stackprotector: better self-test
[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/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29
30 #include <asm/uaccess.h>
31 #include <asm/cacheflush.h>
32 #include <asm/tlb.h>
33 #include <asm/mmu_context.h>
34
35 #ifndef arch_mmap_check
36 #define arch_mmap_check(addr, len, flags)       (0)
37 #endif
38
39 #ifndef arch_rebalance_pgtables
40 #define arch_rebalance_pgtables(addr, len)              (addr)
41 #endif
42
43 static void unmap_region(struct mm_struct *mm,
44                 struct vm_area_struct *vma, struct vm_area_struct *prev,
45                 unsigned long start, unsigned long end);
46
47 /*
48  * WARNING: the debugging will use recursive algorithms so never enable this
49  * unless you know what you are doing.
50  */
51 #undef DEBUG_MM_RB
52
53 /* description of effects of mapping type and prot in current implementation.
54  * this is due to the limited x86 page protection hardware.  The expected
55  * behavior is in parens:
56  *
57  * map_type     prot
58  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
59  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
60  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
61  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
62  *              
63  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
64  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
65  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
66  *
67  */
68 pgprot_t protection_map[16] = {
69         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
70         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
71 };
72
73 pgprot_t vm_get_page_prot(unsigned long vm_flags)
74 {
75         return protection_map[vm_flags &
76                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
77 }
78 EXPORT_SYMBOL(vm_get_page_prot);
79
80 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
81 int sysctl_overcommit_ratio = 50;       /* default is 50% */
82 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
83 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
84
85 /*
86  * Check that a process has enough memory to allocate a new virtual
87  * mapping. 0 means there is enough memory for the allocation to
88  * succeed and -ENOMEM implies there is not.
89  *
90  * We currently support three overcommit policies, which are set via the
91  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
92  *
93  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
94  * Additional code 2002 Jul 20 by Robert Love.
95  *
96  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
97  *
98  * Note this is a helper function intended to be used by LSMs which
99  * wish to use this logic.
100  */
101 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
102 {
103         unsigned long free, allowed;
104
105         vm_acct_memory(pages);
106
107         /*
108          * Sometimes we want to use more memory than we have
109          */
110         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
111                 return 0;
112
113         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
114                 unsigned long n;
115
116                 free = global_page_state(NR_FILE_PAGES);
117                 free += nr_swap_pages;
118
119                 /*
120                  * Any slabs which are created with the
121                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
122                  * which are reclaimable, under pressure.  The dentry
123                  * cache and most inode caches should fall into this
124                  */
125                 free += global_page_state(NR_SLAB_RECLAIMABLE);
126
127                 /*
128                  * Leave the last 3% for root
129                  */
130                 if (!cap_sys_admin)
131                         free -= free / 32;
132
133                 if (free > pages)
134                         return 0;
135
136                 /*
137                  * nr_free_pages() is very expensive on large systems,
138                  * only call if we're about to fail.
139                  */
140                 n = nr_free_pages();
141
142                 /*
143                  * Leave reserved pages. The pages are not for anonymous pages.
144                  */
145                 if (n <= totalreserve_pages)
146                         goto error;
147                 else
148                         n -= totalreserve_pages;
149
150                 /*
151                  * Leave the last 3% for root
152                  */
153                 if (!cap_sys_admin)
154                         n -= n / 32;
155                 free += n;
156
157                 if (free > pages)
158                         return 0;
159
160                 goto error;
161         }
162
163         allowed = (totalram_pages - hugetlb_total_pages())
164                 * sysctl_overcommit_ratio / 100;
165         /*
166          * Leave the last 3% for root
167          */
168         if (!cap_sys_admin)
169                 allowed -= allowed / 32;
170         allowed += total_swap_pages;
171
172         /* Don't let a single process grow too big:
173            leave 3% of the size of this process for other processes */
174         allowed -= mm->total_vm / 32;
175
176         /*
177          * cast `allowed' as a signed long because vm_committed_space
178          * sometimes has a negative value
179          */
180         if (atomic_long_read(&vm_committed_space) < (long)allowed)
181                 return 0;
182 error:
183         vm_unacct_memory(pages);
184
185         return -ENOMEM;
186 }
187
188 /*
189  * Requires inode->i_mapping->i_mmap_lock
190  */
191 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
192                 struct file *file, struct address_space *mapping)
193 {
194         if (vma->vm_flags & VM_DENYWRITE)
195                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
196         if (vma->vm_flags & VM_SHARED)
197                 mapping->i_mmap_writable--;
198
199         flush_dcache_mmap_lock(mapping);
200         if (unlikely(vma->vm_flags & VM_NONLINEAR))
201                 list_del_init(&vma->shared.vm_set.list);
202         else
203                 vma_prio_tree_remove(vma, &mapping->i_mmap);
204         flush_dcache_mmap_unlock(mapping);
205 }
206
207 /*
208  * Unlink a file-based vm structure from its prio_tree, to hide
209  * vma from rmap and vmtruncate before freeing its page tables.
210  */
211 void unlink_file_vma(struct vm_area_struct *vma)
212 {
213         struct file *file = vma->vm_file;
214
215         if (file) {
216                 struct address_space *mapping = file->f_mapping;
217                 spin_lock(&mapping->i_mmap_lock);
218                 __remove_shared_vm_struct(vma, file, mapping);
219                 spin_unlock(&mapping->i_mmap_lock);
220         }
221 }
222
223 /*
224  * Close a vm structure and free it, returning the next.
225  */
226 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
227 {
228         struct vm_area_struct *next = vma->vm_next;
229
230         might_sleep();
231         if (vma->vm_ops && vma->vm_ops->close)
232                 vma->vm_ops->close(vma);
233         if (vma->vm_file) {
234                 fput(vma->vm_file);
235                 if (vma->vm_flags & VM_EXECUTABLE)
236                         removed_exe_file_vma(vma->vm_mm);
237         }
238         mpol_put(vma_policy(vma));
239         kmem_cache_free(vm_area_cachep, vma);
240         return next;
241 }
242
243 asmlinkage unsigned long sys_brk(unsigned long brk)
244 {
245         unsigned long rlim, retval;
246         unsigned long newbrk, oldbrk;
247         struct mm_struct *mm = current->mm;
248         unsigned long min_brk;
249
250         down_write(&mm->mmap_sem);
251
252 #ifdef CONFIG_COMPAT_BRK
253         min_brk = mm->end_code;
254 #else
255         min_brk = mm->start_brk;
256 #endif
257         if (brk < min_brk)
258                 goto out;
259
260         /*
261          * Check against rlimit here. If this check is done later after the test
262          * of oldbrk with newbrk then it can escape the test and let the data
263          * segment grow beyond its set limit the in case where the limit is
264          * not page aligned -Ram Gupta
265          */
266         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
267         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
268                         (mm->end_data - mm->start_data) > rlim)
269                 goto out;
270
271         newbrk = PAGE_ALIGN(brk);
272         oldbrk = PAGE_ALIGN(mm->brk);
273         if (oldbrk == newbrk)
274                 goto set_brk;
275
276         /* Always allow shrinking brk. */
277         if (brk <= mm->brk) {
278                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
279                         goto set_brk;
280                 goto out;
281         }
282
283         /* Check against existing mmap mappings. */
284         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
285                 goto out;
286
287         /* Ok, looks good - let it rip. */
288         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
289                 goto out;
290 set_brk:
291         mm->brk = brk;
292 out:
293         retval = mm->brk;
294         up_write(&mm->mmap_sem);
295         return retval;
296 }
297
298 #ifdef DEBUG_MM_RB
299 static int browse_rb(struct rb_root *root)
300 {
301         int i = 0, j;
302         struct rb_node *nd, *pn = NULL;
303         unsigned long prev = 0, pend = 0;
304
305         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
306                 struct vm_area_struct *vma;
307                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
308                 if (vma->vm_start < prev)
309                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
310                 if (vma->vm_start < pend)
311                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
312                 if (vma->vm_start > vma->vm_end)
313                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
314                 i++;
315                 pn = nd;
316                 prev = vma->vm_start;
317                 pend = vma->vm_end;
318         }
319         j = 0;
320         for (nd = pn; nd; nd = rb_prev(nd)) {
321                 j++;
322         }
323         if (i != j)
324                 printk("backwards %d, forwards %d\n", j, i), i = 0;
325         return i;
326 }
327
328 void validate_mm(struct mm_struct *mm)
329 {
330         int bug = 0;
331         int i = 0;
332         struct vm_area_struct *tmp = mm->mmap;
333         while (tmp) {
334                 tmp = tmp->vm_next;
335                 i++;
336         }
337         if (i != mm->map_count)
