thp: avoid breaking huge pmd invariants in case of vma_adjust failures
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
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 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33
34 #include <asm/uaccess.h>
35 #include <asm/cacheflush.h>
36 #include <asm/tlb.h>
37 #include <asm/mmu_context.h>
38
39 #include "internal.h"
40
41 #ifndef arch_mmap_check
42 #define arch_mmap_check(addr, len, flags)       (0)
43 #endif
44
45 #ifndef arch_rebalance_pgtables
46 #define arch_rebalance_pgtables(addr, len)              (addr)
47 #endif
48
49 static void unmap_region(struct mm_struct *mm,
50                 struct vm_area_struct *vma, struct vm_area_struct *prev,
51                 unsigned long start, unsigned long end);
52
53 /*
54  * WARNING: the debugging will use recursive algorithms so never enable this
55  * unless you know what you are doing.
56  */
57 #undef DEBUG_MM_RB
58
59 /* description of effects of mapping type and prot in current implementation.
60  * this is due to the limited x86 page protection hardware.  The expected
61  * behavior is in parens:
62  *
63  * map_type     prot
64  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
65  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *              
69  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
70  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
71  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
72  *
73  */
74 pgprot_t protection_map[16] = {
75         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77 };
78
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
80 {
81         return __pgprot(pgprot_val(protection_map[vm_flags &
82                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
83                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
84 }
85 EXPORT_SYMBOL(vm_get_page_prot);
86
87 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
88 int sysctl_overcommit_ratio = 50;       /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90 struct percpu_counter vm_committed_as;
91
92 /*
93  * Check that a process has enough memory to allocate a new virtual
94  * mapping. 0 means there is enough memory for the allocation to
95  * succeed and -ENOMEM implies there is not.
96  *
97  * We currently support three overcommit policies, which are set via the
98  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
99  *
100  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
101  * Additional code 2002 Jul 20 by Robert Love.
102  *
103  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
104  *
105  * Note this is a helper function intended to be used by LSMs which
106  * wish to use this logic.
107  */
108 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
109 {
110         unsigned long free, allowed;
111
112         vm_acct_memory(pages);
113
114         /*
115          * Sometimes we want to use more memory than we have
116          */
117         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
118                 return 0;
119
120         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
121                 unsigned long n;
122
123                 free = global_page_state(NR_FILE_PAGES);
124                 free += nr_swap_pages;
125
126                 /*
127                  * Any slabs which are created with the
128                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
129                  * which are reclaimable, under pressure.  The dentry
130                  * cache and most inode caches should fall into this
131                  */
132                 free += global_page_state(NR_SLAB_RECLAIMABLE);
133
134                 /*
135                  * Leave the last 3% for root
136                  */
137                 if (!cap_sys_admin)
138                         free -= free / 32;
139
140                 if (free > pages)
141                         return 0;
142
143                 /*
144                  * nr_free_pages() is very expensive on large systems,
145                  * only call if we're about to fail.
146                  */
147                 n = nr_free_pages();
148
149                 /*
150                  * Leave reserved pages. The pages are not for anonymous pages.
151                  */
152                 if (n <= totalreserve_pages)
153                         goto error;
154                 else
155                         n -= totalreserve_pages;
156
157                 /*
158                  * Leave the last 3% for root
159                  */
160                 if (!cap_sys_admin)
161                         n -= n / 32;
162                 free += n;
163
164                 if (free > pages)
165                         return 0;
166
167                 goto error;
168         }
169
170         allowed = (totalram_pages - hugetlb_total_pages())
171                 * sysctl_overcommit_ratio / 100;
172         /*
173          * Leave the last 3% for root
174          */
175         if (!cap_sys_admin)
176                 allowed -= allowed / 32;
177         allowed += total_swap_pages;
178
179         /* Don't let a single process grow too big:
180            leave 3% of the size of this process for other processes */
181         if (mm)
182                 allowed -= mm->total_vm / 32;
183
184         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
185                 return 0;
186 error:
187         vm_unacct_memory(pages);
188
189         return -ENOMEM;
190 }
191
192 /*
193  * Requires inode->i_mapping->i_mmap_lock
194  */
195 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
196                 struct file *file, struct address_space *mapping)
197 {
198         if (vma->vm_flags & VM_DENYWRITE)
199                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
200         if (vma->vm_flags & VM_SHARED)
201                 mapping->i_mmap_writable--;
202
203         flush_dcache_mmap_lock(mapping);
204         if (unlikely(vma->vm_flags & VM_NONLINEAR))
205                 list_del_init(&vma->shared.vm_set.list);
206         else
207                 vma_prio_tree_remove(vma, &mapping->i_mmap);
208         flush_dcache_mmap_unlock(mapping);
209 }
210
211 /*
212  * Unlink a file-based vm structure from its prio_tree, to hide
213  * vma from rmap and vmtruncate before freeing its page tables.
214  */
215 void unlink_file_vma(struct vm_area_struct *vma)
216 {
217         struct file *file = vma->vm_file;
218
219         if (file) {
220                 struct address_space *mapping = file->f_mapping;
221                 spin_lock(&mapping->i_mmap_lock);
222                 __remove_shared_vm_struct(vma, file, mapping);
223                 spin_unlock(&mapping->i_mmap_lock);
224         }
225 }
226
227 /*
228  * Close a vm structure and free it, returning the next.
229  */
230 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
231 {
232         struct vm_area_struct *next = vma->vm_next;
233
234         might_sleep();
235         if (vma->vm_ops && vma->vm_ops->close)
236                 vma->vm_ops->close(vma);
237         if (vma->vm_file) {
238                 fput(vma->vm_file);
239                 if (vma->vm_flags & VM_EXECUTABLE)
240                         removed_exe_file_vma(vma->vm_mm);
241         }
242         mpol_put(vma_policy(vma));
243         kmem_cache_free(vm_area_cachep, vma);
244         return next;
245 }
246
247 SYSCALL_DEFINE1(brk, unsigned long, brk)
248 {
249         unsigned long rlim, retval;
250         unsigned long newbrk, oldbrk;
251         struct mm_struct *mm = current->mm;
252         unsigned long min_brk;
253
254         down_write(&mm->mmap_sem);
255
256 #ifdef CONFIG_COMPAT_BRK
257         min_brk = mm->end_code;
258 #else
259         min_brk = mm->start_brk;
260 #endif
261         if (brk < min_brk)
262                 goto out;
263
264         /*
265          * Check against rlimit here. If this check is done later after the test
266          * of oldbrk with newbrk then it can escape the test and let the data
267          * segment grow beyond its set limit the in case where the limit is
268          * not page aligned -Ram Gupta
269          */
270         rlim = rlimit(RLIMIT_DATA);
271         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
272                         (mm->end_data - mm->start_data) > rlim)
273                 goto out;
274
275         newbrk = PAGE_ALIGN(brk);
276         oldbrk = PAGE_ALIGN(mm->brk);
277         if (oldbrk == newbrk)
278                 goto set_brk;
279
280         /* Always allow shrinking brk. */
281         if (brk <= mm->brk) {
282                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
283                         goto set_brk;
284                 goto out;
285         }
286
287         /* Check against existing mmap mappings. */
288         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
289                 goto out;
290
291         /* Ok, looks good - let it rip. */
292         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
293                 goto out;
294 set_brk:
295         mm->brk = brk;
296 out:
297         retval = mm->brk;
298         up_write(&mm->mmap_sem);
299         return retval;
300 }
301
302 #ifdef DEBUG_MM_RB
303 static int browse_rb(struct rb_root *root)
304 {
305         int i = 0, j;
306         struct rb_node *nd, *pn = NULL;
307         unsigned long prev = 0, pend = 0;
308
309         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
310                 struct vm_area_struct *vma;
311                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
312                 if (vma->vm_start < prev)
313                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
314                 if (vma->vm_start < pend)
315                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
316                 if (vma->vm_start > vma->vm_end)
317                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
318                 i++;
319                 pn = nd;
320                 prev = vma->vm_start;
321                 pend = vma->vm_end;
322         }
323         j = 0;
324         for (nd = pn; nd; nd = rb_prev(nd)) {
325                 j++;
326         }
327         if (i != j)
328                 printk("backwards %d, forwards %d\n", j, i), i = 0;
329         return i;
330 }
331
332 void validate_mm(struct mm_struct *mm)
333 {
334         int bug = 0;
335         int i = 0;
336         struct vm_area_struct *tmp = mm->mmap;
337         while (tmp) {
338                 tmp = tmp->vm_next;
339                 i++;
340         }
341         if (i != mm->map_count)
342                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
343         i = browse_rb(&mm->mm_rb);
344         if (i != mm->map_count)
345                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
346         BUG_ON(bug);
347 }
348 #else
349 #define validate_mm(mm) do { } while (0)
350 #endif
351
352 static struct vm_area_struct *
353 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
354                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
355                 struct rb_node ** rb_parent)
356 {
357         struct vm_area_struct * vma;
358         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
359
360         __rb_link = &mm->mm_rb.rb_node;
361         rb_prev = __rb_parent = NULL;
362         vma = NULL;
363
364         while (*__rb_link) {
365                 struct vm_area_struct *vma_tmp;
366
367                 __rb_parent = *__rb_link;
368                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
369
370                 if (vma_tmp->vm_end > addr) {
371                         vma = vma_tmp;
372                         if (vma_tmp->vm_start <= addr)
373                                 break;
374                         __rb_link = &__rb_parent->rb_left;
375                 } else {
376                         rb_prev = __rb_parent;
377                         __rb_link = &__rb_parent->rb_right;
378                 }
379         }
380
381         *pprev = NULL;
382         if (rb_prev)
383                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
384         *rb_link = __rb_link;
385         *rb_parent = __rb_parent;
386         return vma;
387 }
388
389 static inline void
390 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
391                 struct vm_area_struct *prev, struct rb_node *rb_parent)
392 {
393         struct vm_area_struct *next;
394
395         vma->vm_prev = prev;
396         if (prev) {
397                 next = prev->vm_next;
398                 prev->vm_next = vma;
399         } else {
400                 mm->mmap = vma;
401                 if (rb_parent)
402                         next = rb_entry(rb_parent,
403                                         struct vm_area_struct, vm_rb);
404                 else
405                         next = NULL;
406         }
407         vma->vm_next = next;
408         if (next)
409                 next->vm_prev = vma;
410 }
411
412 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
413                 struct rb_node **rb_link, struct rb_node *rb_parent)
414 {
415         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
416         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
417 }
418
419 static void __vma_link_file(struct vm_area_struct *vma)
420 {
421         struct file *file;
422
423         file = vma->vm_file;
424         if (file) {
425                 struct address_space *mapping = file->f_mapping;
426
427                 if (vma->vm_flags & VM_DENYWRITE)
428                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
429                 if (vma->vm_flags & VM_SHARED)
430                         mapping->i_mmap_writable++;
431
432                 flush_dcache_mmap_lock(mapping);
433                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
434                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
435                 else
436                         vma_prio_tree_insert(vma, &mapping->i_mmap);
437                 flush_dcache_mmap_unlock(mapping);
438         }
439 }
440
441 static void
442 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
443         struct vm_area_struct *prev, struct rb_node **rb_link,
444         struct rb_node *rb_parent)
445 {
446         __vma_link_list(mm, vma, prev, rb_parent);
447         __vma_link_rb(mm, vma, rb_link, rb_parent);
448 }
449
450 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
451                         struct vm_area_struct *prev, struct rb_node **rb_link,
452                         struct rb_node *rb_parent)
453 {
454         struct address_space *mapping = NULL;
455
456         if (vma->vm_file)
457                 mapping = vma->vm_file->f_mapping;
458
459         if (mapping) {
460                 spin_lock(&mapping->i_mmap_lock);
461                 vma->vm_truncate_count = mapping->truncate_count;
462         }
463
464         __vma_link(mm, vma, prev, rb_link, rb_parent);
465         __vma_link_file(vma);
466
467         if (mapping)
468                 spin_unlock(&mapping->i_mmap_lock);
469
470         mm->map_count++;
471         validate_mm(mm);
472 }
473
474 /*
475  * Helper for vma_adjust in the split_vma insert case:
476  * insert vm structure into list and rbtree and anon_vma,
477  * but it has already been inserted into prio_tree earlier.
