security: Protection for exploiting null dereference using mmap
[linux-2.6.git] / 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 #include <asm/mmu_context.h>
33
34 #ifndef arch_mmap_check
35 #define arch_mmap_check(addr, len, flags)       (0)
36 #endif
37
38 static void unmap_region(struct mm_struct *mm,
39                 struct vm_area_struct *vma, struct vm_area_struct *prev,
40                 unsigned long start, unsigned long end);
41
42 /*
43  * WARNING: the debugging will use recursive algorithms so never enable this
44  * unless you know what you are doing.
45  */
46 #undef DEBUG_MM_RB
47
48 /* description of effects of mapping type and prot in current implementation.
49  * this is due to the limited x86 page protection hardware.  The expected
50  * behavior is in parens:
51  *
52  * map_type     prot
53  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
54  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
55  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
56  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
57  *              
58  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
59  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
60  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
61  *
62  */
63 pgprot_t protection_map[16] = {
64         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
65         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
66 };
67
68 pgprot_t vm_get_page_prot(unsigned long vm_flags)
69 {
70         return protection_map[vm_flags &
71                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
72 }
73 EXPORT_SYMBOL(vm_get_page_prot);
74
75 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
76 int sysctl_overcommit_ratio = 50;       /* default is 50% */
77 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
78 atomic_t vm_committed_space = ATOMIC_INIT(0);
79
80 /*
81  * Check that a process has enough memory to allocate a new virtual
82  * mapping. 0 means there is enough memory for the allocation to
83  * succeed and -ENOMEM implies there is not.
84  *
85  * We currently support three overcommit policies, which are set via the
86  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
87  *
88  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
89  * Additional code 2002 Jul 20 by Robert Love.
90  *
91  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
92  *
93  * Note this is a helper function intended to be used by LSMs which
94  * wish to use this logic.
95  */
96 int __vm_enough_memory(long pages, int cap_sys_admin)
97 {
98         unsigned long free, allowed;
99
100         vm_acct_memory(pages);
101
102         /*
103          * Sometimes we want to use more memory than we have
104          */
105         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
106                 return 0;
107
108         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
109                 unsigned long n;
110
111                 free = global_page_state(NR_FILE_PAGES);
112                 free += nr_swap_pages;
113
114                 /*
115                  * Any slabs which are created with the
116                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
117                  * which are reclaimable, under pressure.  The dentry
118                  * cache and most inode caches should fall into this
119                  */
120                 free += global_page_state(NR_SLAB_RECLAIMABLE);
121
122                 /*
123                  * Leave the last 3% for root
124                  */
125                 if (!cap_sys_admin)
126                         free -= free / 32;
127
128                 if (free > pages)
129                         return 0;
130
131                 /*
132                  * nr_free_pages() is very expensive on large systems,
133                  * only call if we're about to fail.
134                  */
135                 n = nr_free_pages();
136
137                 /*
138                  * Leave reserved pages. The pages are not for anonymous pages.
139                  */
140                 if (n <= totalreserve_pages)
141                         goto error;
142                 else
143                         n -= totalreserve_pages;
144
145                 /*
146                  * Leave the last 3% for root
147                  */
148                 if (!cap_sys_admin)
149                         n -= n / 32;
150                 free += n;
151
152                 if (free > pages)
153                         return 0;
154
155                 goto error;
156         }
157
158         allowed = (totalram_pages - hugetlb_total_pages())
159                 * sysctl_overcommit_ratio / 100;
160         /*
161          * Leave the last 3% for root
162          */
163         if (!cap_sys_admin)
164                 allowed -= allowed / 32;
165         allowed += total_swap_pages;
166
167         /* Don't let a single process grow too big:
168            leave 3% of the size of this process for other processes */
169         allowed -= current->mm->total_vm / 32;
170
171         /*
172          * cast `allowed' as a signed long because vm_committed_space
173          * sometimes has a negative value
174          */
175         if (atomic_read(&vm_committed_space) < (long)allowed)
176                 return 0;
177 error:
178         vm_unacct_memory(pages);
179
180         return -ENOMEM;
181 }
182
183 EXPORT_SYMBOL(__vm_enough_memory);
184
185 /*
186  * Requires inode->i_mapping->i_mmap_lock
187  */
188 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
189                 struct file *file, struct address_space *mapping)
190 {
191         if (vma->vm_flags & VM_DENYWRITE)
192                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
193         if (vma->vm_flags & VM_SHARED)
194                 mapping->i_mmap_writable--;
195
196         flush_dcache_mmap_lock(mapping);
197         if (unlikely(vma->vm_flags & VM_NONLINEAR))
198                 list_del_init(&vma->shared.vm_set.list);
199         else
200                 vma_prio_tree_remove(vma, &mapping->i_mmap);
201         flush_dcache_mmap_unlock(mapping);
202 }
203
204 /*
205  * Unlink a file-based vm structure from its prio_tree, to hide
206  * vma from rmap and vmtruncate before freeing its page tables.
207  */
208 void unlink_file_vma(struct vm_area_struct *vma)
209 {
210         struct file *file = vma->vm_file;
211
212         if (file) {
213                 struct address_space *mapping = file->f_mapping;
214                 spin_lock(&mapping->i_mmap_lock);
215                 __remove_shared_vm_struct(vma, file, mapping);
216                 spin_unlock(&mapping->i_mmap_lock);
217         }
218 }
219
220 /*
221  * Close a vm structure and free it, returning the next.
222  */
223 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
224 {
225         struct vm_area_struct *next = vma->vm_next;
226
227         might_sleep();
228         if (vma->vm_ops && vma->vm_ops->close)
229                 vma->vm_ops->close(vma);
230         if (vma->vm_file)
231                 fput(vma->vm_file);
232         mpol_free(vma_policy(vma));
233         kmem_cache_free(vm_area_cachep, vma);
234         return next;
235 }
236
237 asmlinkage unsigned long sys_brk(unsigned long brk)
238 {
239         unsigned long rlim, retval;
240         unsigned long newbrk, oldbrk;
241         struct mm_struct *mm = current->mm;
242
243         down_write(&mm->mmap_sem);
244
245         if (brk < mm->end_code)
246                 goto out;
247
248         /*
249          * Check against rlimit here. If this check is done later after the test
250          * of oldbrk with newbrk then it can escape the test and let the data
251          * segment grow beyond its set limit the in case where the limit is
252          * not page aligned -Ram Gupta
253          */
254         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
255         if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
256                 goto out;
257
258         newbrk = PAGE_ALIGN(brk);
259         oldbrk = PAGE_ALIGN(mm->brk);
260         if (oldbrk == newbrk)
261                 goto set_brk;
262
263         /* Always allow shrinking brk. */
264         if (brk <= mm->brk) {
265                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
266                         goto set_brk;
267                 goto out;
268         }
269
270         /* Check against existing mmap mappings. */
271         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
272                 goto out;
273
274         /* Ok, looks good - let it rip. */
275         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
276                 goto out;
277 set_brk:
278         mm->brk = brk;
279 out:
280         retval = mm->brk;
281         up_write(&mm->mmap_sem);
282         return retval;
283 }
284
285 #ifdef DEBUG_MM_RB
286 static int browse_rb(struct rb_root *root)
287 {
288         int i = 0, j;
289         struct rb_node *nd, *pn = NULL;
290         unsigned long prev = 0, pend = 0;
291
292         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
293                 struct vm_area_struct *vma;
294                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
295                 if (vma->vm_start < prev)
296                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
297                 if (vma->vm_start < pend)
298                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
299                 if (vma->vm_start > vma->vm_end)
300                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
301                 i++;
302                 pn = nd;
303                 prev = vma->vm_start;
304                 pend = vma->vm_end;
305         }
306         j = 0;
307         for (nd = pn; nd; nd = rb_prev(nd)) {
308                 j++;
309         }
310         if (i != j)
311                 printk("backwards %d, forwards %d\n", j, i), i = 0;
312         return i;
313 }
314
315 void validate_mm(struct mm_struct *mm)
316 {
317         int bug = 0;
318         int i = 0;
319         struct vm_area_struct *tmp = mm->mmap;
320         while (tmp) {
321                 tmp = tmp->vm_next;
322                 i++;
323         }
324         if (i != mm->map_count)
325                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
326         i = browse_rb(&mm->mm_rb);
327         if (i != mm->map_count)
328                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
