[PATCH] Fix 'make headers_check' on biarch architectures
[linux-2.6] / 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/mm.h>
11 #include <linux/shm.h>
12 #include <linux/mman.h>
13 #include <linux/pagemap.h>
14 #include <linux/swap.h>
15 #include <linux/syscalls.h>
16 #include <linux/capability.h>
17 #include <linux/init.h>
18 #include <linux/file.h>
19 #include <linux/fs.h>
20 #include <linux/personality.h>
21 #include <linux/security.h>
22 #include <linux/hugetlb.h>
23 #include <linux/profile.h>
24 #include <linux/module.h>
25 #include <linux/mount.h>
26 #include <linux/mempolicy.h>
27 #include <linux/rmap.h>
28
29 #include <asm/uaccess.h>
30 #include <asm/cacheflush.h>
31 #include <asm/tlb.h>
32
33 #ifndef arch_mmap_check
34 #define arch_mmap_check(addr, len, flags)       (0)
35 #endif
36
37 static void unmap_region(struct mm_struct *mm,
38                 struct vm_area_struct *vma, struct vm_area_struct *prev,
39                 unsigned long start, unsigned long end);
40
41 /*
42  * WARNING: the debugging will use recursive algorithms so never enable this
43  * unless you know what you are doing.
44  */
45 #undef DEBUG_MM_RB
46
47 /* description of effects of mapping type and prot in current implementation.
48  * this is due to the limited x86 page protection hardware.  The expected
49  * behavior is in parens:
50  *
51  * map_type     prot
52  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
53  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
54  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
55  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
56  *              
57  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
58  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
59  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
60  *
61  */
62 pgprot_t protection_map[16] = {
63         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
64         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
65 };
66
67 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
68 int sysctl_overcommit_ratio = 50;       /* default is 50% */
69 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
70 atomic_t vm_committed_space = ATOMIC_INIT(0);
71
72 /*
73  * Check that a process has enough memory to allocate a new virtual
74  * mapping. 0 means there is enough memory for the allocation to
75  * succeed and -ENOMEM implies there is not.
76  *
77  * We currently support three overcommit policies, which are set via the
78  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
79  *
80  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
81  * Additional code 2002 Jul 20 by Robert Love.
82  *
83  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
84  *
85  * Note this is a helper function intended to be used by LSMs which
86  * wish to use this logic.
87  */
88 int __vm_enough_memory(long pages, int cap_sys_admin)
89 {
90         unsigned long free, allowed;
91
92         vm_acct_memory(pages);
93
94         /*
95          * Sometimes we want to use more memory than we have
96          */
97         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
98                 return 0;
99
100         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
101                 unsigned long n;
102
103                 free = global_page_state(NR_FILE_PAGES);
104                 free += nr_swap_pages;
105
106                 /*
107                  * Any slabs which are created with the
108                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
109                  * which are reclaimable, under pressure.  The dentry
110                  * cache and most inode caches should fall into this
111                  */
112                 free += atomic_read(&slab_reclaim_pages);
113
114                 /*
115                  * Leave the last 3% for root
116                  */
117                 if (!cap_sys_admin)
118                         free -= free / 32;
119
120                 if (free > pages)
121                         return 0;
122
123                 /*
124                  * nr_free_pages() is very expensive on large systems,
125                  * only call if we're about to fail.
126                  */
127                 n = nr_free_pages();
128
129                 /*
130                  * Leave reserved pages. The pages are not for anonymous pages.
131                  */
132                 if (n <= totalreserve_pages)
133                         goto error;
134                 else
135                         n -= totalreserve_pages;
136
137                 /*
138                  * Leave the last 3% for root
139                  */
140                 if (!cap_sys_admin)
141                         n -= n / 32;
142                 free += n;
143
144                 if (free > pages)
145                         return 0;
146
147                 goto error;
148         }
149
150         allowed = (totalram_pages - hugetlb_total_pages())
151                 * sysctl_overcommit_ratio / 100;
152         /*
153          * Leave the last 3% for root
154          */
155         if (!cap_sys_admin)
156                 allowed -= allowed / 32;
157         allowed += total_swap_pages;
158
159         /* Don't let a single process grow too big:
160            leave 3% of the size of this process for other processes */
161         allowed -= current->mm->total_vm / 32;
162
163         /*
164          * cast `allowed' as a signed long because vm_committed_space
165          * sometimes has a negative value
166          */
167         if (atomic_read(&vm_committed_space) < (long)allowed)
168                 return 0;
169 error:
170         vm_unacct_memory(pages);
171
172         return -ENOMEM;
173 }
174
175 EXPORT_SYMBOL(__vm_enough_memory);
176
177 /*
178  * Requires inode->i_mapping->i_mmap_lock
179  */
180 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
181                 struct file *file, struct address_space *mapping)
182 {
183         if (vma->vm_flags & VM_DENYWRITE)
184                 atomic_inc(&file->f_dentry->d_inode->i_writecount);
185         if (vma->vm_flags & VM_SHARED)
186                 mapping->i_mmap_writable--;
187
188         flush_dcache_mmap_lock(mapping);
189         if (unlikely(vma->vm_flags & VM_NONLINEAR))
190                 list_del_init(&vma->shared.vm_set.list);
191         else
192                 vma_prio_tree_remove(vma, &mapping->i_mmap);
193         flush_dcache_mmap_unlock(mapping);
194 }
195
196 /*
197  * Unlink a file-based vm structure from its prio_tree, to hide
198  * vma from rmap and vmtruncate before freeing its page tables.
199  */
200 void unlink_file_vma(struct vm_area_struct *vma)
201 {
202         struct file *file = vma->vm_file;
203
204         if (file) {
205                 struct address_space *mapping = file->f_mapping;
206                 spin_lock(&mapping->i_mmap_lock);
207                 __remove_shared_vm_struct(vma, file, mapping);
208                 spin_unlock(&mapping->i_mmap_lock);
209         }
210 }
211
212 /*
213  * Close a vm structure and free it, returning the next.