478  */
479 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
480 {
481         struct vm_area_struct *__vma, *prev;
482         struct rb_node **rb_link, *rb_parent;
483
484         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
485         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
486         __vma_link(mm, vma, prev, rb_link, rb_parent);
487         mm->map_count++;
488 }
489
490 static inline void
491 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
492                 struct vm_area_struct *prev)
493 {
494         struct vm_area_struct *next = vma->vm_next;
495
496         prev->vm_next = next;
497         if (next)
498                 next->vm_prev = prev;
499         rb_erase(&vma->vm_rb, &mm->mm_rb);
500         if (mm->mmap_cache == vma)
501                 mm->mmap_cache = prev;
502 }
503
504 /*
505  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
506  * is already present in an i_mmap tree without adjusting the tree.
507  * The following helper function should be used when such adjustments
508  * are necessary.  The "insert" vma (if any) is to be inserted
509  * before we drop the necessary locks.
510  */
511 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
512         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
513 {
514         struct mm_struct *mm = vma->vm_mm;
515         struct vm_area_struct *next = vma->vm_next;
516         struct vm_area_struct *importer = NULL;
517         struct address_space *mapping = NULL;
518         struct prio_tree_root *root = NULL;
519         struct anon_vma *anon_vma = NULL;
520         struct file *file = vma->vm_file;
521         long adjust_next = 0;
522         int remove_next = 0;
523
524         if (next && !insert) {
525                 struct vm_area_struct *exporter = NULL;
526
527                 if (end >= next->vm_end) {
528                         /*
529                          * vma expands, overlapping all the next, and
530                          * perhaps the one after too (mprotect case 6).
531                          */
532 again:                  remove_next = 1 + (end > next->vm_end);
533                         end = next->vm_end;
534                         exporter = next;
535                         importer = vma;
536                 } else if (end > next->vm_start) {
537                         /*
538                          * vma expands, overlapping part of the next:
539                          * mprotect case 5 shifting the boundary up.
540                          */
541                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
542                         exporter = next;
543                         importer = vma;
544                 } else if (end < vma->vm_end) {
545                         /*
546                          * vma shrinks, and !insert tells it's not
547                          * split_vma inserting another: so it must be
548                          * mprotect case 4 shifting the boundary down.
549                          */
550                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
551                         exporter = vma;
552                         importer = next;
553                 }
554
555                 /*
556                  * Easily overlooked: when mprotect shifts the boundary,
557                  * make sure the expanding vma has anon_vma set if the
558                  * shrinking vma had, to cover any anon pages imported.
559                  */
560                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
561                         if (anon_vma_clone(importer, exporter))
562                                 return -ENOMEM;
563                         importer->anon_vma = exporter->anon_vma;
564                 }
565         }
566
567         if (file) {
568                 mapping = file->f_mapping;
569                 if (!(vma->vm_flags & VM_NONLINEAR))
570                         root = &mapping->i_mmap;
571                 spin_lock(&mapping->i_mmap_lock);
572                 if (importer &&
573                     vma->vm_truncate_count != next->vm_truncate_count) {
574                         /*
575                          * unmap_mapping_range might be in progress:
576                          * ensure that the expanding vma is rescanned.
577                          */
578                         importer->vm_truncate_count = 0;
579                 }
580                 if (insert) {
581                         insert->vm_truncate_count = vma->vm_truncate_count;
582                         /*
583                          * Put into prio_tree now, so instantiated pages
584                          * are visible to arm/parisc __flush_dcache_page
585                          * throughout; but we cannot insert into address
586                          * space until vma start or end is updated.
587                          */
588                         __vma_link_file(insert);
589                 }
590         }
591
592         vma_adjust_trans_huge(vma, start, end, adjust_next);
593
594         /*
595          * When changing only vma->vm_end, we don't really need anon_vma
596          * lock. This is a fairly rare case by itself, but the anon_vma
597          * lock may be shared between many sibling processes.  Skipping
598          * the lock for brk adjustments makes a difference sometimes.
599          */
600         if (vma->anon_vma && (insert || importer || start != vma->vm_start)) {
601                 anon_vma = vma->anon_vma;
602                 anon_vma_lock(anon_vma);
603         }
604
605         if (root) {
606                 flush_dcache_mmap_lock(mapping);
607                 vma_prio_tree_remove(vma, root);
608                 if (adjust_next)
609                         vma_prio_tree_remove(next, root);
610         }
611
612         vma->vm_start = start;
613         vma->vm_end = end;
614         vma->vm_pgoff = pgoff;
615         if (adjust_next) {
616                 next->vm_start += adjust_next << PAGE_SHIFT;
617                 next->vm_pgoff += adjust_next;
618         }
619
620         if (root) {
621                 if (adjust_next)
622                         vma_prio_tree_insert(next, root);
623                 vma_prio_tree_insert(vma, root);
624                 flush_dcache_mmap_unlock(mapping);
625         }
626
627         if (remove_next) {
628                 /*
629                  * vma_merge has merged next into vma, and needs
630                  * us to remove next before dropping the locks.
631                  */
632                 __vma_unlink(mm, next, vma);
633                 if (file)
634                         __remove_shared_vm_struct(next, file, mapping);
635         } else if (insert) {
636                 /*
637                  * split_vma has split insert from vma, and needs
638                  * us to insert it before dropping the locks
639                  * (it may either follow vma or precede it).
640                  */
641                 __insert_vm_struct(mm, insert);
642         }
643
644         if (anon_vma)
645                 anon_vma_unlock(anon_vma);
646         if (mapping)
647                 spin_unlock(&mapping->i_mmap_lock);
648
649         if (remove_next) {
650                 if (file) {
651                         fput(file);
652                         if (next->vm_flags & VM_EXECUTABLE)
653                                 removed_exe_file_vma(mm);
654                 }
655                 if (next->anon_vma)
656                         anon_vma_merge(vma, next);
657                 mm->map_count--;
658                 mpol_put(vma_policy(next));
659                 kmem_cache_free(vm_area_cachep, next);
660                 /*
661                  * In mprotect's case 6 (see comments on vma_merge),
662                  * we must remove another next too. It would clutter
663                  * up the code too much to do both in one go.
664                  */
665                 if (remove_next == 2) {
666                         next = vma->vm_next;
667                         goto again;
668                 }
669         }
670
671         validate_mm(mm);
672
673         return 0;
674 }
675
676 /*
677  * If the vma has a ->close operation then the driver probably needs to release
678  * per-vma resources, so we don't attempt to merge those.
679  */
680 static inline int is_mergeable_vma(struct vm_area_struct *vma,
681                         struct file *file, unsigned long vm_flags)
682 {
683         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
684         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
685                 return 0;
686         if (vma->vm_file != file)
687                 return 0;
688         if (vma->vm_ops && vma->vm_ops->close)
689                 return 0;
690         return 1;
691 }
692
693 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
694                                         struct anon_vma *anon_vma2)
695 {
696         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
697 }
698
699 /*
700  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
701  * in front of (at a lower virtual address and file offset than) the vma.
702  *
703  * We cannot merge two vmas if they have differently assigned (non-NULL)
704  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
705  *
706  * We don't check here for the merged mmap wrapping around the end of pagecache
707  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
708  * wrap, nor mmaps which cover the final page at index -1UL.
709  */
710 static int
711 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
712         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
713 {
714         if (is_mergeable_vma(vma, file, vm_flags) &&
715             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
716                 if (vma->vm_pgoff == vm_pgoff)
717                         return 1;
718         }
719         return 0;
720 }
721
722 /*
723  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
724  * beyond (at a higher virtual address and file offset than) the vma.
725  *
726  * We cannot merge two vmas if they have differently assigned (non-NULL)
727  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
728  */
729 static int
730 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
731         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
732 {
733         if (is_mergeable_vma(vma, file, vm_flags) &&
734             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
735                 pgoff_t vm_pglen;
736                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
737                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
738                         return 1;
739         }
740         return 0;
741 }
742
743 /*
744  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
745  * whether that can be merged with its predecessor or its successor.