329         BUG_ON(bug);
330 }
331 #else
332 #define validate_mm(mm) do { } while (0)
333 #endif
334
335 static struct vm_area_struct *
336 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
337                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
338                 struct rb_node ** rb_parent)
339 {
340         struct vm_area_struct * vma;
341         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
342
343         __rb_link = &mm->mm_rb.rb_node;
344         rb_prev = __rb_parent = NULL;
345         vma = NULL;
346
347         while (*__rb_link) {
348                 struct vm_area_struct *vma_tmp;
349
350                 __rb_parent = *__rb_link;
351                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
352
353                 if (vma_tmp->vm_end > addr) {
354                         vma = vma_tmp;
355                         if (vma_tmp->vm_start <= addr)
356                                 return vma;
357                         __rb_link = &__rb_parent->rb_left;
358                 } else {
359                         rb_prev = __rb_parent;
360                         __rb_link = &__rb_parent->rb_right;
361                 }
362         }
363
364         *pprev = NULL;
365         if (rb_prev)
366                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
367         *rb_link = __rb_link;
368         *rb_parent = __rb_parent;
369         return vma;
370 }
371
372 static inline void
373 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
374                 struct vm_area_struct *prev, struct rb_node *rb_parent)
375 {
376         if (prev) {
377                 vma->vm_next = prev->vm_next;
378                 prev->vm_next = vma;
379         } else {
380                 mm->mmap = vma;
381                 if (rb_parent)
382                         vma->vm_next = rb_entry(rb_parent,
383                                         struct vm_area_struct, vm_rb);
384                 else
385                         vma->vm_next = NULL;
386         }
387 }
388
389 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
390                 struct rb_node **rb_link, struct rb_node *rb_parent)
391 {
392         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
393         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
394 }
395
396 static inline void __vma_link_file(struct vm_area_struct *vma)
397 {
398         struct file * file;
399
400         file = vma->vm_file;
401         if (file) {
402                 struct address_space *mapping = file->f_mapping;
403
404                 if (vma->vm_flags & VM_DENYWRITE)
405                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
406                 if (vma->vm_flags & VM_SHARED)
407                         mapping->i_mmap_writable++;
408
409                 flush_dcache_mmap_lock(mapping);
410                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
411                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
412                 else
413                         vma_prio_tree_insert(vma, &mapping->i_mmap);
414                 flush_dcache_mmap_unlock(mapping);
415         }
416 }
417
418 static void
419 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
420         struct vm_area_struct *prev, struct rb_node **rb_link,
421         struct rb_node *rb_parent)
422 {
423         __vma_link_list(mm, vma, prev, rb_parent);
424         __vma_link_rb(mm, vma, rb_link, rb_parent);
425         __anon_vma_link(vma);
426 }
427
428 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
429                         struct vm_area_struct *prev, struct rb_node **rb_link,
430                         struct rb_node *rb_parent)
431 {
432         struct address_space *mapping = NULL;
433
434         if (vma->vm_file)
435                 mapping = vma->vm_file->f_mapping;
436
437         if (mapping) {
438                 spin_lock(&mapping->i_mmap_lock);
439                 vma->vm_truncate_count = mapping->truncate_count;
440         }
441         anon_vma_lock(vma);
442
443         __vma_link(mm, vma, prev, rb_link, rb_parent);
444         __vma_link_file(vma);
445
446         anon_vma_unlock(vma);
447         if (mapping)
448                 spin_unlock(&mapping->i_mmap_lock);
449
450         mm->map_count++;
451         validate_mm(mm);
452 }
453
454 /*
455  * Helper for vma_adjust in the split_vma insert case:
456  * insert vm structure into list and rbtree and anon_vma,
457  * but it has already been inserted into prio_tree earlier.
458  */
459 static void
460 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
461 {
462         struct vm_area_struct * __vma, * prev;
463         struct rb_node ** rb_link, * rb_parent;
464
465         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467         __vma_link(mm, vma, prev, rb_link, rb_parent);
468         mm->map_count++;
469 }
470
471 static inline void
472 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473                 struct vm_area_struct *prev)
474 {
475         prev->vm_next = vma->vm_next;
476         rb_erase(&vma->vm_rb, &mm->mm_rb);
477         if (mm->mmap_cache == vma)
478                 mm->mmap_cache = prev;
479 }
480
481 /*
482  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
483  * is already present in an i_mmap tree without adjusting the tree.
484  * The following helper function should be used when such adjustments
485  * are necessary.  The "insert" vma (if any) is to be inserted
486  * before we drop the necessary locks.
487  */
488 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
489         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
490 {
491         struct mm_struct *mm = vma->vm_mm;
492         struct vm_area_struct *next = vma->vm_next;
493         struct vm_area_struct *importer = NULL;
494         struct address_space *mapping = NULL;
495         struct prio_tree_root *root = NULL;
496         struct file *file = vma->vm_file;
497         struct anon_vma *anon_vma = NULL;
498         long adjust_next = 0;
499         int remove_next = 0;
500
501         if (next && !insert) {
502                 if (end >= next->vm_end) {
503                         /*
504                          * vma expands, overlapping all the next, and
505                          * perhaps the one after too (mprotect case 6).
506                          */
507 again:                  remove_next = 1 + (end > next->vm_end);
508                         end = next->vm_end;
509                         anon_vma = next->anon_vma;
510                         importer = vma;
511                 } else if (end > next->vm_start) {
512                         /*
513                          * vma expands, overlapping part of the next:
514                          * mprotect case 5 shifting the boundary up.
515                          */
516                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
517                         anon_vma = next->anon_vma;
518                         importer = vma;
519                 } else if (end < vma->vm_end) {
520                         /*
521                          * vma shrinks, and !insert tells it's not
522                          * split_vma inserting another: so it must be
523                          * mprotect case 4 shifting the boundary down.
524                          */
525                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
526                         anon_vma = next->anon_vma;
527                         importer = next;
528                 }
529         }
530
531         if (file) {
532                 mapping = file->f_mapping;
533                 if (!(vma->vm_flags & VM_NONLINEAR))
534                         root = &mapping->i_mmap;
535                 spin_lock(&mapping->i_mmap_lock);
536                 if (importer &&
537                     vma->vm_truncate_count != next->vm_truncate_count) {
538                         /*
539                          * unmap_mapping_range might be in progress:
540                          * ensure that the expanding vma is rescanned.
541                          */
542                         importer->vm_truncate_count = 0;
543                 }
544                 if (insert) {
545                         insert->vm_truncate_count = vma->vm_truncate_count;
546                         /*
547                          * Put into prio_tree now, so instantiated pages
548                          * are visible to arm/parisc __flush_dcache_page
549                          * throughout; but we cannot insert into address
550                          * space until vma start or end is updated.
551                          */
552                         __vma_link_file(insert);
553                 }
554         }
555
556         /*
557          * When changing only vma->vm_end, we don't really need
558          * anon_vma lock: but is that case worth optimizing out?
559          */
560         if (vma->anon_vma)
561                 anon_vma = vma->anon_vma;
562         if (anon_vma) {
563                 spin_lock(&anon_vma->lock);
564                 /*
565                  * Easily overlooked: when mprotect shifts the boundary,
566                  * make sure the expanding vma has anon_vma set if the
567                  * shrinking vma had, to cover any anon pages imported.
568                  */
569                 if (importer && !importer->anon_vma) {
570                         importer->anon_vma = anon_vma;
571                         __anon_vma_link(importer);
572                 }
573         }
574
575         if (root) {
576                 flush_dcache_mmap_lock(mapping);
577                 vma_prio_tree_remove(vma, root);
578                 if (adjust_next)
579                         vma_prio_tree_remove(next, root);
580         }
581
582         vma->vm_start = start;
583         vma->vm_end = end;
584         vma->vm_pgoff = pgoff;
585         if (adjust_next) {
586                 next->vm_start += adjust_next << PAGE_SHIFT;
587                 next->vm_pgoff += adjust_next;
588         }
589
590         if (root) {
591                 if (adjust_next)
592                         vma_prio_tree_insert(next, root);
593                 vma_prio_tree_insert(vma, root);
594                 flush_dcache_mmap_unlock(mapping);
595         }
596
597         if (remove_next) {
598                 /*
599                  * vma_merge has merged next into vma, and needs
600                  * us to remove next before dropping the locks.