338                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
339         i = browse_rb(&mm->mm_rb);
340         if (i != mm->map_count)
341                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
342         BUG_ON(bug);
343 }
344 #else
345 #define validate_mm(mm) do { } while (0)
346 #endif
347
348 static struct vm_area_struct *
349 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
350                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
351                 struct rb_node ** rb_parent)
352 {
353         struct vm_area_struct * vma;
354         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
355
356         __rb_link = &mm->mm_rb.rb_node;
357         rb_prev = __rb_parent = NULL;
358         vma = NULL;
359
360         while (*__rb_link) {
361                 struct vm_area_struct *vma_tmp;
362
363                 __rb_parent = *__rb_link;
364                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
365
366                 if (vma_tmp->vm_end > addr) {
367                         vma = vma_tmp;
368                         if (vma_tmp->vm_start <= addr)
369                                 return vma;
370                         __rb_link = &__rb_parent->rb_left;
371                 } else {
372                         rb_prev = __rb_parent;
373                         __rb_link = &__rb_parent->rb_right;
374                 }
375         }
376
377         *pprev = NULL;
378         if (rb_prev)
379                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
380         *rb_link = __rb_link;
381         *rb_parent = __rb_parent;
382         return vma;
383 }
384
385 static inline void
386 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
387                 struct vm_area_struct *prev, struct rb_node *rb_parent)
388 {
389         if (prev) {
390                 vma->vm_next = prev->vm_next;
391                 prev->vm_next = vma;
392         } else {
393                 mm->mmap = vma;
394                 if (rb_parent)
395                         vma->vm_next = rb_entry(rb_parent,
396                                         struct vm_area_struct, vm_rb);
397                 else
398                         vma->vm_next = NULL;
399         }
400 }
401
402 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
403                 struct rb_node **rb_link, struct rb_node *rb_parent)
404 {
405         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
406         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
407 }
408
409 static inline void __vma_link_file(struct vm_area_struct *vma)
410 {
411         struct file * file;
412
413         file = vma->vm_file;
414         if (file) {
415                 struct address_space *mapping = file->f_mapping;
416
417                 if (vma->vm_flags & VM_DENYWRITE)
418                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
419                 if (vma->vm_flags & VM_SHARED)
420                         mapping->i_mmap_writable++;
421
422                 flush_dcache_mmap_lock(mapping);
423                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
424                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
425                 else
426                         vma_prio_tree_insert(vma, &mapping->i_mmap);
427                 flush_dcache_mmap_unlock(mapping);
428         }
429 }
430
431 static void
432 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
433         struct vm_area_struct *prev, struct rb_node **rb_link,
434         struct rb_node *rb_parent)
435 {
436         __vma_link_list(mm, vma, prev, rb_parent);
437         __vma_link_rb(mm, vma, rb_link, rb_parent);
438         __anon_vma_link(vma);
439 }
440
441 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
442                         struct vm_area_struct *prev, struct rb_node **rb_link,
443                         struct rb_node *rb_parent)
444 {
445         struct address_space *mapping = NULL;
446
447         if (vma->vm_file)
448                 mapping = vma->vm_file->f_mapping;
449
450         if (mapping) {
451                 spin_lock(&mapping->i_mmap_lock);
452                 vma->vm_truncate_count = mapping->truncate_count;
453         }
454         anon_vma_lock(vma);
455
456         __vma_link(mm, vma, prev, rb_link, rb_parent);
457         __vma_link_file(vma);
458
459         anon_vma_unlock(vma);
460         if (mapping)
461                 spin_unlock(&mapping->i_mmap_lock);
462
463         mm->map_count++;
464         validate_mm(mm);
465 }
466
467 /*
468  * Helper for vma_adjust in the split_vma insert case:
469  * insert vm structure into list and rbtree and anon_vma,
470  * but it has already been inserted into prio_tree earlier.
471  */
472 static void
473 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
474 {
475         struct vm_area_struct * __vma, * prev;
476         struct rb_node ** rb_link, * rb_parent;
477
478         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
479         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
480         __vma_link(mm, vma, prev, rb_link, rb_parent);
481         mm->map_count++;
482 }
483
484 static inline void
485 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
486                 struct vm_area_struct *prev)
487 {
488         prev->vm_next = vma->vm_next;
489         rb_erase(&vma->vm_rb, &mm->mm_rb);
490         if (mm->mmap_cache == vma)
491                 mm->mmap_cache = prev;
492 }
493
494 /*
495  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
496  * is already present in an i_mmap tree without adjusting the tree.
497  * The following helper function should be used when such adjustments
498  * are necessary.  The "insert" vma (if any) is to be inserted
499  * before we drop the necessary locks.
500  */
501 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
502         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
503 {
504         struct mm_struct *mm = vma->vm_mm;
505         struct vm_area_struct *next = vma->vm_next;
506         struct vm_area_struct *importer = NULL;
507         struct address_space *mapping = NULL;
508         struct prio_tree_root *root = NULL;
509         struct file *file = vma->vm_file;
510         struct anon_vma *anon_vma = NULL;
511         long adjust_next = 0;
512         int remove_next = 0;
513
514         if (next && !insert) {
515                 if (end >= next->vm_end) {
516                         /*
517                          * vma expands, overlapping all the next, and
518                          * perhaps the one after too (mprotect case 6).
519                          */
520 again:                  remove_next = 1 + (end > next->vm_end);
521                         end = next->vm_end;
522                         anon_vma = next->anon_vma;
523                         importer = vma;
524                 } else if (end > next->vm_start) {
525                         /*
526                          * vma expands, overlapping part of the next:
527                          * mprotect case 5 shifting the boundary up.
528                          */
529                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
530                         anon_vma = next->anon_vma;
531                         importer = vma;
532                 } else if (end < vma->vm_end) {
533                         /*
534                          * vma shrinks, and !insert tells it's not
535                          * split_vma inserting another: so it must be
536                          * mprotect case 4 shifting the boundary down.
537                          */
538                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
539                         anon_vma = next->anon_vma;
540                         importer = next;
541                 }
542         }
543
544         if (file) {
545                 mapping = file->f_mapping;
546                 if (!(vma->vm_flags & VM_NONLINEAR))
547                         root = &mapping->i_mmap;
548                 spin_lock(&mapping->i_mmap_lock);
549                 if (importer &&
550                     vma->vm_truncate_count != next->vm_truncate_count) {
551                         /*
552                          * unmap_mapping_range might be in progress:
553                          * ensure that the expanding vma is rescanned.
554                          */
555                         importer->vm_truncate_count = 0;
556                 }
557                 if (insert) {
558                         insert->vm_truncate_count = vma->vm_truncate_count;
559                         /*
560                          * Put into prio_tree now, so instantiated pages
561                          * are visible to arm/parisc __flush_dcache_page
562                          * throughout; but we cannot insert into address
563                          * space until vma start or end is updated.
564                          */
565                         __vma_link_file(insert);
566                 }
567         }
568
569         /*
570          * When changing only vma->vm_end, we don't really need
571          * anon_vma lock: but is that case worth optimizing out?
572          */
573         if (vma->anon_vma)
574                 anon_vma = vma->anon_vma;
575         if (anon_vma) {
576                 spin_lock(&anon_vma->lock);
577                 /*
578                  * Easily overlooked: when mprotect shifts the boundary,
579                  * make sure the expanding vma has anon_vma set if the
580                  * shrinking vma had, to cover any anon pages imported.
581                  */
582                 if (importer && !importer->anon_vma) {
583                         importer->anon_vma = anon_vma;
584                         __anon_vma_link(importer);
585                 }
586         }
587
588         if (root) {
589                 flush_dcache_mmap_lock(mapping);
590                 vma_prio_tree_remove(vma, root);
591                 if (adjust_next)
592                         vma_prio_tree_remove(next, root);
593         }
594
595         vma->vm_start = start;
596         vma->vm_end = end;
597         vma->vm_pgoff = pgoff;
598         if (adjust_next) {
599                 next->vm_start += adjust_next << PAGE_SHIFT;
600                 next->vm_pgoff += adjust_next;
601         }
602
603         if (root) {
604                 if (adjust_next)
605                         vma_prio_tree_insert(next, root);
606                 vma_prio_tree_insert(vma, root);
607                 flush_dcache_mmap_unlock(mapping);
608         }
609
610         if (remove_next) {
611                 /*
612                  * vma_merge has merged next into vma, and needs
613                  * us to remove next before dropping the locks.