214  */
215 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
216 {
217         struct vm_area_struct *next = vma->vm_next;
218
219         might_sleep();
220         if (vma->vm_ops && vma->vm_ops->close)
221                 vma->vm_ops->close(vma);
222         if (vma->vm_file)
223                 fput(vma->vm_file);
224         mpol_free(vma_policy(vma));
225         kmem_cache_free(vm_area_cachep, vma);
226         return next;
227 }
228
229 asmlinkage unsigned long sys_brk(unsigned long brk)
230 {
231         unsigned long rlim, retval;
232         unsigned long newbrk, oldbrk;
233         struct mm_struct *mm = current->mm;
234
235         down_write(&mm->mmap_sem);
236
237         if (brk < mm->end_code)
238                 goto out;
239
240         /*
241          * Check against rlimit here. If this check is done later after the test
242          * of oldbrk with newbrk then it can escape the test and let the data
243          * segment grow beyond its set limit the in case where the limit is
244          * not page aligned -Ram Gupta
245          */
246         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
247         if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
248                 goto out;
249
250         newbrk = PAGE_ALIGN(brk);
251         oldbrk = PAGE_ALIGN(mm->brk);
252         if (oldbrk == newbrk)
253                 goto set_brk;
254
255         /* Always allow shrinking brk. */
256         if (brk <= mm->brk) {
257                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
258                         goto set_brk;
259                 goto out;
260         }
261
262         /* Check against existing mmap mappings. */
263         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
264                 goto out;
265
266         /* Ok, looks good - let it rip. */
267         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
268                 goto out;
269 set_brk:
270         mm->brk = brk;
271 out:
272         retval = mm->brk;
273         up_write(&mm->mmap_sem);
274         return retval;
275 }
276
277 #ifdef DEBUG_MM_RB
278 static int browse_rb(struct rb_root *root)
279 {
280         int i = 0, j;
281         struct rb_node *nd, *pn = NULL;
282         unsigned long prev = 0, pend = 0;
283
284         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
285                 struct vm_area_struct *vma;
286                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
287                 if (vma->vm_start < prev)
288                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
289                 if (vma->vm_start < pend)
290                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
291                 if (vma->vm_start > vma->vm_end)
292                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
293                 i++;
294                 pn = nd;
295         }
296         j = 0;
297         for (nd = pn; nd; nd = rb_prev(nd)) {
298                 j++;
299         }
300         if (i != j)
301                 printk("backwards %d, forwards %d\n", j, i), i = 0;
302         return i;
303 }
304
305 void validate_mm(struct mm_struct *mm)
306 {
307         int bug = 0;
308         int i = 0;
309         struct vm_area_struct *tmp = mm->mmap;
310         while (tmp) {
311                 tmp = tmp->vm_next;
312                 i++;
313         }
314         if (i != mm->map_count)
315                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
316         i = browse_rb(&mm->mm_rb);
317         if (i != mm->map_count)
318                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
319         BUG_ON(bug);
320 }
321 #else
322 #define validate_mm(mm) do { } while (0)
323 #endif
324
325 static struct vm_area_struct *
326 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
327                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
328                 struct rb_node ** rb_parent)
329 {
330         struct vm_area_struct * vma;
331         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
332
333         __rb_link = &mm->mm_rb.rb_node;
334         rb_prev = __rb_parent = NULL;
335         vma = NULL;
336
337         while (*__rb_link) {
338                 struct vm_area_struct *vma_tmp;
339
340                 __rb_parent = *__rb_link;
341                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
342
343                 if (vma_tmp->vm_end > addr) {
344                         vma = vma_tmp;
345                         if (vma_tmp->vm_start <= addr)
346                                 return vma;
347                         __rb_link = &__rb_parent->rb_left;
348                 } else {
349                         rb_prev = __rb_parent;
350                         __rb_link = &__rb_parent->rb_right;
351                 }
352         }
353
354         *pprev = NULL;
355         if (rb_prev)
356                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
357         *rb_link = __rb_link;
358         *rb_parent = __rb_parent;
359         return vma;
360 }
361
362 static inline void
363 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
364                 struct vm_area_struct *prev, struct rb_node *rb_parent)
365 {
366         if (prev) {
367                 vma->vm_next = prev->vm_next;
368                 prev->vm_next = vma;
369         } else {
370                 mm->mmap = vma;
371                 if (rb_parent)
372                         vma->vm_next = rb_entry(rb_parent,
373                                         struct vm_area_struct, vm_rb);
374                 else
375                         vma->vm_next = NULL;
376         }
377 }
378
379 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
380                 struct rb_node **rb_link, struct rb_node *rb_parent)
381 {
382         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
383         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
384 }
385
386 static inline void __vma_link_file(struct vm_area_struct *vma)
387 {
388         struct file * file;
389
390         file = vma->vm_file;
391         if (file) {
392                 struct address_space *mapping = file->f_mapping;
393
394                 if (vma->vm_flags & VM_DENYWRITE)
395                         atomic_dec(&file->f_dentry->d_inode->i_writecount);
396                 if (vma->vm_flags & VM_SHARED)
397                         mapping->i_mmap_writable++;
398
399                 flush_dcache_mmap_lock(mapping);
400                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
401                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
402                 else
403                         vma_prio_tree_insert(vma, &mapping->i_mmap);
404                 flush_dcache_mmap_unlock(mapping);
405         }
406 }
407
408 static void
409 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
410         struct vm_area_struct *prev, struct rb_node **rb_link,
411         struct rb_node *rb_parent)
412 {
413         __vma_link_list(mm, vma, prev, rb_parent);
414         __vma_link_rb(mm, vma, rb_link, rb_parent);
415         __anon_vma_link(vma);
416 }
417
418 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
419                         struct vm_area_struct *prev, struct rb_node **rb_link,
420                         struct rb_node *rb_parent)
421 {
422         struct address_space *mapping = NULL;
423
424         if (vma->vm_file)
425                 mapping = vma->vm_file->f_mapping;
426
427         if (mapping) {
428                 spin_lock(&mapping->i_mmap_lock);
429                 vma->vm_truncate_count = mapping->truncate_count;
430         }
431         anon_vma_lock(vma);
432
433         __vma_link(mm, vma, prev, rb_link, rb_parent);
434         __vma_link_file(vma);
435
436         anon_vma_unlock(vma);
437         if (mapping)
438                 spin_unlock(&mapping->i_mmap_lock);
439
440         mm->map_count++;
441         validate_mm(mm);
442 }
443
444 /*
445  * Helper for vma_adjust in the split_vma insert case:
446  * insert vm structure into list and rbtree and anon_vma,
447  * but it has already been inserted into prio_tree earlier.
448  */
449 static void
450 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
451 {
452         struct vm_area_struct * __vma, * prev;
453         struct rb_node ** rb_link, * rb_parent;
454
455         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
456         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
457         __vma_link(mm, vma, prev, rb_link, rb_parent);
458         mm->map_count++;
459 }
460
461 static inline void
462 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
463                 struct vm_area_struct *prev)
464 {
465         prev->vm_next = vma->vm_next;
466         rb_erase(&vma->vm_rb, &mm->mm_rb);
467         if (mm->mmap_cache == vma)
468                 mm->mmap_cache = prev;
469 }
470
471 /*
472  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
473  * is already present in an i_mmap tree without adjusting the tree.
474  * The following helper function should be used when such adjustments
475  * are necessary.  The "insert" vma (if any) is to be inserted
476  * before we drop the necessary locks.
477  */
478 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
479         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
480 {
481         struct mm_struct *mm = vma->vm_mm;
482         struct vm_area_struct *next = vma->vm_next;
483         struct vm_area_struct *importer = NULL;
484         struct address_space *mapping = NULL;
485         struct prio_tree_root *root = NULL;
486         struct file *file = vma->vm_file;
487         struct anon_vma *anon_vma = NULL;
488         long adjust_next = 0;
489         int remove_next = 0;
490
491         if (next && !insert) {
492                 if (end >= next->vm_end) {
493                         /*
494                          * vma expands, overlapping all the next, and
495                          * perhaps the one after too (mprotect case 6).
496                          */
497 again:                  remove_next = 1 + (end > next->vm_end);
498                         end = next->vm_end;
499                         anon_vma = next->anon_vma;
500                         importer = vma;
501                 } else if (end > next->vm_start) {
502                         /*
503                          * vma expands, overlapping part of the next:
504                          * mprotect case 5 shifting the boundary up.
505                          */
506                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
507                         anon_vma = next->anon_vma;
508                         importer = vma;
509                 } else if (end < vma->vm_end) {
510                         /*
511                          * vma shrinks, and !insert tells it's not
512                          * split_vma inserting another: so it must be
513                          * mprotect case 4 shifting the boundary down.
514                          */
515                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
516                         anon_vma = next->anon_vma;
517                         importer = next;
518                 }
519         }
520
521         if (file) {
522                 mapping = file->f_mapping;
523                 if (!(vma->vm_flags & VM_NONLINEAR))
524                         root = &mapping->i_mmap;
525                 spin_lock(&mapping->i_mmap_lock);
526                 if (importer &&
527                     vma->vm_truncate_count != next->vm_truncate_count) {
528                         /*
529                          * unmap_mapping_range might be in progress:
530                          * ensure that the expanding vma is rescanned.
531                          */
532                         importer->vm_truncate_count = 0;
533                 }
534                 if (insert) {
535                         insert->vm_truncate_count = vma->vm_truncate_count;
536                         /*
537                          * Put into prio_tree now, so instantiated pages
538                          * are visible to arm/parisc __flush_dcache_page
539                          * throughout; but we cannot insert into address
540                          * space until vma start or end is updated.
541                          */
542                         __vma_link_file(insert);
543                 }
544         }
545
546         /*
547          * When changing only vma->vm_end, we don't really need
548          * anon_vma lock: but is that case worth optimizing out?
549          */
550         if (vma->anon_vma)
551                 anon_vma = vma->anon_vma;
552         if (anon_vma) {
553                 spin_lock(&anon_vma->lock);
554                 /*
555                  * Easily overlooked: when mprotect shifts the boundary,
556                  * make sure the expanding vma has anon_vma set if the
557                  * shrinking vma had, to cover any anon pages imported.