601                  */
602                 __vma_unlink(mm, next, vma);
603                 if (file)
604                         __remove_shared_vm_struct(next, file, mapping);
605                 if (next->anon_vma)
606                         __anon_vma_merge(vma, next);
607         } else if (insert) {
608                 /*
609                  * split_vma has split insert from vma, and needs
610                  * us to insert it before dropping the locks
611                  * (it may either follow vma or precede it).
612                  */
613                 __insert_vm_struct(mm, insert);
614         }
615
616         if (anon_vma)
617                 spin_unlock(&anon_vma->lock);
618         if (mapping)
619                 spin_unlock(&mapping->i_mmap_lock);
620
621         if (remove_next) {
622                 if (file)
623                         fput(file);
624                 mm->map_count--;
625                 mpol_free(vma_policy(next));
626                 kmem_cache_free(vm_area_cachep, next);
627                 /*
628                  * In mprotect's case 6 (see comments on vma_merge),
629                  * we must remove another next too. It would clutter
630                  * up the code too much to do both in one go.
631                  */
632                 if (remove_next == 2) {
633                         next = vma->vm_next;
634                         goto again;
635                 }
636         }
637
638         validate_mm(mm);
639 }
640
641 /*
642  * If the vma has a ->close operation then the driver probably needs to release
643  * per-vma resources, so we don't attempt to merge those.
644  */
645 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
646
647 static inline int is_mergeable_vma(struct vm_area_struct *vma,
648                         struct file *file, unsigned long vm_flags)
649 {
650         if (vma->vm_flags != vm_flags)
651                 return 0;
652         if (vma->vm_file != file)
653                 return 0;
654         if (vma->vm_ops && vma->vm_ops->close)
655                 return 0;
656         return 1;
657 }
658
659 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
660                                         struct anon_vma *anon_vma2)
661 {
662         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
663 }
664
665 /*
666  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
667  * in front of (at a lower virtual address and file offset than) the vma.
668  *
669  * We cannot merge two vmas if they have differently assigned (non-NULL)
670  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
671  *
672  * We don't check here for the merged mmap wrapping around the end of pagecache
673  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
674  * wrap, nor mmaps which cover the final page at index -1UL.
675  */
676 static int
677 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
678         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
679 {
680         if (is_mergeable_vma(vma, file, vm_flags) &&
681             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
682                 if (vma->vm_pgoff == vm_pgoff)
683                         return 1;
684         }
685         return 0;
686 }
687
688 /*
689  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
690  * beyond (at a higher virtual address and file offset than) the vma.
691  *
692  * We cannot merge two vmas if they have differently assigned (non-NULL)
693  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
694  */
695 static int
696 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
697         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
698 {
699         if (is_mergeable_vma(vma, file, vm_flags) &&
700             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
701                 pgoff_t vm_pglen;
702                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
703                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
704                         return 1;
705         }
706         return 0;
707 }
708
709 /*
710  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
711  * whether that can be merged with its predecessor or its successor.
712  * Or both (it neatly fills a hole).
713  *
714  * In most cases - when called for mmap, brk or mremap - [addr,end) is
715  * certain not to be mapped by the time vma_merge is called; but when
716  * called for mprotect, it is certain to be already mapped (either at
717  * an offset within prev, or at the start of next), and the flags of
718  * this area are about to be changed to vm_flags - and the no-change
719  * case has already been eliminated.
720  *
721  * The following mprotect cases have to be considered, where AAAA is
722  * the area passed down from mprotect_fixup, never extending beyond one
723  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
724  *
725  *     AAAA             AAAA                AAAA          AAAA
726  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
727  *    cannot merge    might become    might become    might become
728  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
729  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
730  *    mremap move:                                    PPPPNNNNNNNN 8
731  *        AAAA
732  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
733  *    might become    case 1 below    case 2 below    case 3 below
734  *
735  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
736  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
737  */
738 struct vm_area_struct *vma_merge(struct mm_struct *mm,
739                         struct vm_area_struct *prev, unsigned long addr,
740                         unsigned long end, unsigned long vm_flags,
741                         struct anon_vma *anon_vma, struct file *file,
742                         pgoff_t pgoff, struct mempolicy *policy)
743 {
744         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
745         struct vm_area_struct *area, *next;
746
747         /*
748          * We later require that vma->vm_flags == vm_flags,
749          * so this tests vma->vm_flags & VM_SPECIAL, too.
750          */
751         if (vm_flags & VM_SPECIAL)
752                 return NULL;
753
754         if (prev)
755                 next = prev->vm_next;
756         else
757                 next = mm->mmap;
758         area = next;
759         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
760                 next = next->vm_next;
761
762         /*
763          * Can it merge with the predecessor?
764          */
765         if (prev && prev->vm_end == addr &&
766                         mpol_equal(vma_policy(prev), policy) &&
767                         can_vma_merge_after(prev, vm_flags,
768                                                 anon_vma, file, pgoff)) {
769                 /*
770                  * OK, it can.  Can we now merge in the successor as well?
771                  */
772                 if (next && end == next->vm_start &&
773                                 mpol_equal(policy, vma_policy(next)) &&
774                                 can_vma_merge_before(next, vm_flags,
775                                         anon_vma, file, pgoff+pglen) &&
776                                 is_mergeable_anon_vma(prev->anon_vma,
777                                                       next->anon_vma)) {
778                                                         /* cases 1, 6 */
779                         vma_adjust(prev, prev->vm_start,
780                                 next->vm_end, prev->vm_pgoff, NULL);
781                 } else                                  /* cases 2, 5, 7 */
782                         vma_adjust(prev, prev->vm_start,
783                                 end, prev->vm_pgoff, NULL);
784                 return prev;
785         }
786
787         /*
788          * Can this new request be merged in front of next?
789          */
790         if (next && end == next->vm_start &&
791                         mpol_equal(policy, vma_policy(next)) &&
792                         can_vma_merge_before(next, vm_flags,
793                                         anon_vma, file, pgoff+pglen)) {
794                 if (prev && addr < prev->vm_end)        /* case 4 */
795                         vma_adjust(prev, prev->vm_start,
796                                 addr, prev->vm_pgoff, NULL);
797                 else                                    /* cases 3, 8 */
798                         vma_adjust(area, addr, next->vm_end,
799                                 next->vm_pgoff - pglen, NULL);
800                 return area;
801         }
802
803         return NULL;
804 }
805
806 /*
807  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
808  * neighbouring vmas for a suitable anon_vma, before it goes off
809  * to allocate a new anon_vma.  It checks because a repetitive
810  * sequence of mprotects and faults may otherwise lead to distinct
811  * anon_vmas being allocated, preventing vma merge in subsequent
812  * mprotect.
813  */
814 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
815 {
816         struct vm_area_struct *near;
817         unsigned long vm_flags;
818
819         near = vma->vm_next;
820         if (!near)
821                 goto try_prev;
822
823         /*
824          * Since only mprotect tries to remerge vmas, match flags
825          * which might be mprotected into each other later on.
826          * Neither mlock nor madvise tries to remerge at present,
827          * so leave their flags as obstructing a merge.
828          */
829         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
830         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
831
832         if (near->anon_vma && vma->vm_end == near->vm_start &&
833                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
834                         can_vma_merge_before(near, vm_flags,
835                                 NULL, vma->vm_file, vma->vm_pgoff +
836                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
837                 return near->anon_vma;
838 try_prev:
839         /*
840          * It is potentially slow to have to call find_vma_prev here.
841          * But it's only on the first write fault on the vma, not
842          * every time, and we could devise a way to avoid it later
843          * (e.g. stash info in next's anon_vma_node when assigning
844          * an anon_vma, or when trying vma_merge).  Another time.
845          */
846         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
847         if (!near)
848                 goto none;
849
850         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
851         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
852
853         if (near->anon_vma && near->vm_end == vma->vm_start &&
854                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
855                         can_vma_merge_after(near, vm_flags,
856                                 NULL, vma->vm_file, vma->vm_pgoff))
857                 return near->anon_vma;
858 none:
859         /*
860          * There's no absolute need to look only at touching neighbours:
861          * we could search further afield for "compatible" anon_vmas.
