753f44d17047d80ff1b1cb64e9d53d09fd054cb4
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33
34 #include <asm/uaccess.h>
35 #include <asm/cacheflush.h>
36 #include <asm/tlb.h>
37 #include <asm/mmu_context.h>
38
39 #include "internal.h"
40
41 #ifndef arch_mmap_check
42 #define arch_mmap_check(addr, len, flags)       (0)
43 #endif
44
45 #ifndef arch_rebalance_pgtables
46 #define arch_rebalance_pgtables(addr, len)              (addr)
47 #endif
48
49 static void unmap_region(struct mm_struct *mm,
50                 struct vm_area_struct *vma, struct vm_area_struct *prev,
51                 unsigned long start, unsigned long end);
52
53 /*
54  * WARNING: the debugging will use recursive algorithms so never enable this
55  * unless you know what you are doing.
56  */
57 #undef DEBUG_MM_RB
58
59 /* description of effects of mapping type and prot in current implementation.
60  * this is due to the limited x86 page protection hardware.  The expected
61  * behavior is in parens:
62  *
63  * map_type     prot
64  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
65  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *              
69  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
70  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
71  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
72  *
73  */
74 pgprot_t protection_map[16] = {
75         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77 };
78
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
80 {
81         return __pgprot(pgprot_val(protection_map[vm_flags &
82                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
83                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
84 }
85 EXPORT_SYMBOL(vm_get_page_prot);
86
87 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
88 int sysctl_overcommit_ratio = 50;       /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90 struct percpu_counter vm_committed_as;
91
92 /*
93  * Check that a process has enough memory to allocate a new virtual
94  * mapping. 0 means there is enough memory for the allocation to
95  * succeed and -ENOMEM implies there is not.
96  *
97  * We currently support three overcommit policies, which are set via the
98  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
99  *
100  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
101  * Additional code 2002 Jul 20 by Robert Love.
102  *
103  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
104  *
105  * Note this is a helper function intended to be used by LSMs which
106  * wish to use this logic.
107  */
108 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
109 {
110         unsigned long free, allowed;
111
112         vm_acct_memory(pages);
113
114         /*
115          * Sometimes we want to use more memory than we have
116          */
117         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
118                 return 0;
119
120         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
121                 unsigned long n;
122
123                 free = global_page_state(NR_FILE_PAGES);
124                 free += nr_swap_pages;
125
126                 /*
127                  * Any slabs which are created with the
128                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
129                  * which are reclaimable, under pressure.  The dentry
130                  * cache and most inode caches should fall into this
131                  */
132                 free += global_page_state(NR_SLAB_RECLAIMABLE);
133
134                 /*
135                  * Leave the last 3% for root
136                  */
137                 if (!cap_sys_admin)
138                         free -= free / 32;
139
140                 if (free > pages)
141                         return 0;
142
143                 /*
144                  * nr_free_pages() is very expensive on large systems,
145                  * only call if we're about to fail.
146                  */
147                 n = nr_free_pages();
148
149                 /*
150                  * Leave reserved pages. The pages are not for anonymous pages.
151                  */
152                 if (n <= totalreserve_pages)
153                         goto error;
154                 else
155                         n -= totalreserve_pages;
156
157                 /*
158                  * Leave the last 3% for root
159                  */
160                 if (!cap_sys_admin)
161                         n -= n / 32;
162                 free += n;
163
164                 if (free > pages)
165                         return 0;
166
167                 goto error;
168         }
169
170         allowed = (totalram_pages - hugetlb_total_pages())
171                 * sysctl_overcommit_ratio / 100;
172         /*
173          * Leave the last 3% for root
174          */
175         if (!cap_sys_admin)
176                 allowed -= allowed / 32;
177         allowed += total_swap_pages;
178
179         /* Don't let a single process grow too big:
180            leave 3% of the size of this process for other processes */
181         if (mm)
182                 allowed -= mm->total_vm / 32;
183
184         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
185                 return 0;
186 error:
187         vm_unacct_memory(pages);
188
189         return -ENOMEM;
190 }
191
192 /*
193  * Requires inode->i_mapping->i_mmap_lock
194  */
195 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
196                 struct file *file, struct address_space *mapping)
197 {
198         if (vma->vm_flags & VM_DENYWRITE)
199                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
200         if (vma->vm_flags & VM_SHARED)
201                 mapping->i_mmap_writable--;
202
203         flush_dcache_mmap_lock(mapping);
204         if (unlikely(vma->vm_flags & VM_NONLINEAR))
205                 list_del_init(&vma->shared.vm_set.list);
206         else
207                 vma_prio_tree_remove(vma, &mapping->i_mmap);
208         flush_dcache_mmap_unlock(mapping);
209 }
210
211 /*
212  * Unlink a file-based vm structure from its prio_tree, to hide
213  * vma from rmap and vmtruncate before freeing its page tables.
214  */
215 void unlink_file_vma(struct vm_area_struct *vma)
216 {
217         struct file *file = vma->vm_file;
218
219         if (file) {
220                 struct address_space *mapping = file->f_mapping;
221                 spin_lock(&mapping->i_mmap_lock);
222                 __remove_shared_vm_struct(vma, file, mapping);
223                 spin_unlock(&mapping->i_mmap_lock);
224         }
225 }
226
227 /*
228  * Close a vm structure and free it, returning the next.
229  */
230 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
231 {
232         struct vm_area_struct *next = vma->vm_next;
233
234         might_sleep();
235         if (vma->vm_ops && vma->vm_ops->close)
236                 vma->vm_ops->close(vma);
237         if (vma->vm_file) {
238                 fput(vma->vm_file);
239                 if (vma->vm_flags & VM_EXECUTABLE)
240                         removed_exe_file_vma(vma->vm_mm);
241         }
242         mpol_put(vma_policy(vma));
243         kmem_cache_free(vm_area_cachep, vma);
244         return next;
245 }
246
247 SYSCALL_DEFINE1(brk, unsigned long, brk)
248 {
249         unsigned long rlim, retval;
250         unsigned long newbrk, oldbrk;
251         struct mm_struct *mm = current->mm;
252         unsigned long min_brk;
253
254         down_write(&mm->mmap_sem);
255
256 #ifdef CONFIG_COMPAT_BRK
257         min_brk = mm->end_code;
258 #else
259         min_brk = mm->start_brk;
260 #endif
261         if (brk < min_brk)
262                 goto out;
263
264         /*
265          * Check against rlimit here. If this check is done later after the test
266          * of oldbrk with newbrk then it can escape the test and let the data
267          * segment grow beyond its set limit the in case where the limit is
268          * not page aligned -Ram Gupta
269          */
270         rlim = rlimit(RLIMIT_DATA);
271         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
272                         (mm->end_data - mm->start_data) > rlim)
273                 goto out;
274
275         newbrk = PAGE_ALIGN(brk);
276         oldbrk = PAGE_ALIGN(mm->brk);
277         if (oldbrk == newbrk)
278                 goto set_brk;
279
280         /* Always allow shrinking brk. */
281         if (brk <= mm->brk) {
282                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
283                         goto set_brk;
284                 goto out;
285         }
286
287         /* Check against existing mmap mappings. */
288         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
289                 goto out;
290
291         /* Ok, looks good - let it rip. */
292         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
293                 goto out;
294 set_brk:
295         mm->brk = brk;
296 out:
297         retval = mm->brk;
298         up_write(&mm->mmap_sem);
299         return retval;
300 }
301
302 #ifdef DEBUG_MM_RB
303 static int browse_rb(struct rb_root *root)
304 {
305         int i = 0, j;
306         struct rb_node *nd, *pn = NULL;
307         unsigned long prev = 0, pend = 0;
308
309         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
310                 struct vm_area_struct *vma;
311                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
312                 if (vma->vm_start < prev)
313                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
314                 if (vma->vm_start < pend)
315                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
316                 if (vma->vm_start > vma->vm_end)
317                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
318                 i++;
319                 pn = nd;
320                 prev = vma->vm_start;
321                 pend = vma->vm_end;
322         }
323         j = 0;
324         for (nd = pn; nd; nd = rb_prev(nd)) {
325                 j++;
326         }
327         if (i != j)
328                 printk("backwards %d, forwards %d\n", j, i), i = 0;
329         return i;
330 }
331
332 void validate_mm(struct mm_struct *mm)
333 {
334         int bug = 0;
335         int i = 0;
336         struct vm_area_struct *tmp = mm->mmap;
337         while (tmp) {
338                 tmp = tmp->vm_next;
339                 i++;
340         }
341         if (i != mm->map_count)
342                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
343         i = browse_rb(&mm->mm_rb);
344         if (i != mm->map_count)
345                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
346         BUG_ON(bug);
347 }
348 #else
349 #define validate_mm(mm) do { } while (0)
350 #endif
351
352 static struct vm_area_struct *
353 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
354                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
355                 struct rb_node ** rb_parent)
356 {
357         struct vm_area_struct * vma;