746  * Or both (it neatly fills a hole).
747  *
748  * In most cases - when called for mmap, brk or mremap - [addr,end) is
749  * certain not to be mapped by the time vma_merge is called; but when
750  * called for mprotect, it is certain to be already mapped (either at
751  * an offset within prev, or at the start of next), and the flags of
752  * this area are about to be changed to vm_flags - and the no-change
753  * case has already been eliminated.
754  *
755  * The following mprotect cases have to be considered, where AAAA is
756  * the area passed down from mprotect_fixup, never extending beyond one
757  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
758  *
759  *     AAAA             AAAA                AAAA          AAAA
760  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
761  *    cannot merge    might become    might become    might become
762  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
763  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
764  *    mremap move:                                    PPPPNNNNNNNN 8
765  *        AAAA
766  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
767  *    might become    case 1 below    case 2 below    case 3 below
768  *
769  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
770  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
771  */
772 struct vm_area_struct *vma_merge(struct mm_struct *mm,
773                         struct vm_area_struct *prev, unsigned long addr,
774                         unsigned long end, unsigned long vm_flags,
775                         struct anon_vma *anon_vma, struct file *file,
776                         pgoff_t pgoff, struct mempolicy *policy)
777 {
778         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
779         struct vm_area_struct *area, *next;
780         int err;
781
782         /*
783          * We later require that vma->vm_flags == vm_flags,
784          * so this tests vma->vm_flags & VM_SPECIAL, too.
785          */
786         if (vm_flags & VM_SPECIAL)
787                 return NULL;
788
789         if (prev)
790                 next = prev->vm_next;
791         else
792                 next = mm->mmap;
793         area = next;
794         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
795                 next = next->vm_next;
796
797         /*
798          * Can it merge with the predecessor?
799          */
800         if (prev && prev->vm_end == addr &&
801                         mpol_equal(vma_policy(prev), policy) &&
802                         can_vma_merge_after(prev, vm_flags,
803                                                 anon_vma, file, pgoff)) {
804                 /*
805                  * OK, it can.  Can we now merge in the successor as well?
806                  */
807                 if (next && end == next->vm_start &&
808                                 mpol_equal(policy, vma_policy(next)) &&
809                                 can_vma_merge_before(next, vm_flags,
810                                         anon_vma, file, pgoff+pglen) &&
811                                 is_mergeable_anon_vma(prev->anon_vma,
812                                                       next->anon_vma)) {
813                                                         /* cases 1, 6 */
814                         err = vma_adjust(prev, prev->vm_start,
815                                 next->vm_end, prev->vm_pgoff, NULL);
816                 } else                                  /* cases 2, 5, 7 */
817                         err = vma_adjust(prev, prev->vm_start,
818                                 end, prev->vm_pgoff, NULL);
819                 if (err)
820                         return NULL;
821                 khugepaged_enter_vma_merge(prev);
822                 return prev;
823         }
824
825         /*
826          * Can this new request be merged in front of next?
827          */
828         if (next && end == next->vm_start &&
829                         mpol_equal(policy, vma_policy(next)) &&
830                         can_vma_merge_before(next, vm_flags,
831                                         anon_vma, file, pgoff+pglen)) {
832                 if (prev && addr < prev->vm_end)        /* case 4 */
833                         err = vma_adjust(prev, prev->vm_start,
834                                 addr, prev->vm_pgoff, NULL);
835                 else                                    /* cases 3, 8 */
836                         err = vma_adjust(area, addr, next->vm_end,
837                                 next->vm_pgoff - pglen, NULL);
838                 if (err)
839                         return NULL;
840                 khugepaged_enter_vma_merge(area);
841                 return area;
842         }
843
844         return NULL;
845 }
846
847 /*
848  * Rough compatbility check to quickly see if it's even worth looking
849  * at sharing an anon_vma.
850  *
851  * They need to have the same vm_file, and the flags can only differ
852  * in things that mprotect may change.
853  *
854  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
855  * we can merge the two vma's. For example, we refuse to merge a vma if
856  * there is a vm_ops->close() function, because that indicates that the
857  * driver is doing some kind of reference counting. But that doesn't
858  * really matter for the anon_vma sharing case.
859  */
860 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
861 {
862         return a->vm_end == b->vm_start &&
863                 mpol_equal(vma_policy(a), vma_policy(b)) &&
864                 a->vm_file == b->vm_file &&
865                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
866                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
867 }
868
869 /*
870  * Do some basic sanity checking to see if we can re-use the anon_vma
871  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
872  * the same as 'old', the other will be the new one that is trying
873  * to share the anon_vma.
874  *
875  * NOTE! This runs with mm_sem held for reading, so it is possible that
876  * the anon_vma of 'old' is concurrently in the process of being set up
877  * by another page fault trying to merge _that_. But that's ok: if it
878  * is being set up, that automatically means that it will be a singleton
879  * acceptable for merging, so we can do all of this optimistically. But
880  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
881  *
882  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
883  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
884  * is to return an anon_vma that is "complex" due to having gone through
885  * a fork).
886  *
887  * We also make sure that the two vma's are compatible (adjacent,
888  * and with the same memory policies). That's all stable, even with just
889  * a read lock on the mm_sem.
890  */
891 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
892 {
893         if (anon_vma_compatible(a, b)) {
894                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
895
896                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
897                         return anon_vma;
898         }
899         return NULL;
900 }
901
902 /*
903  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
904  * neighbouring vmas for a suitable anon_vma, before it goes off
905  * to allocate a new anon_vma.  It checks because a repetitive
906  * sequence of mprotects and faults may otherwise lead to distinct
907  * anon_vmas being allocated, preventing vma merge in subsequent
908  * mprotect.
909  */
910 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
911 {
912         struct anon_vma *anon_vma;
913         struct vm_area_struct *near;
914
915         near = vma->vm_next;
916         if (!near)
917                 goto try_prev;
918
919         anon_vma = reusable_anon_vma(near, vma, near);
920         if (anon_vma)
921                 return anon_vma;
922 try_prev:
923         /*
924          * It is potentially slow to have to call find_vma_prev here.
925          * But it's only on the first write fault on the vma, not
926          * every time, and we could devise a way to avoid it later
927          * (e.g. stash info in next's anon_vma_node when assigning
928          * an anon_vma, or when trying vma_merge).  Another time.
929          */
930         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
931         if (!near)
932                 goto none;
933
934         anon_vma = reusable_anon_vma(near, near, vma);
935         if (anon_vma)
936                 return anon_vma;
937 none:
938         /*
939          * There's no absolute need to look only at touching neighbours:
940          * we could search further afield for "compatible" anon_vmas.
941          * But it would probably just be a waste of time searching,
942          * or lead to too many vmas hanging off the same anon_vma.
943          * We're trying to allow mprotect remerging later on,
944          * not trying to minimize memory used for anon_vmas.
945          */
946         return NULL;
947 }
948
949 #ifdef CONFIG_PROC_FS
950 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
951                                                 struct file *file, long pages)
952 {
953         const unsigned long stack_flags
954                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
955
956         if (file) {
957                 mm->shared_vm += pages;
958                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
959                         mm->exec_vm += pages;
960         } else if (flags & stack_flags)
961                 mm->stack_vm += pages;
962         if (flags & (VM_RESERVED|VM_IO))
963                 mm->reserved_vm += pages;
964 }
965 #endif /* CONFIG_PROC_FS */
966
967 /*
968  * The caller must hold down_write(&current->mm->mmap_sem).
969  */
970
971 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
972                         unsigned long len, unsigned long prot,
973                         unsigned long flags, unsigned long pgoff)
974 {
975         struct mm_struct * mm = current->mm;
976         struct inode *inode;
977         unsigned int vm_flags;
978         int error;
979         unsigned long reqprot = prot;
980
981         /*
982          * Does the application expect PROT_READ to imply PROT_EXEC?
983          *
984          * (the exception is when the underlying filesystem is noexec
985          *  mounted, in which case we dont add PROT_EXEC.)
986          */
987         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
988                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
989                         prot |= PROT_EXEC;
990
991         if (!len)
992                 return -EINVAL;
993
994         if (!(flags & MAP_FIXED))
995                 addr = round_hint_to_min(addr);
996
997         /* Careful about overflows.. */
998         len = PAGE_ALIGN(len);
999         if (!len)
1000                 return -ENOMEM;
1001
1002         /* offset overflow? */
1003         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1004                return -EOVERFLOW;
1005
1006         /* Too many mappings? */
1007         if (mm->map_count > sysctl_max_map_count)
1008                 return -ENOMEM;
1009
1010         /* Obtain the address to map to. we verify (or select) it and ensure
1011          * that it represents a valid section of the address space.
1012          */
1013         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1014         if (addr & ~PAGE_MASK)
1015                 return addr;
1016
1017         /* Do simple checking here so the lower-level routines won't have
1018          * to. we assume access permissions have been handled by the open
1019          * of the memory object, so we don't do any here.
1020          */
1021         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1022                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1023
1024         if (flags & MAP_LOCKED)
1025                 if (!can_do_mlock())
1026                         return -EPERM;
1027
1028         /* mlock MCL_FUTURE? */
1029         if (vm_flags & VM_LOCKED) {
1030                 unsigned long locked, lock_limit;
1031                 locked = len >> PAGE_SHIFT;
1032                 locked += mm->locked_vm;
1033                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1034                 lock_limit >>= PAGE_SHIFT;
1035                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1036                         return -EAGAIN;
1037         }
1038
1039         inode = file ? file->f_path.dentry->d_inode : NULL;
1040
1041         if (file) {
1042                 switch (flags & MAP_TYPE) {
1043                 case MAP_SHARED:
1044                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1045                                 return -EACCES;
1046
1047                         /*
1048                          * Make sure we don't allow writing to an append-only
1049                          * file..
1050                          */
1051                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1052                                 return -EACCES;
1053
1054                         /*
1055                          * Make sure there are no mandatory locks on the file.