862          * But it would probably just be a waste of time searching,
863          * or lead to too many vmas hanging off the same anon_vma.
864          * We're trying to allow mprotect remerging later on,
865          * not trying to minimize memory used for anon_vmas.
866          */
867         return NULL;
868 }
869
870 #ifdef CONFIG_PROC_FS
871 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
872                                                 struct file *file, long pages)
873 {
874         const unsigned long stack_flags
875                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
876
877         if (file) {
878                 mm->shared_vm += pages;
879                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
880                         mm->exec_vm += pages;
881         } else if (flags & stack_flags)
882                 mm->stack_vm += pages;
883         if (flags & (VM_RESERVED|VM_IO))
884                 mm->reserved_vm += pages;
885 }
886 #endif /* CONFIG_PROC_FS */
887
888 /*
889  * The caller must hold down_write(current->mm->mmap_sem).
890  */
891
892 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
893                         unsigned long len, unsigned long prot,
894                         unsigned long flags, unsigned long pgoff)
895 {
896         struct mm_struct * mm = current->mm;
897         struct vm_area_struct * vma, * prev;
898         struct inode *inode;
899         unsigned int vm_flags;
900         int correct_wcount = 0;
901         int error;
902         struct rb_node ** rb_link, * rb_parent;
903         int accountable = 1;
904         unsigned long charged = 0, reqprot = prot;
905
906         /*
907          * Does the application expect PROT_READ to imply PROT_EXEC?
908          *
909          * (the exception is when the underlying filesystem is noexec
910          *  mounted, in which case we dont add PROT_EXEC.)
911          */
912         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
913                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
914                         prot |= PROT_EXEC;
915
916         if (!len)
917                 return -EINVAL;
918
919         error = arch_mmap_check(addr, len, flags);
920         if (error)
921                 return error;
922
923         /* Careful about overflows.. */
924         len = PAGE_ALIGN(len);
925         if (!len || len > TASK_SIZE)
926                 return -ENOMEM;
927
928         /* offset overflow? */
929         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
930                return -EOVERFLOW;
931
932         /* Too many mappings? */
933         if (mm->map_count > sysctl_max_map_count)
934                 return -ENOMEM;
935
936         /* Obtain the address to map to. we verify (or select) it and ensure
937          * that it represents a valid section of the address space.
938          */
939         addr = get_unmapped_area(file, addr, len, pgoff, flags);
940         if (addr & ~PAGE_MASK)
941                 return addr;
942
943         /* Do simple checking here so the lower-level routines won't have
944          * to. we assume access permissions have been handled by the open
945          * of the memory object, so we don't do any here.
946          */
947         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
948                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
949
950         if (flags & MAP_LOCKED) {
951                 if (!can_do_mlock())
952                         return -EPERM;
953                 vm_flags |= VM_LOCKED;
954         }
955         /* mlock MCL_FUTURE? */
956         if (vm_flags & VM_LOCKED) {
957                 unsigned long locked, lock_limit;
958                 locked = len >> PAGE_SHIFT;
959                 locked += mm->locked_vm;
960                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
961                 lock_limit >>= PAGE_SHIFT;
962                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
963                         return -EAGAIN;
964         }
965
966         inode = file ? file->f_path.dentry->d_inode : NULL;
967
968         if (file) {
969                 switch (flags & MAP_TYPE) {
970                 case MAP_SHARED:
971                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
972                                 return -EACCES;
973
974                         /*
975                          * Make sure we don't allow writing to an append-only
976                          * file..
977                          */
978                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
979                                 return -EACCES;
980
981                         /*
982                          * Make sure there are no mandatory locks on the file.
983                          */
984                         if (locks_verify_locked(inode))
985                                 return -EAGAIN;
986
987                         vm_flags |= VM_SHARED | VM_MAYSHARE;
988                         if (!(file->f_mode & FMODE_WRITE))
989                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
990
991                         /* fall through */
992                 case MAP_PRIVATE:
993                         if (!(file->f_mode & FMODE_READ))
994                                 return -EACCES;
995                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
996                                 if (vm_flags & VM_EXEC)
997                                         return -EPERM;
998                                 vm_flags &= ~VM_MAYEXEC;
999                         }
1000                         if (is_file_hugepages(file))
1001                                 accountable = 0;
1002
1003                         if (!file->f_op || !file->f_op->mmap)
1004                                 return -ENODEV;
1005                         break;
1006
1007                 default:
1008                         return -EINVAL;
1009                 }
1010         } else {
1011                 switch (flags & MAP_TYPE) {
1012                 case MAP_SHARED:
1013                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1014                         break;
1015                 case MAP_PRIVATE:
1016                         /*
1017                          * Set pgoff according to addr for anon_vma.
1018                          */
1019                         pgoff = addr >> PAGE_SHIFT;
1020                         break;
1021                 default:
1022                         return -EINVAL;
1023                 }
1024         }
1025
1026         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1027         if (error)
1028                 return error;
1029
1030         /* Clear old maps */
1031         error = -ENOMEM;
1032 munmap_back:
1033         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1034         if (vma && vma->vm_start < addr + len) {
1035                 if (do_munmap(mm, addr, len))
1036                         return -ENOMEM;
1037                 goto munmap_back;
1038         }
1039
1040         /* Check against address space limit. */
1041         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1042                 return -ENOMEM;
1043
1044         if (accountable && (!(flags & MAP_NORESERVE) ||
1045                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1046                 if (vm_flags & VM_SHARED) {
1047                         /* Check memory availability in shmem_file_setup? */
1048                         vm_flags |= VM_ACCOUNT;
1049                 } else if (vm_flags & VM_WRITE) {
1050                         /*
1051                          * Private writable mapping: check memory availability
1052                          */
1053                         charged = len >> PAGE_SHIFT;
1054                         if (security_vm_enough_memory(charged))
1055                                 return -ENOMEM;
1056                         vm_flags |= VM_ACCOUNT;
1057                 }
1058         }
1059
1060         /*
1061          * Can we just expand an old private anonymous mapping?
1062          * The VM_SHARED test is necessary because shmem_zero_setup
1063          * will create the file object for a shared anonymous map below.
1064          */
1065         if (!file && !(vm_flags & VM_SHARED) &&
1066             vma_merge(mm, prev, addr, addr + len, vm_flags,
1067                                         NULL, NULL, pgoff, NULL))
1068                 goto out;
1069
1070         /*
1071          * Determine the object being mapped and call the appropriate
1072          * specific mapper. the address has already been validated, but
1073          * not unmapped, but the maps are removed from the list.
1074          */
1075         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1076         if (!vma) {
1077                 error = -ENOMEM;
1078                 goto unacct_error;
1079         }
1080
1081         vma->vm_mm = mm;
1082         vma->vm_start = addr;
1083         vma->vm_end = addr + len;
1084         vma->vm_flags = vm_flags;
1085         vma->vm_page_prot = protection_map[vm_flags &
1086                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1087         vma->vm_pgoff = pgoff;
1088
1089         if (file) {
1090                 error = -EINVAL;
1091                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1092                         goto free_vma;
1093                 if (vm_flags & VM_DENYWRITE) {
1094                         error = deny_write_access(file);
1095                         if (error)
1096                                 goto free_vma;
1097                         correct_wcount = 1;
1098                 }
1099                 vma->vm_file = file;
1100                 get_file(file);
1101                 error = file->f_op->mmap(file, vma);
1102                 if (error)
1103                         goto unmap_and_free_vma;
1104         } else if (vm_flags & VM_SHARED) {
1105                 error = shmem_zero_setup(vma);
1106                 if (error)
1107                         goto free_vma;
1108         }
1109
1110         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1111          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1112          * that memory reservation must be checked; but that reservation
1113          * belongs to shared memory object, not to vma: so now clear it.
1114          */
1115         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1116                 vma->vm_flags &= ~VM_ACCOUNT;
1117
1118         /* Can addr have changed??