558                  */
559                 if (importer && !importer->anon_vma) {
560                         importer->anon_vma = anon_vma;
561                         __anon_vma_link(importer);
562                 }
563         }
564
565         if (root) {
566                 flush_dcache_mmap_lock(mapping);
567                 vma_prio_tree_remove(vma, root);
568                 if (adjust_next)
569                         vma_prio_tree_remove(next, root);
570         }
571
572         vma->vm_start = start;
573         vma->vm_end = end;
574         vma->vm_pgoff = pgoff;
575         if (adjust_next) {
576                 next->vm_start += adjust_next << PAGE_SHIFT;
577                 next->vm_pgoff += adjust_next;
578         }
579
580         if (root) {
581                 if (adjust_next)
582                         vma_prio_tree_insert(next, root);
583                 vma_prio_tree_insert(vma, root);
584                 flush_dcache_mmap_unlock(mapping);
585         }
586
587         if (remove_next) {
588                 /*
589                  * vma_merge has merged next into vma, and needs
590                  * us to remove next before dropping the locks.
591                  */
592                 __vma_unlink(mm, next, vma);
593                 if (file)
594                         __remove_shared_vm_struct(next, file, mapping);
595                 if (next->anon_vma)
596                         __anon_vma_merge(vma, next);
597         } else if (insert) {
598                 /*
599                  * split_vma has split insert from vma, and needs
600                  * us to insert it before dropping the locks
601                  * (it may either follow vma or precede it).
602                  */
603                 __insert_vm_struct(mm, insert);
604         }
605
606         if (anon_vma)
607                 spin_unlock(&anon_vma->lock);
608         if (mapping)
609                 spin_unlock(&mapping->i_mmap_lock);
610
611         if (remove_next) {
612                 if (file)
613                         fput(file);
614                 mm->map_count--;
615                 mpol_free(vma_policy(next));
616                 kmem_cache_free(vm_area_cachep, next);
617                 /*
618                  * In mprotect's case 6 (see comments on vma_merge),
619                  * we must remove another next too. It would clutter
620                  * up the code too much to do both in one go.
621                  */
622                 if (remove_next == 2) {
623                         next = vma->vm_next;
624                         goto again;
625                 }
626         }
627
628         validate_mm(mm);
629 }
630
631 /*
632  * If the vma has a ->close operation then the driver probably needs to release
633  * per-vma resources, so we don't attempt to merge those.
634  */
635 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
636
637 static inline int is_mergeable_vma(struct vm_area_struct *vma,
638                         struct file *file, unsigned long vm_flags)
639 {
640         if (vma->vm_flags != vm_flags)
641                 return 0;
642         if (vma->vm_file != file)
643                 return 0;
644         if (vma->vm_ops && vma->vm_ops->close)
645                 return 0;
646         return 1;
647 }
648
649 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
650                                         struct anon_vma *anon_vma2)
651 {
652         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
653 }
654
655 /*
656  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
657  * in front of (at a lower virtual address and file offset than) the vma.
658  *
659  * We cannot merge two vmas if they have differently assigned (non-NULL)
660  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
661  *
662  * We don't check here for the merged mmap wrapping around the end of pagecache
663  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
664  * wrap, nor mmaps which cover the final page at index -1UL.
665  */
666 static int
667 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
668         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
669 {
670         if (is_mergeable_vma(vma, file, vm_flags) &&
671             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
672                 if (vma->vm_pgoff == vm_pgoff)
673                         return 1;
674         }
675         return 0;
676 }
677
678 /*
679  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
680  * beyond (at a higher virtual address and file offset than) the vma.
681  *
682  * We cannot merge two vmas if they have differently assigned (non-NULL)
683  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
684  */
685 static int
686 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
687         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
688 {
689         if (is_mergeable_vma(vma, file, vm_flags) &&
690             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
691                 pgoff_t vm_pglen;
692                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
693                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
694                         return 1;
695         }
696         return 0;
697 }
698
699 /*
700  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
701  * whether that can be merged with its predecessor or its successor.
702  * Or both (it neatly fills a hole).
703  *
704  * In most cases - when called for mmap, brk or mremap - [addr,end) is
705  * certain not to be mapped by the time vma_merge is called; but when
706  * called for mprotect, it is certain to be already mapped (either at
707  * an offset within prev, or at the start of next), and the flags of
708  * this area are about to be changed to vm_flags - and the no-change
709  * case has already been eliminated.
710  *
711  * The following mprotect cases have to be considered, where AAAA is
712  * the area passed down from mprotect_fixup, never extending beyond one
713  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
714  *
715  *     AAAA             AAAA                AAAA          AAAA
716  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
717  *    cannot merge    might become    might become    might become
718  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
719  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
720  *    mremap move:                                    PPPPNNNNNNNN 8
721  *        AAAA
722  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
723  *    might become    case 1 below    case 2 below    case 3 below
724  *
725  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
726  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
727  */
728 struct vm_area_struct *vma_merge(struct mm_struct *mm,
729                         struct vm_area_struct *prev, unsigned long addr,
730                         unsigned long end, unsigned long vm_flags,
731                         struct anon_vma *anon_vma, struct file *file,
732                         pgoff_t pgoff, struct mempolicy *policy)
733 {
734         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
735         struct vm_area_struct *area, *next;
736
737         /*
738          * We later require that vma->vm_flags == vm_flags,
739          * so this tests vma->vm_flags & VM_SPECIAL, too.
740          */
741         if (vm_flags & VM_SPECIAL)
742                 return NULL;
743
744         if (prev)
745                 next = prev->vm_next;
746         else
747                 next = mm->mmap;
748         area = next;
749         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
750                 next = next->vm_next;
751
752         /*
753          * Can it merge with the predecessor?
754          */
755         if (prev && prev->vm_end == addr &&
756                         mpol_equal(vma_policy(prev), policy) &&
757                         can_vma_merge_after(prev, vm_flags,
758                                                 anon_vma, file, pgoff)) {
759                 /*
760                  * OK, it can.  Can we now merge in the successor as well?
761                  */
762                 if (next && end == next->vm_start &&
763                                 mpol_equal(policy, vma_policy(next)) &&
764                                 can_vma_merge_before(next, vm_flags,
765                                         anon_vma, file, pgoff+pglen) &&
766                                 is_mergeable_anon_vma(prev->anon_vma,
767                                                       next->anon_vma)) {
768                                                         /* cases 1, 6 */
769                         vma_adjust(prev, prev->vm_start,
770                                 next->vm_end, prev->vm_pgoff, NULL);
771                 } else                                  /* cases 2, 5, 7 */
772                         vma_adjust(prev, prev->vm_start,
773                                 end, prev->vm_pgoff, NULL);
774                 return prev;
775         }
776
777         /*
778          * Can this new request be merged in front of next?
779          */
780         if (next && end == next->vm_start &&
781                         mpol_equal(policy, vma_policy(next)) &&
782                         can_vma_merge_before(next, vm_flags,
783                                         anon_vma, file, pgoff+pglen)) {
784                 if (prev && addr < prev->vm_end)        /* case 4 */
785                         vma_adjust(prev, prev->vm_start,
786                                 addr, prev->vm_pgoff, NULL);
787                 else                                    /* cases 3, 8 */
788                         vma_adjust(area, addr, next->vm_end,
789                                 next->vm_pgoff - pglen, NULL);
790                 return area;
791         }
792
793         return NULL;
794 }
795
796 /*
797  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
798  * neighbouring vmas for a suitable anon_vma, before it goes off
799  * to allocate a new anon_vma.  It checks because a repetitive
800  * sequence of mprotects and faults may otherwise lead to distinct
801  * anon_vmas being allocated, preventing vma merge in subsequent
802  * mprotect.
803  */
804 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
805 {
806         struct vm_area_struct *near;
807         unsigned long vm_flags;
808
809         near = vma->vm_next;
810         if (!near)
811                 goto try_prev;
812
813         /*
814          * Since only mprotect tries to remerge vmas, match flags
815          * which might be mprotected into each other later on.