614                  */
615                 __vma_unlink(mm, next, vma);
616                 if (file)
617                         __remove_shared_vm_struct(next, file, mapping);
618                 if (next->anon_vma)
619                         __anon_vma_merge(vma, next);
620         } else if (insert) {
621                 /*
622                  * split_vma has split insert from vma, and needs
623                  * us to insert it before dropping the locks
624                  * (it may either follow vma or precede it).
625                  */
626                 __insert_vm_struct(mm, insert);
627         }
628
629         if (anon_vma)
630                 spin_unlock(&anon_vma->lock);
631         if (mapping)
632                 spin_unlock(&mapping->i_mmap_lock);
633
634         if (remove_next) {
635                 if (file) {
636                         fput(file);
637                         if (next->vm_flags & VM_EXECUTABLE)
638                                 removed_exe_file_vma(mm);
639                 }
640                 mm->map_count--;
641                 mpol_put(vma_policy(next));
642                 kmem_cache_free(vm_area_cachep, next);
643                 /*
644                  * In mprotect's case 6 (see comments on vma_merge),
645                  * we must remove another next too. It would clutter
646                  * up the code too much to do both in one go.
647                  */
648                 if (remove_next == 2) {
649                         next = vma->vm_next;
650                         goto again;
651                 }
652         }
653
654         validate_mm(mm);
655 }
656
657 /*
658  * If the vma has a ->close operation then the driver probably needs to release
659  * per-vma resources, so we don't attempt to merge those.
660  */
661 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
662
663 static inline int is_mergeable_vma(struct vm_area_struct *vma,
664                         struct file *file, unsigned long vm_flags)
665 {
666         if (vma->vm_flags != vm_flags)
667                 return 0;
668         if (vma->vm_file != file)
669                 return 0;
670         if (vma->vm_ops && vma->vm_ops->close)
671                 return 0;
672         return 1;
673 }
674
675 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
676                                         struct anon_vma *anon_vma2)
677 {
678         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
679 }
680
681 /*
682  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
683  * in front of (at a lower virtual address and file offset than) the vma.
684  *
685  * We cannot merge two vmas if they have differently assigned (non-NULL)
686  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
687  *
688  * We don't check here for the merged mmap wrapping around the end of pagecache
689  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
690  * wrap, nor mmaps which cover the final page at index -1UL.
691  */
692 static int
693 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
694         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
695 {
696         if (is_mergeable_vma(vma, file, vm_flags) &&
697             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
698                 if (vma->vm_pgoff == vm_pgoff)
699                         return 1;
700         }
701         return 0;
702 }
703
704 /*
705  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
706  * beyond (at a higher virtual address and file offset than) the vma.
707  *
708  * We cannot merge two vmas if they have differently assigned (non-NULL)
709  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
710  */
711 static int
712 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
713         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
714 {
715         if (is_mergeable_vma(vma, file, vm_flags) &&
716             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
717                 pgoff_t vm_pglen;
718                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
719                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
720                         return 1;
721         }
722         return 0;
723 }
724
725 /*
726  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
727  * whether that can be merged with its predecessor or its successor.
728  * Or both (it neatly fills a hole).
729  *
730  * In most cases - when called for mmap, brk or mremap - [addr,end) is
731  * certain not to be mapped by the time vma_merge is called; but when
732  * called for mprotect, it is certain to be already mapped (either at
733  * an offset within prev, or at the start of next), and the flags of
734  * this area are about to be changed to vm_flags - and the no-change
735  * case has already been eliminated.
736  *
737  * The following mprotect cases have to be considered, where AAAA is
738  * the area passed down from mprotect_fixup, never extending beyond one
739  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
740  *
741  *     AAAA             AAAA                AAAA          AAAA
742  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
743  *    cannot merge    might become    might become    might become
744  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
745  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
746  *    mremap move:                                    PPPPNNNNNNNN 8
747  *        AAAA
748  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
749  *    might become    case 1 below    case 2 below    case 3 below
750  *
751  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
752  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
753  */
754 struct vm_area_struct *vma_merge(struct mm_struct *mm,
755                         struct vm_area_struct *prev, unsigned long addr,
756                         unsigned long end, unsigned long vm_flags,
757                         struct anon_vma *anon_vma, struct file *file,
758                         pgoff_t pgoff, struct mempolicy *policy)
759 {
760         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
761         struct vm_area_struct *area, *next;
762
763         /*
764          * We later require that vma->vm_flags == vm_flags,
765          * so this tests vma->vm_flags & VM_SPECIAL, too.
766          */
767         if (vm_flags & VM_SPECIAL)
768                 return NULL;
769
770         if (prev)
771                 next = prev->vm_next;
772         else
773                 next = mm->mmap;
774         area = next;
775         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
776                 next = next->vm_next;
777
778         /*
779          * Can it merge with the predecessor?
780          */
781         if (prev && prev->vm_end == addr &&
782                         mpol_equal(vma_policy(prev), policy) &&
783                         can_vma_merge_after(prev, vm_flags,
784                                                 anon_vma, file, pgoff)) {
785                 /*
786                  * OK, it can.  Can we now merge in the successor as well?
787                  */
788                 if (next && end == next->vm_start &&
789                                 mpol_equal(policy, vma_policy(next)) &&
790                                 can_vma_merge_before(next, vm_flags,
791                                         anon_vma, file, pgoff+pglen) &&
792                                 is_mergeable_anon_vma(prev->anon_vma,
793                                                       next->anon_vma)) {
794                                                         /* cases 1, 6 */
795                         vma_adjust(prev, prev->vm_start,
796                                 next->vm_end, prev->vm_pgoff, NULL);
797                 } else                                  /* cases 2, 5, 7 */
798                         vma_adjust(prev, prev->vm_start,
799                                 end, prev->vm_pgoff, NULL);
800                 return prev;
801         }
802
803         /*
804          * Can this new request be merged in front of next?
805          */
806         if (next && end == next->vm_start &&
807                         mpol_equal(policy, vma_policy(next)) &&
808                         can_vma_merge_before(next, vm_flags,
809                                         anon_vma, file, pgoff+pglen)) {
810                 if (prev && addr < prev->vm_end)        /* case 4 */
811                         vma_adjust(prev, prev->vm_start,
812                                 addr, prev->vm_pgoff, NULL);
813                 else                                    /* cases 3, 8 */
814                         vma_adjust(area, addr, next->vm_end,
815                                 next->vm_pgoff - pglen, NULL);
816                 return area;
817         }
818
819         return NULL;
820 }
821
822 /*
823  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
824  * neighbouring vmas for a suitable anon_vma, before it goes off
825  * to allocate a new anon_vma.  It checks because a repetitive
826  * sequence of mprotects and faults may otherwise lead to distinct
827  * anon_vmas being allocated, preventing vma merge in subsequent
828  * mprotect.
829  */
830 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
831 {
832         struct vm_area_struct *near;
833         unsigned long vm_flags;
834
835         near = vma->vm_next;
836         if (!near)
837                 goto try_prev;
838
839         /*
840          * Since only mprotect tries to remerge vmas, match flags
841          * which might be mprotected into each other later on.
842          * Neither mlock nor madvise tries to remerge at present,
843          * so leave their flags as obstructing a merge.
844          */
845         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
846         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
847
848         if (near->anon_vma && vma->vm_end == near->vm_start &&
849                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
850                         can_vma_merge_before(near, vm_flags,
851                                 NULL, vma->vm_file, vma->vm_pgoff +
852                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
853                 return near->anon_vma;
854 try_prev:
855         /*
856          * It is potentially slow to have to call find_vma_prev here.
857          * But it's only on the first write fault on the vma, not
858          * every time, and we could devise a way to avoid it later
859          * (e.g. stash info in next's anon_vma_node when assigning
860          * an anon_vma, or when trying vma_merge).  Another time.
861          */
862         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
863         if (!near)
864                 goto none;
865
866         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
867         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
868
869         if (near->anon_vma && near->vm_end == vma->vm_start &&
870                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
871                         can_vma_merge_after(near, vm_flags,
872                                 NULL, vma->vm_file, vma->vm_pgoff))
873                 return near->anon_vma;
874 none:
875         /*
876          * There's no absolute need to look only at touching neighbours:
877          * we could search further afield for "compatible" anon_vmas.
878          * But it would probably just be a waste of time searching,
879          * or lead to too many vmas hanging off the same anon_vma.
880          * We're trying to allow mprotect remerging later on,
881          * not trying to minimize memory used for anon_vmas.