1119          *
1120          * Answer: Yes, several device drivers can do it in their
1121          *         f_op->mmap method. -DaveM
1122          */
1123         addr = vma->vm_start;
1124         pgoff = vma->vm_pgoff;
1125         vm_flags = vma->vm_flags;
1126
1127         if (vma_wants_writenotify(vma))
1128                 vma->vm_page_prot =
1129                         protection_map[vm_flags & (VM_READ|VM_WRITE|VM_EXEC)];
1130
1131         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1132                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1133                 file = vma->vm_file;
1134                 vma_link(mm, vma, prev, rb_link, rb_parent);
1135                 if (correct_wcount)
1136                         atomic_inc(&inode->i_writecount);
1137         } else {
1138                 if (file) {
1139                         if (correct_wcount)
1140                                 atomic_inc(&inode->i_writecount);
1141                         fput(file);
1142                 }
1143                 mpol_free(vma_policy(vma));
1144                 kmem_cache_free(vm_area_cachep, vma);
1145         }
1146 out:    
1147         mm->total_vm += len >> PAGE_SHIFT;
1148         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1149         if (vm_flags & VM_LOCKED) {
1150                 mm->locked_vm += len >> PAGE_SHIFT;
1151                 make_pages_present(addr, addr + len);
1152         }
1153         if (flags & MAP_POPULATE) {
1154                 up_write(&mm->mmap_sem);
1155                 sys_remap_file_pages(addr, len, 0,
1156                                         pgoff, flags & MAP_NONBLOCK);
1157                 down_write(&mm->mmap_sem);
1158         }
1159         return addr;
1160
1161 unmap_and_free_vma:
1162         if (correct_wcount)
1163                 atomic_inc(&inode->i_writecount);
1164         vma->vm_file = NULL;
1165         fput(file);
1166
1167         /* Undo any partial mapping done by a device driver. */
1168         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1169         charged = 0;
1170 free_vma:
1171         kmem_cache_free(vm_area_cachep, vma);
1172 unacct_error:
1173         if (charged)
1174                 vm_unacct_memory(charged);
1175         return error;
1176 }
1177
1178 EXPORT_SYMBOL(do_mmap_pgoff);
1179
1180 /* Get an address range which is currently unmapped.
1181  * For shmat() with addr=0.
1182  *
1183  * Ugly calling convention alert:
1184  * Return value with the low bits set means error value,
1185  * ie
1186  *      if (ret & ~PAGE_MASK)
1187  *              error = ret;
1188  *
1189  * This function "knows" that -ENOMEM has the bits set.
1190  */
1191 #ifndef HAVE_ARCH_UNMAPPED_AREA
1192 unsigned long
1193 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1194                 unsigned long len, unsigned long pgoff, unsigned long flags)
1195 {
1196         struct mm_struct *mm = current->mm;
1197         struct vm_area_struct *vma;
1198         unsigned long start_addr;
1199
1200         if (len > TASK_SIZE)
1201                 return -ENOMEM;
1202
1203         if (flags & MAP_FIXED)
1204                 return addr;
1205
1206         if (addr) {
1207                 addr = PAGE_ALIGN(addr);
1208                 vma = find_vma(mm, addr);
1209                 if (TASK_SIZE - len >= addr &&
1210                     (!vma || addr + len <= vma->vm_start))
1211                         return addr;
1212         }
1213         if (len > mm->cached_hole_size) {
1214                 start_addr = addr = mm->free_area_cache;
1215         } else {
1216                 start_addr = addr = TASK_UNMAPPED_BASE;
1217                 mm->cached_hole_size = 0;
1218         }
1219
1220 full_search:
1221         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1222                 /* At this point:  (!vma || addr < vma->vm_end). */
1223                 if (TASK_SIZE - len < addr) {
1224                         /*
1225                          * Start a new search - just in case we missed
1226                          * some holes.
1227                          */
1228                         if (start_addr != TASK_UNMAPPED_BASE) {
1229                                 addr = TASK_UNMAPPED_BASE;
1230                                 start_addr = addr;
1231                                 mm->cached_hole_size = 0;
1232                                 goto full_search;
1233                         }
1234                         return -ENOMEM;
1235                 }
1236                 if (!vma || addr + len <= vma->vm_start) {
1237                         /*
1238                          * Remember the place where we stopped the search:
1239                          */
1240                         mm->free_area_cache = addr + len;
1241                         return addr;
1242                 }
1243                 if (addr + mm->cached_hole_size < vma->vm_start)
1244                         mm->cached_hole_size = vma->vm_start - addr;
1245                 addr = vma->vm_end;
1246         }
1247 }
1248 #endif  
1249
1250 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1251 {
1252         /*
1253          * Is this a new hole at the lowest possible address?
1254          */
1255         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1256                 mm->free_area_cache = addr;
1257                 mm->cached_hole_size = ~0UL;
1258         }
1259 }
1260
1261 /*
1262  * This mmap-allocator allocates new areas top-down from below the
1263  * stack's low limit (the base):
1264  */
1265 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1266 unsigned long
1267 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1268                           const unsigned long len, const unsigned long pgoff,
1269                           const unsigned long flags)
1270 {
1271         struct vm_area_struct *vma;
1272         struct mm_struct *mm = current->mm;
1273         unsigned long addr = addr0;
1274
1275         /* requested length too big for entire address space */
1276         if (len > TASK_SIZE)
1277                 return -ENOMEM;
1278
1279         if (flags & MAP_FIXED)
1280                 return addr;
1281
1282         /* requesting a specific address */
1283         if (addr) {
1284                 addr = PAGE_ALIGN(addr);
1285                 vma = find_vma(mm, addr);
1286                 if (TASK_SIZE - len >= addr &&
1287                                 (!vma || addr + len <= vma->vm_start))
1288                         return addr;
1289         }
1290
1291         /* check if free_area_cache is useful for us */
1292         if (len <= mm->cached_hole_size) {
1293                 mm->cached_hole_size = 0;
1294                 mm->free_area_cache = mm->mmap_base;
1295         }
1296
1297         /* either no address requested or can't fit in requested address hole */
1298         addr = mm->free_area_cache;
1299
1300         /* make sure it can fit in the remaining address space */
1301         if (addr > len) {
1302                 vma = find_vma(mm, addr-len);
1303                 if (!vma || addr <= vma->vm_start)
1304                         /* remember the address as a hint for next time */
1305                         return (mm->free_area_cache = addr-len);
1306         }
1307
1308         if (mm->mmap_base < len)
1309                 goto bottomup;
1310
1311         addr = mm->mmap_base-len;
1312
1313         do {
1314                 /*
1315                  * Lookup failure means no vma is above this address,
1316                  * else if new region fits below vma->vm_start,
1317                  * return with success:
1318                  */
1319                 vma = find_vma(mm, addr);
1320                 if (!vma || addr+len <= vma->vm_start)
1321                         /* remember the address as a hint for next time */
1322                         return (mm->free_area_cache = addr);
1323
1324                 /* remember the largest hole we saw so far */
1325                 if (addr + mm->cached_hole_size < vma->vm_start)
1326                         mm->cached_hole_size = vma->vm_start - addr;
1327
1328                 /* try just below the current vma->vm_start */
1329                 addr = vma->vm_start-len;
1330         } while (len < vma->vm_start);
1331
1332 bottomup:
1333         /*
1334          * A failed mmap() very likely causes application failure,
1335          * so fall back to the bottom-up function here. This scenario
1336          * can happen with large stack limits and large mmap()
1337          * allocations.
1338          */
1339         mm->cached_hole_size = ~0UL;
1340         mm->free_area_cache = TASK_UNMAPPED_BASE;
1341         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1342         /*
1343          * Restore the topdown base:
1344          */
1345         mm->free_area_cache = mm->mmap_base;
1346         mm->cached_hole_size = ~0UL;
1347
1348         return addr;
1349 }
1350 #endif
1351
1352 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1353 {
1354         /*
1355          * Is this a new hole at the highest possible address?