1056                          */
1057                         if (locks_verify_locked(inode))
1058                                 return -EAGAIN;
1059
1060                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1061                         if (!(file->f_mode & FMODE_WRITE))
1062                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1063
1064                         /* fall through */
1065                 case MAP_PRIVATE:
1066                         if (!(file->f_mode & FMODE_READ))
1067                                 return -EACCES;
1068                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1069                                 if (vm_flags & VM_EXEC)
1070                                         return -EPERM;
1071                                 vm_flags &= ~VM_MAYEXEC;
1072                         }
1073
1074                         if (!file->f_op || !file->f_op->mmap)
1075                                 return -ENODEV;
1076                         break;
1077
1078                 default:
1079                         return -EINVAL;
1080                 }
1081         } else {
1082                 switch (flags & MAP_TYPE) {
1083                 case MAP_SHARED:
1084                         /*
1085                          * Ignore pgoff.
1086                          */
1087                         pgoff = 0;
1088                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1089                         break;
1090                 case MAP_PRIVATE:
1091                         /*
1092                          * Set pgoff according to addr for anon_vma.
1093                          */
1094                         pgoff = addr >> PAGE_SHIFT;
1095                         break;
1096                 default:
1097                         return -EINVAL;
1098                 }
1099         }
1100
1101         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1102         if (error)
1103                 return error;
1104
1105         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1106 }
1107 EXPORT_SYMBOL(do_mmap_pgoff);
1108
1109 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1110                 unsigned long, prot, unsigned long, flags,
1111                 unsigned long, fd, unsigned long, pgoff)
1112 {
1113         struct file *file = NULL;
1114         unsigned long retval = -EBADF;
1115
1116         if (!(flags & MAP_ANONYMOUS)) {
1117                 audit_mmap_fd(fd, flags);
1118                 if (unlikely(flags & MAP_HUGETLB))
1119                         return -EINVAL;
1120                 file = fget(fd);
1121                 if (!file)
1122                         goto out;
1123         } else if (flags & MAP_HUGETLB) {
1124                 struct user_struct *user = NULL;
1125                 /*
1126                  * VM_NORESERVE is used because the reservations will be
1127                  * taken when vm_ops->mmap() is called
1128                  * A dummy user value is used because we are not locking
1129                  * memory so no accounting is necessary
1130                  */
1131                 len = ALIGN(len, huge_page_size(&default_hstate));
1132                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1133                                                 &user, HUGETLB_ANONHUGE_INODE);
1134                 if (IS_ERR(file))
1135                         return PTR_ERR(file);
1136         }
1137
1138         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1139
1140         down_write(&current->mm->mmap_sem);
1141         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1142         up_write(&current->mm->mmap_sem);
1143
1144         if (file)
1145                 fput(file);
1146 out:
1147         return retval;
1148 }
1149
1150 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1151 struct mmap_arg_struct {
1152         unsigned long addr;
1153         unsigned long len;
1154         unsigned long prot;
1155         unsigned long flags;
1156         unsigned long fd;
1157         unsigned long offset;
1158 };
1159
1160 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1161 {
1162         struct mmap_arg_struct a;
1163
1164         if (copy_from_user(&a, arg, sizeof(a)))
1165                 return -EFAULT;
1166         if (a.offset & ~PAGE_MASK)
1167                 return -EINVAL;
1168
1169         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1170                               a.offset >> PAGE_SHIFT);
1171 }
1172 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1173
1174 /*
1175  * Some shared mappigns will want the pages marked read-only
1176  * to track write events. If so, we'll downgrade vm_page_prot
1177  * to the private version (using protection_map[] without the
1178  * VM_SHARED bit).
1179  */
1180 int vma_wants_writenotify(struct vm_area_struct *vma)
1181 {
1182         unsigned int vm_flags = vma->vm_flags;
1183
1184         /* If it was private or non-writable, the write bit is already clear */
1185         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1186                 return 0;
1187
1188         /* The backer wishes to know when pages are first written to? */
1189         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1190                 return 1;
1191
1192         /* The open routine did something to the protections already? */
1193         if (pgprot_val(vma->vm_page_prot) !=
1194             pgprot_val(vm_get_page_prot(vm_flags)))
1195                 return 0;
1196
1197         /* Specialty mapping? */
1198         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1199                 return 0;
1200
1201         /* Can the mapping track the dirty pages? */
1202         return vma->vm_file && vma->vm_file->f_mapping &&
1203                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1204 }
1205
1206 /*
1207  * We account for memory if it's a private writeable mapping,
1208  * not hugepages and VM_NORESERVE wasn't set.
1209  */
1210 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1211 {
1212         /*
1213          * hugetlb has its own accounting separate from the core VM
1214          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1215          */
1216         if (file && is_file_hugepages(file))
1217                 return 0;
1218
1219         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1220 }
1221
1222 unsigned long mmap_region(struct file *file, unsigned long addr,
1223                           unsigned long len, unsigned long flags,
1224                           unsigned int vm_flags, unsigned long pgoff)
1225 {
1226         struct mm_struct *mm = current->mm;
1227         struct vm_area_struct *vma, *prev;
1228         int correct_wcount = 0;
1229         int error;
1230         struct rb_node **rb_link, *rb_parent;
1231         unsigned long charged = 0;
1232         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1233
1234         /* Clear old maps */
1235         error = -ENOMEM;
1236 munmap_back:
1237         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1238         if (vma && vma->vm_start < addr + len) {
1239                 if (do_munmap(mm, addr, len))
1240                         return -ENOMEM;
1241                 goto munmap_back;
1242         }
1243
1244         /* Check against address space limit. */
1245         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1246                 return -ENOMEM;
1247
1248         /*
1249          * Set 'VM_NORESERVE' if we should not account for the
1250          * memory use of this mapping.
1251          */
1252         if ((flags & MAP_NORESERVE)) {
1253                 /* We honor MAP_NORESERVE if allowed to overcommit */
1254                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1255                         vm_flags |= VM_NORESERVE;
1256
1257                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1258                 if (file && is_file_hugepages(file))
1259                         vm_flags |= VM_NORESERVE;
1260         }
1261
1262         /*
1263          * Private writable mapping: check memory availability
1264          */
1265         if (accountable_mapping(file, vm_flags)) {
1266                 charged = len >> PAGE_SHIFT;
1267                 if (security_vm_enough_memory(charged))
1268                         return -ENOMEM;
1269                 vm_flags |= VM_ACCOUNT;
1270         }
1271
1272         /*
1273          * Can we just expand an old mapping?
1274          */
1275         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1276         if (vma)
1277                 goto out;
1278
1279         /*
1280          * Determine the object being mapped and call the appropriate
1281          * specific mapper. the address has already been validated, but
1282          * not unmapped, but the maps are removed from the list.
1283          */
1284         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1285         if (!vma) {
1286                 error = -ENOMEM;
1287                 goto unacct_error;
1288         }
1289
1290         vma->vm_mm = mm;
1291         vma->vm_start = addr;
1292         vma->vm_end = addr + len;
1293         vma->vm_flags = vm_flags;
1294         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1295         vma->vm_pgoff = pgoff;
1296         INIT_LIST_HEAD(&vma->anon_vma_chain);
1297
1298         if (file) {
1299                 error = -EINVAL;
1300                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1301                         goto free_vma;
1302                 if (vm_flags & VM_DENYWRITE) {
1303                         error = deny_write_access(file);
1304                         if (error)
1305                                 goto free_vma;
1306                         correct_wcount = 1;
1307                 }
1308                 vma->vm_file = file;
1309                 get_file(file);
1310                 error = file->f_op->mmap(file, vma);
1311                 if (error)
1312                         goto unmap_and_free_vma;
1313                 if (vm_flags & VM_EXECUTABLE)
1314                         added_exe_file_vma(mm);
1315
1316                 /* Can addr have changed??
1317                  *
1318                  * Answer: Yes, several device drivers can do it in their
1319                  *         f_op->mmap method. -DaveM
1320                  */
1321                 addr = vma->vm_start;
1322                 pgoff = vma->vm_pgoff;
1323                 vm_flags = vma->vm_flags;
1324         } else if (vm_flags & VM_SHARED) {
1325                 error = shmem_zero_setup(vma);
1326                 if (error)
1327                         goto free_vma;
1328         }
1329
1330         if (vma_wants_writenotify(vma)) {
1331                 pgprot_t pprot = vma->vm_page_prot;
1332
1333                 /* Can vma->vm_page_prot have changed??
1334                  *
1335                  * Answer: Yes, drivers may have changed it in their
1336                  *         f_op->mmap method.
1337                  *
1338                  * Ensures that vmas marked as uncached stay that way.
1339                  */
1340                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1341                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1342                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1343         }
1344
1345         vma_link(mm, vma, prev, rb_link, rb_parent);
1346         file = vma->vm_file;
1347
1348         /* Once vma denies write, undo our temporary denial count */
1349         if (correct_wcount)
1350                 atomic_inc(&inode->i_writecount);
1351 out:
1352         perf_event_mmap(vma);
1353
1354         mm->total_vm += len >> PAGE_SHIFT;
1355         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1356         if (vm_flags & VM_LOCKED) {
1357                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1358                         mm->locked_vm += (len >> PAGE_SHIFT);
1359         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1360                 make_pages_present(addr, addr + len);
1361         return addr;
1362
1363 unmap_and_free_vma:
1364         if (correct_wcount)
1365                 atomic_inc(&inode->i_writecount);
1366         vma->vm_file = NULL;
1367         fput(file);
1368
1369         /* Undo any partial mapping done by a device driver. */
1370         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1371         charged = 0;
1372 free_vma:
1373         kmem_cache_free(vm_area_cachep, vma);
1374 unacct_error:
1375         if (charged)
1376                 vm_unacct_memory(charged);
1377         return error;
1378 }
1379
1380 /* Get an address range which is currently unmapped.
1381  * For shmat() with addr=0.