816          * Neither mlock nor madvise tries to remerge at present,
817          * so leave their flags as obstructing a merge.
818          */
819         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
820         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
821
822         if (near->anon_vma && vma->vm_end == near->vm_start &&
823                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
824                         can_vma_merge_before(near, vm_flags,
825                                 NULL, vma->vm_file, vma->vm_pgoff +
826                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
827                 return near->anon_vma;
828 try_prev:
829         /*
830          * It is potentially slow to have to call find_vma_prev here.
831          * But it's only on the first write fault on the vma, not
832          * every time, and we could devise a way to avoid it later
833          * (e.g. stash info in next's anon_vma_node when assigning
834          * an anon_vma, or when trying vma_merge).  Another time.
835          */
836         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
837         if (!near)
838                 goto none;
839
840         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
841         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
842
843         if (near->anon_vma && near->vm_end == vma->vm_start &&
844                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
845                         can_vma_merge_after(near, vm_flags,
846                                 NULL, vma->vm_file, vma->vm_pgoff))
847                 return near->anon_vma;
848 none:
849         /*
850          * There's no absolute need to look only at touching neighbours:
851          * we could search further afield for "compatible" anon_vmas.
852          * But it would probably just be a waste of time searching,
853          * or lead to too many vmas hanging off the same anon_vma.
854          * We're trying to allow mprotect remerging later on,
855          * not trying to minimize memory used for anon_vmas.
856          */
857         return NULL;
858 }
859
860 #ifdef CONFIG_PROC_FS
861 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
862                                                 struct file *file, long pages)
863 {
864         const unsigned long stack_flags
865                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
866
867         if (file) {
868                 mm->shared_vm += pages;
869                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
870                         mm->exec_vm += pages;
871         } else if (flags & stack_flags)
872                 mm->stack_vm += pages;
873         if (flags & (VM_RESERVED|VM_IO))
874                 mm->reserved_vm += pages;
875 }
876 #endif /* CONFIG_PROC_FS */
877
878 /*
879  * The caller must hold down_write(current->mm->mmap_sem).
880  */
881
882 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
883                         unsigned long len, unsigned long prot,
884                         unsigned long flags, unsigned long pgoff)
885 {
886         struct mm_struct * mm = current->mm;
887         struct vm_area_struct * vma, * prev;
888         struct inode *inode;
889         unsigned int vm_flags;
890         int correct_wcount = 0;
891         int error;
892         struct rb_node ** rb_link, * rb_parent;
893         int accountable = 1;
894         unsigned long charged = 0, reqprot = prot;
895
896         if (file) {
897                 if (is_file_hugepages(file))
898                         accountable = 0;
899
900                 if (!file->f_op || !file->f_op->mmap)
901                         return -ENODEV;
902
903                 if ((prot & PROT_EXEC) &&
904                     (file->f_vfsmnt->mnt_flags & MNT_NOEXEC))
905                         return -EPERM;
906         }
907         /*
908          * Does the application expect PROT_READ to imply PROT_EXEC?
909          *
910          * (the exception is when the underlying filesystem is noexec
911          *  mounted, in which case we dont add PROT_EXEC.)
912          */
913         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
914                 if (!(file && (file->f_vfsmnt->mnt_flags & MNT_NOEXEC)))
915                         prot |= PROT_EXEC;
916
917         if (!len)
918                 return -EINVAL;
919
920         error = arch_mmap_check(addr, len, flags);
921         if (error)
922                 return error;
923
924         /* Careful about overflows.. */
925         len = PAGE_ALIGN(len);
926         if (!len || len > TASK_SIZE)
927                 return -ENOMEM;
928
929         /* offset overflow? */
930         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
931                return -EOVERFLOW;
932
933         /* Too many mappings? */
934         if (mm->map_count > sysctl_max_map_count)
935                 return -ENOMEM;
936
937         /* Obtain the address to map to. we verify (or select) it and ensure
938          * that it represents a valid section of the address space.
939          */
940         addr = get_unmapped_area(file, addr, len, pgoff, flags);
941         if (addr & ~PAGE_MASK)
942                 return addr;
943
944         /* Do simple checking here so the lower-level routines won't have
945          * to. we assume access permissions have been handled by the open
946          * of the memory object, so we don't do any here.
947          */
948         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
949                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
950
951         if (flags & MAP_LOCKED) {
952                 if (!can_do_mlock())
953                         return -EPERM;
954                 vm_flags |= VM_LOCKED;
955         }
956         /* mlock MCL_FUTURE? */
957         if (vm_flags & VM_LOCKED) {
958                 unsigned long locked, lock_limit;
959                 locked = len >> PAGE_SHIFT;
960                 locked += mm->locked_vm;
961                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
962                 lock_limit >>= PAGE_SHIFT;
963                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
964                         return -EAGAIN;
965         }
966
967         inode = file ? file->f_dentry->d_inode : NULL;
968
969         if (file) {
970                 switch (flags & MAP_TYPE) {
971                 case MAP_SHARED:
972                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
973                                 return -EACCES;
974
975                         /*
976                          * Make sure we don't allow writing to an append-only
977                          * file..
978                          */
979                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
980                                 return -EACCES;
981
982                         /*
983                          * Make sure there are no mandatory locks on the file.
984                          */
985                         if (locks_verify_locked(inode))
986                                 return -EAGAIN;
987
988                         vm_flags |= VM_SHARED | VM_MAYSHARE;
989                         if (!(file->f_mode & FMODE_WRITE))
990                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
991
992                         /* fall through */
993                 case MAP_PRIVATE:
994                         if (!(file->f_mode & FMODE_READ))
995                                 return -EACCES;
996                         break;
997
998                 default:
999                         return -EINVAL;
1000                 }
1001         } else {
1002                 switch (flags & MAP_TYPE) {
1003                 case MAP_SHARED:
1004                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1005                         break;
1006                 case MAP_PRIVATE:
1007                         /*
1008                          * Set pgoff according to addr for anon_vma.
1009                          */
1010                         pgoff = addr >> PAGE_SHIFT;
1011                         break;
1012                 default:
1013                         return -EINVAL;
1014                 }
1015         }
1016
1017         error = security_file_mmap(file, reqprot, prot, flags);
1018         if (error)
1019                 return error;
1020                 
1021         /* Clear old maps */
1022         error = -ENOMEM;
1023 munmap_back:
1024         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1025         if (vma && vma->vm_start < addr + len) {
1026                 if (do_munmap(mm, addr, len))
1027                         return -ENOMEM;
1028                 goto munmap_back;
1029         }
1030
1031         /* Check against address space limit. */
1032         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1033                 return -ENOMEM;
1034
1035         if (accountable && (!(flags & MAP_NORESERVE) ||
1036                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1037                 if (vm_flags & VM_SHARED) {
1038                         /* Check memory availability in shmem_file_setup? */
1039                         vm_flags |= VM_ACCOUNT;
1040                 } else if (vm_flags & VM_WRITE) {
1041                         /*
1042                          * Private writable mapping: check memory availability
1043                          */
1044                         charged = len >> PAGE_SHIFT;
1045                         if (security_vm_enough_memory(charged))
1046                                 return -ENOMEM;
1047                         vm_flags |= VM_ACCOUNT;
1048                 }
1049         }
1050
1051         /*
1052          * Can we just expand an old private anonymous mapping?
1053          * The VM_SHARED test is necessary because shmem_zero_setup
1054          * will create the file object for a shared anonymous map below.
1055          */
1056         if (!file && !(vm_flags & VM_SHARED) &&
1057             vma_merge(mm, prev, addr, addr + len, vm_flags,
1058                                         NULL, NULL, pgoff, NULL))
1059                 goto out;
1060
1061         /*
1062          * Determine the object being mapped and call the appropriate
1063          * specific mapper. the address has already been validated, but
1064          * not unmapped, but the maps are removed from the list.