358         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
359
360         __rb_link = &mm->mm_rb.rb_node;
361         rb_prev = __rb_parent = NULL;
362         vma = NULL;
363
364         while (*__rb_link) {
365                 struct vm_area_struct *vma_tmp;
366
367                 __rb_parent = *__rb_link;
368                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
369
370                 if (vma_tmp->vm_end > addr) {
371                         vma = vma_tmp;
372                         if (vma_tmp->vm_start <= addr)
373                                 break;
374                         __rb_link = &__rb_parent->rb_left;
375                 } else {
376                         rb_prev = __rb_parent;
377                         __rb_link = &__rb_parent->rb_right;
378                 }
379         }
380
381         *pprev = NULL;
382         if (rb_prev)
383                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
384         *rb_link = __rb_link;
385         *rb_parent = __rb_parent;
386         return vma;
387 }
388
389 static inline void
390 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
391                 struct vm_area_struct *prev, struct rb_node *rb_parent)
392 {
393         struct vm_area_struct *next;
394
395         vma->vm_prev = prev;
396         if (prev) {
397                 next = prev->vm_next;
398                 prev->vm_next = vma;
399         } else {
400                 mm->mmap = vma;
401                 if (rb_parent)
402                         next = rb_entry(rb_parent,
403                                         struct vm_area_struct, vm_rb);
404                 else
405                         next = NULL;
406         }
407         vma->vm_next = next;
408         if (next)
409                 next->vm_prev = vma;
410 }
411
412 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
413                 struct rb_node **rb_link, struct rb_node *rb_parent)
414 {
415         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
416         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
417 }
418
419 static void __vma_link_file(struct vm_area_struct *vma)
420 {
421         struct file *file;
422
423         file = vma->vm_file;
424         if (file) {
425                 struct address_space *mapping = file->f_mapping;
426
427                 if (vma->vm_flags & VM_DENYWRITE)
428                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
429                 if (vma->vm_flags & VM_SHARED)
430                         mapping->i_mmap_writable++;
431
432                 flush_dcache_mmap_lock(mapping);
433                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
434                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
435                 else
436                         vma_prio_tree_insert(vma, &mapping->i_mmap);
437                 flush_dcache_mmap_unlock(mapping);
438         }
439 }
440
441 static void
442 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
443         struct vm_area_struct *prev, struct rb_node **rb_link,
444         struct rb_node *rb_parent)
445 {
446         __vma_link_list(mm, vma, prev, rb_parent);
447         __vma_link_rb(mm, vma, rb_link, rb_parent);
448 }
449
450 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
451                         struct vm_area_struct *prev, struct rb_node **rb_link,
452                         struct rb_node *rb_parent)
453 {
454         struct address_space *mapping = NULL;
455
456         if (vma->vm_file)
457                 mapping = vma->vm_file->f_mapping;
458
459         if (mapping) {
460                 spin_lock(&mapping->i_mmap_lock);
461                 vma->vm_truncate_count = mapping->truncate_count;
462         }
463
464         __vma_link(mm, vma, prev, rb_link, rb_parent);
465         __vma_link_file(vma);
466
467         if (mapping)
468                 spin_unlock(&mapping->i_mmap_lock);
469
470         mm->map_count++;
471         validate_mm(mm);
472 }
473
474 /*
475  * Helper for vma_adjust in the split_vma insert case:
476  * insert vm structure into list and rbtree and anon_vma,
477  * but it has already been inserted into prio_tree earlier.
478  */
479 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
480 {
481         struct vm_area_struct *__vma, *prev;
482         struct rb_node **rb_link, *rb_parent;
483
484         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
485         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
486         __vma_link(mm, vma, prev, rb_link, rb_parent);
487         mm->map_count++;
488 }
489
490 static inline void
491 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
492                 struct vm_area_struct *prev)
493 {
494         struct vm_area_struct *next = vma->vm_next;
495
496         prev->vm_next = next;
497         if (next)
498                 next->vm_prev = prev;
499         rb_erase(&vma->vm_rb, &mm->mm_rb);
500         if (mm->mmap_cache == vma)
501                 mm->mmap_cache = prev;
502 }
503
504 /*
505  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
506  * is already present in an i_mmap tree without adjusting the tree.
507  * The following helper function should be used when such adjustments
508  * are necessary.  The "insert" vma (if any) is to be inserted
509  * before we drop the necessary locks.
510  */
511 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
512         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
513 {
514         struct mm_struct *mm = vma->vm_mm;
515         struct vm_area_struct *next = vma->vm_next;
516         struct vm_area_struct *importer = NULL;
517         struct address_space *mapping = NULL;
518         struct prio_tree_root *root = NULL;
519         struct anon_vma *anon_vma = NULL;
520         struct file *file = vma->vm_file;
521         long adjust_next = 0;
522         int remove_next = 0;
523
524         if (next && !insert) {
525                 struct vm_area_struct *exporter = NULL;
526
527                 if (end >= next->vm_end) {
528                         /*
529                          * vma expands, overlapping all the next, and
530                          * perhaps the one after too (mprotect case 6).
531                          */
532 again:                  remove_next = 1 + (end > next->vm_end);
533                         end = next->vm_end;
534                         exporter = next;
535                         importer = vma;
536                 } else if (end > next->vm_start) {
537                         /*
538                          * vma expands, overlapping part of the next:
539                          * mprotect case 5 shifting the boundary up.
540                          */
541                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
542                         exporter = next;
543                         importer = vma;
544                 } else if (end < vma->vm_end) {
545                         /*
546                          * vma shrinks, and !insert tells it's not
547                          * split_vma inserting another: so it must be
548                          * mprotect case 4 shifting the boundary down.
549                          */
550                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
551                         exporter = vma;
552                         importer = next;
553                 }
554
555                 /*
556                  * Easily overlooked: when mprotect shifts the boundary,
557                  * make sure the expanding vma has anon_vma set if the
558                  * shrinking vma had, to cover any anon pages imported.
559                  */
560                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
561                         if (anon_vma_clone(importer, exporter))
562                                 return -ENOMEM;
563                         importer->anon_vma = exporter->anon_vma;
564                 }
565         }
566
567         if (file) {
568                 mapping = file->f_mapping;
569                 if (!(vma->vm_flags & VM_NONLINEAR))
570                         root = &mapping->i_mmap;
571                 spin_lock(&mapping->i_mmap_lock);
572                 if (importer &&
573                     vma->vm_truncate_count != next->vm_truncate_count) {
574                         /*
575                          * unmap_mapping_range might be in progress:
576                          * ensure that the expanding vma is rescanned.
577                          */
578                         importer->vm_truncate_count = 0;
579                 }
580                 if (insert) {
581                         insert->vm_truncate_count = vma->vm_truncate_count;
582                         /*
583                          * Put into prio_tree now, so instantiated pages
584                          * are visible to arm/parisc __flush_dcache_page
585                          * throughout; but we cannot insert into address
586                          * space until vma start or end is updated.
587                          */
588                         __vma_link_file(insert);
589                 }
590         }
591
592         /*
593          * When changing only vma->vm_end, we don't really need anon_vma
594          * lock. This is a fairly rare case by itself, but the anon_vma
595          * lock may be shared between many sibling processes.  Skipping
596          * the lock for brk adjustments makes a difference sometimes.
597          */
598         if (vma->anon_vma && (insert || importer || start != vma->vm_start)) {
599                 anon_vma = vma->anon_vma;
600                 anon_vma_lock(anon_vma);
601         }
602
603         if (root) {
604                 flush_dcache_mmap_lock(mapping);
605                 vma_prio_tree_remove(vma, root);
606                 if (adjust_next)
607                         vma_prio_tree_remove(next, root);
608         }
609
610         vma->vm_start = start;
611         vma->vm_end = end;
612         vma->vm_pgoff = pgoff;
613         if (adjust_next) {
614                 next->vm_start += adjust_next << PAGE_SHIFT;
615                 next->vm_pgoff += adjust_next;
616         }
617
618         if (root) {
619                 if (adjust_next)
620                         vma_prio_tree_insert(next, root);
621                 vma_prio_tree_insert(vma, root);
622                 flush_dcache_mmap_unlock(mapping);
623         }
624
625         if (remove_next) {
626                 /*
627                  * vma_merge has merged next into vma, and needs
628                  * us to remove next before dropping the locks.
629                  */
630                 __vma_unlink(mm, next, vma);
631                 if (file)
632                         __remove_shared_vm_struct(next, file, mapping);
633         } else if (insert) {
634                 /*
635                  * split_vma has split insert from vma, and needs
636                  * us to insert it before dropping the locks
637                  * (it may either follow vma or precede it).