1356          */
1357         if (addr > mm->free_area_cache)
1358                 mm->free_area_cache = addr;
1359
1360         /* dont allow allocations above current base */
1361         if (mm->free_area_cache > mm->mmap_base)
1362                 mm->free_area_cache = mm->mmap_base;
1363 }
1364
1365 unsigned long
1366 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1367                 unsigned long pgoff, unsigned long flags)
1368 {
1369         unsigned long (*get_area)(struct file *, unsigned long,
1370                                   unsigned long, unsigned long, unsigned long);
1371
1372         get_area = current->mm->get_unmapped_area;
1373         if (file && file->f_op && file->f_op->get_unmapped_area)
1374                 get_area = file->f_op->get_unmapped_area;
1375         addr = get_area(file, addr, len, pgoff, flags);
1376         if (IS_ERR_VALUE(addr))
1377                 return addr;
1378
1379         if (addr > TASK_SIZE - len)
1380                 return -ENOMEM;
1381         if (addr & ~PAGE_MASK)
1382                 return -EINVAL;
1383
1384         return addr;
1385 }
1386
1387 EXPORT_SYMBOL(get_unmapped_area);
1388
1389 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1390 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1391 {
1392         struct vm_area_struct *vma = NULL;
1393
1394         if (mm) {
1395                 /* Check the cache first. */
1396                 /* (Cache hit rate is typically around 35%.) */
1397                 vma = mm->mmap_cache;
1398                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1399                         struct rb_node * rb_node;
1400
1401                         rb_node = mm->mm_rb.rb_node;
1402                         vma = NULL;
1403
1404                         while (rb_node) {
1405                                 struct vm_area_struct * vma_tmp;
1406
1407                                 vma_tmp = rb_entry(rb_node,
1408                                                 struct vm_area_struct, vm_rb);
1409
1410                                 if (vma_tmp->vm_end > addr) {
1411                                         vma = vma_tmp;
1412                                         if (vma_tmp->vm_start <= addr)
1413                                                 break;
1414                                         rb_node = rb_node->rb_left;
1415                                 } else
1416                                         rb_node = rb_node->rb_right;
1417                         }
1418                         if (vma)
1419                                 mm->mmap_cache = vma;
1420                 }
1421         }
1422         return vma;
1423 }
1424
1425 EXPORT_SYMBOL(find_vma);
1426
1427 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1428 struct vm_area_struct *
1429 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1430                         struct vm_area_struct **pprev)
1431 {
1432         struct vm_area_struct *vma = NULL, *prev = NULL;
1433         struct rb_node * rb_node;
1434         if (!mm)
1435                 goto out;
1436
1437         /* Guard against addr being lower than the first VMA */
1438         vma = mm->mmap;
1439
1440         /* Go through the RB tree quickly. */
1441         rb_node = mm->mm_rb.rb_node;
1442
1443         while (rb_node) {
1444                 struct vm_area_struct *vma_tmp;
1445                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1446
1447                 if (addr < vma_tmp->vm_end) {
1448                         rb_node = rb_node->rb_left;
1449                 } else {
1450                         prev = vma_tmp;
1451                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1452                                 break;
1453                         rb_node = rb_node->rb_right;
1454                 }
1455         }
1456
1457 out:
1458         *pprev = prev;
1459         return prev ? prev->vm_next : vma;
1460 }
1461
1462 /*
1463  * Verify that the stack growth is acceptable and
1464  * update accounting. This is shared with both the
1465  * grow-up and grow-down cases.
1466  */
1467 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1468 {
1469         struct mm_struct *mm = vma->vm_mm;
1470         struct rlimit *rlim = current->signal->rlim;
1471         unsigned long new_start;
1472
1473         /* address space limit tests */
1474         if (!may_expand_vm(mm, grow))
1475                 return -ENOMEM;
1476
1477         /* Stack limit test */
1478         if (size > rlim[RLIMIT_STACK].rlim_cur)
1479                 return -ENOMEM;
1480
1481         /* mlock limit tests */
1482         if (vma->vm_flags & VM_LOCKED) {
1483                 unsigned long locked;
1484                 unsigned long limit;
1485                 locked = mm->locked_vm + grow;
1486                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1487                 if (locked > limit && !capable(CAP_IPC_LOCK))
1488                         return -ENOMEM;
1489         }
1490
1491         /* Check to ensure the stack will not grow into a hugetlb-only region */
1492         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1493                         vma->vm_end - size;
1494         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1495                 return -EFAULT;
1496
1497         /*
1498          * Overcommit..  This must be the final test, as it will
1499          * update security statistics.
1500          */
1501         if (security_vm_enough_memory(grow))
1502                 return -ENOMEM;
1503
1504         /* Ok, everything looks good - let it rip */
1505         mm->total_vm += grow;
1506         if (vma->vm_flags & VM_LOCKED)
1507                 mm->locked_vm += grow;
1508         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1509         return 0;
1510 }
1511
1512 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1513 /*
1514  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1515  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1516  */
1517 #ifndef CONFIG_IA64
1518 static inline
1519 #endif
1520 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1521 {
1522         int error;
1523
1524         if (!(vma->vm_flags & VM_GROWSUP))
1525                 return -EFAULT;
1526
1527         /*
1528          * We must make sure the anon_vma is allocated
1529          * so that the anon_vma locking is not a noop.
1530          */
1531         if (unlikely(anon_vma_prepare(vma)))
1532                 return -ENOMEM;
1533         anon_vma_lock(vma);
1534
1535         /*
1536          * vma->vm_start/vm_end cannot change under us because the caller
1537          * is required to hold the mmap_sem in read mode.  We need the
1538          * anon_vma lock to serialize against concurrent expand_stacks.
1539          * Also guard against wrapping around to address 0.
1540          */
1541         if (address < PAGE_ALIGN(address+4))
1542                 address = PAGE_ALIGN(address+4);
1543         else {
1544                 anon_vma_unlock(vma);
1545                 return -ENOMEM;
1546         }
1547         error = 0;
1548
1549         /* Somebody else might have raced and expanded it already */
1550         if (address > vma->vm_end) {
1551                 unsigned long size, grow;
1552
1553                 size = address - vma->vm_start;
1554                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1555
1556                 error = acct_stack_growth(vma, size, grow);
1557                 if (!error)
1558                         vma->vm_end = address;
1559         }
1560         anon_vma_unlock(vma);
1561         return error;
1562 }
1563 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1564
1565 #ifdef CONFIG_STACK_GROWSUP
1566 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1567 {
1568         return expand_upwards(vma, address);
1569 }
1570
1571 struct vm_area_struct *
1572 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1573 {
1574         struct vm_area_struct *vma, *prev;
1575
1576         addr &= PAGE_MASK;
1577         vma = find_vma_prev(mm, addr, &prev);
1578         if (vma && (vma->vm_start <= addr))
1579                 return vma;
1580         if (!prev || expand_stack(prev, addr))
1581                 return NULL;
1582         if (prev->vm_flags & VM_LOCKED) {
1583                 make_pages_present(addr, prev->vm_end);
1584         }
1585         return prev;
1586 }
1587 #else
1588 /*
1589  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1590  */
1591 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1592 {
1593         int error;
1594
1595         /*
1596          * We must make sure the anon_vma is allocated
1597          * so that the anon_vma locking is not a noop.
1598          */
1599         if (unlikely(anon_vma_prepare(vma)))
1600                 return -ENOMEM;
1601         anon_vma_lock(vma);
1602
1603         /*
1604          * vma->vm_start/vm_end cannot change under us because the caller
1605          * is required to hold the mmap_sem in read mode.  We need the
1606          * anon_vma lock to serialize against concurrent expand_stacks.
1607          */
1608         address &= PAGE_MASK;
1609         error = 0;
1610
1611         /* Somebody else might have raced and expanded it already */
1612         if (address < vma->vm_start) {
1613                 unsigned long size, grow;
1614
1615                 size = vma->vm_end - address;
1616                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1617
1618                 error = acct_stack_growth(vma, size, grow);
1619                 if (!error) {
1620                         vma->vm_start = address;
1621                         vma->vm_pgoff -= grow;
1622                 }
1623         }
1624         anon_vma_unlock(vma);
1625         return error;
1626 }
1627
1628 struct vm_area_struct *
1629 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1630 {
1631         struct vm_area_struct * vma;
1632         unsigned long start;
1633
1634         addr &= PAGE_MASK;
1635         vma = find_vma(mm,addr);
1636         if (!vma)
1637                 return NULL;
1638         if (vma->vm_start <= addr)
1639                 return vma;
1640         if (!(vma->vm_flags & VM_GROWSDOWN))
1641                 return NULL;
1642         start = vma->vm_start;
1643         if (expand_stack(vma, addr))
1644                 return NULL;
1645         if (vma->vm_flags & VM_LOCKED) {
1646                 make_pages_present(addr, start);
1647         }
1648         return vma;
1649 }
1650 #endif
1651
1652 /*
1653  * Ok - we have the memory areas we should free on the vma list,
1654  * so release them, and do the vma updates.