882          */
883         return NULL;
884 }
885
886 #ifdef CONFIG_PROC_FS
887 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
888                                                 struct file *file, long pages)
889 {
890         const unsigned long stack_flags
891                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
892
893         if (file) {
894                 mm->shared_vm += pages;
895                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
896                         mm->exec_vm += pages;
897         } else if (flags & stack_flags)
898                 mm->stack_vm += pages;
899         if (flags & (VM_RESERVED|VM_IO))
900                 mm->reserved_vm += pages;
901 }
902 #endif /* CONFIG_PROC_FS */
903
904 /*
905  * The caller must hold down_write(current->mm->mmap_sem).
906  */
907
908 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
909                         unsigned long len, unsigned long prot,
910                         unsigned long flags, unsigned long pgoff)
911 {
912         struct mm_struct * mm = current->mm;
913         struct inode *inode;
914         unsigned int vm_flags;
915         int error;
916         int accountable = 1;
917         unsigned long reqprot = prot;
918
919         /*
920          * Does the application expect PROT_READ to imply PROT_EXEC?
921          *
922          * (the exception is when the underlying filesystem is noexec
923          *  mounted, in which case we dont add PROT_EXEC.)
924          */
925         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
926                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
927                         prot |= PROT_EXEC;
928
929         if (!len)
930                 return -EINVAL;
931
932         if (!(flags & MAP_FIXED))
933                 addr = round_hint_to_min(addr);
934
935         error = arch_mmap_check(addr, len, flags);
936         if (error)
937                 return error;
938
939         /* Careful about overflows.. */
940         len = PAGE_ALIGN(len);
941         if (!len || len > TASK_SIZE)
942                 return -ENOMEM;
943
944         /* offset overflow? */
945         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
946                return -EOVERFLOW;
947
948         /* Too many mappings? */
949         if (mm->map_count > sysctl_max_map_count)
950                 return -ENOMEM;
951
952         /* Obtain the address to map to. we verify (or select) it and ensure
953          * that it represents a valid section of the address space.
954          */
955         addr = get_unmapped_area(file, addr, len, pgoff, flags);
956         if (addr & ~PAGE_MASK)
957                 return addr;
958
959         /* Do simple checking here so the lower-level routines won't have
960          * to. we assume access permissions have been handled by the open
961          * of the memory object, so we don't do any here.
962          */
963         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
964                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
965
966         if (flags & MAP_LOCKED) {
967                 if (!can_do_mlock())
968                         return -EPERM;
969                 vm_flags |= VM_LOCKED;
970         }
971         /* mlock MCL_FUTURE? */
972         if (vm_flags & VM_LOCKED) {
973                 unsigned long locked, lock_limit;
974                 locked = len >> PAGE_SHIFT;
975                 locked += mm->locked_vm;
976                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
977                 lock_limit >>= PAGE_SHIFT;
978                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
979                         return -EAGAIN;
980         }
981
982         inode = file ? file->f_path.dentry->d_inode : NULL;
983
984         if (file) {
985                 switch (flags & MAP_TYPE) {
986                 case MAP_SHARED:
987                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
988                                 return -EACCES;
989
990                         /*
991                          * Make sure we don't allow writing to an append-only
992                          * file..
993                          */
994                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
995                                 return -EACCES;
996
997                         /*
998                          * Make sure there are no mandatory locks on the file.
999                          */
1000                         if (locks_verify_locked(inode))
1001                                 return -EAGAIN;
1002
1003                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1004                         if (!(file->f_mode & FMODE_WRITE))
1005                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1006
1007                         /* fall through */
1008                 case MAP_PRIVATE:
1009                         if (!(file->f_mode & FMODE_READ))
1010                                 return -EACCES;
1011                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1012                                 if (vm_flags & VM_EXEC)
1013                                         return -EPERM;
1014                                 vm_flags &= ~VM_MAYEXEC;
1015                         }
1016                         if (is_file_hugepages(file))
1017                                 accountable = 0;
1018
1019                         if (!file->f_op || !file->f_op->mmap)
1020                                 return -ENODEV;
1021                         break;
1022
1023                 default:
1024                         return -EINVAL;
1025                 }
1026         } else {
1027                 switch (flags & MAP_TYPE) {
1028                 case MAP_SHARED:
1029                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1030                         break;
1031                 case MAP_PRIVATE:
1032                         /*
1033                          * Set pgoff according to addr for anon_vma.
1034                          */
1035                         pgoff = addr >> PAGE_SHIFT;
1036                         break;
1037                 default:
1038                         return -EINVAL;
1039                 }
1040         }
1041
1042         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1043         if (error)
1044                 return error;
1045
1046         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1047                            accountable);
1048 }
1049 EXPORT_SYMBOL(do_mmap_pgoff);
1050
1051 /*
1052  * Some shared mappigns will want the pages marked read-only
1053  * to track write events. If so, we'll downgrade vm_page_prot
1054  * to the private version (using protection_map[] without the
1055  * VM_SHARED bit).
1056  */
1057 int vma_wants_writenotify(struct vm_area_struct *vma)
1058 {
1059         unsigned int vm_flags = vma->vm_flags;
1060
1061         /* If it was private or non-writable, the write bit is already clear */
1062         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1063                 return 0;
1064
1065         /* The backer wishes to know when pages are first written to? */
1066         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1067                 return 1;
1068
1069         /* The open routine did something to the protections already? */
1070         if (pgprot_val(vma->vm_page_prot) !=
1071             pgprot_val(vm_get_page_prot(vm_flags)))
1072                 return 0;
1073
1074         /* Specialty mapping? */
1075         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1076                 return 0;
1077
1078         /* Can the mapping track the dirty pages? */
1079         return vma->vm_file && vma->vm_file->f_mapping &&
1080                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1081 }
1082
1083 unsigned long mmap_region(struct file *file, unsigned long addr,
1084                           unsigned long len, unsigned long flags,
1085                           unsigned int vm_flags, unsigned long pgoff,
1086                           int accountable)
1087 {
1088         struct mm_struct *mm = current->mm;
1089         struct vm_area_struct *vma, *prev;
1090         int correct_wcount = 0;
1091         int error;
1092         struct rb_node **rb_link, *rb_parent;
1093         unsigned long charged = 0;
1094         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1095
1096         /* Clear old maps */
1097         error = -ENOMEM;
1098 munmap_back:
1099         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1100         if (vma && vma->vm_start < addr + len) {
1101                 if (do_munmap(mm, addr, len))
1102                         return -ENOMEM;
1103                 goto munmap_back;
1104         }
1105
1106         /* Check against address space limit. */
1107         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1108                 return -ENOMEM;
1109
1110         if (accountable && (!(flags & MAP_NORESERVE) ||
1111                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1112                 if (vm_flags & VM_SHARED) {
1113                         /* Check memory availability in shmem_file_setup? */
1114                         vm_flags |= VM_ACCOUNT;
1115                 } else if (vm_flags & VM_WRITE) {
1116                         /*
1117                          * Private writable mapping: check memory availability
1118                          */
1119                         charged = len >> PAGE_SHIFT;
1120                         if (security_vm_enough_memory(charged))
1121                                 return -ENOMEM;
1122                         vm_flags |= VM_ACCOUNT;
1123                 }
1124         }
1125
1126         /*
1127          * Can we just expand an old private anonymous mapping?
1128          * The VM_SHARED test is necessary because shmem_zero_setup
1129          * will create the file object for a shared anonymous map below.
1130          */
1131         if (!file && !(vm_flags & VM_SHARED) &&
1132             vma_merge(mm, prev, addr, addr + len, vm_flags,
1133                                         NULL, NULL, pgoff, NULL))
1134                 goto out;
1135
1136         /*
1137          * Determine the object being mapped and call the appropriate
1138          * specific mapper. the address has already been validated, but
1139          * not unmapped, but the maps are removed from the list.
1140          */
1141         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1142         if (!vma) {
1143                 error = -ENOMEM;
1144                 goto unacct_error;
1145         }
1146
1147         vma->vm_mm = mm;
1148         vma->vm_start = addr;
1149         vma->vm_end = addr + len;
1150         vma->vm_flags = vm_flags;
1151         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1152         vma->vm_pgoff = pgoff;
1153
1154         if (file) {
1155                 error = -EINVAL;
1156                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1157                         goto free_vma;
1158                 if (vm_flags & VM_DENYWRITE) {
1159                         error = deny_write_access(file);
1160                         if (error)
1161                                 goto free_vma;
1162                         correct_wcount = 1;
1163                 }
1164                 vma->vm_file = file;
1165                 get_file(file);
1166                 error = file->f_op->mmap(file, vma);
1167                 if (error)
1168                         goto unmap_and_free_vma;
1169                 if (vm_flags & VM_EXECUTABLE)
1170                         added_exe_file_vma(mm);
1171         } else if (vm_flags & VM_SHARED) {
1172                 error = shmem_zero_setup(vma);
1173                 if (error)
1174                         goto free_vma;
1175         }
1176
1177         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1178          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1179          * that memory reservation must be checked; but that reservation
1180          * belongs to shared memory object, not to vma: so now clear it.