638                  */
639                 __insert_vm_struct(mm, insert);
640         }
641
642         if (anon_vma)
643                 anon_vma_unlock(anon_vma);
644         if (mapping)
645                 spin_unlock(&mapping->i_mmap_lock);
646
647         if (remove_next) {
648                 if (file) {
649                         fput(file);
650                         if (next->vm_flags & VM_EXECUTABLE)
651                                 removed_exe_file_vma(mm);
652                 }
653                 if (next->anon_vma)
654                         anon_vma_merge(vma, next);
655                 mm->map_count--;
656                 mpol_put(vma_policy(next));
657                 kmem_cache_free(vm_area_cachep, next);
658                 /*
659                  * In mprotect's case 6 (see comments on vma_merge),
660                  * we must remove another next too. It would clutter
661                  * up the code too much to do both in one go.
662                  */
663                 if (remove_next == 2) {
664                         next = vma->vm_next;
665                         goto again;
666                 }
667         }
668
669         validate_mm(mm);
670
671         return 0;
672 }
673
674 /*
675  * If the vma has a ->close operation then the driver probably needs to release
676  * per-vma resources, so we don't attempt to merge those.
677  */
678 static inline int is_mergeable_vma(struct vm_area_struct *vma,
679                         struct file *file, unsigned long vm_flags)
680 {
681         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
682         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
683                 return 0;
684         if (vma->vm_file != file)
685                 return 0;
686         if (vma->vm_ops && vma->vm_ops->close)
687                 return 0;
688         return 1;
689 }
690
691 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
692                                         struct anon_vma *anon_vma2)
693 {
694         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
695 }
696
697 /*
698  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
699  * in front of (at a lower virtual address and file offset than) the vma.
700  *
701  * We cannot merge two vmas if they have differently assigned (non-NULL)
702  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
703  *
704  * We don't check here for the merged mmap wrapping around the end of pagecache
705  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
706  * wrap, nor mmaps which cover the final page at index -1UL.
707  */
708 static int
709 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
710         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
711 {
712         if (is_mergeable_vma(vma, file, vm_flags) &&
713             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
714                 if (vma->vm_pgoff == vm_pgoff)
715                         return 1;
716         }
717         return 0;
718 }
719
720 /*
721  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
722  * beyond (at a higher virtual address and file offset than) the vma.
723  *
724  * We cannot merge two vmas if they have differently assigned (non-NULL)
725  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
726  */
727 static int
728 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
729         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
730 {
731         if (is_mergeable_vma(vma, file, vm_flags) &&
732             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
733                 pgoff_t vm_pglen;
734                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
735                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
736                         return 1;
737         }
738         return 0;
739 }
740
741 /*
742  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
743  * whether that can be merged with its predecessor or its successor.
744  * Or both (it neatly fills a hole).
745  *
746  * In most cases - when called for mmap, brk or mremap - [addr,end) is
747  * certain not to be mapped by the time vma_merge is called; but when
748  * called for mprotect, it is certain to be already mapped (either at
749  * an offset within prev, or at the start of next), and the flags of
750  * this area are about to be changed to vm_flags - and the no-change
751  * case has already been eliminated.
752  *
753  * The following mprotect cases have to be considered, where AAAA is
754  * the area passed down from mprotect_fixup, never extending beyond one
755  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
756  *
757  *     AAAA             AAAA                AAAA          AAAA
758  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
759  *    cannot merge    might become    might become    might become
760  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
761  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
762  *    mremap move:                                    PPPPNNNNNNNN 8
763  *        AAAA
764  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
765  *    might become    case 1 below    case 2 below    case 3 below
766  *
767  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
768  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
769  */
770 struct vm_area_struct *vma_merge(struct mm_struct *mm,
771                         struct vm_area_struct *prev, unsigned long addr,
772                         unsigned long end, unsigned long vm_flags,
773                         struct anon_vma *anon_vma, struct file *file,
774                         pgoff_t pgoff, struct mempolicy *policy)
775 {
776         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
777         struct vm_area_struct *area, *next;
778         int err;
779
780         /*
781          * We later require that vma->vm_flags == vm_flags,
782          * so this tests vma->vm_flags & VM_SPECIAL, too.
783          */
784         if (vm_flags & VM_SPECIAL)
785                 return NULL;
786
787         if (prev)
788                 next = prev->vm_next;
789         else
790                 next = mm->mmap;
791         area = next;
792         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
793                 next = next->vm_next;
794
795         /*
796          * Can it merge with the predecessor?
797          */
798         if (prev && prev->vm_end == addr &&
799                         mpol_equal(vma_policy(prev), policy) &&
800                         can_vma_merge_after(prev, vm_flags,
801                                                 anon_vma, file, pgoff)) {
802                 /*
803                  * OK, it can.  Can we now merge in the successor as well?
804                  */
805                 if (next && end == next->vm_start &&
806                                 mpol_equal(policy, vma_policy(next)) &&
807                                 can_vma_merge_before(next, vm_flags,
808                                         anon_vma, file, pgoff+pglen) &&
809                                 is_mergeable_anon_vma(prev->anon_vma,
810                                                       next->anon_vma)) {
811                                                         /* cases 1, 6 */
812                         err = vma_adjust(prev, prev->vm_start,
813                                 next->vm_end, prev->vm_pgoff, NULL);
814                 } else                                  /* cases 2, 5, 7 */
815                         err = vma_adjust(prev, prev->vm_start,
816                                 end, prev->vm_pgoff, NULL);
817                 if (err)
818                         return NULL;
819                 khugepaged_enter_vma_merge(prev);
820                 return prev;
821         }
822
823         /*
824          * Can this new request be merged in front of next?
825          */
826         if (next && end == next->vm_start &&
827                         mpol_equal(policy, vma_policy(next)) &&
828                         can_vma_merge_before(next, vm_flags,
829                                         anon_vma, file, pgoff+pglen)) {
830                 if (prev && addr < prev->vm_end)        /* case 4 */
831                         err = vma_adjust(prev, prev->vm_start,
832                                 addr, prev->vm_pgoff, NULL);
833                 else                                    /* cases 3, 8 */
834                         err = vma_adjust(area, addr, next->vm_end,
835                                 next->vm_pgoff - pglen, NULL);
836                 if (err)
837                         return NULL;
838                 khugepaged_enter_vma_merge(area);
839                 return area;
840         }
841
842         return NULL;
843 }
844
845 /*
846  * Rough compatbility check to quickly see if it's even worth looking
847  * at sharing an anon_vma.
848  *
849  * They need to have the same vm_file, and the flags can only differ
850  * in things that mprotect may change.
851  *
852  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
853  * we can merge the two vma's. For example, we refuse to merge a vma if
854  * there is a vm_ops->close() function, because that indicates that the
855  * driver is doing some kind of reference counting. But that doesn't
856  * really matter for the anon_vma sharing case.
857  */
858 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
859 {
860         return a->vm_end == b->vm_start &&
861                 mpol_equal(vma_policy(a), vma_policy(b)) &&
862                 a->vm_file == b->vm_file &&
863                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
864                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
865 }
866
867 /*
868  * Do some basic sanity checking to see if we can re-use the anon_vma
869  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
870  * the same as 'old', the other will be the new one that is trying
871  * to share the anon_vma.
872  *
873  * NOTE! This runs with mm_sem held for reading, so it is possible that
874  * the anon_vma of 'old' is concurrently in the process of being set up
875  * by another page fault trying to merge _that_. But that's ok: if it
876  * is being set up, that automatically means that it will be a singleton
877  * acceptable for merging, so we can do all of this optimistically. But
878  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
879  *
880  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
881  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
882  * is to return an anon_vma that is "complex" due to having gone through
883  * a fork).
884  *
885  * We also make sure that the two vma's are compatible (adjacent,
886  * and with the same memory policies). That's all stable, even with just
887  * a read lock on the mm_sem.
888  */
889 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
890 {
891         if (anon_vma_compatible(a, b)) {
892                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
893
894                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
895                         return anon_vma;
896         }
897         return NULL;
898 }
899
900 /*
901  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
902  * neighbouring vmas for a suitable anon_vma, before it goes off
903  * to allocate a new anon_vma.  It checks because a repetitive
904  * sequence of mprotects and faults may otherwise lead to distinct
905  * anon_vmas being allocated, preventing vma merge in subsequent
906  * mprotect.
907  */
908 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
909 {
910         struct anon_vma *anon_vma;
911         struct vm_area_struct *near;
912
913         near = vma->vm_next;
914         if (!near)
915                 goto try_prev;
916
917         anon_vma = reusable_anon_vma(near, vma, near);
918         if (anon_vma)
919                 return anon_vma;
920 try_prev:
921         /*
922          * It is potentially slow to have to call find_vma_prev here.
923          * But it's only on the first write fault on the vma, not
924          * every time, and we could devise a way to avoid it later
925          * (e.g. stash info in next's anon_vma_node when assigning
926          * an anon_vma, or when trying vma_merge).  Another time.
927          */
928         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
929         if (!near)
930                 goto none;
931
932         anon_vma = reusable_anon_vma(near, near, vma);
933         if (anon_vma)
934                 return anon_vma;
935 none:
936         /*
937          * There's no absolute need to look only at touching neighbours:
938          * we could search further afield for "compatible" anon_vmas.
