brk: fix min_brk lower bound computation for COMPAT_BRK
[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         /*
258          * CONFIG_COMPAT_BRK can still be overridden by setting
259          * randomize_va_space to 2, which will still cause mm->start_brk
260          * to be arbitrarily shifted
261          */
262         if (mm->start_brk > PAGE_ALIGN(mm->end_data))
263                 min_brk = mm->start_brk;
264         else
265                 min_brk = mm->end_data;
266 #else
267         min_brk = mm->start_brk;
268 #endif
269         if (brk < min_brk)
270                 goto out;
271
272         /*
273          * Check against rlimit here. If this check is done later after the test
274          * of oldbrk with newbrk then it can escape the test and let the data
275          * segment grow beyond its set limit the in case where the limit is
276          * not page aligned -Ram Gupta
277          */
278         rlim = rlimit(RLIMIT_DATA);
279         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
280                         (mm->end_data - mm->start_data) > rlim)
281                 goto out;
282
283         newbrk = PAGE_ALIGN(brk);
284         oldbrk = PAGE_ALIGN(mm->brk);
285         if (oldbrk == newbrk)
286                 goto set_brk;
287
288         /* Always allow shrinking brk. */
289         if (brk <= mm->brk) {
290                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
291                         goto set_brk;
292                 goto out;
293         }
294
295         /* Check against existing mmap mappings. */
296         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
297                 goto out;
298
299         /* Ok, looks good - let it rip. */
300         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
301                 goto out;
302 set_brk:
303         mm->brk = brk;
304 out:
305         retval = mm->brk;
306         up_write(&mm->mmap_sem);
307         return retval;
308 }
309
310 #ifdef DEBUG_MM_RB
311 static int browse_rb(struct rb_root *root)
312 {
313         int i = 0, j;
314         struct rb_node *nd, *pn = NULL;
315         unsigned long prev = 0, pend = 0;
316
317         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
318                 struct vm_area_struct *vma;
319                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
320                 if (vma->vm_start < prev)
321                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
322                 if (vma->vm_start < pend)
323                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
324                 if (vma->vm_start > vma->vm_end)
325                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
326                 i++;
327                 pn = nd;
328                 prev = vma->vm_start;
329                 pend = vma->vm_end;
330         }
331         j = 0;
332         for (nd = pn; nd; nd = rb_prev(nd)) {
333                 j++;
334         }
335         if (i != j)
336                 printk("backwards %d, forwards %d\n", j, i), i = 0;
337         return i;
338 }
339
340 void validate_mm(struct mm_struct *mm)
341 {
342         int bug = 0;
343         int i = 0;
344         struct vm_area_struct *tmp = mm->mmap;
345         while (tmp) {
346                 tmp = tmp->vm_next;
347                 i++;
348         }
349         if (i != mm->map_count)
350                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
351         i = browse_rb(&mm->mm_rb);
352         if (i != mm->map_count)
353                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
354         BUG_ON(bug);
355 }
356 #else
357 #define validate_mm(mm) do { } while (0)
358 #endif
359
360 static struct vm_area_struct *
361 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
362                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
363                 struct rb_node ** rb_parent)
364 {
365         struct vm_area_struct * vma;
366         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
367
368         __rb_link = &mm->mm_rb.rb_node;
369         rb_prev = __rb_parent = NULL;
370         vma = NULL;
371
372         while (*__rb_link) {
373                 struct vm_area_struct *vma_tmp;
374
375                 __rb_parent = *__rb_link;
376                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
377
378                 if (vma_tmp->vm_end > addr) {
379                         vma = vma_tmp;
380                         if (vma_tmp->vm_start <= addr)
381                                 break;
382                         __rb_link = &__rb_parent->rb_left;
383                 } else {
384                         rb_prev = __rb_parent;
385                         __rb_link = &__rb_parent->rb_right;
386                 }
387         }
388
389         *pprev = NULL;
390         if (rb_prev)
391                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
392         *rb_link = __rb_link;
393         *rb_parent = __rb_parent;
394         return vma;
395 }
396
397 static inline void
398 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
399                 struct vm_area_struct *prev, struct rb_node *rb_parent)
400 {
401         struct vm_area_struct *next;
402
403         vma->vm_prev = prev;
404         if (prev) {
405                 next = prev->vm_next;
406                 prev->vm_next = vma;
407         } else {
408                 mm->mmap = vma;
409                 if (rb_parent)
410                         next = rb_entry(rb_parent,
411                                         struct vm_area_struct, vm_rb);
412                 else
413                         next = NULL;
414         }
415         vma->vm_next = next;
416         if (next)
417                 next->vm_prev = vma;
418 }
419
420 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
421                 struct rb_node **rb_link, struct rb_node *rb_parent)
422 {
423         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
424         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
425 }
426
427 static void __vma_link_file(struct vm_area_struct *vma)
428 {
429         struct file *file;
430
431         file = vma->vm_file;
432         if (file) {
433                 struct address_space *mapping = file->f_mapping;
434
435                 if (vma->vm_flags & VM_DENYWRITE)
436                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
437                 if (vma->vm_flags & VM_SHARED)
438                         mapping->i_mmap_writable++;
439
440                 flush_dcache_mmap_lock(mapping);
441                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
442                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
443                 else
444                         vma_prio_tree_insert(vma, &mapping->i_mmap);
445                 flush_dcache_mmap_unlock(mapping);
446         }
447 }
448
449 static void
450 __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         __vma_link_list(mm, vma, prev, rb_parent);
455         __vma_link_rb(mm, vma, rb_link, rb_parent);
456 }
457
458 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
459                         struct vm_area_struct *prev, struct rb_node **rb_link,
460                         struct rb_node *rb_parent)
461 {
462         struct address_space *mapping = NULL;
463
464         if (vma->vm_file)
465                 mapping = vma->vm_file->f_mapping;
466
467         if (mapping) {
468                 spin_lock(&mapping->i_mmap_lock);
469                 vma->vm_truncate_count = mapping->truncate_count;
470         }
471
472         __vma_link(mm, vma, prev, rb_link, rb_parent);
473         __vma_link_file(vma);
474
475         if (mapping)
476                 spin_unlock(&mapping->i_mmap_lock);
477
478         mm->map_count++;
479         validate_mm(mm);
480 }
481
482 /*
483  * Helper for vma_adjust in the split_vma insert case:
484  * insert vm structure into list and rbtree and anon_vma,
485  * but it has already been inserted into prio_tree earlier.
486  */
487 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
488 {
489         struct vm_area_struct *__vma, *prev;
490         struct rb_node **rb_link, *rb_parent;
491
492         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
493         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
494         __vma_link(mm, vma, prev, rb_link, rb_parent);
495         mm->map_count++;
496 }
497
498 static inline void
499 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
500                 struct vm_area_struct *prev)
501 {
502         struct vm_area_struct *next = vma->vm_next;
503
504         prev->vm_next = next;
505         if (next)
506                 next->vm_prev = prev;
507         rb_erase(&vma->vm_rb, &mm->mm_rb);
508         if (mm->mmap_cache == vma)
509                 mm->mmap_cache = prev;
510 }
511
512 /*
513  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
514  * is already present in an i_mmap tree without adjusting the tree.
515  * The following helper function should be used when such adjustments
516  * are necessary.  The "insert" vma (if any) is to be inserted
517  * before we drop the necessary locks.
518  */
519 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
520         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
521 {
522         struct mm_struct *mm = vma->vm_mm;
523         struct vm_area_struct *next = vma->vm_next;
524         struct vm_area_struct *importer = NULL;
525         struct address_space *mapping = NULL;
526         struct prio_tree_root *root = NULL;
527         struct anon_vma *anon_vma = NULL;
528         struct file *file = vma->vm_file;
529         long adjust_next = 0;
530         int remove_next = 0;
531
532         if (next && !insert) {
533                 struct vm_area_struct *exporter = NULL;
534
535                 if (end >= next->vm_end) {
536                         /*
537                          * vma expands, overlapping all the next, and
538                          * perhaps the one after too (mprotect case 6).
539                          */
540 again:                  remove_next = 1 + (end > next->vm_end);
541                         end = next->vm_end;
542                         exporter = next;
543                         importer = vma;
544                 } else if (end > next->vm_start) {
545                         /*
546                          * vma expands, overlapping part of the next:
547                          * mprotect case 5 shifting the boundary up.
548                          */
549                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
550                         exporter = next;
551                         importer = vma;
552                 } else if (end < vma->vm_end) {
553                         /*
554                          * vma shrinks, and !insert tells it's not
555                          * split_vma inserting another: so it must be
556                          * mprotect case 4 shifting the boundary down.
557                          */
558                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
559                         exporter = vma;
560                         importer = next;
561                 }
562
563                 /*
564                  * Easily overlooked: when mprotect shifts the boundary,
565                  * make sure the expanding vma has anon_vma set if the
566                  * shrinking vma had, to cover any anon pages imported.
567                  */
568                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
569                         if (anon_vma_clone(importer, exporter))
570                                 return -ENOMEM;
571                         importer->anon_vma = exporter->anon_vma;
572                 }
573         }
574
575         if (file) {
576                 mapping = file->f_mapping;
577                 if (!(vma->vm_flags & VM_NONLINEAR))
578                         root = &mapping->i_mmap;
579                 spin_lock(&mapping->i_mmap_lock);
580                 if (importer &&
581                     vma->vm_truncate_count != next->vm_truncate_count) {
582                         /*
583                          * unmap_mapping_range might be in progress:
584                          * ensure that the expanding vma is rescanned.
