mm/mmap.c.orig

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