939          * But it would probably just be a waste of time searching,
940          * or lead to too many vmas hanging off the same anon_vma.
941          * We're trying to allow mprotect remerging later on,
942          * not trying to minimize memory used for anon_vmas.
943          */
944         return NULL;
945 }
946
947 #ifdef CONFIG_PROC_FS
948 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
949                                                 struct file *file, long pages)
950 {
951         const unsigned long stack_flags
952                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
953
954         if (file) {
955                 mm->shared_vm += pages;
956                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
957                         mm->exec_vm += pages;
958         } else if (flags & stack_flags)
959                 mm->stack_vm += pages;
960         if (flags & (VM_RESERVED|VM_IO))
961                 mm->reserved_vm += pages;
962 }
963 #endif /* CONFIG_PROC_FS */
964
965 /*
966  * The caller must hold down_write(&current->mm->mmap_sem).
967  */
968
969 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
970                         unsigned long len, unsigned long prot,
971                         unsigned long flags, unsigned long pgoff)
972 {
973         struct mm_struct * mm = current->mm;
974         struct inode *inode;
975         unsigned int vm_flags;
976         int error;
977         unsigned long reqprot = prot;
978
979         /*
980          * Does the application expect PROT_READ to imply PROT_EXEC?
981          *
982          * (the exception is when the underlying filesystem is noexec
983          *  mounted, in which case we dont add PROT_EXEC.)
984          */
985         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
986                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
987                         prot |= PROT_EXEC;
988
989         if (!len)
990                 return -EINVAL;
991
992         if (!(flags & MAP_FIXED))
993                 addr = round_hint_to_min(addr);
994
995         /* Careful about overflows.. */
996         len = PAGE_ALIGN(len);
997         if (!len)
998                 return -ENOMEM;
999
1000         /* offset overflow? */
1001         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1002                return -EOVERFLOW;
1003
1004         /* Too many mappings? */
1005         if (mm->map_count > sysctl_max_map_count)
1006                 return -ENOMEM;
1007
1008         /* Obtain the address to map to. we verify (or select) it and ensure
1009          * that it represents a valid section of the address space.
1010          */
1011         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1012         if (addr & ~PAGE_MASK)
1013                 return addr;
1014
1015         /* Do simple checking here so the lower-level routines won't have
1016          * to. we assume access permissions have been handled by the open
1017          * of the memory object, so we don't do any here.
1018          */
1019         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1020                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1021
1022         if (flags & MAP_LOCKED)
1023                 if (!can_do_mlock())
1024                         return -EPERM;
1025
1026         /* mlock MCL_FUTURE? */
1027         if (vm_flags & VM_LOCKED) {
1028                 unsigned long locked, lock_limit;
1029                 locked = len >> PAGE_SHIFT;
1030                 locked += mm->locked_vm;
1031                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1032                 lock_limit >>= PAGE_SHIFT;
1033                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1034                         return -EAGAIN;
1035         }
1036
1037         inode = file ? file->f_path.dentry->d_inode : NULL;
1038
1039         if (file) {
1040                 switch (flags & MAP_TYPE) {
1041                 case MAP_SHARED:
1042                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1043                                 return -EACCES;
1044
1045                         /*
1046                          * Make sure we don't allow writing to an append-only
1047                          * file..
1048                          */
1049                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1050                                 return -EACCES;
1051
1052                         /*
1053                          * Make sure there are no mandatory locks on the file.
1054                          */
1055                         if (locks_verify_locked(inode))
1056                                 return -EAGAIN;
1057
1058                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1059                         if (!(file->f_mode & FMODE_WRITE))
1060                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1061
1062                         /* fall through */
1063                 case MAP_PRIVATE:
1064                         if (!(file->f_mode & FMODE_READ))
1065                                 return -EACCES;
1066                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1067                                 if (vm_flags & VM_EXEC)
1068                                         return -EPERM;
1069                                 vm_flags &= ~VM_MAYEXEC;
1070                         }
1071
1072                         if (!file->f_op || !file->f_op->mmap)
1073                                 return -ENODEV;
1074                         break;
1075
1076                 default:
1077                         return -EINVAL;
1078                 }
1079         } else {
1080                 switch (flags & MAP_TYPE) {
1081                 case MAP_SHARED:
1082                         /*
1083                          * Ignore pgoff.
1084                          */
1085                         pgoff = 0;
1086                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1087                         break;
1088                 case MAP_PRIVATE:
1089                         /*
1090                          * Set pgoff according to addr for anon_vma.
1091                          */
1092                         pgoff = addr >> PAGE_SHIFT;
1093                         break;
1094                 default:
1095                         return -EINVAL;
1096                 }
1097         }
1098
1099         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1100         if (error)
1101                 return error;
1102
1103         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1104 }
1105 EXPORT_SYMBOL(do_mmap_pgoff);
1106
1107 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1108                 unsigned long, prot, unsigned long, flags,
1109                 unsigned long, fd, unsigned long, pgoff)
1110 {
1111         struct file *file = NULL;
1112         unsigned long retval = -EBADF;
1113
1114         if (!(flags & MAP_ANONYMOUS)) {
1115                 audit_mmap_fd(fd, flags);
1116                 if (unlikely(flags & MAP_HUGETLB))
1117                         return -EINVAL;
1118                 file = fget(fd);
1119                 if (!file)
1120                         goto out;
1121         } else if (flags & MAP_HUGETLB) {
1122                 struct user_struct *user = NULL;
1123                 /*
1124                  * VM_NORESERVE is used because the reservations will be
1125                  * taken when vm_ops->mmap() is called
1126                  * A dummy user value is used because we are not locking
1127                  * memory so no accounting is necessary
1128                  */
1129                 len = ALIGN(len, huge_page_size(&default_hstate));
1130                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1131                                                 &user, HUGETLB_ANONHUGE_INODE);
1132                 if (IS_ERR(file))
1133                         return PTR_ERR(file);
1134         }
1135
1136         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1137
1138         down_write(&current->mm->mmap_sem);
1139         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1140         up_write(&current->mm->mmap_sem);
1141
1142         if (file)
1143                 fput(file);
1144 out:
1145         return retval;
1146 }
1147
1148 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1149 struct mmap_arg_struct {
1150         unsigned long addr;
1151         unsigned long len;
1152         unsigned long prot;
1153         unsigned long flags;
1154         unsigned long fd;
1155         unsigned long offset;
1156 };
1157
1158 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1159 {
1160         struct mmap_arg_struct a;
1161
1162         if (copy_from_user(&a, arg, sizeof(a)))
1163                 return -EFAULT;
1164         if (a.offset & ~PAGE_MASK)
1165                 return -EINVAL;
1166
1167         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1168                               a.offset >> PAGE_SHIFT);
1169 }
1170 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1171
1172 /*
1173  * Some shared mappigns will want the pages marked read-only
1174  * to track write events. If so, we'll downgrade vm_page_prot
1175  * to the private version (using protection_map[] without the
1176  * VM_SHARED bit).
1177  */
1178 int vma_wants_writenotify(struct vm_area_struct *vma)
1179 {
1180         unsigned int vm_flags = vma->vm_flags;
1181
1182         /* If it was private or non-writable, the write bit is already clear */
1183         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1184                 return 0;
1185
1186         /* The backer wishes to know when pages are first written to? */
1187         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1188                 return 1;
1189
1190         /* The open routine did something to the protections already? */
1191         if (pgprot_val(vma->vm_page_prot) !=
1192             pgprot_val(vm_get_page_prot(vm_flags)))
1193                 return 0;
1194
1195         /* Specialty mapping? */
1196         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1197                 return 0;
1198
1199         /* Can the mapping track the dirty pages? */
1200         return vma->vm_file && vma->vm_file->f_mapping &&
1201                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1202 }
1203
1204 /*
1205  * We account for memory if it's a private writeable mapping,
1206  * not hugepages and VM_NORESERVE wasn't set.
1207  */
1208 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1209 {
1210         /*
1211          * hugetlb has its own accounting separate from the core VM
1212          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1213          */
1214         if (file && is_file_hugepages(file))
1215                 return 0;
1216
1217         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1218 }
1219
1220 unsigned long mmap_region(struct file *file, unsigned long addr,
1221                           unsigned long len, unsigned long flags,
1222                           unsigned int vm_flags, unsigned long pgoff)
1223 {
1224         struct mm_struct *mm = current->mm;
1225         struct vm_area_struct *vma, *prev;
1226         int correct_wcount = 0;
1227         int error;
1228         struct rb_node **rb_link, *rb_parent;
1229         unsigned long charged = 0;
1230         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1231
1232         /* Clear old maps */
1233         error = -ENOMEM;
1234 munmap_back:
1235         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1236         if (vma && vma->vm_start < addr + len) {
1237                 if (do_munmap(mm, addr, len))
1238                         return -ENOMEM;
1239                 goto munmap_back;
1240         }
1241
1242         /* Check against address space limit. */
1243         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1244                 return -ENOMEM;
1245
1246         /*
1247          * Set 'VM_NORESERVE' if we should not account for the
1248          * memory use of this mapping.