585                          */
586                         importer->vm_truncate_count = 0;
587                 }
588                 if (insert) {
589                         insert->vm_truncate_count = vma->vm_truncate_count;
590                         /*
591                          * Put into prio_tree now, so instantiated pages
592                          * are visible to arm/parisc __flush_dcache_page
593                          * throughout; but we cannot insert into address
594                          * space until vma start or end is updated.
595                          */
596                         __vma_link_file(insert);
597                 }
598         }
599
600         vma_adjust_trans_huge(vma, start, end, adjust_next);
601
602         /*
603          * When changing only vma->vm_end, we don't really need anon_vma
604          * lock. This is a fairly rare case by itself, but the anon_vma
605          * lock may be shared between many sibling processes.  Skipping
606          * the lock for brk adjustments makes a difference sometimes.
607          */
608         if (vma->anon_vma && (insert || importer || start != vma->vm_start)) {
609                 anon_vma = vma->anon_vma;
610                 anon_vma_lock(anon_vma);
611         }
612
613         if (root) {
614                 flush_dcache_mmap_lock(mapping);
615                 vma_prio_tree_remove(vma, root);
616                 if (adjust_next)
617                         vma_prio_tree_remove(next, root);
618         }
619
620         vma->vm_start = start;
621         vma->vm_end = end;
622         vma->vm_pgoff = pgoff;
623         if (adjust_next) {
624                 next->vm_start += adjust_next << PAGE_SHIFT;
625                 next->vm_pgoff += adjust_next;
626         }
627
628         if (root) {
629                 if (adjust_next)
630                         vma_prio_tree_insert(next, root);
631                 vma_prio_tree_insert(vma, root);
632                 flush_dcache_mmap_unlock(mapping);
633         }
634
635         if (remove_next) {
636                 /*
637                  * vma_merge has merged next into vma, and needs
638                  * us to remove next before dropping the locks.
639                  */
640                 __vma_unlink(mm, next, vma);
641                 if (file)
642                         __remove_shared_vm_struct(next, file, mapping);
643         } else if (insert) {
644                 /*
645                  * split_vma has split insert from vma, and needs
646                  * us to insert it before dropping the locks
647                  * (it may either follow vma or precede it).
648                  */
649                 __insert_vm_struct(mm, insert);
650         }
651
652         if (anon_vma)
653                 anon_vma_unlock(anon_vma);
654         if (mapping)
655                 spin_unlock(&mapping->i_mmap_lock);
656
657         if (remove_next) {
658                 if (file) {
659                         fput(file);
660                         if (next->vm_flags & VM_EXECUTABLE)
661                                 removed_exe_file_vma(mm);
662                 }
663                 if (next->anon_vma)
664                         anon_vma_merge(vma, next);
665                 mm->map_count--;
666                 mpol_put(vma_policy(next));
667                 kmem_cache_free(vm_area_cachep, next);
668                 /*
669                  * In mprotect's case 6 (see comments on vma_merge),
670                  * we must remove another next too. It would clutter
671                  * up the code too much to do both in one go.
672                  */
673                 if (remove_next == 2) {
674                         next = vma->vm_next;
675                         goto again;
676                 }
677         }
678
679         validate_mm(mm);
680
681         return 0;
682 }
683
684 /*
685  * If the vma has a ->close operation then the driver probably needs to release
686  * per-vma resources, so we don't attempt to merge those.
687  */
688 static inline int is_mergeable_vma(struct vm_area_struct *vma,
689                         struct file *file, unsigned long vm_flags)
690 {
691         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
692         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
693                 return 0;
694         if (vma->vm_file != file)
695                 return 0;
696         if (vma->vm_ops && vma->vm_ops->close)
697                 return 0;
698         return 1;
699 }
700
701 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
702                                         struct anon_vma *anon_vma2)
703 {
704         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
705 }
706
707 /*
708  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
709  * in front of (at a lower virtual address and file offset than) the vma.
710  *
711  * We cannot merge two vmas if they have differently assigned (non-NULL)
712  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
713  *
714  * We don't check here for the merged mmap wrapping around the end of pagecache
715  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
716  * wrap, nor mmaps which cover the final page at index -1UL.
717  */
718 static int
719 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
720         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
721 {
722         if (is_mergeable_vma(vma, file, vm_flags) &&
723             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
724                 if (vma->vm_pgoff == vm_pgoff)
725                         return 1;
726         }
727         return 0;
728 }
729
730 /*
731  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
732  * beyond (at a higher virtual address and file offset than) the vma.
733  *
734  * We cannot merge two vmas if they have differently assigned (non-NULL)
735  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
736  */
737 static int
738 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
739         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
740 {
741         if (is_mergeable_vma(vma, file, vm_flags) &&
742             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
743                 pgoff_t vm_pglen;
744                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
745                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
746                         return 1;
747         }
748         return 0;
749 }
750
751 /*
752  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
753  * whether that can be merged with its predecessor or its successor.
754  * Or both (it neatly fills a hole).
755  *
756  * In most cases - when called for mmap, brk or mremap - [addr,end) is
757  * certain not to be mapped by the time vma_merge is called; but when
758  * called for mprotect, it is certain to be already mapped (either at
759  * an offset within prev, or at the start of next), and the flags of
760  * this area are about to be changed to vm_flags - and the no-change
761  * case has already been eliminated.
762  *
763  * The following mprotect cases have to be considered, where AAAA is
764  * the area passed down from mprotect_fixup, never extending beyond one
765  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
766  *
767  *     AAAA             AAAA                AAAA          AAAA
768  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
769  *    cannot merge    might become    might become    might become
770  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
771  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
772  *    mremap move:                                    PPPPNNNNNNNN 8
773  *        AAAA
774  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
775  *    might become    case 1 below    case 2 below    case 3 below
776  *
777  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
778  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
779  */
780 struct vm_area_struct *vma_merge(struct mm_struct *mm,
781                         struct vm_area_struct *prev, unsigned long addr,
782                         unsigned long end, unsigned long vm_flags,
783                         struct anon_vma *anon_vma, struct file *file,
784                         pgoff_t pgoff, struct mempolicy *policy)
785 {
786         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
787         struct vm_area_struct *area, *next;
788         int err;
789
790         /*
791          * We later require that vma->vm_flags == vm_flags,
792          * so this tests vma->vm_flags & VM_SPECIAL, too.
793          */
794         if (vm_flags & VM_SPECIAL)
795                 return NULL;
796
797         if (prev)
798                 next = prev->vm_next;
799         else
800                 next = mm->mmap;
801         area = next;
802         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
803                 next = next->vm_next;
804
805         /*
806          * Can it merge with the predecessor?
807          */
808         if (prev && prev->vm_end == addr &&
809                         mpol_equal(vma_policy(prev), policy) &&
810                         can_vma_merge_after(prev, vm_flags,
811                                                 anon_vma, file, pgoff)) {
812                 /*
813                  * OK, it can.  Can we now merge in the successor as well?
814                  */
815                 if (next && end == next->vm_start &&
816                                 mpol_equal(policy, vma_policy(next)) &&
817                                 can_vma_merge_before(next, vm_flags,
818                                         anon_vma, file, pgoff+pglen) &&
819                                 is_mergeable_anon_vma(prev->anon_vma,
820                                                       next->anon_vma)) {
821                                                         /* cases 1, 6 */
822                         err = vma_adjust(prev, prev->vm_start,
823                                 next->vm_end, prev->vm_pgoff, NULL);
824                 } else                                  /* cases 2, 5, 7 */
825                         err = vma_adjust(prev, prev->vm_start,
826                                 end, prev->vm_pgoff, NULL);
827                 if (err)
828                         return NULL;
829                 khugepaged_enter_vma_merge(prev);
830                 return prev;
831         }
832
833         /*
834          * Can this new request be merged in front of next?
835          */
836         if (next && end == next->vm_start &&
837                         mpol_equal(policy, vma_policy(next)) &&
838                         can_vma_merge_before(next, vm_flags,
839                                         anon_vma, file, pgoff+pglen)) {
840                 if (prev && addr < prev->vm_end)        /* case 4 */
841                         err = vma_adjust(prev, prev->vm_start,
842                                 addr, prev->vm_pgoff, NULL);
843                 else                                    /* cases 3, 8 */
844                         err = vma_adjust(area, addr, next->vm_end,
845                                 next->vm_pgoff - pglen, NULL);
846                 if (err)
847                         return NULL;
848                 khugepaged_enter_vma_merge(area);
849                 return area;
850         }
851
852         return NULL;
853 }
854
855 /*
856  * Rough compatbility check to quickly see if it's even worth looking
857  * at sharing an anon_vma.
858  *
859  * They need to have the same vm_file, and the flags can only differ
860  * in things that mprotect may change.
861  *
862  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
863  * we can merge the two vma's. For example, we refuse to merge a vma if
864  * there is a vm_ops->close() function, because that indicates that the
865  * driver is doing some kind of reference counting. But that doesn't
866  * really matter for the anon_vma sharing case.