1382  *
1383  * Ugly calling convention alert:
1384  * Return value with the low bits set means error value,
1385  * ie
1386  *      if (ret & ~PAGE_MASK)
1387  *              error = ret;
1388  *
1389  * This function "knows" that -ENOMEM has the bits set.
1390  */
1391 #ifndef HAVE_ARCH_UNMAPPED_AREA
1392 unsigned long
1393 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1394                 unsigned long len, unsigned long pgoff, unsigned long flags)
1395 {
1396         struct mm_struct *mm = current->mm;
1397         struct vm_area_struct *vma;
1398         unsigned long start_addr;
1399
1400         if (len > TASK_SIZE)
1401                 return -ENOMEM;
1402
1403         if (flags & MAP_FIXED)
1404                 return addr;
1405
1406         if (addr) {
1407                 addr = PAGE_ALIGN(addr);
1408                 vma = find_vma(mm, addr);
1409                 if (TASK_SIZE - len >= addr &&
1410                     (!vma || addr + len <= vma->vm_start))
1411                         return addr;
1412         }
1413         if (len > mm->cached_hole_size) {
1414                 start_addr = addr = mm->free_area_cache;
1415         } else {
1416                 start_addr = addr = TASK_UNMAPPED_BASE;
1417                 mm->cached_hole_size = 0;
1418         }
1419
1420 full_search:
1421         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1422                 /* At this point:  (!vma || addr < vma->vm_end). */
1423                 if (TASK_SIZE - len < addr) {
1424                         /*
1425                          * Start a new search - just in case we missed
1426                          * some holes.
1427                          */
1428                         if (start_addr != TASK_UNMAPPED_BASE) {
1429                                 addr = TASK_UNMAPPED_BASE;
1430                                 start_addr = addr;
1431                                 mm->cached_hole_size = 0;
1432                                 goto full_search;
1433                         }
1434                         return -ENOMEM;
1435                 }
1436                 if (!vma || addr + len <= vma->vm_start) {
1437                         /*
1438                          * Remember the place where we stopped the search:
1439                          */
1440                         mm->free_area_cache = addr + len;
1441                         return addr;
1442                 }
1443                 if (addr + mm->cached_hole_size < vma->vm_start)
1444                         mm->cached_hole_size = vma->vm_start - addr;
1445                 addr = vma->vm_end;
1446         }
1447 }
1448 #endif  
1449
1450 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1451 {
1452         /*
1453          * Is this a new hole at the lowest possible address?
1454          */
1455         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1456                 mm->free_area_cache = addr;
1457                 mm->cached_hole_size = ~0UL;
1458         }
1459 }
1460
1461 /*
1462  * This mmap-allocator allocates new areas top-down from below the
1463  * stack's low limit (the base):
1464  */
1465 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1466 unsigned long
1467 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1468                           const unsigned long len, const unsigned long pgoff,
1469                           const unsigned long flags)
1470 {
1471         struct vm_area_struct *vma;
1472         struct mm_struct *mm = current->mm;
1473         unsigned long addr = addr0;
1474
1475         /* requested length too big for entire address space */
1476         if (len > TASK_SIZE)
1477                 return -ENOMEM;
1478
1479         if (flags & MAP_FIXED)
1480                 return addr;
1481
1482         /* requesting a specific address */
1483         if (addr) {
1484                 addr = PAGE_ALIGN(addr);
1485                 vma = find_vma(mm, addr);
1486                 if (TASK_SIZE - len >= addr &&
1487                                 (!vma || addr + len <= vma->vm_start))
1488                         return addr;
1489         }
1490
1491         /* check if free_area_cache is useful for us */
1492         if (len <= mm->cached_hole_size) {
1493                 mm->cached_hole_size = 0;
1494                 mm->free_area_cache = mm->mmap_base;
1495         }
1496
1497         /* either no address requested or can't fit in requested address hole */
1498         addr = mm->free_area_cache;
1499
1500         /* make sure it can fit in the remaining address space */
1501         if (addr > len) {
1502                 vma = find_vma(mm, addr-len);
1503                 if (!vma || addr <= vma->vm_start)
1504                         /* remember the address as a hint for next time */
1505                         return (mm->free_area_cache = addr-len);
1506         }
1507
1508         if (mm->mmap_base < len)
1509                 goto bottomup;
1510
1511         addr = mm->mmap_base-len;
1512
1513         do {
1514                 /*
1515                  * Lookup failure means no vma is above this address,
1516                  * else if new region fits below vma->vm_start,
1517                  * return with success:
1518                  */
1519                 vma = find_vma(mm, addr);
1520                 if (!vma || addr+len <= vma->vm_start)
1521                         /* remember the address as a hint for next time */
1522                         return (mm->free_area_cache = addr);
1523
1524                 /* remember the largest hole we saw so far */
1525                 if (addr + mm->cached_hole_size < vma->vm_start)
1526                         mm->cached_hole_size = vma->vm_start - addr;
1527
1528                 /* try just below the current vma->vm_start */
1529                 addr = vma->vm_start-len;
1530         } while (len < vma->vm_start);
1531
1532 bottomup:
1533         /*
1534          * A failed mmap() very likely causes application failure,
1535          * so fall back to the bottom-up function here. This scenario
1536          * can happen with large stack limits and large mmap()
1537          * allocations.
1538          */
1539         mm->cached_hole_size = ~0UL;
1540         mm->free_area_cache = TASK_UNMAPPED_BASE;
1541         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1542         /*
1543          * Restore the topdown base:
1544          */
1545         mm->free_area_cache = mm->mmap_base;
1546         mm->cached_hole_size = ~0UL;
1547
1548         return addr;
1549 }
1550 #endif
1551
1552 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1553 {
1554         /*
1555          * Is this a new hole at the highest possible address?
1556          */
1557         if (addr > mm->free_area_cache)
1558                 mm->free_area_cache = addr;
1559
1560         /* dont allow allocations above current base */
1561         if (mm->free_area_cache > mm->mmap_base)
1562                 mm->free_area_cache = mm->mmap_base;
1563 }
1564
1565 unsigned long
1566 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1567                 unsigned long pgoff, unsigned long flags)
1568 {
1569         unsigned long (*get_area)(struct file *, unsigned long,
1570                                   unsigned long, unsigned long, unsigned long);
1571
1572         unsigned long error = arch_mmap_check(addr, len, flags);
1573         if (error)
1574                 return error;
1575
1576         /* Careful about overflows.. */
1577         if (len > TASK_SIZE)
1578                 return -ENOMEM;
1579
1580         get_area = current->mm->get_unmapped_area;
1581         if (file && file->f_op && file->f_op->get_unmapped_area)
1582                 get_area = file->f_op->get_unmapped_area;
1583         addr = get_area(file, addr, len, pgoff, flags);
1584         if (IS_ERR_VALUE(addr))
1585                 return addr;
1586
1587         if (addr > TASK_SIZE - len)
1588                 return -ENOMEM;
1589         if (addr & ~PAGE_MASK)
1590                 return -EINVAL;
1591
1592         return arch_rebalance_pgtables(addr, len);
1593 }
1594
1595 EXPORT_SYMBOL(get_unmapped_area);
1596
1597 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1598 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1599 {
1600         struct vm_area_struct *vma = NULL;
1601
1602         if (mm) {
1603                 /* Check the cache first. */
1604                 /* (Cache hit rate is typically around 35%.) */
1605                 vma = mm->mmap_cache;
1606                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1607                         struct rb_node * rb_node;
1608
1609                         rb_node = mm->mm_rb.rb_node;
1610                         vma = NULL;
1611
1612                         while (rb_node) {
1613                                 struct vm_area_struct * vma_tmp;
1614
1615                                 vma_tmp = rb_entry(rb_node,
1616                                                 struct vm_area_struct, vm_rb);
1617
1618                                 if (vma_tmp->vm_end > addr) {
1619                                         vma = vma_tmp;
1620                                         if (vma_tmp->vm_start <= addr)
1621                                                 break;
1622                                         rb_node = rb_node->rb_left;
1623                                 } else
1624                                         rb_node = rb_node->rb_right;
1625                         }
1626                         if (vma)
1627                                 mm->mmap_cache = vma;
1628                 }
1629         }
1630         return vma;
1631 }
1632
1633 EXPORT_SYMBOL(find_vma);
1634
1635 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1636 struct vm_area_struct *
1637 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1638                         struct vm_area_struct **pprev)
1639 {
1640         struct vm_area_struct *vma = NULL, *prev = NULL;
1641         struct rb_node *rb_node;
1642         if (!mm)
1643                 goto out;
1644
1645         /* Guard against addr being lower than the first VMA */
1646         vma = mm->mmap;
1647
1648         /* Go through the RB tree quickly. */
1649         rb_node = mm->mm_rb.rb_node;
1650
1651         while (rb_node) {
1652                 struct vm_area_struct *vma_tmp;
1653                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1654
1655                 if (addr < vma_tmp->vm_end) {
1656                         rb_node = rb_node->rb_left;
1657                 } else {
1658                         prev = vma_tmp;
1659                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1660                                 break;
1661                         rb_node = rb_node->rb_right;
1662                 }
1663         }
1664
1665 out:
1666         *pprev = prev;
1667         return prev ? prev->vm_next : vma;
1668 }
1669
1670 /*
1671  * Verify that the stack growth is acceptable and
1672  * update accounting. This is shared with both the
1673  * grow-up and grow-down cases.
1674  */
1675 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1676 {
1677         struct mm_struct *mm = vma->vm_mm;
1678         struct rlimit *rlim = current->signal->rlim;
1679         unsigned long new_start;
1680
1681         /* address space limit tests */
1682         if (!may_expand_vm(mm, grow))
1683                 return -ENOMEM;
1684
1685         /* Stack limit test */
1686         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1687                 return -ENOMEM;
1688
1689         /* mlock limit tests */
1690         if (vma->vm_flags & VM_LOCKED) {
1691                 unsigned long locked;
1692                 unsigned long limit;
1693                 locked = mm->locked_vm + grow;
1694                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1695                 limit >>= PAGE_SHIFT;
1696                 if (locked > limit && !capable(CAP_IPC_LOCK))
1697                         return -ENOMEM;
1698         }
1699
1700         /* Check to ensure the stack will not grow into a hugetlb-only region */
1701         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1702                         vma->vm_end - size;
1703         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1704                 return -EFAULT;
1705
1706         /*
1707          * Overcommit..  This must be the final test, as it will
1708          * update security statistics.