1065          */
1066         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1067         if (!vma) {
1068                 error = -ENOMEM;
1069                 goto unacct_error;
1070         }
1071
1072         vma->vm_mm = mm;
1073         vma->vm_start = addr;
1074         vma->vm_end = addr + len;
1075         vma->vm_flags = vm_flags;
1076         vma->vm_page_prot = protection_map[vm_flags &
1077                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1078         vma->vm_pgoff = pgoff;
1079
1080         if (file) {
1081                 error = -EINVAL;
1082                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1083                         goto free_vma;
1084                 if (vm_flags & VM_DENYWRITE) {
1085                         error = deny_write_access(file);
1086                         if (error)
1087                                 goto free_vma;
1088                         correct_wcount = 1;
1089                 }
1090                 vma->vm_file = file;
1091                 get_file(file);
1092                 error = file->f_op->mmap(file, vma);
1093                 if (error)
1094                         goto unmap_and_free_vma;
1095         } else if (vm_flags & VM_SHARED) {
1096                 error = shmem_zero_setup(vma);
1097                 if (error)
1098                         goto free_vma;
1099         }
1100
1101         /* Don't make the VMA automatically writable if it's shared, but the
1102          * backer wishes to know when pages are first written to */
1103         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1104                 vma->vm_page_prot =
1105                         protection_map[vm_flags & (VM_READ|VM_WRITE|VM_EXEC)];
1106
1107         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1108          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1109          * that memory reservation must be checked; but that reservation
1110          * belongs to shared memory object, not to vma: so now clear it.
1111          */
1112         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1113                 vma->vm_flags &= ~VM_ACCOUNT;
1114
1115         /* Can addr have changed??
1116          *
1117          * Answer: Yes, several device drivers can do it in their
1118          *         f_op->mmap method. -DaveM
1119          */
1120         addr = vma->vm_start;
1121         pgoff = vma->vm_pgoff;
1122         vm_flags = vma->vm_flags;
1123
1124         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1125                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1126                 file = vma->vm_file;
1127                 vma_link(mm, vma, prev, rb_link, rb_parent);
1128                 if (correct_wcount)
1129                         atomic_inc(&inode->i_writecount);
1130         } else {
1131                 if (file) {
1132                         if (correct_wcount)
1133                                 atomic_inc(&inode->i_writecount);
1134                         fput(file);
1135                 }
1136                 mpol_free(vma_policy(vma));
1137                 kmem_cache_free(vm_area_cachep, vma);
1138         }
1139 out:    
1140         mm->total_vm += len >> PAGE_SHIFT;
1141         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1142         if (vm_flags & VM_LOCKED) {
1143                 mm->locked_vm += len >> PAGE_SHIFT;
1144                 make_pages_present(addr, addr + len);
1145         }
1146         if (flags & MAP_POPULATE) {
1147                 up_write(&mm->mmap_sem);
1148                 sys_remap_file_pages(addr, len, 0,
1149                                         pgoff, flags & MAP_NONBLOCK);
1150                 down_write(&mm->mmap_sem);
1151         }
1152         return addr;
1153
1154 unmap_and_free_vma:
1155         if (correct_wcount)
1156                 atomic_inc(&inode->i_writecount);
1157         vma->vm_file = NULL;
1158         fput(file);
1159
1160         /* Undo any partial mapping done by a device driver. */
1161         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1162         charged = 0;
1163 free_vma:
1164         kmem_cache_free(vm_area_cachep, vma);
1165 unacct_error:
1166         if (charged)
1167                 vm_unacct_memory(charged);
1168         return error;
1169 }
1170
1171 EXPORT_SYMBOL(do_mmap_pgoff);
1172
1173 /* Get an address range which is currently unmapped.
1174  * For shmat() with addr=0.
1175  *
1176  * Ugly calling convention alert:
1177  * Return value with the low bits set means error value,
1178  * ie
1179  *      if (ret & ~PAGE_MASK)
1180  *              error = ret;
1181  *
1182  * This function "knows" that -ENOMEM has the bits set.
1183  */
1184 #ifndef HAVE_ARCH_UNMAPPED_AREA
1185 unsigned long
1186 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1187                 unsigned long len, unsigned long pgoff, unsigned long flags)
1188 {
1189         struct mm_struct *mm = current->mm;
1190         struct vm_area_struct *vma;
1191         unsigned long start_addr;
1192
1193         if (len > TASK_SIZE)
1194                 return -ENOMEM;
1195
1196         if (addr) {
1197                 addr = PAGE_ALIGN(addr);
1198                 vma = find_vma(mm, addr);
1199                 if (TASK_SIZE - len >= addr &&
1200                     (!vma || addr + len <= vma->vm_start))
1201                         return addr;
1202         }
1203         if (len > mm->cached_hole_size) {
1204                 start_addr = addr = mm->free_area_cache;
1205         } else {
1206                 start_addr = addr = TASK_UNMAPPED_BASE;
1207                 mm->cached_hole_size = 0;
1208         }
1209
1210 full_search:
1211         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1212                 /* At this point:  (!vma || addr < vma->vm_end). */
1213                 if (TASK_SIZE - len < addr) {
1214                         /*
1215                          * Start a new search - just in case we missed
1216                          * some holes.
1217                          */
1218                         if (start_addr != TASK_UNMAPPED_BASE) {
1219                                 addr = TASK_UNMAPPED_BASE;
1220                                 start_addr = addr;
1221                                 mm->cached_hole_size = 0;
1222                                 goto full_search;
1223                         }
1224                         return -ENOMEM;
1225                 }
1226                 if (!vma || addr + len <= vma->vm_start) {
1227                         /*
1228                          * Remember the place where we stopped the search:
1229                          */
1230                         mm->free_area_cache = addr + len;
1231                         return addr;
1232                 }
1233                 if (addr + mm->cached_hole_size < vma->vm_start)
1234                         mm->cached_hole_size = vma->vm_start - addr;
1235                 addr = vma->vm_end;
1236         }
1237 }
1238 #endif  
1239
1240 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1241 {
1242         /*
1243          * Is this a new hole at the lowest possible address?
1244          */
1245         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1246                 mm->free_area_cache = addr;
1247                 mm->cached_hole_size = ~0UL;
1248         }
1249 }
1250
1251 /*
1252  * This mmap-allocator allocates new areas top-down from below the
1253  * stack's low limit (the base):
1254  */
1255 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1256 unsigned long
1257 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1258                           const unsigned long len, const unsigned long pgoff,
1259                           const unsigned long flags)
1260 {
1261         struct vm_area_struct *vma;
1262         struct mm_struct *mm = current->mm;
1263         unsigned long addr = addr0;
1264
1265         /* requested length too big for entire address space */
1266         if (len > TASK_SIZE)
1267                 return -ENOMEM;
1268
1269         /* requesting a specific address */
1270         if (addr) {
1271                 addr = PAGE_ALIGN(addr);
1272                 vma = find_vma(mm, addr);
1273                 if (TASK_SIZE - len >= addr &&
1274                                 (!vma || addr + len <= vma->vm_start))
1275                         return addr;
1276         }
1277
1278         /* check if free_area_cache is useful for us */
1279         if (len <= mm->cached_hole_size) {
1280                 mm->cached_hole_size = 0;
1281                 mm->free_area_cache = mm->mmap_base;
1282         }
1283
1284         /* either no address requested or can't fit in requested address hole */
1285         addr = mm->free_area_cache;
1286
1287         /* make sure it can fit in the remaining address space */
1288         if (addr > len) {
1289                 vma = find_vma(mm, addr-len);
1290                 if (!vma || addr <= vma->vm_start)
1291                         /* remember the address as a hint for next time */
1292                         return (mm->free_area_cache = addr-len);
1293         }
1294
1295         if (mm->mmap_base < len)
1296                 goto bottomup;
1297
1298         addr = mm->mmap_base-len;
1299
1300         do {
1301                 /*
1302                  * Lookup failure means no vma is above this address,
1303                  * else if new region fits below vma->vm_start,
1304                  * return with success:
1305                  */
1306                 vma = find_vma(mm, addr);
1307                 if (!vma || addr+len <= vma->vm_start)
1308                         /* remember the address as a hint for next time */
1309                         return (mm->free_area_cache = addr);
1310
1311                 /* remember the largest hole we saw so far */
1312                 if (addr + mm->cached_hole_size < vma->vm_start)
1313                         mm->cached_hole_size = vma->vm_start - addr;
1314
1315                 /* try just below the current vma->vm_start */
1316                 addr = vma->vm_start-len;
1317         } while (len < vma->vm_start);
1318
1319 bottomup:
1320         /*
1321          * A failed mmap() very likely causes application failure,
1322          * so fall back to the bottom-up function here. This scenario
1323          * can happen with large stack limits and large mmap()
1324          * allocations.