867  */
868 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
869 {
870         return a->vm_end == b->vm_start &&
871                 mpol_equal(vma_policy(a), vma_policy(b)) &&
872                 a->vm_file == b->vm_file &&
873                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
874                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
875 }
876
877 /*
878  * Do some basic sanity checking to see if we can re-use the anon_vma
879  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
880  * the same as 'old', the other will be the new one that is trying
881  * to share the anon_vma.
882  *
883  * NOTE! This runs with mm_sem held for reading, so it is possible that
884  * the anon_vma of 'old' is concurrently in the process of being set up
885  * by another page fault trying to merge _that_. But that's ok: if it
886  * is being set up, that automatically means that it will be a singleton
887  * acceptable for merging, so we can do all of this optimistically. But
888  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
889  *
890  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
891  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
892  * is to return an anon_vma that is "complex" due to having gone through
893  * a fork).
894  *
895  * We also make sure that the two vma's are compatible (adjacent,
896  * and with the same memory policies). That's all stable, even with just
897  * a read lock on the mm_sem.
898  */
899 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
900 {
901         if (anon_vma_compatible(a, b)) {
902                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
903
904                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
905                         return anon_vma;
906         }
907         return NULL;
908 }
909
910 /*
911  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
912  * neighbouring vmas for a suitable anon_vma, before it goes off
913  * to allocate a new anon_vma.  It checks because a repetitive
914  * sequence of mprotects and faults may otherwise lead to distinct
915  * anon_vmas being allocated, preventing vma merge in subsequent
916  * mprotect.
917  */
918 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
919 {
920         struct anon_vma *anon_vma;
921         struct vm_area_struct *near;
922
923         near = vma->vm_next;
924         if (!near)
925                 goto try_prev;
926
927         anon_vma = reusable_anon_vma(near, vma, near);
928         if (anon_vma)
929                 return anon_vma;
930 try_prev:
931         /*
932          * It is potentially slow to have to call find_vma_prev here.
933          * But it's only on the first write fault on the vma, not
934          * every time, and we could devise a way to avoid it later
935          * (e.g. stash info in next's anon_vma_node when assigning
936          * an anon_vma, or when trying vma_merge).  Another time.
937          */
938         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
939         if (!near)
940                 goto none;
941
942         anon_vma = reusable_anon_vma(near, near, vma);
943         if (anon_vma)
944                 return anon_vma;
945 none:
946         /*
947          * There's no absolute need to look only at touching neighbours:
948          * we could search further afield for "compatible" anon_vmas.
949          * But it would probably just be a waste of time searching,
950          * or lead to too many vmas hanging off the same anon_vma.
951          * We're trying to allow mprotect remerging later on,
952          * not trying to minimize memory used for anon_vmas.
953          */
954         return NULL;
955 }
956
957 #ifdef CONFIG_PROC_FS
958 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
959                                                 struct file *file, long pages)
960 {
961         const unsigned long stack_flags
962                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
963
964         if (file) {
965                 mm->shared_vm += pages;
966                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
967                         mm->exec_vm += pages;
968         } else if (flags & stack_flags)
969                 mm->stack_vm += pages;
970         if (flags & (VM_RESERVED|VM_IO))
971                 mm->reserved_vm += pages;
972 }
973 #endif /* CONFIG_PROC_FS */
974
975 /*
976  * The caller must hold down_write(&current->mm->mmap_sem).
977  */
978
979 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
980                         unsigned long len, unsigned long prot,
981                         unsigned long flags, unsigned long pgoff)
982 {
983         struct mm_struct * mm = current->mm;
984         struct inode *inode;
985         unsigned int vm_flags;
986         int error;
987         unsigned long reqprot = prot;
988
989         /*
990          * Does the application expect PROT_READ to imply PROT_EXEC?
991          *
992          * (the exception is when the underlying filesystem is noexec
993          *  mounted, in which case we dont add PROT_EXEC.)
994          */
995         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
996                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
997                         prot |= PROT_EXEC;
998
999         if (!len)
1000                 return -EINVAL;
1001
1002         if (!(flags & MAP_FIXED))
1003                 addr = round_hint_to_min(addr);
1004
1005         /* Careful about overflows.. */
1006         len = PAGE_ALIGN(len);
1007         if (!len)
1008                 return -ENOMEM;
1009
1010         /* offset overflow? */
1011         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1012                return -EOVERFLOW;
1013
1014         /* Too many mappings? */
1015         if (mm->map_count > sysctl_max_map_count)
1016                 return -ENOMEM;
1017
1018         /* Obtain the address to map to. we verify (or select) it and ensure
1019          * that it represents a valid section of the address space.
1020          */
1021         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1022         if (addr & ~PAGE_MASK)
1023                 return addr;
1024
1025         /* Do simple checking here so the lower-level routines won't have
1026          * to. we assume access permissions have been handled by the open
1027          * of the memory object, so we don't do any here.
1028          */
1029         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1030                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1031
1032         if (flags & MAP_LOCKED)
1033                 if (!can_do_mlock())
1034                         return -EPERM;
1035
1036         /* mlock MCL_FUTURE? */
1037         if (vm_flags & VM_LOCKED) {
1038                 unsigned long locked, lock_limit;
1039                 locked = len >> PAGE_SHIFT;
1040                 locked += mm->locked_vm;
1041                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1042                 lock_limit >>= PAGE_SHIFT;
1043                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1044                         return -EAGAIN;
1045         }
1046
1047         inode = file ? file->f_path.dentry->d_inode : NULL;
1048
1049         if (file) {
1050                 switch (flags & MAP_TYPE) {
1051                 case MAP_SHARED:
1052                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1053                                 return -EACCES;
1054
1055                         /*
1056                          * Make sure we don't allow writing to an append-only
1057                          * file..
1058                          */
1059                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1060                                 return -EACCES;
1061
1062                         /*
1063                          * Make sure there are no mandatory locks on the file.
1064                          */
1065                         if (locks_verify_locked(inode))
1066                                 return -EAGAIN;
1067
1068                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1069                         if (!(file->f_mode & FMODE_WRITE))
1070                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1071
1072                         /* fall through */
1073                 case MAP_PRIVATE:
1074                         if (!(file->f_mode & FMODE_READ))
1075                                 return -EACCES;
1076                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1077                                 if (vm_flags & VM_EXEC)
1078                                         return -EPERM;
1079                                 vm_flags &= ~VM_MAYEXEC;
1080                         }
1081
1082                         if (!file->f_op || !file->f_op->mmap)
1083                                 return -ENODEV;
1084                         break;
1085
1086                 default:
1087                         return -EINVAL;
1088                 }
1089         } else {
1090                 switch (flags & MAP_TYPE) {
1091                 case MAP_SHARED:
1092                         /*
1093                          * Ignore pgoff.
1094                          */
1095                         pgoff = 0;
1096                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1097                         break;
1098                 case MAP_PRIVATE:
1099                         /*
1100                          * Set pgoff according to addr for anon_vma.
1101                          */
1102                         pgoff = addr >> PAGE_SHIFT;
1103                         break;
1104                 default:
1105                         return -EINVAL;
1106                 }
1107         }
1108
1109         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1110         if (error)
1111                 return error;
1112
1113         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1114 }
1115 EXPORT_SYMBOL(do_mmap_pgoff);
1116
1117 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1118                 unsigned long, prot, unsigned long, flags,
1119                 unsigned long, fd, unsigned long, pgoff)
1120 {
1121         struct file *file = NULL;
1122         unsigned long retval = -EBADF;
1123
1124         if (!(flags & MAP_ANONYMOUS)) {
1125                 audit_mmap_fd(fd, flags);
1126                 if (unlikely(flags & MAP_HUGETLB))
1127                         return -EINVAL;
1128                 file = fget(fd);
1129                 if (!file)
1130                         goto out;
1131         } else if (flags & MAP_HUGETLB) {
1132                 struct user_struct *user = NULL;
1133                 /*
1134                  * VM_NORESERVE is used because the reservations will be
1135                  * taken when vm_ops->mmap() is called
1136                  * A dummy user value is used because we are not locking
1137                  * memory so no accounting is necessary
1138                  */
1139                 len = ALIGN(len, huge_page_size(&default_hstate));
1140                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1141                                                 &user, HUGETLB_ANONHUGE_INODE);
1142                 if (IS_ERR(file))
1143                         return PTR_ERR(file);
1144         }
1145
1146         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1147
1148         down_write(&current->mm->mmap_sem);
1149         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1150         up_write(&current->mm->mmap_sem);
1151
1152         if (file)
1153                 fput(file);
1154 out:
1155         return retval;
1156 }
1157
1158 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1159 struct mmap_arg_struct {
1160         unsigned long addr;
1161         unsigned long len;
1162         unsigned long prot;
1163         unsigned long flags;
1164         unsigned long fd;
1165         unsigned long offset;
1166 };
1167
1168 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1169 {
1170         struct mmap_arg_struct a;
1171
1172         if (copy_from_user(&a, arg, sizeof(a)))
1173                 return -EFAULT;
1174         if (a.offset & ~PAGE_MASK)
1175                 return -EINVAL;
1176
1177         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1178                               a.offset >> PAGE_SHIFT);
1179 }
1180 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1181
1182 /*
1183  * Some shared mappigns will want the pages marked read-only
1184  * to track write events. If so, we'll downgrade vm_page_prot
1185  * to the private version (using protection_map[] without the
1186  * VM_SHARED bit).