1249          */
1250         if ((flags & MAP_NORESERVE)) {
1251                 /* We honor MAP_NORESERVE if allowed to overcommit */
1252                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1253                         vm_flags |= VM_NORESERVE;
1254
1255                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1256                 if (file && is_file_hugepages(file))
1257                         vm_flags |= VM_NORESERVE;
1258         }
1259
1260         /*
1261          * Private writable mapping: check memory availability
1262          */
1263         if (accountable_mapping(file, vm_flags)) {
1264                 charged = len >> PAGE_SHIFT;
1265                 if (security_vm_enough_memory(charged))
1266                         return -ENOMEM;
1267                 vm_flags |= VM_ACCOUNT;
1268         }
1269
1270         /*
1271          * Can we just expand an old mapping?
1272          */
1273         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1274         if (vma)
1275                 goto out;
1276
1277         /*
1278          * Determine the object being mapped and call the appropriate
1279          * specific mapper. the address has already been validated, but
1280          * not unmapped, but the maps are removed from the list.
1281          */
1282         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1283         if (!vma) {
1284                 error = -ENOMEM;
1285                 goto unacct_error;
1286         }
1287
1288         vma->vm_mm = mm;
1289         vma->vm_start = addr;
1290         vma->vm_end = addr + len;
1291         vma->vm_flags = vm_flags;
1292         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1293         vma->vm_pgoff = pgoff;
1294         INIT_LIST_HEAD(&vma->anon_vma_chain);
1295
1296         if (file) {
1297                 error = -EINVAL;
1298                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1299                         goto free_vma;
1300                 if (vm_flags & VM_DENYWRITE) {
1301                         error = deny_write_access(file);
1302                         if (error)
1303                                 goto free_vma;
1304                         correct_wcount = 1;
1305                 }
1306                 vma->vm_file = file;
1307                 get_file(file);
1308                 error = file->f_op->mmap(file, vma);
1309                 if (error)
1310                         goto unmap_and_free_vma;
1311                 if (vm_flags & VM_EXECUTABLE)
1312                         added_exe_file_vma(mm);
1313
1314                 /* Can addr have changed??
1315                  *
1316                  * Answer: Yes, several device drivers can do it in their
1317                  *         f_op->mmap method. -DaveM
1318                  */
1319                 addr = vma->vm_start;
1320                 pgoff = vma->vm_pgoff;
1321                 vm_flags = vma->vm_flags;
1322         } else if (vm_flags & VM_SHARED) {
1323                 error = shmem_zero_setup(vma);
1324                 if (error)
1325                         goto free_vma;
1326         }
1327
1328         if (vma_wants_writenotify(vma)) {
1329                 pgprot_t pprot = vma->vm_page_prot;
1330
1331                 /* Can vma->vm_page_prot have changed??
1332                  *
1333                  * Answer: Yes, drivers may have changed it in their
1334                  *         f_op->mmap method.
1335                  *
1336                  * Ensures that vmas marked as uncached stay that way.
1337                  */
1338                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1339                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1340                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1341         }
1342
1343         vma_link(mm, vma, prev, rb_link, rb_parent);
1344         file = vma->vm_file;
1345
1346         /* Once vma denies write, undo our temporary denial count */
1347         if (correct_wcount)
1348                 atomic_inc(&inode->i_writecount);
1349 out:
1350         perf_event_mmap(vma);
1351
1352         mm->total_vm += len >> PAGE_SHIFT;
1353         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1354         if (vm_flags & VM_LOCKED) {
1355                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1356                         mm->locked_vm += (len >> PAGE_SHIFT);
1357         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1358                 make_pages_present(addr, addr + len);
1359         return addr;
1360
1361 unmap_and_free_vma:
1362         if (correct_wcount)
1363                 atomic_inc(&inode->i_writecount);
1364         vma->vm_file = NULL;
1365         fput(file);
1366
1367         /* Undo any partial mapping done by a device driver. */
1368         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1369         charged = 0;
1370 free_vma:
1371         kmem_cache_free(vm_area_cachep, vma);
1372 unacct_error:
1373         if (charged)
1374                 vm_unacct_memory(charged);
1375         return error;
1376 }
1377
1378 /* Get an address range which is currently unmapped.
1379  * For shmat() with addr=0.
1380  *
1381  * Ugly calling convention alert:
1382  * Return value with the low bits set means error value,
1383  * ie
1384  *      if (ret & ~PAGE_MASK)
1385  *              error = ret;
1386  *
1387  * This function "knows" that -ENOMEM has the bits set.
1388  */
1389 #ifndef HAVE_ARCH_UNMAPPED_AREA
1390 unsigned long
1391 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1392                 unsigned long len, unsigned long pgoff, unsigned long flags)
1393 {
1394         struct mm_struct *mm = current->mm;
1395         struct vm_area_struct *vma;
1396         unsigned long start_addr;
1397
1398         if (len > TASK_SIZE)
1399                 return -ENOMEM;
1400
1401         if (flags & MAP_FIXED)
1402                 return addr;
1403
1404         if (addr) {
1405                 addr = PAGE_ALIGN(addr);
1406                 vma = find_vma(mm, addr);
1407                 if (TASK_SIZE - len >= addr &&
1408                     (!vma || addr + len <= vma->vm_start))
1409                         return addr;
1410         }
1411         if (len > mm->cached_hole_size) {
1412                 start_addr = addr = mm->free_area_cache;
1413         } else {
1414                 start_addr = addr = TASK_UNMAPPED_BASE;
1415                 mm->cached_hole_size = 0;
1416         }
1417
1418 full_search:
1419         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1420                 /* At this point:  (!vma || addr < vma->vm_end). */
1421                 if (TASK_SIZE - len < addr) {
1422                         /*
1423                          * Start a new search - just in case we missed
1424                          * some holes.
1425                          */
1426                         if (start_addr != TASK_UNMAPPED_BASE) {
1427                                 addr = TASK_UNMAPPED_BASE;
1428                                 start_addr = addr;
1429                                 mm->cached_hole_size = 0;
1430                                 goto full_search;
1431                         }
1432                         return -ENOMEM;
1433                 }
1434                 if (!vma || addr + len <= vma->vm_start) {
1435                         /*
1436                          * Remember the place where we stopped the search:
1437                          */
1438                         mm->free_area_cache = addr + len;
1439                         return addr;
1440                 }
1441                 if (addr + mm->cached_hole_size < vma->vm_start)
1442                         mm->cached_hole_size = vma->vm_start - addr;
1443                 addr = vma->vm_end;
1444         }
1445 }
1446 #endif  
1447
1448 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1449 {
1450         /*
1451          * Is this a new hole at the lowest possible address?
1452          */
1453         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1454                 mm->free_area_cache = addr;
1455                 mm->cached_hole_size = ~0UL;
1456         }
1457 }
1458
1459 /*
1460  * This mmap-allocator allocates new areas top-down from below the
1461  * stack's low limit (the base):
1462  */
1463 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1464 unsigned long
1465 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1466                           const unsigned long len, const unsigned long pgoff,
1467                           const unsigned long flags)
1468 {
1469         struct vm_area_struct *vma;
1470         struct mm_struct *mm = current->mm;
1471         unsigned long addr = addr0;
1472
1473         /* requested length too big for entire address space */
1474         if (len > TASK_SIZE)
1475                 return -ENOMEM;
1476
1477         if (flags & MAP_FIXED)
1478                 return addr;
1479
1480         /* requesting a specific address */
1481         if (addr) {
1482                 addr = PAGE_ALIGN(addr);
1483                 vma = find_vma(mm, addr);
1484                 if (TASK_SIZE - len >= addr &&
1485                                 (!vma || addr + len <= vma->vm_start))
1486                         return addr;
1487         }
1488
1489         /* check if free_area_cache is useful for us */
1490         if (len <= mm->cached_hole_size) {
1491                 mm->cached_hole_size = 0;
1492                 mm->free_area_cache = mm->mmap_base;
1493         }
1494
1495         /* either no address requested or can't fit in requested address hole */
1496         addr = mm->free_area_cache;
1497
1498         /* make sure it can fit in the remaining address space */
1499         if (addr > len) {
1500                 vma = find_vma(mm, addr-len);
1501                 if (!vma || addr <= vma->vm_start)
1502                         /* remember the address as a hint for next time */
1503                         return (mm->free_area_cache = addr-len);
1504         }
1505
1506         if (mm->mmap_base < len)
1507                 goto bottomup;
1508
1509         addr = mm->mmap_base-len;
1510
1511         do {
1512                 /*
1513                  * Lookup failure means no vma is above this address,
1514                  * else if new region fits below vma->vm_start,
1515                  * return with success:
1516                  */
1517                 vma = find_vma(mm, addr);
1518                 if (!vma || addr+len <= vma->vm_start)
1519                         /* remember the address as a hint for next time */
1520                         return (mm->free_area_cache = addr);
1521
1522                 /* remember the largest hole we saw so far */
1523                 if (addr + mm->cached_hole_size < vma->vm_start)
1524                         mm->cached_hole_size = vma->vm_start - addr;
1525
1526                 /* try just below the current vma->vm_start */
1527                 addr = vma->vm_start-len;
1528         } while (len < vma->vm_start);
1529
1530 bottomup:
1531         /*
1532          * A failed mmap() very likely causes application failure,
1533          * so fall back to the bottom-up function here. This scenario
1534          * can happen with large stack limits and large mmap()
1535          * allocations.