1655  *
1656  * Called with the mm semaphore held.
1657  */
1658 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1659 {
1660         /* Update high watermark before we lower total_vm */
1661         update_hiwater_vm(mm);
1662         do {
1663                 long nrpages = vma_pages(vma);
1664
1665                 mm->total_vm -= nrpages;
1666                 if (vma->vm_flags & VM_LOCKED)
1667                         mm->locked_vm -= nrpages;
1668                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1669                 vma = remove_vma(vma);
1670         } while (vma);
1671         validate_mm(mm);
1672 }
1673
1674 /*
1675  * Get rid of page table information in the indicated region.
1676  *
1677  * Called with the mm semaphore held.
1678  */
1679 static void unmap_region(struct mm_struct *mm,
1680                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1681                 unsigned long start, unsigned long end)
1682 {
1683         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1684         struct mmu_gather *tlb;
1685         unsigned long nr_accounted = 0;
1686
1687         lru_add_drain();
1688         tlb = tlb_gather_mmu(mm, 0);
1689         update_hiwater_rss(mm);
1690         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1691         vm_unacct_memory(nr_accounted);
1692         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1693                                  next? next->vm_start: 0);
1694         tlb_finish_mmu(tlb, start, end);
1695 }
1696
1697 /*
1698  * Create a list of vma's touched by the unmap, removing them from the mm's
1699  * vma list as we go..
1700  */
1701 static void
1702 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1703         struct vm_area_struct *prev, unsigned long end)
1704 {
1705         struct vm_area_struct **insertion_point;
1706         struct vm_area_struct *tail_vma = NULL;
1707         unsigned long addr;
1708
1709         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1710         do {
1711                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1712                 mm->map_count--;
1713                 tail_vma = vma;
1714                 vma = vma->vm_next;
1715         } while (vma && vma->vm_start < end);
1716         *insertion_point = vma;
1717         tail_vma->vm_next = NULL;
1718         if (mm->unmap_area == arch_unmap_area)
1719                 addr = prev ? prev->vm_end : mm->mmap_base;
1720         else
1721                 addr = vma ?  vma->vm_start : mm->mmap_base;
1722         mm->unmap_area(mm, addr);
1723         mm->mmap_cache = NULL;          /* Kill the cache. */
1724 }
1725
1726 /*
1727  * Split a vma into two pieces at address 'addr', a new vma is allocated
1728  * either for the first part or the tail.
1729  */
1730 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1731               unsigned long addr, int new_below)
1732 {
1733         struct mempolicy *pol;
1734         struct vm_area_struct *new;
1735
1736         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1737                 return -EINVAL;
1738
1739         if (mm->map_count >= sysctl_max_map_count)
1740                 return -ENOMEM;
1741
1742         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1743         if (!new)
1744                 return -ENOMEM;
1745
1746         /* most fields are the same, copy all, and then fixup */
1747         *new = *vma;
1748
1749         if (new_below)
1750                 new->vm_end = addr;
1751         else {
1752                 new->vm_start = addr;
1753                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1754         }
1755
1756         pol = mpol_copy(vma_policy(vma));
1757         if (IS_ERR(pol)) {
1758                 kmem_cache_free(vm_area_cachep, new);
1759                 return PTR_ERR(pol);
1760         }
1761         vma_set_policy(new, pol);
1762
1763         if (new->vm_file)
1764                 get_file(new->vm_file);
1765
1766         if (new->vm_ops && new->vm_ops->open)
1767                 new->vm_ops->open(new);
1768
1769         if (new_below)
1770                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1771                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1772         else
1773                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1774
1775         return 0;
1776 }
1777
1778 /* Munmap is split into 2 main parts -- this part which finds
1779  * what needs doing, and the areas themselves, which do the
1780  * work.  This now handles partial unmappings.
1781  * Jeremy Fitzhardinge <jeremy@goop.org>
1782  */
1783 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1784 {
1785         unsigned long end;
1786         struct vm_area_struct *vma, *prev, *last;
1787
1788         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1789                 return -EINVAL;
1790
1791         if ((len = PAGE_ALIGN(len)) == 0)
1792                 return -EINVAL;
1793
1794         /* Find the first overlapping VMA */
1795         vma = find_vma_prev(mm, start, &prev);
1796         if (!vma)
1797                 return 0;
1798         /* we have  start < vma->vm_end  */
1799
1800         /* if it doesn't overlap, we have nothing.. */
1801         end = start + len;
1802         if (vma->vm_start >= end)
1803                 return 0;
1804
1805         /*
1806          * If we need to split any vma, do it now to save pain later.
1807          *
1808          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1809          * unmapped vm_area_struct will remain in use: so lower split_vma
1810          * places tmp vma above, and higher split_vma places tmp vma below.
1811          */
1812         if (start > vma->vm_start) {
1813                 int error = split_vma(mm, vma, start, 0);
1814                 if (error)
1815                         return error;
1816                 prev = vma;
1817         }
1818
1819         /* Does it split the last one? */
1820         last = find_vma(mm, end);
1821         if (last && end > last->vm_start) {
1822                 int error = split_vma(mm, last, end, 1);
1823                 if (error)
1824                         return error;
1825         }
1826         vma = prev? prev->vm_next: mm->mmap;
1827
1828         /*
1829          * Remove the vma's, and unmap the actual pages
1830          */
1831         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1832         unmap_region(mm, vma, prev, start, end);
1833
1834         /* Fix up all other VM information */
1835         remove_vma_list(mm, vma);
1836
1837         return 0;
1838 }
1839
1840 EXPORT_SYMBOL(do_munmap);
1841
1842 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1843 {
1844         int ret;
1845         struct mm_struct *mm = current->mm;
1846
1847         profile_munmap(addr);
1848
1849         down_write(&mm->mmap_sem);
1850         ret = do_munmap(mm, addr, len);
1851         up_write(&mm->mmap_sem);
1852         return ret;
1853 }
1854
1855 static inline void verify_mm_writelocked(struct mm_struct *mm)
1856 {
1857 #ifdef CONFIG_DEBUG_VM
1858         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1859                 WARN_ON(1);
1860                 up_read(&mm->mmap_sem);
1861         }
1862 #endif
1863 }
1864
1865 /*
1866  *  this is really a simplified "do_mmap".  it only handles
1867  *  anonymous maps.  eventually we may be able to do some
1868  *  brk-specific accounting here.
1869  */
1870 unsigned long do_brk(unsigned long addr, unsigned long len)
1871 {
1872         struct mm_struct * mm = current->mm;
1873         struct vm_area_struct * vma, * prev;
1874         unsigned long flags;
1875         struct rb_node ** rb_link, * rb_parent;
1876         pgoff_t pgoff = addr >> PAGE_SHIFT;
1877         int error;
1878
1879         len = PAGE_ALIGN(len);
1880         if (!len)
1881                 return addr;
1882
1883         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1884                 return -EINVAL;
1885
1886         if (is_hugepage_only_range(mm, addr, len))
1887                 return -EINVAL;
1888
1889         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1890
1891         error = arch_mmap_check(addr, len, flags);
1892         if (error)
1893                 return error;
1894
1895         /*
1896          * mlock MCL_FUTURE?
1897          */
1898         if (mm->def_flags & VM_LOCKED) {
1899                 unsigned long locked, lock_limit;
1900                 locked = len >> PAGE_SHIFT;
1901                 locked += mm->locked_vm;
1902                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1903                 lock_limit >>= PAGE_SHIFT;
1904                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1905                         return -EAGAIN;
1906         }
1907
1908         /*
1909          * mm->mmap_sem is required to protect against another thread
1910          * changing the mappings in case we sleep.