1187  */
1188 int vma_wants_writenotify(struct vm_area_struct *vma)
1189 {
1190         unsigned int vm_flags = vma->vm_flags;
1191
1192         /* If it was private or non-writable, the write bit is already clear */
1193         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1194                 return 0;
1195
1196         /* The backer wishes to know when pages are first written to? */
1197         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1198                 return 1;
1199
1200         /* The open routine did something to the protections already? */
1201         if (pgprot_val(vma->vm_page_prot) !=
1202             pgprot_val(vm_get_page_prot(vm_flags)))
1203                 return 0;
1204
1205         /* Specialty mapping? */
1206         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1207                 return 0;
1208
1209         /* Can the mapping track the dirty pages? */
1210         return vma->vm_file && vma->vm_file->f_mapping &&
1211                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1212 }
1213
1214 /*
1215  * We account for memory if it's a private writeable mapping,
1216  * not hugepages and VM_NORESERVE wasn't set.
1217  */
1218 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1219 {
1220         /*
1221          * hugetlb has its own accounting separate from the core VM
1222          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1223          */
1224         if (file && is_file_hugepages(file))
1225                 return 0;
1226
1227         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1228 }
1229
1230 unsigned long mmap_region(struct file *file, unsigned long addr,
1231                           unsigned long len, unsigned long flags,
1232                           unsigned int vm_flags, unsigned long pgoff)
1233 {
1234         struct mm_struct *mm = current->mm;
1235         struct vm_area_struct *vma, *prev;
1236         int correct_wcount = 0;
1237         int error;
1238         struct rb_node **rb_link, *rb_parent;
1239         unsigned long charged = 0;
1240         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1241
1242         /* Clear old maps */
1243         error = -ENOMEM;
1244 munmap_back:
1245         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1246         if (vma && vma->vm_start < addr + len) {
1247                 if (do_munmap(mm, addr, len))
1248                         return -ENOMEM;
1249                 goto munmap_back;
1250         }
1251
1252         /* Check against address space limit. */
1253         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1254                 return -ENOMEM;
1255
1256         /*
1257          * Set 'VM_NORESERVE' if we should not account for the
1258          * memory use of this mapping.
1259          */
1260         if ((flags & MAP_NORESERVE)) {
1261                 /* We honor MAP_NORESERVE if allowed to overcommit */
1262                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1263                         vm_flags |= VM_NORESERVE;
1264
1265                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1266                 if (file && is_file_hugepages(file))
1267                         vm_flags |= VM_NORESERVE;
1268         }
1269
1270         /*
1271          * Private writable mapping: check memory availability
1272          */
1273         if (accountable_mapping(file, vm_flags)) {
1274                 charged = len >> PAGE_SHIFT;
1275                 if (security_vm_enough_memory(charged))
1276                         return -ENOMEM;
1277                 vm_flags |= VM_ACCOUNT;
1278         }
1279
1280         /*
1281          * Can we just expand an old mapping?
1282          */
1283         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1284         if (vma)
1285                 goto out;
1286
1287         /*
1288          * Determine the object being mapped and call the appropriate
1289          * specific mapper. the address has already been validated, but
1290          * not unmapped, but the maps are removed from the list.
1291          */
1292         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1293         if (!vma) {
1294                 error = -ENOMEM;
1295                 goto unacct_error;
1296         }
1297
1298         vma->vm_mm = mm;
1299         vma->vm_start = addr;
1300         vma->vm_end = addr + len;
1301         vma->vm_flags = vm_flags;
1302         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1303         vma->vm_pgoff = pgoff;
1304         INIT_LIST_HEAD(&vma->anon_vma_chain);
1305
1306         if (file) {
1307                 error = -EINVAL;
1308                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1309                         goto free_vma;
1310                 if (vm_flags & VM_DENYWRITE) {
1311                         error = deny_write_access(file);
1312                         if (error)
1313                                 goto free_vma;
1314                         correct_wcount = 1;
1315                 }
1316                 vma->vm_file = file;
1317                 get_file(file);
1318                 error = file->f_op->mmap(file, vma);
1319                 if (error)
1320                         goto unmap_and_free_vma;
1321                 if (vm_flags & VM_EXECUTABLE)
1322                         added_exe_file_vma(mm);
1323
1324                 /* Can addr have changed??
1325                  *
1326                  * Answer: Yes, several device drivers can do it in their
1327                  *         f_op->mmap method. -DaveM
1328                  */
1329                 addr = vma->vm_start;
1330                 pgoff = vma->vm_pgoff;
1331                 vm_flags = vma->vm_flags;
1332         } else if (vm_flags & VM_SHARED) {
1333                 error = shmem_zero_setup(vma);
1334                 if (error)
1335                         goto free_vma;
1336         }
1337
1338         if (vma_wants_writenotify(vma)) {
1339                 pgprot_t pprot = vma->vm_page_prot;
1340
1341                 /* Can vma->vm_page_prot have changed??
1342                  *
1343                  * Answer: Yes, drivers may have changed it in their
1344                  *         f_op->mmap method.
1345                  *
1346                  * Ensures that vmas marked as uncached stay that way.
1347                  */
1348                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1349                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1350                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1351         }
1352
1353         vma_link(mm, vma, prev, rb_link, rb_parent);
1354         file = vma->vm_file;
1355
1356         /* Once vma denies write, undo our temporary denial count */
1357         if (correct_wcount)
1358                 atomic_inc(&inode->i_writecount);
1359 out:
1360         perf_event_mmap(vma);
1361
1362         mm->total_vm += len >> PAGE_SHIFT;
1363         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1364         if (vm_flags & VM_LOCKED) {
1365                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1366                         mm->locked_vm += (len >> PAGE_SHIFT);
1367         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1368                 make_pages_present(addr, addr + len);
1369         return addr;
1370
1371 unmap_and_free_vma:
1372         if (correct_wcount)
1373                 atomic_inc(&inode->i_writecount);
1374         vma->vm_file = NULL;
1375         fput(file);
1376
1377         /* Undo any partial mapping done by a device driver. */
1378         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1379         charged = 0;
1380 free_vma:
1381         kmem_cache_free(vm_area_cachep, vma);
1382 unacct_error:
1383         if (charged)
1384                 vm_unacct_memory(charged);
1385         return error;
1386 }
1387
1388 /* Get an address range which is currently unmapped.
1389  * For shmat() with addr=0.
1390  *
1391  * Ugly calling convention alert:
1392  * Return value with the low bits set means error value,
1393  * ie
1394  *      if (ret & ~PAGE_MASK)
1395  *              error = ret;
1396  *
1397  * This function "knows" that -ENOMEM has the bits set.
1398  */
1399 #ifndef HAVE_ARCH_UNMAPPED_AREA
1400 unsigned long
1401 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1402                 unsigned long len, unsigned long pgoff, unsigned long flags)
1403 {
1404         struct mm_struct *mm = current->mm;
1405         struct vm_area_struct *vma;
1406         unsigned long start_addr;
1407
1408         if (len > TASK_SIZE)
1409                 return -ENOMEM;
1410
1411         if (flags & MAP_FIXED)
1412                 return addr;
1413
1414         if (addr) {
1415                 addr = PAGE_ALIGN(addr);
1416                 vma = find_vma(mm, addr);
1417                 if (TASK_SIZE - len >= addr &&
1418                     (!vma || addr + len <= vma->vm_start))
1419                         return addr;
1420         }
1421         if (len > mm->cached_hole_size) {
1422                 start_addr = addr = mm->free_area_cache;
1423         } else {
1424                 start_addr = addr = TASK_UNMAPPED_BASE;
1425                 mm->cached_hole_size = 0;
1426         }
1427
1428 full_search:
1429         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1430                 /* At this point:  (!vma || addr < vma->vm_end). */
1431                 if (TASK_SIZE - len < addr) {
1432                         /*
1433                          * Start a new search - just in case we missed
1434                          * some holes.
1435                          */
1436                         if (start_addr != TASK_UNMAPPED_BASE) {
1437                                 addr = TASK_UNMAPPED_BASE;
1438                                 start_addr = addr;
1439                                 mm->cached_hole_size = 0;
1440                                 goto full_search;
1441                         }
1442                         return -ENOMEM;
1443                 }
1444                 if (!vma || addr + len <= vma->vm_start) {
1445                         /*
1446                          * Remember the place where we stopped the search:
1447                          */
1448                         mm->free_area_cache = addr + len;
1449                         return addr;
1450                 }
1451                 if (addr + mm->cached_hole_size < vma->vm_start)
1452                         mm->cached_hole_size = vma->vm_start - addr;
1453                 addr = vma->vm_end;
1454         }
1455 }
1456 #endif  
1457
1458 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1459 {
1460         /*
1461          * Is this a new hole at the lowest possible address?