1536          */
1537         mm->cached_hole_size = ~0UL;
1538         mm->free_area_cache = TASK_UNMAPPED_BASE;
1539         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1540         /*
1541          * Restore the topdown base:
1542          */
1543         mm->free_area_cache = mm->mmap_base;
1544         mm->cached_hole_size = ~0UL;
1545
1546         return addr;
1547 }
1548 #endif
1549
1550 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1551 {
1552         /*
1553          * Is this a new hole at the highest possible address?
1554          */
1555         if (addr > mm->free_area_cache)
1556                 mm->free_area_cache = addr;
1557
1558         /* dont allow allocations above current base */
1559         if (mm->free_area_cache > mm->mmap_base)
1560                 mm->free_area_cache = mm->mmap_base;
1561 }
1562
1563 unsigned long
1564 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1565                 unsigned long pgoff, unsigned long flags)
1566 {
1567         unsigned long (*get_area)(struct file *, unsigned long,
1568                                   unsigned long, unsigned long, unsigned long);
1569
1570         unsigned long error = arch_mmap_check(addr, len, flags);
1571         if (error)
1572                 return error;
1573
1574         /* Careful about overflows.. */
1575         if (len > TASK_SIZE)
1576                 return -ENOMEM;
1577
1578         get_area = current->mm->get_unmapped_area;
1579         if (file && file->f_op && file->f_op->get_unmapped_area)
1580                 get_area = file->f_op->get_unmapped_area;
1581         addr = get_area(file, addr, len, pgoff, flags);
1582         if (IS_ERR_VALUE(addr))
1583                 return addr;
1584
1585         if (addr > TASK_SIZE - len)
1586                 return -ENOMEM;
1587         if (addr & ~PAGE_MASK)
1588                 return -EINVAL;
1589
1590         return arch_rebalance_pgtables(addr, len);
1591 }
1592
1593 EXPORT_SYMBOL(get_unmapped_area);
1594
1595 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1596 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1597 {
1598         struct vm_area_struct *vma = NULL;
1599
1600         if (mm) {
1601                 /* Check the cache first. */
1602                 /* (Cache hit rate is typically around 35%.) */
1603                 vma = mm->mmap_cache;
1604                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1605                         struct rb_node * rb_node;
1606
1607                         rb_node = mm->mm_rb.rb_node;
1608                         vma = NULL;
1609
1610                         while (rb_node) {
1611                                 struct vm_area_struct * vma_tmp;
1612
1613                                 vma_tmp = rb_entry(rb_node,
1614                                                 struct vm_area_struct, vm_rb);
1615
1616                                 if (vma_tmp->vm_end > addr) {
1617                                         vma = vma_tmp;
1618                                         if (vma_tmp->vm_start <= addr)
1619                                                 break;
1620                                         rb_node = rb_node->rb_left;
1621                                 } else
1622                                         rb_node = rb_node->rb_right;
1623                         }
1624                         if (vma)
1625                                 mm->mmap_cache = vma;
1626                 }
1627         }
1628         return vma;
1629 }
1630
1631 EXPORT_SYMBOL(find_vma);
1632
1633 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1634 struct vm_area_struct *
1635 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1636                         struct vm_area_struct **pprev)
1637 {
1638         struct vm_area_struct *vma = NULL, *prev = NULL;
1639         struct rb_node *rb_node;
1640         if (!mm)
1641                 goto out;
1642
1643         /* Guard against addr being lower than the first VMA */
1644         vma = mm->mmap;
1645
1646         /* Go through the RB tree quickly. */
1647         rb_node = mm->mm_rb.rb_node;
1648
1649         while (rb_node) {
1650                 struct vm_area_struct *vma_tmp;
1651                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1652
1653                 if (addr < vma_tmp->vm_end) {
1654                         rb_node = rb_node->rb_left;
1655                 } else {
1656                         prev = vma_tmp;
1657                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1658                                 break;
1659                         rb_node = rb_node->rb_right;
1660                 }
1661         }
1662
1663 out:
1664         *pprev = prev;
1665         return prev ? prev->vm_next : vma;
1666 }
1667
1668 /*
1669  * Verify that the stack growth is acceptable and
1670  * update accounting. This is shared with both the
1671  * grow-up and grow-down cases.
1672  */
1673 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1674 {
1675         struct mm_struct *mm = vma->vm_mm;
1676         struct rlimit *rlim = current->signal->rlim;
1677         unsigned long new_start;
1678
1679         /* address space limit tests */
1680         if (!may_expand_vm(mm, grow))
1681                 return -ENOMEM;
1682
1683         /* Stack limit test */
1684         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1685                 return -ENOMEM;
1686
1687         /* mlock limit tests */
1688         if (vma->vm_flags & VM_LOCKED) {
1689                 unsigned long locked;
1690                 unsigned long limit;
1691                 locked = mm->locked_vm + grow;
1692                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1693                 limit >>= PAGE_SHIFT;
1694                 if (locked > limit && !capable(CAP_IPC_LOCK))
1695                         return -ENOMEM;
1696         }
1697
1698         /* Check to ensure the stack will not grow into a hugetlb-only region */
1699         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1700                         vma->vm_end - size;
1701         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1702                 return -EFAULT;
1703
1704         /*
1705          * Overcommit..  This must be the final test, as it will
1706          * update security statistics.
1707          */
1708         if (security_vm_enough_memory_mm(mm, grow))
1709                 return -ENOMEM;
1710
1711         /* Ok, everything looks good - let it rip */
1712         mm->total_vm += grow;
1713         if (vma->vm_flags & VM_LOCKED)
1714                 mm->locked_vm += grow;
1715         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1716         return 0;
1717 }
1718
1719 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1720 /*
1721  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1722  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1723  */
1724 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1725 {
1726         int error;
1727
1728         if (!(vma->vm_flags & VM_GROWSUP))
1729                 return -EFAULT;
1730
1731         /*
1732          * We must make sure the anon_vma is allocated
1733          * so that the anon_vma locking is not a noop.
1734          */
1735         if (unlikely(anon_vma_prepare(vma)))
1736                 return -ENOMEM;
1737         vma_lock_anon_vma(vma);
1738
1739         /*
1740          * vma->vm_start/vm_end cannot change under us because the caller
1741          * is required to hold the mmap_sem in read mode.  We need the
1742          * anon_vma lock to serialize against concurrent expand_stacks.
1743          * Also guard against wrapping around to address 0.
1744          */
1745         if (address < PAGE_ALIGN(address+4))
1746                 address = PAGE_ALIGN(address+4);
1747         else {
1748                 vma_unlock_anon_vma(vma);
1749                 return -ENOMEM;
1750         }
1751         error = 0;
1752
1753         /* Somebody else might have raced and expanded it already */
1754         if (address > vma->vm_end) {
1755                 unsigned long size, grow;
1756
1757                 size = address - vma->vm_start;
1758                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1759
1760                 error = acct_stack_growth(vma, size, grow);
1761                 if (!error) {
1762                         vma->vm_end = address;
1763                         perf_event_mmap(vma);
1764                 }
1765         }
1766         vma_unlock_anon_vma(vma);
1767         khugepaged_enter_vma_merge(vma);
1768         return error;
1769 }
1770 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1771
1772 /*
1773  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1774  */
1775 static int expand_downwards(struct vm_area_struct *vma,
1776                                    unsigned long address)
1777 {
1778         int error;
1779
1780         /*
1781          * We must make sure the anon_vma is allocated
1782          * so that the anon_vma locking is not a noop.
1783          */
1784         if (unlikely(anon_vma_prepare(vma)))
1785                 return -ENOMEM;
1786
1787         address &= PAGE_MASK;
1788         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1789         if (error)
1790                 return error;
1791
1792         vma_lock_anon_vma(vma);
1793
1794         /*
1795          * vma->vm_start/vm_end cannot change under us because the caller
1796          * is required to hold the mmap_sem in read mode.  We need the
1797          * anon_vma lock to serialize against concurrent expand_stacks.