1181          */
1182         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1183                 vma->vm_flags &= ~VM_ACCOUNT;
1184
1185         /* Can addr have changed??
1186          *
1187          * Answer: Yes, several device drivers can do it in their
1188          *         f_op->mmap method. -DaveM
1189          */
1190         addr = vma->vm_start;
1191         pgoff = vma->vm_pgoff;
1192         vm_flags = vma->vm_flags;
1193
1194         if (vma_wants_writenotify(vma))
1195                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1196
1197         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1198                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1199                 mpol_put(vma_policy(vma));
1200                 kmem_cache_free(vm_area_cachep, vma);
1201                 fput(file);
1202                 if (vm_flags & VM_EXECUTABLE)
1203                         removed_exe_file_vma(mm);
1204         } else {
1205                 vma_link(mm, vma, prev, rb_link, rb_parent);
1206                 file = vma->vm_file;
1207         }
1208
1209         /* Once vma denies write, undo our temporary denial count */
1210         if (correct_wcount)
1211                 atomic_inc(&inode->i_writecount);
1212 out:
1213         mm->total_vm += len >> PAGE_SHIFT;
1214         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1215         if (vm_flags & VM_LOCKED) {
1216                 mm->locked_vm += len >> PAGE_SHIFT;
1217                 make_pages_present(addr, addr + len);
1218         }
1219         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1220                 make_pages_present(addr, addr + len);
1221         return addr;
1222
1223 unmap_and_free_vma:
1224         if (correct_wcount)
1225                 atomic_inc(&inode->i_writecount);
1226         vma->vm_file = NULL;
1227         fput(file);
1228
1229         /* Undo any partial mapping done by a device driver. */
1230         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1231         charged = 0;
1232 free_vma:
1233         kmem_cache_free(vm_area_cachep, vma);
1234 unacct_error:
1235         if (charged)
1236                 vm_unacct_memory(charged);
1237         return error;
1238 }
1239
1240 /* Get an address range which is currently unmapped.
1241  * For shmat() with addr=0.
1242  *
1243  * Ugly calling convention alert:
1244  * Return value with the low bits set means error value,
1245  * ie
1246  *      if (ret & ~PAGE_MASK)
1247  *              error = ret;
1248  *
1249  * This function "knows" that -ENOMEM has the bits set.
1250  */
1251 #ifndef HAVE_ARCH_UNMAPPED_AREA
1252 unsigned long
1253 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1254                 unsigned long len, unsigned long pgoff, unsigned long flags)
1255 {
1256         struct mm_struct *mm = current->mm;
1257         struct vm_area_struct *vma;
1258         unsigned long start_addr;
1259
1260         if (len > TASK_SIZE)
1261                 return -ENOMEM;
1262
1263         if (flags & MAP_FIXED)
1264                 return addr;
1265
1266         if (addr) {
1267                 addr = PAGE_ALIGN(addr);
1268                 vma = find_vma(mm, addr);
1269                 if (TASK_SIZE - len >= addr &&
1270                     (!vma || addr + len <= vma->vm_start))
1271                         return addr;
1272         }
1273         if (len > mm->cached_hole_size) {
1274                 start_addr = addr = mm->free_area_cache;
1275         } else {
1276                 start_addr = addr = TASK_UNMAPPED_BASE;
1277                 mm->cached_hole_size = 0;
1278         }
1279
1280 full_search:
1281         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1282                 /* At this point:  (!vma || addr < vma->vm_end). */
1283                 if (TASK_SIZE - len < addr) {
1284                         /*
1285                          * Start a new search - just in case we missed
1286                          * some holes.
1287                          */
1288                         if (start_addr != TASK_UNMAPPED_BASE) {
1289                                 addr = TASK_UNMAPPED_BASE;
1290                                 start_addr = addr;
1291                                 mm->cached_hole_size = 0;
1292                                 goto full_search;
1293                         }
1294                         return -ENOMEM;
1295                 }
1296                 if (!vma || addr + len <= vma->vm_start) {
1297                         /*
1298                          * Remember the place where we stopped the search:
1299                          */
1300                         mm->free_area_cache = addr + len;
1301                         return addr;
1302                 }
1303                 if (addr + mm->cached_hole_size < vma->vm_start)
1304                         mm->cached_hole_size = vma->vm_start - addr;
1305                 addr = vma->vm_end;
1306         }
1307 }
1308 #endif  
1309
1310 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1311 {
1312         /*
1313          * Is this a new hole at the lowest possible address?
1314          */
1315         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1316                 mm->free_area_cache = addr;
1317                 mm->cached_hole_size = ~0UL;
1318         }
1319 }
1320
1321 /*
1322  * This mmap-allocator allocates new areas top-down from below the
1323  * stack's low limit (the base):
1324  */
1325 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1326 unsigned long
1327 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1328                           const unsigned long len, const unsigned long pgoff,
1329                           const unsigned long flags)
1330 {
1331         struct vm_area_struct *vma;
1332         struct mm_struct *mm = current->mm;
1333         unsigned long addr = addr0;
1334
1335         /* requested length too big for entire address space */
1336         if (len > TASK_SIZE)
1337                 return -ENOMEM;
1338
1339         if (flags & MAP_FIXED)
1340                 return addr;
1341
1342         /* requesting a specific address */
1343         if (addr) {
1344                 addr = PAGE_ALIGN(addr);
1345                 vma = find_vma(mm, addr);
1346                 if (TASK_SIZE - len >= addr &&
1347                                 (!vma || addr + len <= vma->vm_start))
1348                         return addr;
1349         }
1350
1351         /* check if free_area_cache is useful for us */
1352         if (len <= mm->cached_hole_size) {
1353                 mm->cached_hole_size = 0;
1354                 mm->free_area_cache = mm->mmap_base;
1355         }
1356
1357         /* either no address requested or can't fit in requested address hole */
1358         addr = mm->free_area_cache;
1359
1360         /* make sure it can fit in the remaining address space */
1361         if (addr > len) {
1362                 vma = find_vma(mm, addr-len);
1363                 if (!vma || addr <= vma->vm_start)
1364                         /* remember the address as a hint for next time */
1365                         return (mm->free_area_cache = addr-len);
1366         }
1367
1368         if (mm->mmap_base < len)
1369                 goto bottomup;
1370
1371         addr = mm->mmap_base-len;
1372
1373         do {
1374                 /*
1375                  * Lookup failure means no vma is above this address,
1376                  * else if new region fits below vma->vm_start,
1377                  * return with success:
1378                  */
1379                 vma = find_vma(mm, addr);
1380                 if (!vma || addr+len <= vma->vm_start)
1381                         /* remember the address as a hint for next time */
1382                         return (mm->free_area_cache = addr);
1383
1384                 /* remember the largest hole we saw so far */
1385                 if (addr + mm->cached_hole_size < vma->vm_start)
1386                         mm->cached_hole_size = vma->vm_start - addr;
1387
1388                 /* try just below the current vma->vm_start */
1389                 addr = vma->vm_start-len;
1390         } while (len < vma->vm_start);
1391
1392 bottomup:
1393         /*
1394          * A failed mmap() very likely causes application failure,
1395          * so fall back to the bottom-up function here. This scenario
1396          * can happen with large stack limits and large mmap()
1397          * allocations.
1398          */
1399         mm->cached_hole_size = ~0UL;
1400         mm->free_area_cache = TASK_UNMAPPED_BASE;
1401         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1402         /*
1403          * Restore the topdown base:
1404          */
1405         mm->free_area_cache = mm->mmap_base;
1406         mm->cached_hole_size = ~0UL;
1407
1408         return addr;
1409 }
1410 #endif
1411
1412 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1413 {
1414         /*
1415          * Is this a new hole at the highest possible address?