1462          */
1463         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1464                 mm->free_area_cache = addr;
1465                 mm->cached_hole_size = ~0UL;
1466         }
1467 }
1468
1469 /*
1470  * This mmap-allocator allocates new areas top-down from below the
1471  * stack's low limit (the base):
1472  */
1473 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1474 unsigned long
1475 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1476                           const unsigned long len, const unsigned long pgoff,
1477                           const unsigned long flags)
1478 {
1479         struct vm_area_struct *vma;
1480         struct mm_struct *mm = current->mm;
1481         unsigned long addr = addr0;
1482
1483         /* requested length too big for entire address space */
1484         if (len > TASK_SIZE)
1485                 return -ENOMEM;
1486
1487         if (flags & MAP_FIXED)
1488                 return addr;
1489
1490         /* requesting a specific address */
1491         if (addr) {
1492                 addr = PAGE_ALIGN(addr);
1493                 vma = find_vma(mm, addr);
1494                 if (TASK_SIZE - len >= addr &&
1495                                 (!vma || addr + len <= vma->vm_start))
1496                         return addr;
1497         }
1498
1499         /* check if free_area_cache is useful for us */
1500         if (len <= mm->cached_hole_size) {
1501                 mm->cached_hole_size = 0;
1502                 mm->free_area_cache = mm->mmap_base;
1503         }
1504
1505         /* either no address requested or can't fit in requested address hole */
1506         addr = mm->free_area_cache;
1507
1508         /* make sure it can fit in the remaining address space */
1509         if (addr > len) {
1510                 vma = find_vma(mm, addr-len);
1511                 if (!vma || addr <= vma->vm_start)
1512                         /* remember the address as a hint for next time */
1513                         return (mm->free_area_cache = addr-len);
1514         }
1515
1516         if (mm->mmap_base < len)
1517                 goto bottomup;
1518
1519         addr = mm->mmap_base-len;
1520
1521         do {
1522                 /*
1523                  * Lookup failure means no vma is above this address,
1524                  * else if new region fits below vma->vm_start,
1525                  * return with success:
1526                  */
1527                 vma = find_vma(mm, addr);
1528                 if (!vma || addr+len <= vma->vm_start)
1529                         /* remember the address as a hint for next time */
1530                         return (mm->free_area_cache = addr);
1531
1532                 /* remember the largest hole we saw so far */
1533                 if (addr + mm->cached_hole_size < vma->vm_start)
1534                         mm->cached_hole_size = vma->vm_start - addr;
1535
1536                 /* try just below the current vma->vm_start */
1537                 addr = vma->vm_start-len;
1538         } while (len < vma->vm_start);
1539
1540 bottomup:
1541         /*
1542          * A failed mmap() very likely causes application failure,
1543          * so fall back to the bottom-up function here. This scenario
1544          * can happen with large stack limits and large mmap()
1545          * allocations.
1546          */
1547         mm->cached_hole_size = ~0UL;
1548         mm->free_area_cache = TASK_UNMAPPED_BASE;
1549         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1550         /*
1551          * Restore the topdown base:
1552          */
1553         mm->free_area_cache = mm->mmap_base;
1554         mm->cached_hole_size = ~0UL;
1555
1556         return addr;
1557 }
1558 #endif
1559
1560 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1561 {
1562         /*
1563          * Is this a new hole at the highest possible address?
1564          */
1565         if (addr > mm->free_area_cache)
1566                 mm->free_area_cache = addr;
1567
1568         /* dont allow allocations above current base */
1569         if (mm->free_area_cache > mm->mmap_base)
1570                 mm->free_area_cache = mm->mmap_base;
1571 }
1572
1573 unsigned long
1574 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1575                 unsigned long pgoff, unsigned long flags)
1576 {
1577         unsigned long (*get_area)(struct file *, unsigned long,
1578                                   unsigned long, unsigned long, unsigned long);
1579
1580         unsigned long error = arch_mmap_check(addr, len, flags);
1581         if (error)
1582                 return error;
1583
1584         /* Careful about overflows.. */
1585         if (len > TASK_SIZE)
1586                 return -ENOMEM;
1587
1588         get_area = current->mm->get_unmapped_area;
1589         if (file && file->f_op && file->f_op->get_unmapped_area)
1590                 get_area = file->f_op->get_unmapped_area;
1591         addr = get_area(file, addr, len, pgoff, flags);
1592         if (IS_ERR_VALUE(addr))
1593                 return addr;
1594
1595         if (addr > TASK_SIZE - len)
1596                 return -ENOMEM;
1597         if (addr & ~PAGE_MASK)
1598                 return -EINVAL;
1599
1600         return arch_rebalance_pgtables(addr, len);
1601 }
1602
1603 EXPORT_SYMBOL(get_unmapped_area);
1604
1605 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1606 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1607 {
1608         struct vm_area_struct *vma = NULL;
1609
1610         if (mm) {
1611                 /* Check the cache first. */
1612                 /* (Cache hit rate is typically around 35%.) */
1613                 vma = mm->mmap_cache;
1614                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1615                         struct rb_node * rb_node;
1616
1617                         rb_node = mm->mm_rb.rb_node;
1618                         vma = NULL;
1619
1620                         while (rb_node) {
1621                                 struct vm_area_struct * vma_tmp;
1622
1623                                 vma_tmp = rb_entry(rb_node,
1624                                                 struct vm_area_struct, vm_rb);
1625
1626                                 if (vma_tmp->vm_end > addr) {
1627                                         vma = vma_tmp;
1628                                         if (vma_tmp->vm_start <= addr)
1629                                                 break;
1630                                         rb_node = rb_node->rb_left;
1631                                 } else
1632                                         rb_node = rb_node->rb_right;
1633                         }
1634                         if (vma)
1635                                 mm->mmap_cache = vma;
1636                 }
1637         }
1638         return vma;
1639 }
1640
1641 EXPORT_SYMBOL(find_vma);
1642
1643 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1644 struct vm_area_struct *
1645 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1646                         struct vm_area_struct **pprev)
1647 {
1648         struct vm_area_struct *vma = NULL, *prev = NULL;
1649         struct rb_node *rb_node;
1650         if (!mm)
1651                 goto out;
1652
1653         /* Guard against addr being lower than the first VMA */
1654         vma = mm->mmap;
1655
1656         /* Go through the RB tree quickly. */
1657         rb_node = mm->mm_rb.rb_node;
1658
1659         while (rb_node) {
1660                 struct vm_area_struct *vma_tmp;
1661                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1662
1663                 if (addr < vma_tmp->vm_end) {
1664                         rb_node = rb_node->rb_left;
1665                 } else {
1666                         prev = vma_tmp;
1667                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1668                                 break;
1669                         rb_node = rb_node->rb_right;
1670                 }
1671         }
1672
1673 out:
1674         *pprev = prev;
1675         return prev ? prev->vm_next : vma;
1676 }
1677
1678 /*
1679  * Verify that the stack growth is acceptable and
1680  * update accounting. This is shared with both the
1681  * grow-up and grow-down cases.
1682  */
1683 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1684 {
1685         struct mm_struct *mm = vma->vm_mm;
1686         struct rlimit *rlim = current->signal->rlim;
1687         unsigned long new_start;
1688
1689         /* address space limit tests */
1690         if (!may_expand_vm(mm, grow))
1691                 return -ENOMEM;
1692
1693         /* Stack limit test */
1694         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1695                 return -ENOMEM;
1696
1697         /* mlock limit tests */
1698         if (vma->vm_flags & VM_LOCKED) {
1699                 unsigned long locked;
1700                 unsigned long limit;
1701                 locked = mm->locked_vm + grow;
1702                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1703                 limit >>= PAGE_SHIFT;
1704                 if (locked > limit && !capable(CAP_IPC_LOCK))
1705                         return -ENOMEM;
1706         }
1707
1708         /* Check to ensure the stack will not grow into a hugetlb-only region */
1709         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1710                         vma->vm_end - size;
1711         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1712                 return -EFAULT;
1713
1714         /*
1715          * Overcommit..  This must be the final test, as it will
1716          * update security statistics.
1717          */
1718         if (security_vm_enough_memory_mm(mm, grow))
1719                 return -ENOMEM;
1720
1721         /* Ok, everything looks good - let it rip */
1722         mm->total_vm += grow;
1723         if (vma->vm_flags & VM_LOCKED)
1724                 mm->locked_vm += grow;
1725         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1726         return 0;
1727 }
1728
1729 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1730 /*
1731  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1732  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1733  */
1734 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1735 {
1736         int error;
1737
1738         if (!(vma->vm_flags & VM_GROWSUP))
1739                 return -EFAULT;
1740
1741         /*
1742          * We must make sure the anon_vma is allocated
1743          * so that the anon_vma locking is not a noop.
1744          */
1745         if (unlikely(anon_vma_prepare(vma)))
1746                 return -ENOMEM;
1747         vma_lock_anon_vma(vma);
1748
1749         /*
1750          * vma->vm_start/vm_end cannot change under us because the caller
1751          * is required to hold the mmap_sem in read mode.  We need the
1752          * anon_vma lock to serialize against concurrent expand_stacks.
1753          * Also guard against wrapping around to address 0.
1754          */
1755         if (address < PAGE_ALIGN(address+4))
1756                 address = PAGE_ALIGN(address+4);
1757         else {
1758                 vma_unlock_anon_vma(vma);
1759                 return -ENOMEM;
1760         }
1761         error = 0;
1762
1763         /* Somebody else might have raced and expanded it already */
1764         if (address > vma->vm_end) {
1765                 unsigned long size, grow;
1766
1767                 size = address - vma->vm_start;
1768                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1769
1770                 error = acct_stack_growth(vma, size, grow);
1771                 if (!error) {
1772                         vma->vm_end = address;
1773                         perf_event_mmap(vma);
1774                 }
1775         }
1776         vma_unlock_anon_vma(vma);
1777         khugepaged_enter_vma_merge(vma);
1778         return error;
1779 }
1780 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1781
1782 /*
1783  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1784  */
1785 static int expand_downwards(struct vm_area_struct *vma,
1786                                    unsigned long address)
1787 {
1788         int error;
1789
1790         /*
1791          * We must make sure the anon_vma is allocated
1792          * so that the anon_vma locking is not a noop.