1798          */
1799
1800         /* Somebody else might have raced and expanded it already */
1801         if (address < vma->vm_start) {
1802                 unsigned long size, grow;
1803
1804                 size = vma->vm_end - address;
1805                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1806
1807                 error = acct_stack_growth(vma, size, grow);
1808                 if (!error) {
1809                         vma->vm_start = address;
1810                         vma->vm_pgoff -= grow;
1811                         perf_event_mmap(vma);
1812                 }
1813         }
1814         vma_unlock_anon_vma(vma);
1815         khugepaged_enter_vma_merge(vma);
1816         return error;
1817 }
1818
1819 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1820 {
1821         return expand_downwards(vma, address);
1822 }
1823
1824 #ifdef CONFIG_STACK_GROWSUP
1825 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1826 {
1827         return expand_upwards(vma, address);
1828 }
1829
1830 struct vm_area_struct *
1831 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1832 {
1833         struct vm_area_struct *vma, *prev;
1834
1835         addr &= PAGE_MASK;
1836         vma = find_vma_prev(mm, addr, &prev);
1837         if (vma && (vma->vm_start <= addr))
1838                 return vma;
1839         if (!prev || expand_stack(prev, addr))
1840                 return NULL;
1841         if (prev->vm_flags & VM_LOCKED) {
1842                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1843         }
1844         return prev;
1845 }
1846 #else
1847 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1848 {
1849         return expand_downwards(vma, address);
1850 }
1851
1852 struct vm_area_struct *
1853 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1854 {
1855         struct vm_area_struct * vma;
1856         unsigned long start;
1857
1858         addr &= PAGE_MASK;
1859         vma = find_vma(mm,addr);
1860         if (!vma)
1861                 return NULL;
1862         if (vma->vm_start <= addr)
1863                 return vma;
1864         if (!(vma->vm_flags & VM_GROWSDOWN))
1865                 return NULL;
1866         start = vma->vm_start;
1867         if (expand_stack(vma, addr))
1868                 return NULL;
1869         if (vma->vm_flags & VM_LOCKED) {
1870                 mlock_vma_pages_range(vma, addr, start);
1871         }
1872         return vma;
1873 }
1874 #endif
1875
1876 /*
1877  * Ok - we have the memory areas we should free on the vma list,
1878  * so release them, and do the vma updates.
1879  *
1880  * Called with the mm semaphore held.
1881  */
1882 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1883 {
1884         /* Update high watermark before we lower total_vm */
1885         update_hiwater_vm(mm);
1886         do {
1887                 long nrpages = vma_pages(vma);
1888
1889                 mm->total_vm -= nrpages;
1890                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1891                 vma = remove_vma(vma);
1892         } while (vma);
1893         validate_mm(mm);
1894 }
1895
1896 /*
1897  * Get rid of page table information in the indicated region.
1898  *
1899  * Called with the mm semaphore held.
1900  */
1901 static void unmap_region(struct mm_struct *mm,
1902                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1903                 unsigned long start, unsigned long end)
1904 {
1905         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1906         struct mmu_gather *tlb;
1907         unsigned long nr_accounted = 0;
1908
1909         lru_add_drain();
1910         tlb = tlb_gather_mmu(mm, 0);
1911         update_hiwater_rss(mm);
1912         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1913         vm_unacct_memory(nr_accounted);
1914         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1915                                  next? next->vm_start: 0);
1916         tlb_finish_mmu(tlb, start, end);
1917 }
1918
1919 /*
1920  * Create a list of vma's touched by the unmap, removing them from the mm's
1921  * vma list as we go..
1922  */
1923 static void
1924 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1925         struct vm_area_struct *prev, unsigned long end)
1926 {
1927         struct vm_area_struct **insertion_point;
1928         struct vm_area_struct *tail_vma = NULL;
1929         unsigned long addr;
1930
1931         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1932         vma->vm_prev = NULL;
1933         do {
1934                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1935                 mm->map_count--;
1936                 tail_vma = vma;
1937                 vma = vma->vm_next;
1938         } while (vma && vma->vm_start < end);
1939         *insertion_point = vma;
1940         if (vma)
1941                 vma->vm_prev = prev;
1942         tail_vma->vm_next = NULL;
1943         if (mm->unmap_area == arch_unmap_area)
1944                 addr = prev ? prev->vm_end : mm->mmap_base;
1945         else
1946                 addr = vma ?  vma->vm_start : mm->mmap_base;
1947         mm->unmap_area(mm, addr);
1948         mm->mmap_cache = NULL;          /* Kill the cache. */
1949 }
1950
1951 /*
1952  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1953  * munmap path where it doesn't make sense to fail.
1954  */
1955 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1956               unsigned long addr, int new_below)
1957 {
1958         struct mempolicy *pol;
1959         struct vm_area_struct *new;
1960         int err = -ENOMEM;
1961
1962         if (is_vm_hugetlb_page(vma) && (addr &
1963                                         ~(huge_page_mask(hstate_vma(vma)))))
1964                 return -EINVAL;
1965
1966         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1967         if (!new)
1968                 goto out_err;
1969
1970         /* most fields are the same, copy all, and then fixup */
1971         *new = *vma;
1972
1973         INIT_LIST_HEAD(&new->anon_vma_chain);
1974
1975         if (new_below)
1976                 new->vm_end = addr;
1977         else {
1978                 new->vm_start = addr;
1979                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1980         }
1981
1982         pol = mpol_dup(vma_policy(vma));
1983         if (IS_ERR(pol)) {
1984                 err = PTR_ERR(pol);
1985                 goto out_free_vma;
1986         }
1987         vma_set_policy(new, pol);
1988
1989         if (anon_vma_clone(new, vma))
1990                 goto out_free_mpol;
1991
1992         if (new->vm_file) {
1993                 get_file(new->vm_file);
1994                 if (vma->vm_flags & VM_EXECUTABLE)
1995                         added_exe_file_vma(mm);
1996         }
1997
1998         if (new->vm_ops && new->vm_ops->open)
1999                 new->vm_ops->open(new);
2000
2001         if (new_below)
2002                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2003                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2004         else
2005                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2006
2007         /* Success. */
2008         if (!err)
2009                 return 0;
2010
2011         /* Clean everything up if vma_adjust failed. */
2012         if (new->vm_ops && new->vm_ops->close)
2013                 new->vm_ops->close(new);
2014         if (new->vm_file) {
2015                 if (vma->vm_flags & VM_EXECUTABLE)
2016                         removed_exe_file_vma(mm);
2017                 fput(new->vm_file);
2018         }
2019         unlink_anon_vmas(new);
2020  out_free_mpol:
2021         mpol_put(pol);
2022  out_free_vma:
2023         kmem_cache_free(vm_area_cachep, new);
2024  out_err:
2025         return err;
2026 }
2027
2028 /*
2029  * Split a vma into two pieces at address 'addr', a new vma is allocated
2030  * either for the first part or the tail.
2031  */
2032 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2033               unsigned long addr, int new_below)
2034 {
2035         if (mm->map_count >= sysctl_max_map_count)
2036                 return -ENOMEM;
2037
2038         return __split_vma(mm, vma, addr, new_below);
2039 }
2040
2041 /* Munmap is split into 2 main parts -- this part which finds
2042  * what needs doing, and the areas themselves, which do the
2043  * work.  This now handles partial unmappings.
2044  * Jeremy Fitzhardinge <jeremy@goop.org>
2045  */
2046 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2047 {
2048         unsigned long end;
2049         struct vm_area_struct *vma, *prev, *last;
2050
2051         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2052                 return -EINVAL;
2053
2054         if ((len = PAGE_ALIGN(len)) == 0)
2055                 return -EINVAL;
2056
2057         /* Find the first overlapping VMA */
2058         vma = find_vma_prev(mm, start, &prev);
2059         if (!vma)
2060                 return 0;
2061         /* we have  start < vma->vm_end  */
2062
2063         /* if it doesn't overlap, we have nothing.. */
2064         end = start + len;
2065         if (vma->vm_start >= end)
2066                 return 0;
2067
2068         /*
2069          * If we need to split any vma, do it now to save pain later.
2070          *
2071          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2072          * unmapped vm_area_struct will remain in use: so lower split_vma
2073          * places tmp vma above, and higher split_vma places tmp vma below.
2074          */
2075         if (start > vma->vm_start) {
2076                 int error;
2077
2078                 /*
2079                  * Make sure that map_count on return from munmap() will
2080                  * not exceed its limit; but let map_count go just above
2081                  * its limit temporarily, to help free resources as expected.