1709          */
1710         if (security_vm_enough_memory_mm(mm, grow))
1711                 return -ENOMEM;
1712
1713         /* Ok, everything looks good - let it rip */
1714         mm->total_vm += grow;
1715         if (vma->vm_flags & VM_LOCKED)
1716                 mm->locked_vm += grow;
1717         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1718         return 0;
1719 }
1720
1721 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1722 /*
1723  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1724  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1725  */
1726 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1727 {
1728         int error;
1729
1730         if (!(vma->vm_flags & VM_GROWSUP))
1731                 return -EFAULT;
1732
1733         /*
1734          * We must make sure the anon_vma is allocated
1735          * so that the anon_vma locking is not a noop.
1736          */
1737         if (unlikely(anon_vma_prepare(vma)))
1738                 return -ENOMEM;
1739         vma_lock_anon_vma(vma);
1740
1741         /*
1742          * vma->vm_start/vm_end cannot change under us because the caller
1743          * is required to hold the mmap_sem in read mode.  We need the
1744          * anon_vma lock to serialize against concurrent expand_stacks.
1745          * Also guard against wrapping around to address 0.
1746          */
1747         if (address < PAGE_ALIGN(address+4))
1748                 address = PAGE_ALIGN(address+4);
1749         else {
1750                 vma_unlock_anon_vma(vma);
1751                 return -ENOMEM;
1752         }
1753         error = 0;
1754
1755         /* Somebody else might have raced and expanded it already */
1756         if (address > vma->vm_end) {
1757                 unsigned long size, grow;
1758
1759                 size = address - vma->vm_start;
1760                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1761
1762                 error = acct_stack_growth(vma, size, grow);
1763                 if (!error) {
1764                         vma->vm_end = address;
1765                         perf_event_mmap(vma);
1766                 }
1767         }
1768         vma_unlock_anon_vma(vma);
1769         khugepaged_enter_vma_merge(vma);
1770         return error;
1771 }
1772 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1773
1774 /*
1775  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1776  */
1777 static int expand_downwards(struct vm_area_struct *vma,
1778                                    unsigned long address)
1779 {
1780         int error;
1781
1782         /*
1783          * We must make sure the anon_vma is allocated
1784          * so that the anon_vma locking is not a noop.
1785          */
1786         if (unlikely(anon_vma_prepare(vma)))
1787                 return -ENOMEM;
1788
1789         address &= PAGE_MASK;
1790         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1791         if (error)
1792                 return error;
1793
1794         vma_lock_anon_vma(vma);
1795
1796         /*
1797          * vma->vm_start/vm_end cannot change under us because the caller
1798          * is required to hold the mmap_sem in read mode.  We need the
1799          * anon_vma lock to serialize against concurrent expand_stacks.
1800          */
1801
1802         /* Somebody else might have raced and expanded it already */
1803         if (address < vma->vm_start) {
1804                 unsigned long size, grow;
1805
1806                 size = vma->vm_end - address;
1807                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1808
1809                 error = acct_stack_growth(vma, size, grow);
1810                 if (!error) {
1811                         vma->vm_start = address;
1812                         vma->vm_pgoff -= grow;
1813                         perf_event_mmap(vma);
1814                 }
1815         }
1816         vma_unlock_anon_vma(vma);
1817         khugepaged_enter_vma_merge(vma);
1818         return error;
1819 }
1820
1821 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1822 {
1823         return expand_downwards(vma, address);
1824 }
1825
1826 #ifdef CONFIG_STACK_GROWSUP
1827 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1828 {
1829         return expand_upwards(vma, address);
1830 }
1831
1832 struct vm_area_struct *
1833 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1834 {
1835         struct vm_area_struct *vma, *prev;
1836
1837         addr &= PAGE_MASK;
1838         vma = find_vma_prev(mm, addr, &prev);
1839         if (vma && (vma->vm_start <= addr))
1840                 return vma;
1841         if (!prev || expand_stack(prev, addr))
1842                 return NULL;
1843         if (prev->vm_flags & VM_LOCKED) {
1844                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1845         }
1846         return prev;
1847 }
1848 #else
1849 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1850 {
1851         return expand_downwards(vma, address);
1852 }
1853
1854 struct vm_area_struct *
1855 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1856 {
1857         struct vm_area_struct * vma;
1858         unsigned long start;
1859
1860         addr &= PAGE_MASK;
1861         vma = find_vma(mm,addr);
1862         if (!vma)
1863                 return NULL;
1864         if (vma->vm_start <= addr)
1865                 return vma;
1866         if (!(vma->vm_flags & VM_GROWSDOWN))
1867                 return NULL;
1868         start = vma->vm_start;
1869         if (expand_stack(vma, addr))
1870                 return NULL;
1871         if (vma->vm_flags & VM_LOCKED) {
1872                 mlock_vma_pages_range(vma, addr, start);
1873         }
1874         return vma;
1875 }
1876 #endif
1877
1878 /*
1879  * Ok - we have the memory areas we should free on the vma list,
1880  * so release them, and do the vma updates.
1881  *
1882  * Called with the mm semaphore held.
1883  */
1884 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1885 {
1886         /* Update high watermark before we lower total_vm */
1887         update_hiwater_vm(mm);
1888         do {
1889                 long nrpages = vma_pages(vma);
1890
1891                 mm->total_vm -= nrpages;
1892                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1893                 vma = remove_vma(vma);
1894         } while (vma);
1895         validate_mm(mm);
1896 }
1897
1898 /*
1899  * Get rid of page table information in the indicated region.
1900  *
1901  * Called with the mm semaphore held.
1902  */
1903 static void unmap_region(struct mm_struct *mm,
1904                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1905                 unsigned long start, unsigned long end)
1906 {
1907         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1908         struct mmu_gather *tlb;
1909         unsigned long nr_accounted = 0;
1910
1911         lru_add_drain();
1912         tlb = tlb_gather_mmu(mm, 0);
1913         update_hiwater_rss(mm);
1914         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1915         vm_unacct_memory(nr_accounted);
1916         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1917                                  next? next->vm_start: 0);
1918         tlb_finish_mmu(tlb, start, end);
1919 }
1920
1921 /*
1922  * Create a list of vma's touched by the unmap, removing them from the mm's
1923  * vma list as we go..
1924  */
1925 static void
1926 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1927         struct vm_area_struct *prev, unsigned long end)
1928 {
1929         struct vm_area_struct **insertion_point;
1930         struct vm_area_struct *tail_vma = NULL;
1931         unsigned long addr;
1932
1933         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1934         vma->vm_prev = NULL;
1935         do {
1936                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1937                 mm->map_count--;
1938                 tail_vma = vma;
1939                 vma = vma->vm_next;
1940         } while (vma && vma->vm_start < end);
1941         *insertion_point = vma;
1942         if (vma)
1943                 vma->vm_prev = prev;
1944         tail_vma->vm_next = NULL;
1945         if (mm->unmap_area == arch_unmap_area)
1946                 addr = prev ? prev->vm_end : mm->mmap_base;
1947         else
1948                 addr = vma ?  vma->vm_start : mm->mmap_base;
1949         mm->unmap_area(mm, addr);
1950         mm->mmap_cache = NULL;          /* Kill the cache. */
1951 }
1952
1953 /*
1954  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1955  * munmap path where it doesn't make sense to fail.
1956  */
1957 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1958               unsigned long addr, int new_below)
1959 {
1960         struct mempolicy *pol;
1961         struct vm_area_struct *new;
1962         int err = -ENOMEM;
1963
1964         if (is_vm_hugetlb_page(vma) && (addr &
1965                                         ~(huge_page_mask(hstate_vma(vma)))))
1966                 return -EINVAL;
1967
1968         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1969         if (!new)
1970                 goto out_err;
1971
1972         /* most fields are the same, copy all, and then fixup */
1973         *new = *vma;
1974
1975         INIT_LIST_HEAD(&new->anon_vma_chain);
1976
1977         if (new_below)
1978                 new->vm_end = addr;
1979         else {
1980                 new->vm_start = addr;
1981                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1982         }
1983
1984         pol = mpol_dup(vma_policy(vma));
1985         if (IS_ERR(pol)) {
1986                 err = PTR_ERR(pol);
1987                 goto out_free_vma;
1988         }
1989         vma_set_policy(new, pol);
1990
1991         if (anon_vma_clone(new, vma))
1992                 goto out_free_mpol;
1993
1994         if (new->vm_file) {
1995                 get_file(new->vm_file);
1996                 if (vma->vm_flags & VM_EXECUTABLE)
1997                         added_exe_file_vma(mm);
1998         }
1999
2000         if (new->vm_ops && new->vm_ops->open)
2001                 new->vm_ops->open(new);
2002
2003         if (new_below)
2004                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2005                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2006         else
2007                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2008
2009         /* Success. */
2010         if (!err)
2011                 return 0;
2012
2013         /* Clean everything up if vma_adjust failed. */
2014         if (new->vm_ops && new->vm_ops->close)
2015                 new->vm_ops->close(new);
2016         if (new->vm_file) {
2017                 if (vma->vm_flags & VM_EXECUTABLE)
2018                         removed_exe_file_vma(mm);
2019                 fput(new->vm_file);
2020         }
2021         unlink_anon_vmas(new);
2022  out_free_mpol:
2023         mpol_put(pol);
2024  out_free_vma:
2025         kmem_cache_free(vm_area_cachep, new);
2026  out_err:
2027         return err;
2028 }
2029
2030 /*
2031  * Split a vma into two pieces at address 'addr', a new vma is allocated
2032  * either for the first part or the tail.
2033  */
2034 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2035               unsigned long addr, int new_below)
2036 {
2037         if (mm->map_count >= sysctl_max_map_count)
2038                 return -ENOMEM;
2039
2040         return __split_vma(mm, vma, addr, new_below);
2041 }
2042
2043 /* Munmap is split into 2 main parts -- this part which finds
2044  * what needs doing, and the areas themselves, which do the
2045  * work.  This now handles partial unmappings.