1416          */
1417         if (addr > mm->free_area_cache)
1418                 mm->free_area_cache = addr;
1419
1420         /* dont allow allocations above current base */
1421         if (mm->free_area_cache > mm->mmap_base)
1422                 mm->free_area_cache = mm->mmap_base;
1423 }
1424
1425 unsigned long
1426 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1427                 unsigned long pgoff, unsigned long flags)
1428 {
1429         unsigned long (*get_area)(struct file *, unsigned long,
1430                                   unsigned long, unsigned long, unsigned long);
1431
1432         get_area = current->mm->get_unmapped_area;
1433         if (file && file->f_op && file->f_op->get_unmapped_area)
1434                 get_area = file->f_op->get_unmapped_area;
1435         addr = get_area(file, addr, len, pgoff, flags);
1436         if (IS_ERR_VALUE(addr))
1437                 return addr;
1438
1439         if (addr > TASK_SIZE - len)
1440                 return -ENOMEM;
1441         if (addr & ~PAGE_MASK)
1442                 return -EINVAL;
1443
1444         return arch_rebalance_pgtables(addr, len);
1445 }
1446
1447 EXPORT_SYMBOL(get_unmapped_area);
1448
1449 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1450 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1451 {
1452         struct vm_area_struct *vma = NULL;
1453
1454         if (mm) {
1455                 /* Check the cache first. */
1456                 /* (Cache hit rate is typically around 35%.) */
1457                 vma = mm->mmap_cache;
1458                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1459                         struct rb_node * rb_node;
1460
1461                         rb_node = mm->mm_rb.rb_node;
1462                         vma = NULL;
1463
1464                         while (rb_node) {
1465                                 struct vm_area_struct * vma_tmp;
1466
1467                                 vma_tmp = rb_entry(rb_node,
1468                                                 struct vm_area_struct, vm_rb);
1469
1470                                 if (vma_tmp->vm_end > addr) {
1471                                         vma = vma_tmp;
1472                                         if (vma_tmp->vm_start <= addr)
1473                                                 break;
1474                                         rb_node = rb_node->rb_left;
1475                                 } else
1476                                         rb_node = rb_node->rb_right;
1477                         }
1478                         if (vma)
1479                                 mm->mmap_cache = vma;
1480                 }
1481         }
1482         return vma;
1483 }
1484
1485 EXPORT_SYMBOL(find_vma);
1486
1487 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1488 struct vm_area_struct *
1489 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1490                         struct vm_area_struct **pprev)
1491 {
1492         struct vm_area_struct *vma = NULL, *prev = NULL;
1493         struct rb_node * rb_node;
1494         if (!mm)
1495                 goto out;
1496
1497         /* Guard against addr being lower than the first VMA */
1498         vma = mm->mmap;
1499
1500         /* Go through the RB tree quickly. */
1501         rb_node = mm->mm_rb.rb_node;
1502
1503         while (rb_node) {
1504                 struct vm_area_struct *vma_tmp;
1505                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1506
1507                 if (addr < vma_tmp->vm_end) {
1508                         rb_node = rb_node->rb_left;
1509                 } else {
1510                         prev = vma_tmp;
1511                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1512                                 break;
1513                         rb_node = rb_node->rb_right;
1514                 }
1515         }
1516
1517 out:
1518         *pprev = prev;
1519         return prev ? prev->vm_next : vma;
1520 }
1521
1522 /*
1523  * Verify that the stack growth is acceptable and
1524  * update accounting. This is shared with both the
1525  * grow-up and grow-down cases.
1526  */
1527 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1528 {
1529         struct mm_struct *mm = vma->vm_mm;
1530         struct rlimit *rlim = current->signal->rlim;
1531         unsigned long new_start;
1532
1533         /* address space limit tests */
1534         if (!may_expand_vm(mm, grow))
1535                 return -ENOMEM;
1536
1537         /* Stack limit test */
1538         if (size > rlim[RLIMIT_STACK].rlim_cur)
1539                 return -ENOMEM;
1540
1541         /* mlock limit tests */
1542         if (vma->vm_flags & VM_LOCKED) {
1543                 unsigned long locked;
1544                 unsigned long limit;
1545                 locked = mm->locked_vm + grow;
1546                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1547                 if (locked > limit && !capable(CAP_IPC_LOCK))
1548                         return -ENOMEM;
1549         }
1550
1551         /* Check to ensure the stack will not grow into a hugetlb-only region */
1552         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1553                         vma->vm_end - size;
1554         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1555                 return -EFAULT;
1556
1557         /*
1558          * Overcommit..  This must be the final test, as it will
1559          * update security statistics.
1560          */
1561         if (security_vm_enough_memory(grow))
1562                 return -ENOMEM;
1563
1564         /* Ok, everything looks good - let it rip */
1565         mm->total_vm += grow;
1566         if (vma->vm_flags & VM_LOCKED)
1567                 mm->locked_vm += grow;
1568         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1569         return 0;
1570 }
1571
1572 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1573 /*
1574  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1575  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1576  */
1577 #ifndef CONFIG_IA64
1578 static inline
1579 #endif
1580 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1581 {
1582         int error;
1583
1584         if (!(vma->vm_flags & VM_GROWSUP))
1585                 return -EFAULT;
1586
1587         /*
1588          * We must make sure the anon_vma is allocated
1589          * so that the anon_vma locking is not a noop.
1590          */
1591         if (unlikely(anon_vma_prepare(vma)))
1592                 return -ENOMEM;
1593         anon_vma_lock(vma);
1594
1595         /*
1596          * vma->vm_start/vm_end cannot change under us because the caller
1597          * is required to hold the mmap_sem in read mode.  We need the
1598          * anon_vma lock to serialize against concurrent expand_stacks.
1599          * Also guard against wrapping around to address 0.
1600          */
1601         if (address < PAGE_ALIGN(address+4))
1602                 address = PAGE_ALIGN(address+4);
1603         else {
1604                 anon_vma_unlock(vma);
1605                 return -ENOMEM;
1606         }
1607         error = 0;
1608
1609         /* Somebody else might have raced and expanded it already */
1610         if (address > vma->vm_end) {
1611                 unsigned long size, grow;
1612
1613                 size = address - vma->vm_start;
1614                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1615
1616                 error = acct_stack_growth(vma, size, grow);
1617                 if (!error)
1618                         vma->vm_end = address;
1619         }
1620         anon_vma_unlock(vma);
1621         return error;
1622 }
1623 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1624
1625 /*
1626  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1627  */
1628 static inline int expand_downwards(struct vm_area_struct *vma,
1629                                    unsigned long address)
1630 {
1631         int error;
1632
1633         /*
1634          * We must make sure the anon_vma is allocated
1635          * so that the anon_vma locking is not a noop.
1636          */
1637         if (unlikely(anon_vma_prepare(vma)))
1638                 return -ENOMEM;
1639
1640         address &= PAGE_MASK;
1641         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1642         if (error)
1643                 return error;
1644
1645         anon_vma_lock(vma);
1646
1647         /*
1648          * vma->vm_start/vm_end cannot change under us because the caller
1649          * is required to hold the mmap_sem in read mode.  We need the
1650          * anon_vma lock to serialize against concurrent expand_stacks.
1651          */
1652
1653         /* Somebody else might have raced and expanded it already */
1654         if (address < vma->vm_start) {
1655                 unsigned long size, grow;
1656
1657                 size = vma->vm_end - address;
1658                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1659
1660                 error = acct_stack_growth(vma, size, grow);
1661                 if (!error) {
1662                         vma->vm_start = address;
1663                         vma->vm_pgoff -= grow;
1664                 }
1665         }
1666         anon_vma_unlock(vma);
1667         return error;
1668 }
1669
1670 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1671 {
1672         return expand_downwards(vma, address);
1673 }
1674
1675 #ifdef CONFIG_STACK_GROWSUP
1676 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1677 {
1678         return expand_upwards(vma, address);
1679 }
1680
1681 struct vm_area_struct *
1682 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1683 {
1684         struct vm_area_struct *vma, *prev;
1685
1686         addr &= PAGE_MASK;
1687         vma = find_vma_prev(mm, addr, &prev);
1688         if (vma && (vma->vm_start <= addr))
1689                 return vma;
1690         if (!prev || expand_stack(prev, addr))
1691                 return NULL;
1692         if (prev->vm_flags & VM_LOCKED)
1693                 make_pages_present(addr, prev->vm_end);
1694         return prev;
1695 }
1696 #else
1697 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1698 {
1699         return expand_downwards(vma, address);
1700 }
1701
1702 struct vm_area_struct *
1703 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1704 {
1705         struct vm_area_struct * vma;
1706         unsigned long start;
1707
1708         addr &= PAGE_MASK;
1709         vma = find_vma(mm,addr);
1710         if (!vma)
1711                 return NULL;
1712         if (vma->vm_start <= addr)
1713                 return vma;
1714         if (!(vma->vm_flags & VM_GROWSDOWN))
1715                 return NULL;
1716         start = vma->vm_start;
1717         if (expand_stack(vma, addr))
1718                 return NULL;
1719         if (vma->vm_flags & VM_LOCKED)
1720                 make_pages_present(addr, start);
1721         return vma;
1722 }
1723 #endif
1724
1725 /*
1726  * Ok - we have the memory areas we should free on the vma list,
1727  * so release them, and do the vma updates.
