end = addr + len;
1670         vma->vm_pgoff = pgoff;
1671         vma->vm_flags = flags;
1672         vma->vm_page_prot = protection_map[flags & 0x0f];
1673         vma_link(mm, vma, prev, rb_link, rb_parent);
1674 out:
1675         mm->total_vm += len >> PAGE_SHIFT;
1676         if (flags & VM_LOCKED) {
1677                 mm->locked_vm += len >> PAGE_SHIFT;
1678                 make_pages_present(addr, addr + len);
1679         }
1680         return addr;
1681 }
1682
1683 EXPORT_SYMBOL(do_brk);
1684
1685 /* Release all mmaps. */
1686 void exit_mmap(struct mm_struct *mm)
1687 {
1688         struct mmu_gather *tlb;
1689         struct vm_area_struct *vma;
1690         unsigned long nr_accounted = 0;
1691
1692         profile_exit_mmap(mm);
1693
1694         lru_add_drain();
1695
1696         spin_lock(&mm->page_table_lock);
1697
1698         tlb = tlb_gather_mmu(mm, 1);
1699         flush_cache_mm(mm);
1700         /* Use ~0UL here to ensure all VMAs in the mm are unmapped */
1701         mm->map_count -= unmap_vmas(&tlb, mm, mm->mmap, 0,
1702                                         ~0UL, &nr_accounted, NULL);
1703         vm_unacct_memory(nr_accounted);
1704         BUG_ON(mm->map_count);  /* This is just debugging */
1705         clear_page_tables(tlb, FIRST_USER_PGD_NR, USER_PTRS_PER_PGD);
1706         tlb_finish_mmu(tlb, 0, MM_VM_SIZE(mm));
1707
1708         vma = mm->mmap;
1709         mm->mmap = mm->mmap_cache = NULL;
1710         mm->mm_rb = RB_ROOT;
1711         mm->rss = 0;
1712         mm->total_vm = 0;
1713         mm->locked_vm = 0;
1714         arch_flush_exec_range(mm);
1715
1716         spin_unlock(&mm->page_table_lock);
1717
1718         /*
1719          * Walk the list again, actually closing and freeing it
1720          * without holding any MM locks.
1721          */
1722         while (vma) {
1723                 struct vm_area_struct *next = vma->vm_next;
1724                 remove_vm_struct(vma);
1725                 vma = next;
1726         }
1727 }
1728
1729 /* Insert vm structure into process list sorted by address
1730  * and into the inode's i_mmap tree.  If vm_file is non-NULL
1731  * then i_mmap_lock is taken here.
1732  */
1733 void insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
1734 {
1735         struct vm_area_struct * __vma, * prev;
1736         struct rb_node ** rb_link, * rb_parent;
1737
1738         /*
1739          * The vm_pgoff of a purely anonymous vma should be irrelevant
1740          * until its first write fault, when page's anon_vma and index
1741          * are set.  But now set the vm_pgoff it will almost certainly
1742          * end up with (unless mremap moves it elsewhere before that
1743          * first wfault), so /proc/pid/maps tells a consistent story.
1744          *
1745          * By setting it to reflect the virtual start address of the
1746          * vma, merges and splits can happen in a seamless way, just
1747          * using the existing file pgoff checks and manipulations.
1748          * Similarly in do_mmap_pgoff and in do_brk.
1749          */
1750         if (!vma->vm_file) {
1751                 BUG_ON(vma->anon_vma);
1752                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
1753         }
1754         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
1755         if (__vma && __vma->vm_start < vma->vm_end)
1756                 BUG();
1757         vma_link(mm, vma, prev, rb_link, rb_parent);
1758 }
1759
1760 /*
1761  * Copy the vma structure to a new location in the same mm,
1762  * prior to moving page table entries, to effect an mremap move.
1763  */
1764 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
1765         unsigned long addr, unsigned long len, pgoff_t pgoff)
1766 {
1767         struct vm_area_struct *vma = *vmap;
1768         unsigned long vma_start = vma->vm_start;
1769         struct mm_struct *mm = vma->vm_mm;
1770         struct vm_area_struct *new_vma, *prev;
1771         struct rb_node **rb_link, *rb_parent;
1772         struct mempolicy *pol;
1773
1774         /*
1775          * If anonymous vma has not yet been faulted, update new pgoff
1776          * to match new location, to increase its chance of merging.
1777          */
1778         if (!vma->vm_file && !vma->anon_vma)
1779                 pgoff = addr >> PAGE_SHIFT;
1780
1781         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1782         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
1783                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
1784         if (new_vma) {
1785                 /*
1786                  * Source vma may have been merged into new_vma
1787                  */
1788                 if (vma_start >= new_vma->vm_start &&
1789                     vma_start < new_vma->vm_end)
1790                         *vmap = new_vma;
1791         } else {
1792                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1793                 if (new_vma) {
1794                         *new_vma = *vma;
1795                         pol = mpol_copy(vma_policy(vma));
1796                         if (IS_ERR(pol)) {
1797                                 kmem_cache_free(vm_area_cachep, new_vma);
1798                                 return NULL;
1799                         }
1800                         vma_set_policy(new_vma, pol);
1801                         new_vma->vm_start = addr;
1802                         new_vma->vm_end = addr + len;
1803                         new_vma->vm_pgoff = pgoff;
1804                         if (new_vma->vm_file)
1805                                 get_file(new_vma->vm_file);
1806                         if (new_vma->vm_ops && new_vma->vm_ops->open)
1807                                 new_vma->vm_ops->open(new_vma);
1808                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
1809                 }
1810         }
1811         return new_vma;
1812 }