2046  * Jeremy Fitzhardinge <jeremy@goop.org>
2047  */
2048 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2049 {
2050         unsigned long end;
2051         struct vm_area_struct *vma, *prev, *last;
2052
2053         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2054                 return -EINVAL;
2055
2056         if ((len = PAGE_ALIGN(len)) == 0)
2057                 return -EINVAL;
2058
2059         /* Find the first overlapping VMA */
2060         vma = find_vma_prev(mm, start, &prev);
2061         if (!vma)
2062                 return 0;
2063         /* we have  start < vma->vm_end  */
2064
2065         /* if it doesn't overlap, we have nothing.. */
2066         end = start + len;
2067         if (vma->vm_start >= end)
2068                 return 0;
2069
2070         /*
2071          * If we need to split any vma, do it now to save pain later.
2072          *
2073          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2074          * unmapped vm_area_struct will remain in use: so lower split_vma
2075          * places tmp vma above, and higher split_vma places tmp vma below.
2076          */
2077         if (start > vma->vm_start) {
2078                 int error;
2079
2080                 /*
2081                  * Make sure that map_count on return from munmap() will
2082                  * not exceed its limit; but let map_count go just above
2083                  * its limit temporarily, to help free resources as expected.
2084                  */
2085                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2086                         return -ENOMEM;
2087
2088                 error = __split_vma(mm, vma, start, 0);
2089                 if (error)
2090                         return error;
2091                 prev = vma;
2092         }
2093
2094         /* Does it split the last one? */
2095         last = find_vma(mm, end);
2096         if (last && end > last->vm_start) {
2097                 int error = __split_vma(mm, last, end, 1);
2098                 if (error)
2099                         return error;
2100         }
2101         vma = prev? prev->vm_next: mm->mmap;
2102
2103         /*
2104          * unlock any mlock()ed ranges before detaching vmas
2105          */
2106         if (mm->locked_vm) {
2107                 struct vm_area_struct *tmp = vma;
2108                 while (tmp && tmp->vm_start < end) {
2109                         if (tmp->vm_flags & VM_LOCKED) {
2110                                 mm->locked_vm -= vma_pages(tmp);
2111                                 munlock_vma_pages_all(tmp);
2112                         }
2113                         tmp = tmp->vm_next;
2114                 }
2115         }
2116
2117         /*
2118          * Remove the vma's, and unmap the actual pages
2119          */
2120         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2121         unmap_region(mm, vma, prev, start, end);
2122
2123         /* Fix up all other VM information */
2124         remove_vma_list(mm, vma);
2125
2126         return 0;
2127 }
2128
2129 EXPORT_SYMBOL(do_munmap);
2130
2131 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2132 {
2133         int ret;
2134         struct mm_struct *mm = current->mm;
2135
2136         profile_munmap(addr);
2137
2138         down_write(&mm->mmap_sem);
2139         ret = do_munmap(mm, addr, len);
2140         up_write(&mm->mmap_sem);
2141         return ret;
2142 }
2143
2144 static inline void verify_mm_writelocked(struct mm_struct *mm)
2145 {
2146 #ifdef CONFIG_DEBUG_VM
2147         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2148                 WARN_ON(1);
2149                 up_read(&mm->mmap_sem);
2150         }
2151 #endif
2152 }
2153
2154 /*
2155  *  this is really a simplified "do_mmap".  it only handles
2156  *  anonymous maps.  eventually we may be able to do some
2157  *  brk-specific accounting here.
2158  */
2159 unsigned long do_brk(unsigned long addr, unsigned long len)
2160 {
2161         struct mm_struct * mm = current->mm;
2162         struct vm_area_struct * vma, * prev;
2163         unsigned long flags;
2164         struct rb_node ** rb_link, * rb_parent;
2165         pgoff_t pgoff = addr >> PAGE_SHIFT;
2166         int error;
2167
2168         len = PAGE_ALIGN(len);
2169         if (!len)
2170                 return addr;
2171
2172         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2173         if (error)
2174                 return error;
2175
2176         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2177
2178         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2179         if (error & ~PAGE_MASK)
2180                 return error;
2181
2182         /*
2183          * mlock MCL_FUTURE?
2184          */
2185         if (mm->def_flags & VM_LOCKED) {
2186                 unsigned long locked, lock_limit;
2187                 locked = len >> PAGE_SHIFT;
2188                 locked += mm->locked_vm;
2189                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2190                 lock_limit >>= PAGE_SHIFT;
2191                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2192                         return -EAGAIN;
2193         }
2194
2195         /*
2196          * mm->mmap_sem is required to protect against another thread
2197          * changing the mappings in case we sleep.
2198          */
2199         verify_mm_writelocked(mm);
2200
2201         /*
2202          * Clear old maps.  this also does some error checking for us
2203          */
2204  munmap_back:
2205         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2206         if (vma && vma->vm_start < addr + len) {
2207                 if (do_munmap(mm, addr, len))
2208                         return -ENOMEM;
2209                 goto munmap_back;
2210         }
2211
2212         /* Check against address space limits *after* clearing old maps... */
2213         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2214                 return -ENOMEM;
2215
2216         if (mm->map_count > sysctl_max_map_count)
2217                 return -ENOMEM;
2218
2219         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2220                 return -ENOMEM;
2221
2222         /* Can we just expand an old private anonymous mapping? */
2223         vma = vma_merge(mm, prev, addr, addr + len, flags,
2224                                         NULL, NULL, pgoff, NULL);
2225         if (vma)
2226                 goto out;
2227
2228         /*
2229          * create a vma struct for an anonymous mapping
2230          */
2231         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2232         if (!vma) {
2233                 vm_unacct_memory(len >> PAGE_SHIFT);
2234                 return -ENOMEM;
2235         }
2236
2237         INIT_LIST_HEAD(&vma->anon_vma_chain);
2238         vma->vm_mm = mm;
2239         vma->vm_start = addr;
2240         vma->vm_end = addr + len;
2241         vma->vm_pgoff = pgoff;
2242         vma->vm_flags = flags;
2243         vma->vm_page_prot = vm_get_page_prot(flags);
2244         vma_link(mm, vma, prev, rb_link, rb_parent);
2245 out:
2246         perf_event_mmap(vma);
2247         mm->total_vm += len >> PAGE_SHIFT;
2248         if (flags & VM_LOCKED) {
2249                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2250                         mm->locked_vm += (len >> PAGE_SHIFT);
2251         }
2252         return addr;
2253 }
2254
2255 EXPORT_SYMBOL(do_brk);
2256
2257 /* Release all mmaps. */
2258 void exit_mmap(struct mm_struct *mm)
2259 {
2260         struct mmu_gather *tlb;
2261         struct vm_area_struct *vma;
2262         unsigned long nr_accounted = 0;
2263         unsigned long end;
2264
2265         /* mm's last user has gone, and its about to be pulled down */
2266         mmu_notifier_release(mm);
2267
2268         if (mm->locked_vm) {
2269                 vma = mm->mmap;
2270                 while (vma) {
2271                         if (vma->vm_flags & VM_LOCKED)
2272                                 munlock_vma_pages_all(vma);
2273                         vma = vma->vm_next;
2274                 }
2275         }
2276
2277         arch_exit_mmap(mm);
2278
2279         vma = mm->mmap;
2280         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2281                 return;
2282
2283         lru_add_drain();
2284         flush_cache_mm(mm);
2285         tlb = tlb_gather_mmu(mm, 1);
2286         /* update_hiwater_rss(mm) here? but nobody should be looking */
2287         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2288         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2289         vm_unacct_memory(nr_accounted);
2290
2291         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2292         tlb_finish_mmu(tlb, 0, end);
2293
2294         /*
2295          * Walk the list again, actually closing and freeing it,
2296          * with preemption enabled, without holding any MM locks.
2297          */
2298         while (vma)
2299                 vma = remove_vma(vma);
2300
2301         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2302 }
2303
2304 /* Insert vm structure into process list sorted by address
2305  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2306  * then i_mmap_lock is taken here.
2307  */
2308 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2309 {
2310         struct vm_area_struct * __vma, * prev;
2311         struct rb_node ** rb_link, * rb_parent;
2312
2313         /*
2314          * The vm_pgoff of a purely anonymous vma should be irrelevant
2315          * until its first write fault, when page's anon_vma and index
2316          * are set.  But now set the vm_pgoff it will almost certainly
2317          * end up with (unless mremap moves it elsewhere before that
2318          * first wfault), so /proc/pid/maps tells a consistent story.
2319          *
2320          * By setting it to reflect the virtual start address of the
2321          * vma, merges and splits can happen in a seamless way, just
2322          * using the existing file pgoff checks and manipulations.
2323          * Similarly in do_mmap_pgoff and in do_brk.
2324          */
2325         if (!vma->vm_file) {
2326                 BUG_ON(vma->anon_vma);
2327                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2328         }
2329         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2330         if (__vma && __vma->vm_start < vma->vm_end)
2331                 return -ENOMEM;
2332         if ((vma->vm_flags & VM_ACCOUNT) &&
2333              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2334                 return -ENOMEM;
2335         vma_link(mm, vma, prev, rb_link, rb_parent);
2336         return 0;
2337 }
2338
2339 /*
2340  * Copy the vma structure to a new location in the same mm,
2341  * prior to moving page table entries, to effect an mremap move.
2342  */
2343 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2344         unsigned long addr, unsigned long len, pgoff_t pgoff)
2345 {
2346         struct vm_area_struct *vma = *vmap;
2347         unsigned long vma_start = vma->vm_start;
2348         struct mm_struct *mm = vma->vm_mm;
2349         struct vm_area_struct *new_vma, *prev;
2350         struct rb_node **rb_link, *rb_parent;
2351         struct mempolicy *pol;
2352
2353         /*
2354          * If anonymous vma has not yet been faulted, update new pgoff
2355          * to match new location, to increase its chance of merging.