1728  *
1729  * Called with the mm semaphore held.
1730  */
1731 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1732 {
1733         /* Update high watermark before we lower total_vm */
1734         update_hiwater_vm(mm);
1735         do {
1736                 long nrpages = vma_pages(vma);
1737
1738                 mm->total_vm -= nrpages;
1739                 if (vma->vm_flags & VM_LOCKED)
1740                         mm->locked_vm -= nrpages;
1741                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1742                 vma = remove_vma(vma);
1743         } while (vma);
1744         validate_mm(mm);
1745 }
1746
1747 /*
1748  * Get rid of page table information in the indicated region.
1749  *
1750  * Called with the mm semaphore held.
1751  */
1752 static void unmap_region(struct mm_struct *mm,
1753                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1754                 unsigned long start, unsigned long end)
1755 {
1756         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1757         struct mmu_gather *tlb;
1758         unsigned long nr_accounted = 0;
1759
1760         lru_add_drain();
1761         tlb = tlb_gather_mmu(mm, 0);
1762         update_hiwater_rss(mm);
1763         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1764         vm_unacct_memory(nr_accounted);
1765         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1766                                  next? next->vm_start: 0);
1767         tlb_finish_mmu(tlb, start, end);
1768 }
1769
1770 /*
1771  * Create a list of vma's touched by the unmap, removing them from the mm's
1772  * vma list as we go..
1773  */
1774 static void
1775 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1776         struct vm_area_struct *prev, unsigned long end)
1777 {
1778         struct vm_area_struct **insertion_point;
1779         struct vm_area_struct *tail_vma = NULL;
1780         unsigned long addr;
1781
1782         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1783         do {
1784                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1785                 mm->map_count--;
1786                 tail_vma = vma;
1787                 vma = vma->vm_next;
1788         } while (vma && vma->vm_start < end);
1789         *insertion_point = vma;
1790         tail_vma->vm_next = NULL;
1791         if (mm->unmap_area == arch_unmap_area)
1792                 addr = prev ? prev->vm_end : mm->mmap_base;
1793         else
1794                 addr = vma ?  vma->vm_start : mm->mmap_base;
1795         mm->unmap_area(mm, addr);
1796         mm->mmap_cache = NULL;          /* Kill the cache. */
1797 }
1798
1799 /*
1800  * Split a vma into two pieces at address 'addr', a new vma is allocated
1801  * either for the first part or the tail.
1802  */
1803 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1804               unsigned long addr, int new_below)
1805 {
1806         struct mempolicy *pol;
1807         struct vm_area_struct *new;
1808
1809         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1810                 return -EINVAL;
1811
1812         if (mm->map_count >= sysctl_max_map_count)
1813                 return -ENOMEM;
1814
1815         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1816         if (!new)
1817                 return -ENOMEM;
1818
1819         /* most fields are the same, copy all, and then fixup */
1820         *new = *vma;
1821
1822         if (new_below)
1823                 new->vm_end = addr;
1824         else {
1825                 new->vm_start = addr;
1826                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1827         }
1828
1829         pol = mpol_dup(vma_policy(vma));
1830         if (IS_ERR(pol)) {
1831                 kmem_cache_free(vm_area_cachep, new);
1832                 return PTR_ERR(pol);
1833         }
1834         vma_set_policy(new, pol);
1835
1836         if (new->vm_file) {
1837                 get_file(new->vm_file);
1838                 if (vma->vm_flags & VM_EXECUTABLE)
1839                         added_exe_file_vma(mm);
1840         }
1841
1842         if (new->vm_ops && new->vm_ops->open)
1843                 new->vm_ops->open(new);
1844
1845         if (new_below)
1846                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1847                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1848         else
1849                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1850
1851         return 0;
1852 }
1853
1854 /* Munmap is split into 2 main parts -- this part which finds
1855  * what needs doing, and the areas themselves, which do the
1856  * work.  This now handles partial unmappings.
1857  * Jeremy Fitzhardinge <jeremy@goop.org>
1858  */
1859 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1860 {
1861         unsigned long end;
1862         struct vm_area_struct *vma, *prev, *last;
1863
1864         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1865                 return -EINVAL;
1866
1867         if ((len = PAGE_ALIGN(len)) == 0)
1868                 return -EINVAL;
1869
1870         /* Find the first overlapping VMA */
1871         vma = find_vma_prev(mm, start, &prev);
1872         if (!vma)
1873                 return 0;
1874         /* we have  start < vma->vm_end  */
1875
1876         /* if it doesn't overlap, we have nothing.. */
1877         end = start + len;
1878         if (vma->vm_start >= end)
1879                 return 0;
1880
1881         /*
1882          * If we need to split any vma, do it now to save pain later.
1883          *
1884          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1885          * unmapped vm_area_struct will remain in use: so lower split_vma
1886          * places tmp vma above, and higher split_vma places tmp vma below.
1887          */
1888         if (start > vma->vm_start) {
1889                 int error = split_vma(mm, vma, start, 0);
1890                 if (error)
1891                         return error;
1892                 prev = vma;
1893         }
1894
1895         /* Does it split the last one? */
1896         last = find_vma(mm, end);
1897         if (last && end > last->vm_start) {
1898                 int error = split_vma(mm, last, end, 1);
1899                 if (error)
1900                         return error;
1901         }
1902         vma = prev? prev->vm_next: mm->mmap;
1903
1904         /*
1905          * Remove the vma's, and unmap the actual pages
1906          */
1907         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1908         unmap_region(mm, vma, prev, start, end);
1909
1910         /* Fix up all other VM information */
1911         remove_vma_list(mm, vma);
1912
1913         return 0;
1914 }
1915
1916 EXPORT_SYMBOL(do_munmap);
1917
1918 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1919 {
1920         int ret;
1921         struct mm_struct *mm = current->mm;
1922
1923         profile_munmap(addr);
1924
1925         down_write(&mm->mmap_sem);
1926         ret = do_munmap(mm, addr, len);
1927         up_write(&mm->mmap_sem);
1928         return ret;
1929 }
1930
1931 static inline void verify_mm_writelocked(struct mm_struct *mm)
1932 {
1933 #ifdef CONFIG_DEBUG_VM
1934         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1935                 WARN_ON(1);
1936                 up_read(&mm->mmap_sem);
1937         }
1938 #endif
1939 }
1940
1941 /*
1942  *  this is really a simplified "do_mmap".  it only handles
1943  *  anonymous maps.  eventually we may be able to do some
1944  *  brk-specific accounting here.
1945  */
1946 unsigned long do_brk(unsigned long addr, unsigned long len)
1947 {
1948         struct mm_struct * mm = current->mm;
1949         struct vm_area_struct * vma, * prev;
1950         unsigned long flags;
1951         struct rb_node ** rb_link, * rb_parent;
1952         pgoff_t pgoff = addr >> PAGE_SHIFT;
1953         int error;
1954
1955         len = PAGE_ALIGN(len);
1956         if (!len)
1957                 return addr;
1958
1959         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1960                 return -EINVAL;
1961
1962         if (is_hugepage_only_range(mm, addr, len))
1963                 return -EINVAL;
1964
1965         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1966         if (error)
1967                 return error;
1968
1969         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1970
1971         error = arch_mmap_check(addr, len, flags);
1972         if (error)
1973                 return error;
1974
1975         /*
1976          * mlock MCL_FUTURE?
1977          */
1978         if (mm->def_flags & VM_LOCKED) {
1979                 unsigned long locked, lock_limit;
1980                 locked = len >> PAGE_SHIFT;
1981                 locked += mm->locked_vm;
1982                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1983                 lock_limit >>= PAGE_SHIFT;
1984                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1985                         return -EAGAIN;
1986         }
1987
1988         /*
1989          * mm->mmap_sem is required to protect against another thread
1990          * changing the mappings in case we sleep.