1793          */
1794         if (unlikely(anon_vma_prepare(vma)))
1795                 return -ENOMEM;
1796
1797         address &= PAGE_MASK;
1798         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1799         if (error)
1800                 return error;
1801
1802         vma_lock_anon_vma(vma);
1803
1804         /*
1805          * vma->vm_start/vm_end cannot change under us because the caller
1806          * is required to hold the mmap_sem in read mode.  We need the
1807          * anon_vma lock to serialize against concurrent expand_stacks.
1808          */
1809
1810         /* Somebody else might have raced and expanded it already */
1811         if (address < vma->vm_start) {
1812                 unsigned long size, grow;
1813
1814                 size = vma->vm_end - address;
1815                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1816
1817                 error = acct_stack_growth(vma, size, grow);
1818                 if (!error) {
1819                         vma->vm_start = address;
1820                         vma->vm_pgoff -= grow;
1821                         perf_event_mmap(vma);
1822                 }
1823         }
1824         vma_unlock_anon_vma(vma);
1825         khugepaged_enter_vma_merge(vma);
1826         return error;
1827 }
1828
1829 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1830 {
1831         return expand_downwards(vma, address);
1832 }
1833
1834 #ifdef CONFIG_STACK_GROWSUP
1835 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1836 {
1837         return expand_upwards(vma, address);
1838 }
1839
1840 struct vm_area_struct *
1841 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1842 {
1843         struct vm_area_struct *vma, *prev;
1844
1845         addr &= PAGE_MASK;
1846         vma = find_vma_prev(mm, addr, &prev);
1847         if (vma && (vma->vm_start <= addr))
1848                 return vma;
1849         if (!prev || expand_stack(prev, addr))
1850                 return NULL;
1851         if (prev->vm_flags & VM_LOCKED) {
1852                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1853         }
1854         return prev;
1855 }
1856 #else
1857 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1858 {
1859         return expand_downwards(vma, address);
1860 }
1861
1862 struct vm_area_struct *
1863 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1864 {
1865         struct vm_area_struct * vma;
1866         unsigned long start;
1867
1868         addr &= PAGE_MASK;
1869         vma = find_vma(mm,addr);
1870         if (!vma)
1871                 return NULL;
1872         if (vma->vm_start <= addr)
1873                 return vma;
1874         if (!(vma->vm_flags & VM_GROWSDOWN))
1875                 return NULL;
1876         start = vma->vm_start;
1877         if (expand_stack(vma, addr))
1878                 return NULL;
1879         if (vma->vm_flags & VM_LOCKED) {
1880                 mlock_vma_pages_range(vma, addr, start);
1881         }
1882         return vma;
1883 }
1884 #endif
1885
1886 /*
1887  * Ok - we have the memory areas we should free on the vma list,
1888  * so release them, and do the vma updates.
1889  *
1890  * Called with the mm semaphore held.
1891  */
1892 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1893 {
1894         /* Update high watermark before we lower total_vm */
1895         update_hiwater_vm(mm);
1896         do {
1897                 long nrpages = vma_pages(vma);
1898
1899                 mm->total_vm -= nrpages;
1900                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1901                 vma = remove_vma(vma);
1902         } while (vma);
1903         validate_mm(mm);
1904 }
1905
1906 /*
1907  * Get rid of page table information in the indicated region.
1908  *
1909  * Called with the mm semaphore held.
1910  */
1911 static void unmap_region(struct mm_struct *mm,
1912                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1913                 unsigned long start, unsigned long end)
1914 {
1915         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1916         struct mmu_gather *tlb;
1917         unsigned long nr_accounted = 0;
1918
1919         lru_add_drain();
1920         tlb = tlb_gather_mmu(mm, 0);
1921         update_hiwater_rss(mm);
1922         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1923         vm_unacct_memory(nr_accounted);
1924         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1925                                  next? next->vm_start: 0);
1926         tlb_finish_mmu(tlb, start, end);
1927 }
1928
1929 /*
1930  * Create a list of vma's touched by the unmap, removing them from the mm's
1931  * vma list as we go..
1932  */
1933 static void
1934 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1935         struct vm_area_struct *prev, unsigned long end)
1936 {
1937         struct vm_area_struct **insertion_point;
1938         struct vm_area_struct *tail_vma = NULL;
1939         unsigned long addr;
1940
1941         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1942         vma->vm_prev = NULL;
1943         do {
1944                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1945                 mm->map_count--;
1946                 tail_vma = vma;
1947                 vma = vma->vm_next;
1948         } while (vma && vma->vm_start < end);
1949         *insertion_point = vma;
1950         if (vma)
1951                 vma->vm_prev = prev;
1952         tail_vma->vm_next = NULL;
1953         if (mm->unmap_area == arch_unmap_area)
1954                 addr = prev ? prev->vm_end : mm->mmap_base;
1955         else
1956                 addr = vma ?  vma->vm_start : mm->mmap_base;
1957         mm->unmap_area(mm, addr);
1958         mm->mmap_cache = NULL;          /* Kill the cache. */
1959 }
1960
1961 /*
1962  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1963  * munmap path where it doesn't make sense to fail.
1964  */
1965 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1966               unsigned long addr, int new_below)
1967 {
1968         struct mempolicy *pol;
1969         struct vm_area_struct *new;
1970         int err = -ENOMEM;
1971
1972         if (is_vm_hugetlb_page(vma) && (addr &
1973                                         ~(huge_page_mask(hstate_vma(vma)))))
1974                 return -EINVAL;
1975
1976         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1977         if (!new)
1978                 goto out_err;
1979
1980         /* most fields are the same, copy all, and then fixup */
1981         *new = *vma;
1982
1983         INIT_LIST_HEAD(&new->anon_vma_chain);
1984
1985         if (new_below)
1986                 new->vm_end = addr;
1987         else {
1988                 new->vm_start = addr;
1989                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1990         }
1991
1992         pol = mpol_dup(vma_policy(vma));
1993         if (IS_ERR(pol)) {
1994                 err = PTR_ERR(pol);
1995                 goto out_free_vma;
1996         }
1997         vma_set_policy(new, pol);
1998
1999         if (anon_vma_clone(new, vma))
2000                 goto out_free_mpol;
2001
2002         if (new->vm_file) {
2003                 get_file(new->vm_file);
2004                 if (vma->vm_flags & VM_EXECUTABLE)
2005                         added_exe_file_vma(mm);
2006         }
2007
2008         if (new->vm_ops && new->vm_ops->open)
2009                 new->vm_ops->open(new);
2010
2011         if (new_below)
2012                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2013                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2014         else
2015                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2016
2017         /* Success. */
2018         if (!err)
2019                 return 0;
2020
2021         /* Clean everything up if vma_adjust failed. */
2022         if (new->vm_ops && new->vm_ops->close)
2023                 new->vm_ops->close(new);
2024         if (new->vm_file) {
2025                 if (vma->vm_flags & VM_EXECUTABLE)
2026                         removed_exe_file_vma(mm);
2027                 fput(new->vm_file);
2028         }
2029         unlink_anon_vmas(new);
2030  out_free_mpol:
2031         mpol_put(pol);
2032  out_free_vma:
2033         kmem_cache_free(vm_area_cachep, new);
2034  out_err:
2035         return err;
2036 }
2037
2038 /*
2039  * Split a vma into two pieces at address 'addr', a new vma is allocated
2040  * either for the first part or the tail.
2041  */
2042 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2043               unsigned long addr, int new_below)
2044 {
2045         if (mm->map_count >= sysctl_max_map_count)
2046                 return -ENOMEM;
2047
2048         return __split_vma(mm, vma, addr, new_below);
2049 }
2050
2051 /* Munmap is split into 2 main parts -- this part which finds
2052  * what needs doing, and the areas themselves, which do the
2053  * work.  This now handles partial unmappings.
2054  * Jeremy Fitzhardinge <jeremy@goop.org>
2055  */
2056 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2057 {
2058         unsigned long end;
2059         struct vm_area_struct *vma, *prev, *last;
2060
2061         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2062                 return -EINVAL;
2063
2064         if ((len = PAGE_ALIGN(len)) == 0)
2065                 return -EINVAL;
2066
2067         /* Find the first overlapping VMA */
2068         vma = find_vma_prev(mm, start, &prev);
2069         if (!vma)
2070                 return 0;
2071         /* we have  start < vma->vm_end  */
2072
2073         /* if it doesn't overlap, we have nothing.. */
2074         end = start + len;
2075         if (vma->vm_start >= end)
2076                 return 0;
2077
2078         /*
2079          * If we need to split any vma, do it now to save pain later.
2080          *
2081          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2082          * unmapped vm_area_struct will remain in use: so lower split_vma
2083          * places tmp vma above, and higher split_vma places tmp vma below.
2084          */
2085         if (start > vma->vm_start) {
2086                 int error;
2087
2088                 /*
2089                  * Make sure that map_count on return from munmap() will
2090                  * not exceed its limit; but let map_count go just above
2091                  * its limit temporarily, to help free resources as expected.