1325          */
1326         mm->cached_hole_size = ~0UL;
1327         mm->free_area_cache = TASK_UNMAPPED_BASE;
1328         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1329         /*
1330          * Restore the topdown base:
1331          */
1332         mm->free_area_cache = mm->mmap_base;
1333         mm->cached_hole_size = ~0UL;
1334
1335         return addr;
1336 }
1337 #endif
1338
1339 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1340 {
1341         /*
1342          * Is this a new hole at the highest possible address?
1343          */
1344         if (addr > mm->free_area_cache)
1345                 mm->free_area_cache = addr;
1346
1347         /* dont allow allocations above current base */
1348         if (mm->free_area_cache > mm->mmap_base)
1349                 mm->free_area_cache = mm->mmap_base;
1350 }
1351
1352 unsigned long
1353 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1354                 unsigned long pgoff, unsigned long flags)
1355 {
1356         unsigned long ret;
1357
1358         if (!(flags & MAP_FIXED)) {
1359                 unsigned long (*get_area)(struct file *, unsigned long, unsigned long, unsigned long, unsigned long);
1360
1361                 get_area = current->mm->get_unmapped_area;
1362                 if (file && file->f_op && file->f_op->get_unmapped_area)
1363                         get_area = file->f_op->get_unmapped_area;
1364                 addr = get_area(file, addr, len, pgoff, flags);
1365                 if (IS_ERR_VALUE(addr))
1366                         return addr;
1367         }
1368
1369         if (addr > TASK_SIZE - len)
1370                 return -ENOMEM;
1371         if (addr & ~PAGE_MASK)
1372                 return -EINVAL;
1373         if (file && is_file_hugepages(file))  {
1374                 /*
1375                  * Check if the given range is hugepage aligned, and
1376                  * can be made suitable for hugepages.
1377                  */
1378                 ret = prepare_hugepage_range(addr, len);
1379         } else {
1380                 /*
1381                  * Ensure that a normal request is not falling in a
1382                  * reserved hugepage range.  For some archs like IA-64,
1383                  * there is a separate region for hugepages.
1384                  */
1385                 ret = is_hugepage_only_range(current->mm, addr, len);
1386         }
1387         if (ret)
1388                 return -EINVAL;
1389         return addr;
1390 }
1391
1392 EXPORT_SYMBOL(get_unmapped_area);
1393
1394 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1395 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1396 {
1397         struct vm_area_struct *vma = NULL;
1398
1399         if (mm) {
1400                 /* Check the cache first. */
1401                 /* (Cache hit rate is typically around 35%.) */
1402                 vma = mm->mmap_cache;
1403                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1404                         struct rb_node * rb_node;
1405
1406                         rb_node = mm->mm_rb.rb_node;
1407                         vma = NULL;
1408
1409                         while (rb_node) {
1410                                 struct vm_area_struct * vma_tmp;
1411
1412                                 vma_tmp = rb_entry(rb_node,
1413                                                 struct vm_area_struct, vm_rb);
1414
1415                                 if (vma_tmp->vm_end > addr) {
1416                                         vma = vma_tmp;
1417                                         if (vma_tmp->vm_start <= addr)
1418                                                 break;
1419                                         rb_node = rb_node->rb_left;
1420                                 } else
1421                                         rb_node = rb_node->rb_right;
1422                         }
1423                         if (vma)
1424                                 mm->mmap_cache = vma;
1425                 }
1426         }
1427         return vma;
1428 }
1429
1430 EXPORT_SYMBOL(find_vma);
1431
1432 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1433 struct vm_area_struct *
1434 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1435                         struct vm_area_struct **pprev)
1436 {
1437         struct vm_area_struct *vma = NULL, *prev = NULL;
1438         struct rb_node * rb_node;
1439         if (!mm)
1440                 goto out;
1441
1442         /* Guard against addr being lower than the first VMA */
1443         vma = mm->mmap;
1444
1445         /* Go through the RB tree quickly. */
1446         rb_node = mm->mm_rb.rb_node;
1447
1448         while (rb_node) {
1449                 struct vm_area_struct *vma_tmp;
1450                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1451
1452                 if (addr < vma_tmp->vm_end) {
1453                         rb_node = rb_node->rb_left;
1454                 } else {
1455                         prev = vma_tmp;
1456                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1457                                 break;
1458                         rb_node = rb_node->rb_right;
1459                 }
1460         }
1461
1462 out:
1463         *pprev = prev;
1464         return prev ? prev->vm_next : vma;
1465 }
1466
1467 /*
1468  * Verify that the stack growth is acceptable and
1469  * update accounting. This is shared with both the
1470  * grow-up and grow-down cases.
1471  */
1472 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1473 {
1474         struct mm_struct *mm = vma->vm_mm;
1475         struct rlimit *rlim = current->signal->rlim;
1476
1477         /* address space limit tests */
1478         if (!may_expand_vm(mm, grow))
1479                 return -ENOMEM;
1480
1481         /* Stack limit test */
1482         if (size > rlim[RLIMIT_STACK].rlim_cur)
1483                 return -ENOMEM;
1484
1485         /* mlock limit tests */
1486         if (vma->vm_flags & VM_LOCKED) {
1487                 unsigned long locked;
1488                 unsigned long limit;
1489                 locked = mm->locked_vm + grow;
1490                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1491                 if (locked > limit && !capable(CAP_IPC_LOCK))
1492                         return -ENOMEM;
1493         }
1494
1495         /*
1496          * Overcommit..  This must be the final test, as it will
1497          * update security statistics.
1498          */
1499         if (security_vm_enough_memory(grow))
1500                 return -ENOMEM;
1501
1502         /* Ok, everything looks good - let it rip */
1503         mm->total_vm += grow;
1504         if (vma->vm_flags & VM_LOCKED)
1505                 mm->locked_vm += grow;
1506         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1507         return 0;
1508 }
1509
1510 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1511 /*
1512  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1513  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1514  */
1515 #ifndef CONFIG_IA64
1516 static inline
1517 #endif
1518 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1519 {
1520         int error;
1521
1522         if (!(vma->vm_flags & VM_GROWSUP))
1523                 return -EFAULT;
1524
1525         /*
1526          * We must make sure the anon_vma is allocated
1527          * so that the anon_vma locking is not a noop.
1528          */
1529         if (unlikely(anon_vma_prepare(vma)))
1530                 return -ENOMEM;
1531         anon_vma_lock(vma);
1532
1533         /*
1534          * vma->vm_start/vm_end cannot change under us because the caller
1535          * is required to hold the mmap_sem in read mode.  We need the
1536          * anon_vma lock to serialize against concurrent expand_stacks.
1537          */
1538         address += 4 + PAGE_SIZE - 1;
1539         address &= PAGE_MASK;
1540         error = 0;
1541
1542         /* Somebody else might have raced and expanded it already */
1543         if (address > vma->vm_end) {
1544                 unsigned long size, grow;
1545
1546                 size = address - vma->vm_start;
1547                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1548
1549                 error = acct_stack_growth(vma, size, grow);
1550                 if (!error)
1551                         vma->vm_end = address;
1552         }
1553         anon_vma_unlock(vma);
1554         return error;
1555 }
1556 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1557
1558 #ifdef CONFIG_STACK_GROWSUP
1559 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1560 {
1561         return expand_upwards(vma, address);
1562 }
1563
1564 struct vm_area_struct *
1565 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1566 {
1567         struct vm_area_struct *vma, *prev;
1568
1569         addr &= PAGE_MASK;
1570         vma = find_vma_prev(mm, addr, &prev);
1571         if (vma && (vma->vm_start <= addr))
1572                 return vma;
1573         if (!prev || expand_stack(prev, addr))
1574                 return NULL;
1575         if (prev->vm_flags & VM_LOCKED) {
1576                 make_pages_present(addr, prev->vm_end);
1577         }
1578         return prev;
1579 }
1580 #else
1581 /*
1582  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1583  */
1584 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1585 {
1586         int error;
1587
1588         /*
1589          * We must make sure the anon_vma is allocated
1590          * so that the anon_vma locking is not a noop.