1991          */
1992         verify_mm_writelocked(mm);
1993
1994         /*
1995          * Clear old maps.  this also does some error checking for us
1996          */
1997  munmap_back:
1998         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1999         if (vma && vma->vm_start < addr + len) {
2000                 if (do_munmap(mm, addr, len))
2001                         return -ENOMEM;
2002                 goto munmap_back;
2003         }
2004
2005         /* Check against address space limits *after* clearing old maps... */
2006         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2007                 return -ENOMEM;
2008
2009         if (mm->map_count > sysctl_max_map_count)
2010                 return -ENOMEM;
2011
2012         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2013                 return -ENOMEM;
2014
2015         /* Can we just expand an old private anonymous mapping? */
2016         if (vma_merge(mm, prev, addr, addr + len, flags,
2017                                         NULL, NULL, pgoff, NULL))
2018                 goto out;
2019
2020         /*
2021          * create a vma struct for an anonymous mapping
2022          */
2023         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2024         if (!vma) {
2025                 vm_unacct_memory(len >> PAGE_SHIFT);
2026                 return -ENOMEM;
2027         }
2028
2029         vma->vm_mm = mm;
2030         vma->vm_start = addr;
2031         vma->vm_end = addr + len;
2032         vma->vm_pgoff = pgoff;
2033         vma->vm_flags = flags;
2034         vma->vm_page_prot = vm_get_page_prot(flags);
2035         vma_link(mm, vma, prev, rb_link, rb_parent);
2036 out:
2037         mm->total_vm += len >> PAGE_SHIFT;
2038         if (flags & VM_LOCKED) {
2039                 mm->locked_vm += len >> PAGE_SHIFT;
2040                 make_pages_present(addr, addr + len);
2041         }
2042         return addr;
2043 }
2044
2045 EXPORT_SYMBOL(do_brk);
2046
2047 /* Release all mmaps. */
2048 void exit_mmap(struct mm_struct *mm)
2049 {
2050         struct mmu_gather *tlb;
2051         struct vm_area_struct *vma = mm->mmap;
2052         unsigned long nr_accounted = 0;
2053         unsigned long end;
2054
2055         /* mm's last user has gone, and its about to be pulled down */
2056         arch_exit_mmap(mm);
2057
2058         lru_add_drain();
2059         flush_cache_mm(mm);
2060         tlb = tlb_gather_mmu(mm, 1);
2061         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2062         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2063         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2064         vm_unacct_memory(nr_accounted);
2065         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2066         tlb_finish_mmu(tlb, 0, end);
2067
2068         /*
2069          * Walk the list again, actually closing and freeing it,
2070          * with preemption enabled, without holding any MM locks.
2071          */
2072         while (vma)
2073                 vma = remove_vma(vma);
2074
2075         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2076 }
2077
2078 /* Insert vm structure into process list sorted by address
2079  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2080  * then i_mmap_lock is taken here.
2081  */
2082 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2083 {
2084         struct vm_area_struct * __vma, * prev;
2085         struct rb_node ** rb_link, * rb_parent;
2086
2087         /*
2088          * The vm_pgoff of a purely anonymous vma should be irrelevant
2089          * until its first write fault, when page's anon_vma and index
2090          * are set.  But now set the vm_pgoff it will almost certainly
2091          * end up with (unless mremap moves it elsewhere before that
2092          * first wfault), so /proc/pid/maps tells a consistent story.
2093          *
2094          * By setting it to reflect the virtual start address of the
2095          * vma, merges and splits can happen in a seamless way, just
2096          * using the existing file pgoff checks and manipulations.
2097          * Similarly in do_mmap_pgoff and in do_brk.
2098          */
2099         if (!vma->vm_file) {
2100                 BUG_ON(vma->anon_vma);
2101                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2102         }
2103         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2104         if (__vma && __vma->vm_start < vma->vm_end)
2105                 return -ENOMEM;
2106         if ((vma->vm_flags & VM_ACCOUNT) &&
2107              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2108                 return -ENOMEM;
2109         vma_link(mm, vma, prev, rb_link, rb_parent);
2110         return 0;
2111 }
2112
2113 /*
2114  * Copy the vma structure to a new location in the same mm,
2115  * prior to moving page table entries, to effect an mremap move.
2116  */
2117 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2118         unsigned long addr, unsigned long len, pgoff_t pgoff)
2119 {
2120         struct vm_area_struct *vma = *vmap;
2121         unsigned long vma_start = vma->vm_start;
2122         struct mm_struct *mm = vma->vm_mm;
2123         struct vm_area_struct *new_vma, *prev;
2124         struct rb_node **rb_link, *rb_parent;
2125         struct mempolicy *pol;
2126
2127         /*
2128          * If anonymous vma has not yet been faulted, update new pgoff
2129          * to match new location, to increase its chance of merging.
2130          */
2131         if (!vma->vm_file && !vma->anon_vma)
2132                 pgoff = addr >> PAGE_SHIFT;
2133
2134         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2135         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2136                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2137         if (new_vma) {
2138                 /*
2139                  * Source vma may have been merged into new_vma
2140                  */
2141                 if (vma_start >= new_vma->vm_start &&
2142                     vma_start < new_vma->vm_end)
2143                         *vmap = new_vma;
2144         } else {
2145                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2146                 if (new_vma) {
2147                         *new_vma = *vma;
2148                         pol = mpol_dup(vma_policy(vma));
2149                         if (IS_ERR(pol)) {
2150                                 kmem_cache_free(vm_area_cachep, new_vma);
2151                                 return NULL;
2152                         }
2153                         vma_set_policy(new_vma, pol);
2154                         new_vma->vm_start = addr;
2155                         new_vma->vm_end = addr + len;
2156                         new_vma->vm_pgoff = pgoff;
2157                         if (new_vma->vm_file) {
2158                                 get_file(new_vma->vm_file);
2159                                 if (vma->vm_flags & VM_EXECUTABLE)
2160                                         added_exe_file_vma(mm);
2161                         }
2162                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2163                                 new_vma->vm_ops->open(new_vma);
2164                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2165                 }
2166         }
2167         return new_vma;
2168 }
2169
2170 /*
2171  * Return true if the calling process may expand its vm space by the passed
2172  * number of pages
2173  */
2174 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2175 {
2176         unsigned long cur = mm->total_vm;       /* pages */
2177         unsigned long lim;
2178
2179         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2180
2181         if (cur + npages > lim)
2182                 return 0;
2183         return 1;
2184 }
2185
2186
2187 static int special_mapping_fault(struct vm_area_struct *vma,
2188                                 struct vm_fault *vmf)
2189 {
2190         pgoff_t pgoff;
2191         struct page **pages;
2192
2193         /*
2194          * special mappings have no vm_file, and in that case, the mm
2195          * uses vm_pgoff internally. So we have to subtract it from here.
2196          * We are allowed to do this because we are the mm; do not copy
2197          * this code into drivers!
2198          */
2199         pgoff = vmf->pgoff - vma->vm_pgoff;
2200
2201         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2202                 pgoff--;
2203
2204         if (*pages) {
2205                 struct page *page = *pages;
2206                 get_page(page);
2207                 vmf->page = page;
2208                 return 0;
2209         }
2210
2211         return VM_FAULT_SIGBUS;
2212 }
2213
2214 /*
2215  * Having a close hook prevents vma merging regardless of flags.
2216  */
2217 static void special_mapping_close(struct vm_area_struct *vma)
2218 {
2219 }
2220
2221 static struct vm_operations_struct special_mapping_vmops = {
2222         .close = special_mapping_close,
2223         .fault = special_mapping_fault,
2224 };
2225
2226 /*
2227  * Called with mm->mmap_sem held for writing.
2228  * Insert a new vma covering the given region, with the given flags.
2229  * Its pages are supplied by the given array of struct page *.
2230  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2231  * The region past the last page supplied will always produce SIGBUS.
2232  * The array pointer and the pages it points to are assumed to stay alive
2233  * for as long as this mapping might exist.
2234  */
2235 int install_special_mapping(struct mm_struct *mm,
2236                             unsigned long addr, unsigned long len,
2237                             unsigned long vm_flags, struct page **pages)
2238 {
2239         struct vm_area_struct *vma;
2240
2241         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2242         if (unlikely(vma == NULL))
2243                 return -ENOMEM;
2244
2245         vma->vm_mm = mm;
2246         vma->vm_start = addr;
2247         vma->vm_end = addr + len;
2248
2249         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2250         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2251
2252         vma->vm_ops = &special_mapping_vmops;
2253         vma->vm_private_data = pages;
2254
2255         if (unlikely(insert_vm_struct(mm, vma))) {
2256                 kmem_cache_free(vm_area_cachep, vma);
2257                 return -ENOMEM;
2258         }
2259
2260         mm->total_vm += len >> PAGE_SHIFT;
2261
2262         return 0;
2263 }