2092                  */
2093                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2094                         return -ENOMEM;
2095
2096                 error = __split_vma(mm, vma, start, 0);
2097                 if (error)
2098                         return error;
2099                 prev = vma;
2100         }
2101
2102         /* Does it split the last one? */
2103         last = find_vma(mm, end);
2104         if (last && end > last->vm_start) {
2105                 int error = __split_vma(mm, last, end, 1);
2106                 if (error)
2107                         return error;
2108         }
2109         vma = prev? prev->vm_next: mm->mmap;
2110
2111         /*
2112          * unlock any mlock()ed ranges before detaching vmas
2113          */
2114         if (mm->locked_vm) {
2115                 struct vm_area_struct *tmp = vma;
2116                 while (tmp && tmp->vm_start < end) {
2117                         if (tmp->vm_flags & VM_LOCKED) {
2118                                 mm->locked_vm -= vma_pages(tmp);
2119                                 munlock_vma_pages_all(tmp);
2120                         }
2121                         tmp = tmp->vm_next;
2122                 }
2123         }
2124
2125         /*
2126          * Remove the vma's, and unmap the actual pages
2127          */
2128         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2129         unmap_region(mm, vma, prev, start, end);
2130
2131         /* Fix up all other VM information */
2132         remove_vma_list(mm, vma);
2133
2134         return 0;
2135 }
2136
2137 EXPORT_SYMBOL(do_munmap);
2138
2139 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2140 {
2141         int ret;
2142         struct mm_struct *mm = current->mm;
2143
2144         profile_munmap(addr);
2145
2146         down_write(&mm->mmap_sem);
2147         ret = do_munmap(mm, addr, len);
2148         up_write(&mm->mmap_sem);
2149         return ret;
2150 }
2151
2152 static inline void verify_mm_writelocked(struct mm_struct *mm)
2153 {
2154 #ifdef CONFIG_DEBUG_VM
2155         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2156                 WARN_ON(1);
2157                 up_read(&mm->mmap_sem);
2158         }
2159 #endif
2160 }
2161
2162 /*
2163  *  this is really a simplified "do_mmap".  it only handles
2164  *  anonymous maps.  eventually we may be able to do some
2165  *  brk-specific accounting here.
2166  */
2167 unsigned long do_brk(unsigned long addr, unsigned long len)
2168 {
2169         struct mm_struct * mm = current->mm;
2170         struct vm_area_struct * vma, * prev;
2171         unsigned long flags;
2172         struct rb_node ** rb_link, * rb_parent;
2173         pgoff_t pgoff = addr >> PAGE_SHIFT;
2174         int error;
2175
2176         len = PAGE_ALIGN(len);
2177         if (!len)
2178                 return addr;
2179
2180         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2181         if (error)
2182                 return error;
2183
2184         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2185
2186         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2187         if (error & ~PAGE_MASK)
2188                 return error;
2189
2190         /*
2191          * mlock MCL_FUTURE?
2192          */
2193         if (mm->def_flags & VM_LOCKED) {
2194                 unsigned long locked, lock_limit;
2195                 locked = len >> PAGE_SHIFT;
2196                 locked += mm->locked_vm;
2197                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2198                 lock_limit >>= PAGE_SHIFT;
2199                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2200                         return -EAGAIN;
2201         }
2202
2203         /*
2204          * mm->mmap_sem is required to protect against another thread
2205          * changing the mappings in case we sleep.
2206          */
2207         verify_mm_writelocked(mm);
2208
2209         /*
2210          * Clear old maps.  this also does some error checking for us
2211          */
2212  munmap_back:
2213         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2214         if (vma && vma->vm_start < addr + len) {
2215                 if (do_munmap(mm, addr, len))
2216                         return -ENOMEM;
2217                 goto munmap_back;
2218         }
2219
2220         /* Check against address space limits *after* clearing old maps... */
2221         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2222                 return -ENOMEM;
2223
2224         if (mm->map_count > sysctl_max_map_count)
2225                 return -ENOMEM;
2226
2227         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2228                 return -ENOMEM;
2229
2230         /* Can we just expand an old private anonymous mapping? */
2231         vma = vma_merge(mm, prev, addr, addr + len, flags,
2232                                         NULL, NULL, pgoff, NULL);
2233         if (vma)
2234                 goto out;
2235
2236         /*
2237          * create a vma struct for an anonymous mapping
2238          */
2239         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2240         if (!vma) {
2241                 vm_unacct_memory(len >> PAGE_SHIFT);
2242                 return -ENOMEM;
2243         }
2244
2245         INIT_LIST_HEAD(&vma->anon_vma_chain);
2246         vma->vm_mm = mm;
2247         vma->vm_start = addr;
2248         vma->vm_end = addr + len;
2249         vma->vm_pgoff = pgoff;
2250         vma->vm_flags = flags;
2251         vma->vm_page_prot = vm_get_page_prot(flags);
2252         vma_link(mm, vma, prev, rb_link, rb_parent);
2253 out:
2254         perf_event_mmap(vma);
2255         mm->total_vm += len >> PAGE_SHIFT;
2256         if (flags & VM_LOCKED) {
2257                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2258                         mm->locked_vm += (len >> PAGE_SHIFT);
2259         }
2260         return addr;
2261 }
2262
2263 EXPORT_SYMBOL(do_brk);
2264
2265 /* Release all mmaps. */
2266 void exit_mmap(struct mm_struct *mm)
2267 {
2268         struct mmu_gather *tlb;
2269         struct vm_area_struct *vma;
2270         unsigned long nr_accounted = 0;
2271         unsigned long end;
2272
2273         /* mm's last user has gone, and its about to be pulled down */
2274         mmu_notifier_release(mm);
2275
2276         if (mm->locked_vm) {
2277                 vma = mm->mmap;
2278                 while (vma) {
2279                         if (vma->vm_flags & VM_LOCKED)
2280                                 munlock_vma_pages_all(vma);
2281                         vma = vma->vm_next;
2282                 }
2283         }
2284
2285         arch_exit_mmap(mm);
2286
2287         vma = mm->mmap;
2288         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2289                 return;
2290
2291         lru_add_drain();
2292         flush_cache_mm(mm);
2293         tlb = tlb_gather_mmu(mm, 1);
2294         /* update_hiwater_rss(mm) here? but nobody should be looking */
2295         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2296         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2297         vm_unacct_memory(nr_accounted);
2298
2299         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2300         tlb_finish_mmu(tlb, 0, end);
2301
2302         /*
2303          * Walk the list again, actually closing and freeing it,
2304          * with preemption enabled, without holding any MM locks.
2305          */
2306         while (vma)
2307                 vma = remove_vma(vma);
2308
2309         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2310 }
2311
2312 /* Insert vm structure into process list sorted by address
2313  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2314  * then i_mmap_lock is taken here.
2315  */
2316 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2317 {
2318         struct vm_area_struct * __vma, * prev;
2319         struct rb_node ** rb_link, * rb_parent;
2320
2321         /*
2322          * The vm_pgoff of a purely anonymous vma should be irrelevant
2323          * until its first write fault, when page's anon_vma and index
2324          * are set.  But now set the vm_pgoff it will almost certainly
2325          * end up with (unless mremap moves it elsewhere before that
2326          * first wfault), so /proc/pid/maps tells a consistent story.
2327          *
2328          * By setting it to reflect the virtual start address of the
2329          * vma, merges and splits can happen in a seamless way, just
2330          * using the existing file pgoff checks and manipulations.
2331          * Similarly in do_mmap_pgoff and in do_brk.
2332          */
2333         if (!vma->vm_file) {
2334                 BUG_ON(vma->anon_vma);
2335                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2336         }
2337         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2338         if (__vma && __vma->vm_start < vma->vm_end)
2339                 return -ENOMEM;
2340         if ((vma->vm_flags & VM_ACCOUNT) &&
2341              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2342                 return -ENOMEM;
2343         vma_link(mm, vma, prev, rb_link, rb_parent);
2344         return 0;
2345 }
2346
2347 /*
2348  * Copy the vma structure to a new location in the same mm,
2349  * prior to moving page table entries, to effect an mremap move.
2350  */
2351 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2352         unsigned long addr, unsigned long len, pgoff_t pgoff)
2353 {
2354         struct vm_area_struct *vma = *vmap;
2355         unsigned long vma_start = vma->vm_start;
2356         struct mm_struct *mm = vma->vm_mm;
2357         struct vm_area_struct *new_vma, *prev;
2358         struct rb_node **rb_link, *rb_parent;
2359         struct mempolicy *pol;
2360
2361         /*
2362          * If anonymous vma has not yet been faulted, update new pgoff
2363          * to match new location, to increase its chance of merging.