1591          */
1592         if (unlikely(anon_vma_prepare(vma)))
1593                 return -ENOMEM;
1594         anon_vma_lock(vma);
1595
1596         /*
1597          * vma->vm_start/vm_end cannot change under us because the caller
1598          * is required to hold the mmap_sem in read mode.  We need the
1599          * anon_vma lock to serialize against concurrent expand_stacks.
1600          */
1601         address &= PAGE_MASK;
1602         error = 0;
1603
1604         /* Somebody else might have raced and expanded it already */
1605         if (address < vma->vm_start) {
1606                 unsigned long size, grow;
1607
1608                 size = vma->vm_end - address;
1609                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1610
1611                 error = acct_stack_growth(vma, size, grow);
1612                 if (!error) {
1613                         vma->vm_start = address;
1614                         vma->vm_pgoff -= grow;
1615                 }
1616         }
1617         anon_vma_unlock(vma);
1618         return error;
1619 }
1620
1621 struct vm_area_struct *
1622 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1623 {
1624         struct vm_area_struct * vma;
1625         unsigned long start;
1626
1627         addr &= PAGE_MASK;
1628         vma = find_vma(mm,addr);
1629         if (!vma)
1630                 return NULL;
1631         if (vma->vm_start <= addr)
1632                 return vma;
1633         if (!(vma->vm_flags & VM_GROWSDOWN))
1634                 return NULL;
1635         start = vma->vm_start;
1636         if (expand_stack(vma, addr))
1637                 return NULL;
1638         if (vma->vm_flags & VM_LOCKED) {
1639                 make_pages_present(addr, start);
1640         }
1641         return vma;
1642 }
1643 #endif
1644
1645 /*
1646  * Ok - we have the memory areas we should free on the vma list,
1647  * so release them, and do the vma updates.
1648  *
1649  * Called with the mm semaphore held.
1650  */
1651 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1652 {
1653         /* Update high watermark before we lower total_vm */
1654         update_hiwater_vm(mm);
1655         do {
1656                 long nrpages = vma_pages(vma);
1657
1658                 mm->total_vm -= nrpages;
1659                 if (vma->vm_flags & VM_LOCKED)
1660                         mm->locked_vm -= nrpages;
1661                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1662                 vma = remove_vma(vma);
1663         } while (vma);
1664         validate_mm(mm);
1665 }
1666
1667 /*
1668  * Get rid of page table information in the indicated region.
1669  *
1670  * Called with the mm semaphore held.
1671  */
1672 static void unmap_region(struct mm_struct *mm,
1673                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1674                 unsigned long start, unsigned long end)
1675 {
1676         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1677         struct mmu_gather *tlb;
1678         unsigned long nr_accounted = 0;
1679
1680         lru_add_drain();
1681         tlb = tlb_gather_mmu(mm, 0);
1682         update_hiwater_rss(mm);
1683         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1684         vm_unacct_memory(nr_accounted);
1685         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1686                                  next? next->vm_start: 0);
1687         tlb_finish_mmu(tlb, start, end);
1688 }
1689
1690 /*
1691  * Create a list of vma's touched by the unmap, removing them from the mm's
1692  * vma list as we go..
1693  */
1694 static void
1695 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1696         struct vm_area_struct *prev, unsigned long end)
1697 {
1698         struct vm_area_struct **insertion_point;
1699         struct vm_area_struct *tail_vma = NULL;
1700         unsigned long addr;
1701
1702         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1703         do {
1704                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1705                 mm->map_count--;
1706                 tail_vma = vma;
1707                 vma = vma->vm_next;
1708         } while (vma && vma->vm_start < end);
1709         *insertion_point = vma;
1710         tail_vma->vm_next = NULL;
1711         if (mm->unmap_area == arch_unmap_area)
1712                 addr = prev ? prev->vm_end : mm->mmap_base;
1713         else
1714                 addr = vma ?  vma->vm_start : mm->mmap_base;
1715         mm->unmap_area(mm, addr);
1716         mm->mmap_cache = NULL;          /* Kill the cache. */
1717 }
1718
1719 /*
1720  * Split a vma into two pieces at address 'addr', a new vma is allocated
1721  * either for the first part or the the tail.
1722  */
1723 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1724               unsigned long addr, int new_below)
1725 {
1726         struct mempolicy *pol;
1727         struct vm_area_struct *new;
1728
1729         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1730                 return -EINVAL;
1731
1732         if (mm->map_count >= sysctl_max_map_count)
1733                 return -ENOMEM;
1734
1735         new = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1736         if (!new)
1737                 return -ENOMEM;
1738
1739         /* most fields are the same, copy all, and then fixup */
1740         *new = *vma;
1741
1742         if (new_below)
1743                 new->vm_end = addr;
1744         else {
1745                 new->vm_start = addr;
1746                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1747         }
1748
1749         pol = mpol_copy(vma_policy(vma));
1750         if (IS_ERR(pol)) {
1751                 kmem_cache_free(vm_area_cachep, new);
1752                 return PTR_ERR(pol);
1753         }
1754         vma_set_policy(new, pol);
1755
1756         if (new->vm_file)
1757                 get_file(new->vm_file);
1758
1759         if (new->vm_ops && new->vm_ops->open)
1760                 new->vm_ops->open(new);
1761
1762         if (new_below)
1763                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1764                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1765         else
1766                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1767
1768         return 0;
1769 }
1770
1771 /* Munmap is split into 2 main parts -- this part which finds
1772  * what needs doing, and the areas themselves, which do the
1773  * work.  This now handles partial unmappings.
1774  * Jeremy Fitzhardinge <jeremy@goop.org>
1775  */
1776 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1777 {
1778         unsigned long end;
1779         struct vm_area_struct *vma, *prev, *last;
1780
1781         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1782                 return -EINVAL;
1783
1784         if ((len = PAGE_ALIGN(len)) == 0)
1785                 return -EINVAL;
1786
1787         /* Find the first overlapping VMA */
1788         vma = find_vma_prev(mm, start, &prev);
1789         if (!vma)
1790                 return 0;
1791         /* we have  start < vma->vm_end  */
1792
1793         /* if it doesn't overlap, we have nothing.. */
1794         end = start + len;
1795         if (vma->vm_start >= end)
1796                 return 0;
1797
1798         /*
1799          * If we need to split any vma, do it now to save pain later.
1800          *
1801          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1802          * unmapped vm_area_struct will remain in use: so lower split_vma
1803          * places tmp vma above, and higher split_vma places tmp vma below.
1804          */
1805         if (start > vma->vm_start) {
1806                 int error = split_vma(mm, vma, start, 0);
1807                 if (error)
1808                         return error;
1809                 prev = vma;
1810         }
1811
1812         /* Does it split the last one? */
1813         last = find_vma(mm, end);
1814         if (last && end > last->vm_start) {
1815                 int error = split_vma(mm, last, end, 1);
1816                 if (error)
1817                         return error;
1818         }
1819         vma = prev? prev->vm_next: mm->mmap;
1820
1821         /*
1822          * Remove the vma's, and unmap the actual pages
1823          */
1824         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1825         unmap_region(mm, vma, prev, start, end);
1826
1827         /* Fix up all other VM information */
1828         remove_vma_list(mm, vma);
1829
1830         return 0;
1831 }
1832
1833 EXPORT_SYMBOL(do_munmap);
1834
1835 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1836 {
1837         int ret;
1838         struct mm_struct *mm = current->mm;
1839
1840         profile_munmap(addr);
1841
1842         down_write(&mm->mmap_sem);
1843         ret = do_munmap(mm, addr, len);
1844         up_write(&mm->mmap_sem);
1845         return ret;
1846 }
1847
1848 static inline void verify_mm_writelocked(struct mm_struct *mm)
1849 {
1850 #ifdef CONFIG_DEBUG_VM
1851         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1852                 WARN_ON(1);
1853                 up_read(&mm->mmap_sem);
1854         }
1855 #endif
1856 }
1857
1858 /*
1859  *  this is really a simplified "do_mmap".  it only handles
1860  *  anonymous maps.  eventually we may be able to do some
1861  *  brk-specific accounting here.