1911          */
1912         verify_mm_writelocked(mm);
1913
1914         /*
1915          * Clear old maps.  this also does some error checking for us
1916          */
1917  munmap_back:
1918         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1919         if (vma && vma->vm_start < addr + len) {
1920                 if (do_munmap(mm, addr, len))
1921                         return -ENOMEM;
1922                 goto munmap_back;
1923         }
1924
1925         /* Check against address space limits *after* clearing old maps... */
1926         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1927                 return -ENOMEM;
1928
1929         if (mm->map_count > sysctl_max_map_count)
1930                 return -ENOMEM;
1931
1932         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1933                 return -ENOMEM;
1934
1935         /* Can we just expand an old private anonymous mapping? */
1936         if (vma_merge(mm, prev, addr, addr + len, flags,
1937                                         NULL, NULL, pgoff, NULL))
1938                 goto out;
1939
1940         /*
1941          * create a vma struct for an anonymous mapping
1942          */
1943         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1944         if (!vma) {
1945                 vm_unacct_memory(len >> PAGE_SHIFT);
1946                 return -ENOMEM;
1947         }
1948
1949         vma->vm_mm = mm;
1950         vma->vm_start = addr;
1951         vma->vm_end = addr + len;
1952         vma->vm_pgoff = pgoff;
1953         vma->vm_flags = flags;
1954         vma->vm_page_prot = protection_map[flags &
1955                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1956         vma_link(mm, vma, prev, rb_link, rb_parent);
1957 out:
1958         mm->total_vm += len >> PAGE_SHIFT;
1959         if (flags & VM_LOCKED) {
1960                 mm->locked_vm += len >> PAGE_SHIFT;
1961                 make_pages_present(addr, addr + len);
1962         }
1963         return addr;
1964 }
1965
1966 EXPORT_SYMBOL(do_brk);
1967
1968 /* Release all mmaps. */
1969 void exit_mmap(struct mm_struct *mm)
1970 {
1971         struct mmu_gather *tlb;
1972         struct vm_area_struct *vma = mm->mmap;
1973         unsigned long nr_accounted = 0;
1974         unsigned long end;
1975
1976         /* mm's last user has gone, and its about to be pulled down */
1977         arch_exit_mmap(mm);
1978
1979         lru_add_drain();
1980         flush_cache_mm(mm);
1981         tlb = tlb_gather_mmu(mm, 1);
1982         /* Don't update_hiwater_rss(mm) here, do_exit already did */
1983         /* Use -1 here to ensure all VMAs in the mm are unmapped */
1984         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
1985         vm_unacct_memory(nr_accounted);
1986         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
1987         tlb_finish_mmu(tlb, 0, end);
1988
1989         /*
1990          * Walk the list again, actually closing and freeing it,
1991          * with preemption enabled, without holding any MM locks.
1992          */
1993         while (vma)
1994                 vma = remove_vma(vma);
1995
1996         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
1997 }
1998
1999 /* Insert vm structure into process list sorted by address
2000  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2001  * then i_mmap_lock is taken here.
2002  */
2003 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2004 {
2005         struct vm_area_struct * __vma, * prev;
2006         struct rb_node ** rb_link, * rb_parent;
2007
2008         /*
2009          * The vm_pgoff of a purely anonymous vma should be irrelevant
2010          * until its first write fault, when page's anon_vma and index
2011          * are set.  But now set the vm_pgoff it will almost certainly
2012          * end up with (unless mremap moves it elsewhere before that
2013          * first wfault), so /proc/pid/maps tells a consistent story.
2014          *
2015          * By setting it to reflect the virtual start address of the
2016          * vma, merges and splits can happen in a seamless way, just
2017          * using the existing file pgoff checks and manipulations.
2018          * Similarly in do_mmap_pgoff and in do_brk.
2019          */
2020         if (!vma->vm_file) {
2021                 BUG_ON(vma->anon_vma);
2022                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2023         }
2024         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2025         if (__vma && __vma->vm_start < vma->vm_end)
2026                 return -ENOMEM;
2027         if ((vma->vm_flags & VM_ACCOUNT) &&
2028              security_vm_enough_memory(vma_pages(vma)))
2029                 return -ENOMEM;
2030         vma_link(mm, vma, prev, rb_link, rb_parent);
2031         return 0;
2032 }
2033
2034 /*
2035  * Copy the vma structure to a new location in the same mm,
2036  * prior to moving page table entries, to effect an mremap move.
2037  */
2038 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2039         unsigned long addr, unsigned long len, pgoff_t pgoff)
2040 {
2041         struct vm_area_struct *vma = *vmap;
2042         unsigned long vma_start = vma->vm_start;
2043         struct mm_struct *mm = vma->vm_mm;
2044         struct vm_area_struct *new_vma, *prev;
2045         struct rb_node **rb_link, *rb_parent;
2046         struct mempolicy *pol;
2047
2048         /*
2049          * If anonymous vma has not yet been faulted, update new pgoff
2050          * to match new location, to increase its chance of merging.
2051          */
2052         if (!vma->vm_file && !vma->anon_vma)
2053                 pgoff = addr >> PAGE_SHIFT;
2054
2055         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2056         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2057                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2058         if (new_vma) {
2059                 /*
2060                  * Source vma may have been merged into new_vma
2061                  */
2062                 if (vma_start >= new_vma->vm_start &&
2063                     vma_start < new_vma->vm_end)
2064                         *vmap = new_vma;
2065         } else {
2066                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2067                 if (new_vma) {
2068                         *new_vma = *vma;
2069                         pol = mpol_copy(vma_policy(vma));
2070                         if (IS_ERR(pol)) {
2071                                 kmem_cache_free(vm_area_cachep, new_vma);
2072                                 return NULL;
2073                         }
2074                         vma_set_policy(new_vma, pol);
2075                         new_vma->vm_start = addr;
2076                         new_vma->vm_end = addr + len;
2077                         new_vma->vm_pgoff = pgoff;
2078                         if (new_vma->vm_file)
2079                                 get_file(new_vma->vm_file);
2080                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2081                                 new_vma->vm_ops->open(new_vma);
2082                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2083                 }
2084         }
2085         return new_vma;
2086 }
2087
2088 /*
2089  * Return true if the calling process may expand its vm space by the passed
2090  * number of pages
2091  */
2092 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2093 {
2094         unsigned long cur = mm->total_vm;       /* pages */
2095         unsigned long lim;
2096
2097         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2098
2099         if (cur + npages > lim)
2100                 return 0;
2101         return 1;
2102 }
2103
2104
2105 static struct page *special_mapping_nopage(struct vm_area_struct *vma,
2106                                            unsigned long address, int *type)
2107 {
2108         struct page **pages;
2109
2110         BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2111
2112         address -= vma->vm_start;
2113         for (pages = vma->vm_private_data; address > 0 && *pages; ++pages)
2114                 address -= PAGE_SIZE;
2115
2116         if (*pages) {
2117                 struct page *page = *pages;
2118                 get_page(page);
2119                 return page;
2120         }
2121
2122         return NOPAGE_SIGBUS;
2123 }
2124
2125 /*
2126  * Having a close hook prevents vma merging regardless of flags.
2127  */
2128 static void special_mapping_close(struct vm_area_struct *vma)
2129 {
2130 }
2131
2132 static struct vm_operations_struct special_mapping_vmops = {
2133         .close = special_mapping_close,
2134         .nopage = special_mapping_nopage,
2135 };
2136
2137 /*
2138  * Called with mm->mmap_sem held for writing.
2139  * Insert a new vma covering the given region, with the given flags.
2140  * Its pages are supplied by the given array of struct page *.
2141  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2142  * The region past the last page supplied will always produce SIGBUS.
2143  * The array pointer and the pages it points to are assumed to stay alive
2144  * for as long as this mapping might exist.
2145  */
2146 int install_special_mapping(struct mm_struct *mm,
2147                             unsigned long addr, unsigned long len,
2148                             unsigned long vm_flags, struct page **pages)
2149 {
2150         struct vm_area_struct *vma;
2151
2152         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2153         if (unlikely(vma == NULL))
2154                 return -ENOMEM;
2155
2156         vma->vm_mm = mm;
2157         vma->vm_start = addr;
2158         vma->vm_end = addr + len;
2159
2160         vma->vm_flags = vm_flags | mm->def_flags;
2161         vma->vm_page_prot = protection_map[vma->vm_flags & 7];
2162
2163         vma->vm_ops = &special_mapping_vmops;
2164         vma->vm_private_data = pages;
2165
2166         if (unlikely(insert_vm_struct(mm, vma))) {
2167                 kmem_cache_free(vm_area_cachep, vma);
2168                 return -ENOMEM;
2169         }
2170
2171         mm->total_vm += len >> PAGE_SHIFT;
2172
2173         return 0;
2174 }