2364          */
2365         if (!vma->vm_file && !vma->anon_vma)
2366                 pgoff = addr >> PAGE_SHIFT;
2367
2368         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2369         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2370                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2371         if (new_vma) {
2372                 /*
2373                  * Source vma may have been merged into new_vma
2374                  */
2375                 if (vma_start >= new_vma->vm_start &&
2376                     vma_start < new_vma->vm_end)
2377                         *vmap = new_vma;
2378         } else {
2379                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2380                 if (new_vma) {
2381                         *new_vma = *vma;
2382                         pol = mpol_dup(vma_policy(vma));
2383                         if (IS_ERR(pol))
2384                                 goto out_free_vma;
2385                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2386                         if (anon_vma_clone(new_vma, vma))
2387                                 goto out_free_mempol;
2388                         vma_set_policy(new_vma, pol);
2389                         new_vma->vm_start = addr;
2390                         new_vma->vm_end = addr + len;
2391                         new_vma->vm_pgoff = pgoff;
2392                         if (new_vma->vm_file) {
2393                                 get_file(new_vma->vm_file);
2394                                 if (vma->vm_flags & VM_EXECUTABLE)
2395                                         added_exe_file_vma(mm);
2396                         }
2397                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2398                                 new_vma->vm_ops->open(new_vma);
2399                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2400                 }
2401         }
2402         return new_vma;
2403
2404  out_free_mempol:
2405         mpol_put(pol);
2406  out_free_vma:
2407         kmem_cache_free(vm_area_cachep, new_vma);
2408         return NULL;
2409 }
2410
2411 /*
2412  * Return true if the calling process may expand its vm space by the passed
2413  * number of pages
2414  */
2415 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2416 {
2417         unsigned long cur = mm->total_vm;       /* pages */
2418         unsigned long lim;
2419
2420         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2421
2422         if (cur + npages > lim)
2423                 return 0;
2424         return 1;
2425 }
2426
2427
2428 static int special_mapping_fault(struct vm_area_struct *vma,
2429                                 struct vm_fault *vmf)
2430 {
2431         pgoff_t pgoff;
2432         struct page **pages;
2433
2434         /*
2435          * special mappings have no vm_file, and in that case, the mm
2436          * uses vm_pgoff internally. So we have to subtract it from here.
2437          * We are allowed to do this because we are the mm; do not copy
2438          * this code into drivers!
2439          */
2440         pgoff = vmf->pgoff - vma->vm_pgoff;
2441
2442         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2443                 pgoff--;
2444
2445         if (*pages) {
2446                 struct page *page = *pages;
2447                 get_page(page);
2448                 vmf->page = page;
2449                 return 0;
2450         }
2451
2452         return VM_FAULT_SIGBUS;
2453 }
2454
2455 /*
2456  * Having a close hook prevents vma merging regardless of flags.
2457  */
2458 static void special_mapping_close(struct vm_area_struct *vma)
2459 {
2460 }
2461
2462 static const struct vm_operations_struct special_mapping_vmops = {
2463         .close = special_mapping_close,
2464         .fault = special_mapping_fault,
2465 };
2466
2467 /*
2468  * Called with mm->mmap_sem held for writing.
2469  * Insert a new vma covering the given region, with the given flags.
2470  * Its pages are supplied by the given array of struct page *.
2471  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2472  * The region past the last page supplied will always produce SIGBUS.
2473  * The array pointer and the pages it points to are assumed to stay alive
2474  * for as long as this mapping might exist.
2475  */
2476 int install_special_mapping(struct mm_struct *mm,
2477                             unsigned long addr, unsigned long len,
2478                             unsigned long vm_flags, struct page **pages)
2479 {
2480         int ret;
2481         struct vm_area_struct *vma;
2482
2483         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2484         if (unlikely(vma == NULL))
2485                 return -ENOMEM;
2486
2487         INIT_LIST_HEAD(&vma->anon_vma_chain);
2488         vma->vm_mm = mm;
2489         vma->vm_start = addr;
2490         vma->vm_end = addr + len;
2491
2492         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2493         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2494
2495         vma->vm_ops = &special_mapping_vmops;
2496         vma->vm_private_data = pages;
2497
2498         ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2499         if (ret)
2500                 goto out;
2501
2502         ret = insert_vm_struct(mm, vma);
2503         if (ret)
2504                 goto out;
2505
2506         mm->total_vm += len >> PAGE_SHIFT;
2507
2508         perf_event_mmap(vma);
2509
2510         return 0;
2511
2512 out:
2513         kmem_cache_free(vm_area_cachep, vma);
2514         return ret;
2515 }
2516
2517 static DEFINE_MUTEX(mm_all_locks_mutex);
2518
2519 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2520 {
2521         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2522                 /*
2523                  * The LSB of head.next can't change from under us
2524                  * because we hold the mm_all_locks_mutex.
2525                  */
2526                 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem);
2527                 /*
2528                  * We can safely modify head.next after taking the
2529                  * anon_vma->root->lock. If some other vma in this mm shares
2530                  * the same anon_vma we won't take it again.
2531                  *
2532                  * No need of atomic instructions here, head.next
2533                  * can't change from under us thanks to the
2534                  * anon_vma->root->lock.
2535                  */
2536                 if (__test_and_set_bit(0, (unsigned long *)
2537                                        &anon_vma->root->head.next))
2538                         BUG();
2539         }
2540 }
2541
2542 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2543 {
2544         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2545                 /*
2546                  * AS_MM_ALL_LOCKS can't change from under us because
2547                  * we hold the mm_all_locks_mutex.
2548                  *
2549                  * Operations on ->flags have to be atomic because
2550                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2551                  * mm_all_locks_mutex, there may be other cpus
2552                  * changing other bitflags in parallel to us.
2553                  */
2554                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2555                         BUG();
2556                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2557         }
2558 }
2559
2560 /*
2561  * This operation locks against the VM for all pte/vma/mm related
2562  * operations that could ever happen on a certain mm. This includes
2563  * vmtruncate, try_to_unmap, and all page faults.
2564  *
2565  * The caller must take the mmap_sem in write mode before calling
2566  * mm_take_all_locks(). The caller isn't allowed to release the
2567  * mmap_sem until mm_drop_all_locks() returns.
2568  *
2569  * mmap_sem in write mode is required in order to block all operations
2570  * that could modify pagetables and free pages without need of
2571  * altering the vma layout (for example populate_range() with
2572  * nonlinear vmas). It's also needed in write mode to avoid new
2573  * anon_vmas to be associated with existing vmas.
2574  *
2575  * A single task can't take more than one mm_take_all_locks() in a row
2576  * or it would deadlock.
2577  *
2578  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2579  * mapping->flags avoid to take the same lock twice, if more than one
2580  * vma in this mm is backed by the same anon_vma or address_space.
2581  *
2582  * We can take all the locks in random order because the VM code
2583  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2584  * takes more than one of them in a row. Secondly we're protected
2585  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2586  *
2587  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2588  * that may have to take thousand of locks.
2589  *
2590  * mm_take_all_locks() can fail if it's interrupted by signals.
2591  */
2592 int mm_take_all_locks(struct mm_struct *mm)
2593 {
2594         struct vm_area_struct *vma;
2595         struct anon_vma_chain *avc;
2596         int ret = -EINTR;
2597
2598         BUG_ON(down_read_trylock(&mm->mmap_sem));
2599
2600         mutex_lock(&mm_all_locks_mutex);
2601
2602         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2603                 if (signal_pending(current))
2604                         goto out_unlock;
2605                 if (vma->vm_file && vma->vm_file->f_mapping)
2606                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2607         }
2608
2609         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2610                 if (signal_pending(current))
2611                         goto out_unlock;
2612                 if (vma->anon_vma)
2613                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2614                                 vm_lock_anon_vma(mm, avc->anon_vma);
2615         }
2616
2617         ret = 0;
2618
2619 out_unlock:
2620         if (ret)
2621                 mm_drop_all_locks(mm);
2622
2623         return ret;
2624 }
2625
2626 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2627 {
2628         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2629                 /*
2630                  * The LSB of head.next can't change to 0 from under
2631                  * us because we hold the mm_all_locks_mutex.
2632                  *
2633                  * We must however clear the bitflag before unlocking
2634                  * the vma so the users using the anon_vma->head will
2635                  * never see our bitflag.
2636                  *
2637                  * No need of atomic instructions here, head.next
2638                  * can't change from under us until we release the
2639                  * anon_vma->root->lock.
2640                  */
2641                 if (!__test_and_clear_bit(0, (unsigned long *)
2642                                           &anon_vma->root->head.next))
2643                         BUG();
2644                 anon_vma_unlock(anon_vma);
2645         }
2646 }
2647
2648 static void vm_unlock_mapping(struct address_space *mapping)
2649 {
2650         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2651                 /*
2652                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2653                  * because we hold the mm_all_locks_mutex.
2654                  */
2655                 spin_unlock(&mapping->i_mmap_lock);
2656                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2657                                         &mapping->flags))
2658                         BUG();
2659         }
2660 }
2661
2662 /*
2663  * The mmap_sem cannot be released by the caller until
2664  * mm_drop_all_locks() returns.
2665  */
2666 void mm_drop_all_locks(struct mm_struct *mm)
2667 {
2668         struct vm_area_struct *vma;
2669         struct anon_vma_chain *avc;
2670
2671         BUG_ON(down_read_trylock(&mm->mmap_sem));
2672         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2673
2674         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2675                 if (vma->anon_vma)
2676                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2677                                 vm_unlock_anon_vma(avc->anon_vma);
2678                 if (vma->vm_file && vma->vm_file->f_mapping)
2679                         vm_unlock_mapping(vma->vm_file->f_mapping);
2680         }
2681
2682         mutex_unlock(&mm_all_locks_mutex);
2683 }
2684
2685 /*
2686  * initialise the VMA slab
2687  */
2688 void __init mmap_init(void)
2689 {
2690         int ret;
2691
2692         ret = percpu_counter_init(&vm_committed_as, 0);
2693         VM_BUG_ON(ret);
2694 }