1862  */
1863 unsigned long do_brk(unsigned long addr, unsigned long len)
1864 {
1865         struct mm_struct * mm = current->mm;
1866         struct vm_area_struct * vma, * prev;
1867         unsigned long flags;
1868         struct rb_node ** rb_link, * rb_parent;
1869         pgoff_t pgoff = addr >> PAGE_SHIFT;
1870         int error;
1871
1872         len = PAGE_ALIGN(len);
1873         if (!len)
1874                 return addr;
1875
1876         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1877                 return -EINVAL;
1878
1879         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1880
1881         error = arch_mmap_check(addr, len, flags);
1882         if (error)
1883                 return error;
1884
1885         /*
1886          * mlock MCL_FUTURE?
1887          */
1888         if (mm->def_flags & VM_LOCKED) {
1889                 unsigned long locked, lock_limit;
1890                 locked = len >> PAGE_SHIFT;
1891                 locked += mm->locked_vm;
1892                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1893                 lock_limit >>= PAGE_SHIFT;
1894                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1895                         return -EAGAIN;
1896         }
1897
1898         /*
1899          * mm->mmap_sem is required to protect against another thread
1900          * changing the mappings in case we sleep.
1901          */
1902         verify_mm_writelocked(mm);
1903
1904         /*
1905          * Clear old maps.  this also does some error checking for us
1906          */
1907  munmap_back:
1908         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1909         if (vma && vma->vm_start < addr + len) {
1910                 if (do_munmap(mm, addr, len))
1911                         return -ENOMEM;
1912                 goto munmap_back;
1913         }
1914
1915         /* Check against address space limits *after* clearing old maps... */
1916         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1917                 return -ENOMEM;
1918
1919         if (mm->map_count > sysctl_max_map_count)
1920                 return -ENOMEM;
1921
1922         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1923                 return -ENOMEM;
1924
1925         /* Can we just expand an old private anonymous mapping? */
1926         if (vma_merge(mm, prev, addr, addr + len, flags,
1927                                         NULL, NULL, pgoff, NULL))
1928                 goto out;
1929
1930         /*
1931          * create a vma struct for an anonymous mapping
1932          */
1933         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1934         if (!vma) {
1935                 vm_unacct_memory(len >> PAGE_SHIFT);
1936                 return -ENOMEM;
1937         }
1938
1939         vma->vm_mm = mm;
1940         vma->vm_start = addr;
1941         vma->vm_end = addr + len;
1942         vma->vm_pgoff = pgoff;
1943         vma->vm_flags = flags;
1944         vma->vm_page_prot = protection_map[flags &
1945                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1946         vma_link(mm, vma, prev, rb_link, rb_parent);
1947 out:
1948         mm->total_vm += len >> PAGE_SHIFT;
1949         if (flags & VM_LOCKED) {
1950                 mm->locked_vm += len >> PAGE_SHIFT;
1951                 make_pages_present(addr, addr + len);
1952         }
1953         return addr;
1954 }
1955
1956 EXPORT_SYMBOL(do_brk);
1957
1958 /* Release all mmaps. */
1959 void exit_mmap(struct mm_struct *mm)
1960 {
1961         struct mmu_gather *tlb;
1962         struct vm_area_struct *vma = mm->mmap;
1963         unsigned long nr_accounted = 0;
1964         unsigned long end;
1965
1966         lru_add_drain();
1967         flush_cache_mm(mm);
1968         tlb = tlb_gather_mmu(mm, 1);
1969         /* Don't update_hiwater_rss(mm) here, do_exit already did */
1970         /* Use -1 here to ensure all VMAs in the mm are unmapped */
1971         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
1972         vm_unacct_memory(nr_accounted);
1973         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
1974         tlb_finish_mmu(tlb, 0, end);
1975
1976         /*
1977          * Walk the list again, actually closing and freeing it,
1978          * with preemption enabled, without holding any MM locks.
1979          */
1980         while (vma)
1981                 vma = remove_vma(vma);
1982
1983         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
1984 }
1985
1986 /* Insert vm structure into process list sorted by address
1987  * and into the inode's i_mmap tree.  If vm_file is non-NULL
1988  * then i_mmap_lock is taken here.
1989  */
1990 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
1991 {
1992         struct vm_area_struct * __vma, * prev;
1993         struct rb_node ** rb_link, * rb_parent;
1994
1995         /*
1996          * The vm_pgoff of a purely anonymous vma should be irrelevant
1997          * until its first write fault, when page's anon_vma and index
1998          * are set.  But now set the vm_pgoff it will almost certainly
1999          * end up with (unless mremap moves it elsewhere before that
2000          * first wfault), so /proc/pid/maps tells a consistent story.
2001          *
2002          * By setting it to reflect the virtual start address of the
2003          * vma, merges and splits can happen in a seamless way, just
2004          * using the existing file pgoff checks and manipulations.
2005          * Similarly in do_mmap_pgoff and in do_brk.
2006          */
2007         if (!vma->vm_file) {
2008                 BUG_ON(vma->anon_vma);
2009                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2010         }
2011         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2012         if (__vma && __vma->vm_start < vma->vm_end)
2013                 return -ENOMEM;
2014         if ((vma->vm_flags & VM_ACCOUNT) &&
2015              security_vm_enough_memory(vma_pages(vma)))
2016                 return -ENOMEM;
2017         vma_link(mm, vma, prev, rb_link, rb_parent);
2018         return 0;
2019 }
2020
2021 /*
2022  * Copy the vma structure to a new location in the same mm,
2023  * prior to moving page table entries, to effect an mremap move.
2024  */
2025 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2026         unsigned long addr, unsigned long len, pgoff_t pgoff)
2027 {
2028         struct vm_area_struct *vma = *vmap;
2029         unsigned long vma_start = vma->vm_start;
2030         struct mm_struct *mm = vma->vm_mm;
2031         struct vm_area_struct *new_vma, *prev;
2032         struct rb_node **rb_link, *rb_parent;
2033         struct mempolicy *pol;
2034
2035         /*
2036          * If anonymous vma has not yet been faulted, update new pgoff
2037          * to match new location, to increase its chance of merging.
2038          */
2039         if (!vma->vm_file && !vma->anon_vma)
2040                 pgoff = addr >> PAGE_SHIFT;
2041
2042         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2043         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2044                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2045         if (new_vma) {
2046                 /*
2047                  * Source vma may have been merged into new_vma
2048                  */
2049                 if (vma_start >= new_vma->vm_start &&
2050                     vma_start < new_vma->vm_end)
2051                         *vmap = new_vma;
2052         } else {
2053                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2054                 if (new_vma) {
2055                         *new_vma = *vma;
2056                         pol = mpol_copy(vma_policy(vma));
2057                         if (IS_ERR(pol)) {
2058                                 kmem_cache_free(vm_area_cachep, new_vma);
2059                                 return NULL;
2060                         }
2061                         vma_set_policy(new_vma, pol);
2062                         new_vma->vm_start = addr;
2063                         new_vma->vm_end = addr + len;
2064                         new_vma->vm_pgoff = pgoff;
2065                         if (new_vma->vm_file)
2066                                 get_file(new_vma->vm_file);
2067                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2068                                 new_vma->vm_ops->open(new_vma);
2069                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2070                 }
2071         }
2072         return new_vma;
2073 }
2074
2075 /*
2076  * Return true if the calling process may expand its vm space by the passed
2077  * number of pages
2078  */
2079 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2080 {
2081         unsigned long cur = mm->total_vm;       /* pages */
2082         unsigned long lim;
2083
2084         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2085
2086         if (cur + npages > lim)
2087                 return 0;
2088         return 1;
2089 }