2356          */
2357         if (!vma->vm_file && !vma->anon_vma)
2358                 pgoff = addr >> PAGE_SHIFT;
2359
2360         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2361         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2362                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2363         if (new_vma) {
2364                 /*
2365                  * Source vma may have been merged into new_vma
2366                  */
2367                 if (vma_start >= new_vma->vm_start &&
2368                     vma_start < new_vma->vm_end)
2369                         *vmap = new_vma;
2370         } else {
2371                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2372                 if (new_vma) {
2373                         *new_vma = *vma;
2374                         pol = mpol_dup(vma_policy(vma));
2375                         if (IS_ERR(pol))
2376                                 goto out_free_vma;
2377                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2378                         if (anon_vma_clone(new_vma, vma))
2379                                 goto out_free_mempol;
2380                         vma_set_policy(new_vma, pol);
2381                         new_vma->vm_start = addr;
2382                         new_vma->vm_end = addr + len;
2383                         new_vma->vm_pgoff = pgoff;
2384                         if (new_vma->vm_file) {
2385                                 get_file(new_vma->vm_file);
2386                                 if (vma->vm_flags & VM_EXECUTABLE)
2387                                         added_exe_file_vma(mm);
2388                         }
2389                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2390                                 new_vma->vm_ops->open(new_vma);
2391                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2392                 }
2393         }
2394         return new_vma;
2395
2396  out_free_mempol:
2397         mpol_put(pol);
2398  out_free_vma:
2399         kmem_cache_free(vm_area_cachep, new_vma);
2400         return NULL;
2401 }
2402
2403 /*
2404  * Return true if the calling process may expand its vm space by the passed
2405  * number of pages
2406  */
2407 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2408 {
2409         unsigned long cur = mm->total_vm;       /* pages */
2410         unsigned long lim;
2411
2412         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2413
2414         if (cur + npages > lim)
2415                 return 0;
2416         return 1;
2417 }
2418
2419
2420 static int special_mapping_fault(struct vm_area_struct *vma,
2421                                 struct vm_fault *vmf)
2422 {
2423         pgoff_t pgoff;
2424         struct page **pages;
2425
2426         /*
2427          * special mappings have no vm_file, and in that case, the mm
2428          * uses vm_pgoff internally. So we have to subtract it from here.
2429          * We are allowed to do this because we are the mm; do not copy
2430          * this code into drivers!
2431          */
2432         pgoff = vmf->pgoff - vma->vm_pgoff;
2433
2434         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2435                 pgoff--;
2436
2437         if (*pages) {
2438                 struct page *page = *pages;
2439                 get_page(page);
2440                 vmf->page = page;
2441                 return 0;
2442         }
2443
2444         return VM_FAULT_SIGBUS;
2445 }
2446
2447 /*
2448  * Having a close hook prevents vma merging regardless of flags.
2449  */
2450 static void special_mapping_close(struct vm_area_struct *vma)
2451 {
2452 }
2453
2454 static const struct vm_operations_struct special_mapping_vmops = {
2455         .close = special_mapping_close,
2456         .fault = special_mapping_fault,
2457 };
2458
2459 /*
2460  * Called with mm->mmap_sem held for writing.
2461  * Insert a new vma covering the given region, with the given flags.
2462  * Its pages are supplied by the given array of struct page *.
2463  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2464  * The region past the last page supplied will always produce SIGBUS.
2465  * The array pointer and the pages it points to are assumed to stay alive
2466  * for as long as this mapping might exist.
2467  */
2468 int install_special_mapping(struct mm_struct *mm,
2469                             unsigned long addr, unsigned long len,
2470                             unsigned long vm_flags, struct page **pages)
2471 {
2472         int ret;
2473         struct vm_area_struct *vma;
2474
2475         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2476         if (unlikely(vma == NULL))
2477                 return -ENOMEM;
2478
2479         INIT_LIST_HEAD(&vma->anon_vma_chain);
2480         vma->vm_mm = mm;
2481         vma->vm_start = addr;
2482         vma->vm_end = addr + len;
2483
2484         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2485         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2486
2487         vma->vm_ops = &special_mapping_vmops;
2488         vma->vm_private_data = pages;
2489
2490         ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2491         if (ret)
2492                 goto out;
2493
2494         ret = insert_vm_struct(mm, vma);
2495         if (ret)
2496                 goto out;
2497
2498         mm->total_vm += len >> PAGE_SHIFT;
2499
2500         perf_event_mmap(vma);
2501
2502         return 0;
2503
2504 out:
2505         kmem_cache_free(vm_area_cachep, vma);
2506         return ret;
2507 }
2508
2509 static DEFINE_MUTEX(mm_all_locks_mutex);
2510
2511 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2512 {
2513         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2514                 /*
2515                  * The LSB of head.next can't change from under us
2516                  * because we hold the mm_all_locks_mutex.
2517                  */
2518                 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem);
2519                 /*
2520                  * We can safely modify head.next after taking the
2521                  * anon_vma->root->lock. If some other vma in this mm shares
2522                  * the same anon_vma we won't take it again.
2523                  *
2524                  * No need of atomic instructions here, head.next
2525                  * can't change from under us thanks to the
2526                  * anon_vma->root->lock.
2527                  */
2528                 if (__test_and_set_bit(0, (unsigned long *)
2529                                        &anon_vma->root->head.next))
2530                         BUG();
2531         }
2532 }
2533
2534 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2535 {
2536         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2537                 /*
2538                  * AS_MM_ALL_LOCKS can't change from under us because
2539                  * we hold the mm_all_locks_mutex.
2540                  *
2541                  * Operations on ->flags have to be atomic because
2542                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2543                  * mm_all_locks_mutex, there may be other cpus
2544                  * changing other bitflags in parallel to us.
2545                  */
2546                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2547                         BUG();
2548                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2549         }
2550 }
2551
2552 /*
2553  * This operation locks against the VM for all pte/vma/mm related
2554  * operations that could ever happen on a certain mm. This includes
2555  * vmtruncate, try_to_unmap, and all page faults.
2556  *
2557  * The caller must take the mmap_sem in write mode before calling
2558  * mm_take_all_locks(). The caller isn't allowed to release the
2559  * mmap_sem until mm_drop_all_locks() returns.
2560  *
2561  * mmap_sem in write mode is required in order to block all operations
2562  * that could modify pagetables and free pages without need of
2563  * altering the vma layout (for example populate_range() with
2564  * nonlinear vmas). It's also needed in write mode to avoid new
2565  * anon_vmas to be associated with existing vmas.
2566  *
2567  * A single task can't take more than one mm_take_all_locks() in a row
2568  * or it would deadlock.
2569  *
2570  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2571  * mapping->flags avoid to take the same lock twice, if more than one
2572  * vma in this mm is backed by the same anon_vma or address_space.
2573  *
2574  * We can take all the locks in random order because the VM code
2575  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2576  * takes more than one of them in a row. Secondly we're protected
2577  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2578  *
2579  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2580  * that may have to take thousand of locks.
2581  *
2582  * mm_take_all_locks() can fail if it's interrupted by signals.
2583  */
2584 int mm_take_all_locks(struct mm_struct *mm)
2585 {
2586         struct vm_area_struct *vma;
2587         struct anon_vma_chain *avc;
2588         int ret = -EINTR;
2589
2590         BUG_ON(down_read_trylock(&mm->mmap_sem));
2591
2592         mutex_lock(&mm_all_locks_mutex);
2593
2594         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2595                 if (signal_pending(current))
2596                         goto out_unlock;
2597                 if (vma->vm_file && vma->vm_file->f_mapping)
2598                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2599         }
2600
2601         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2602                 if (signal_pending(current))
2603                         goto out_unlock;
2604                 if (vma->anon_vma)
2605                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2606                                 vm_lock_anon_vma(mm, avc->anon_vma);
2607         }
2608
2609         ret = 0;
2610
2611 out_unlock:
2612         if (ret)
2613                 mm_drop_all_locks(mm);
2614
2615         return ret;
2616 }
2617
2618 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2619 {
2620         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2621                 /*
2622                  * The LSB of head.next can't change to 0 from under
2623                  * us because we hold the mm_all_locks_mutex.
2624                  *
2625                  * We must however clear the bitflag before unlocking
2626                  * the vma so the users using the anon_vma->head will
2627                  * never see our bitflag.
2628                  *
2629                  * No need of atomic instructions here, head.next
2630                  * can't change from under us until we release the
2631                  * anon_vma->root->lock.
2632                  */
2633                 if (!__test_and_clear_bit(0, (unsigned long *)
2634                                           &anon_vma->root->head.next))
2635                         BUG();
2636                 anon_vma_unlock(anon_vma);
2637         }
2638 }
2639
2640 static void vm_unlock_mapping(struct address_space *mapping)
2641 {
2642         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2643                 /*
2644                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2645                  * because we hold the mm_all_locks_mutex.
2646                  */
2647                 spin_unlock(&mapping->i_mmap_lock);
2648                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2649                                         &mapping->flags))
2650                         BUG();
2651         }
2652 }
2653
2654 /*
2655  * The mmap_sem cannot be released by the caller until
2656  * mm_drop_all_locks() returns.
2657  */
2658 void mm_drop_all_locks(struct mm_struct *mm)
2659 {
2660         struct vm_area_struct *vma;
2661         struct anon_vma_chain *avc;
2662
2663         BUG_ON(down_read_trylock(&mm->mmap_sem));
2664         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2665
2666         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2667                 if (vma->anon_vma)
2668                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2669                                 vm_unlock_anon_vma(avc->anon_vma);
2670                 if (vma->vm_file && vma->vm_file->f_mapping)
2671                         vm_unlock_mapping(vma->vm_file->f_mapping);
2672         }
2673
2674         mutex_unlock(&mm_all_locks_mutex);
2675 }
2676
2677 /*
2678  * initialise the VMA slab
2679  */
2680 void __init mmap_init(void)
2681 {
2682         int ret;
2683
2684         ret = percpu_counter_init(&vm_committed_as, 0);
2685         VM_BUG_ON(ret);
2686 }