2082                  */
2083                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2084                         return -ENOMEM;
2085
2086                 error = __split_vma(mm, vma, start, 0);
2087                 if (error)
2088                         return error;
2089                 prev = vma;
2090         }
2091
2092         /* Does it split the last one? */
2093         last = find_vma(mm, end);
2094         if (last && end > last->vm_start) {
2095                 int error = __split_vma(mm, last, end, 1);
2096                 if (error)
2097                         return error;
2098         }
2099         vma = prev? prev->vm_next: mm->mmap;
2100
2101         /*
2102          * unlock any mlock()ed ranges before detaching vmas
2103          */
2104         if (mm->locked_vm) {
2105                 struct vm_area_struct *tmp = vma;
2106                 while (tmp && tmp->vm_start < end) {
2107                         if (tmp->vm_flags & VM_LOCKED) {
2108                                 mm->locked_vm -= vma_pages(tmp);
2109                                 munlock_vma_pages_all(tmp);
2110                         }
2111                         tmp = tmp->vm_next;
2112                 }
2113         }
2114
2115         /*
2116          * Remove the vma's, and unmap the actual pages
2117          */
2118         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2119         unmap_region(mm, vma, prev, start, end);
2120
2121         /* Fix up all other VM information */
2122         remove_vma_list(mm, vma);
2123
2124         return 0;
2125 }
2126
2127 EXPORT_SYMBOL(do_munmap);
2128
2129 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2130 {
2131         int ret;
2132         struct mm_struct *mm = current->mm;
2133
2134         profile_munmap(addr);
2135
2136         down_write(&mm->mmap_sem);
2137         ret = do_munmap(mm, addr, len);
2138         up_write(&mm->mmap_sem);
2139         return ret;
2140 }
2141
2142 static inline void verify_mm_writelocked(struct mm_struct *mm)
2143 {
2144 #ifdef CONFIG_DEBUG_VM
2145         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2146                 WARN_ON(1);
2147                 up_read(&mm->mmap_sem);
2148         }
2149 #endif
2150 }
2151
2152 /*
2153  *  this is really a simplified "do_mmap".  it only handles
2154  *  anonymous maps.  eventually we may be able to do some
2155  *  brk-specific accounting here.
2156  */
2157 unsigned long do_brk(unsigned long addr, unsigned long len)
2158 {
2159         struct mm_struct * mm = current->mm;
2160         struct vm_area_struct * vma, * prev;
2161         unsigned long flags;
2162         struct rb_node ** rb_link, * rb_parent;
2163         pgoff_t pgoff = addr >> PAGE_SHIFT;
2164         int error;
2165
2166         len = PAGE_ALIGN(len);
2167         if (!len)
2168                 return addr;
2169
2170         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2171         if (error)
2172                 return error;
2173
2174         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2175
2176         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2177         if (error & ~PAGE_MASK)
2178                 return error;
2179
2180         /*
2181          * mlock MCL_FUTURE?
2182          */
2183         if (mm->def_flags & VM_LOCKED) {
2184                 unsigned long locked, lock_limit;
2185                 locked = len >> PAGE_SHIFT;
2186                 locked += mm->locked_vm;
2187                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2188                 lock_limit >>= PAGE_SHIFT;
2189                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2190                         return -EAGAIN;
2191         }
2192
2193         /*
2194          * mm->mmap_sem is required to protect against another thread
2195          * changing the mappings in case we sleep.
2196          */
2197         verify_mm_writelocked(mm);
2198
2199         /*
2200          * Clear old maps.  this also does some error checking for us
2201          */
2202  munmap_back:
2203         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2204         if (vma && vma->vm_start < addr + len) {
2205                 if (do_munmap(mm, addr, len))
2206                         return -ENOMEM;
2207                 goto munmap_back;
2208         }
2209
2210         /* Check against address space limits *after* clearing old maps... */
2211         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2212                 return -ENOMEM;
2213
2214         if (mm->map_count > sysctl_max_map_count)
2215                 return -ENOMEM;
2216
2217         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2218                 return -ENOMEM;
2219
2220         /* Can we just expand an old private anonymous mapping? */
2221         vma = vma_merge(mm, prev, addr, addr + len, flags,
2222                                         NULL, NULL, pgoff, NULL);
2223         if (vma)
2224                 goto out;
2225
2226         /*
2227          * create a vma struct for an anonymous mapping
2228          */
2229         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2230         if (!vma) {
2231                 vm_unacct_memory(len >> PAGE_SHIFT);
2232                 return -ENOMEM;
2233         }
2234
2235         INIT_LIST_HEAD(&vma->anon_vma_chain);
2236         vma->vm_mm = mm;
2237         vma->vm_start = addr;
2238         vma->vm_end = addr + len;
2239         vma->vm_pgoff = pgoff;
2240         vma->vm_flags = flags;
2241         vma->vm_page_prot = vm_get_page_prot(flags);
2242         vma_link(mm, vma, prev, rb_link, rb_parent);
2243 out:
2244         perf_event_mmap(vma);
2245         mm->total_vm += len >> PAGE_SHIFT;
2246         if (flags & VM_LOCKED) {
2247                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2248                         mm->locked_vm += (len >> PAGE_SHIFT);
2249         }
2250         return addr;
2251 }
2252
2253 EXPORT_SYMBOL(do_brk);
2254
2255 /* Release all mmaps. */
2256 void exit_mmap(struct mm_struct *mm)
2257 {
2258         struct mmu_gather *tlb;
2259         struct vm_area_struct *vma;
2260         unsigned long nr_accounted = 0;
2261         unsigned long end;
2262
2263         /* mm's last user has gone, and its about to be pulled down */
2264         mmu_notifier_release(mm);
2265
2266         if (mm->locked_vm) {
2267                 vma = mm->mmap;
2268                 while (vma) {
2269                         if (vma->vm_flags & VM_LOCKED)
2270                                 munlock_vma_pages_all(vma);
2271                         vma = vma->vm_next;
2272                 }
2273         }
2274
2275         arch_exit_mmap(mm);
2276
2277         vma = mm->mmap;
2278         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2279                 return;
2280
2281         lru_add_drain();
2282         flush_cache_mm(mm);
2283         tlb = tlb_gather_mmu(mm, 1);
2284         /* update_hiwater_rss(mm) here? but nobody should be looking */
2285         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2286         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2287         vm_unacct_memory(nr_accounted);
2288
2289         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2290         tlb_finish_mmu(tlb, 0, end);
2291
2292         /*
2293          * Walk the list again, actually closing and freeing it,
2294          * with preemption enabled, without holding any MM locks.
2295          */
2296         while (vma)
2297                 vma = remove_vma(vma);
2298
2299         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2300 }
2301
2302 /* Insert vm structure into process list sorted by address
2303  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2304  * then i_mmap_lock is taken here.
2305  */
2306 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2307 {
2308         struct vm_area_struct * __vma, * prev;
2309         struct rb_node ** rb_link, * rb_parent;
2310
2311         /*
2312          * The vm_pgoff of a purely anonymous vma should be irrelevant
2313          * until its first write fault, when page's anon_vma and index
2314          * are set.  But now set the vm_pgoff it will almost certainly
2315          * end up with (unless mremap moves it elsewhere before that
2316          * first wfault), so /proc/pid/maps tells a consistent story.
2317          *
2318          * By setting it to reflect the virtual start address of the
2319          * vma, merges and splits can happen in a seamless way, just
2320          * using the existing file pgoff checks and manipulations.
2321          * Similarly in do_mmap_pgoff and in do_brk.
2322          */
2323         if (!vma->vm_file) {
2324                 BUG_ON(vma->anon_vma);
2325                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2326         }
2327         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2328         if (__vma && __vma->vm_start < vma->vm_end)
2329                 return -ENOMEM;
2330         if ((vma->vm_flags & VM_ACCOUNT) &&
2331              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2332                 return -ENOMEM;
2333         vma_link(mm, vma, prev, rb_link, rb_parent);
2334         return 0;
2335 }
2336
2337 /*
2338  * Copy the vma structure to a new location in the same mm,
2339  * prior to moving page table entries, to effect an mremap move.
2340  */
2341 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2342         unsigned long addr, unsigned long len, pgoff_t pgoff)
2343 {
2344         struct vm_area_struct *vma = *vmap;
2345         unsigned long vma_start = vma->vm_start;
2346         struct mm_struct *mm = vma->vm_mm;
2347         struct vm_area_struct *new_vma, *prev;
2348         struct rb_node **rb_link, *rb_parent;
2349         struct mempolicy *pol;
2350
2351         /*
2352          * If anonymous vma has not yet been faulted, update new pgoff
2353          * to match new location, to increase its chance of merging.
2354          */
2355         if (!vma->vm_file && !vma->anon_vma)
2356                 pgoff = addr >> PAGE_SHIFT;
2357
2358         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2359         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2360                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2361         if (new_vma) {
2362                 /*
2363                  * Source vma may have been merged into new_vma
2364                  */
2365                 if (vma_start >= new_vma->vm_start &&
2366                     vma_start < new_vma->vm_end)
2367                         *vmap = new_vma;
2368         } else {
2369                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2370                 if (new_vma) {
2371                         *new_vma = *vma;
2372                         pol = mpol_dup(vma_policy(vma));
2373                         if (IS_ERR(pol))
2374                                 goto out_free_vma;
2375                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2376                         if (anon_vma_clone(new_vma, vma))
2377                                 goto out_free_mempol;
2378                         vma_set_policy(new_vma, pol);
2379                         new_vma->vm_start = addr;
2380                         new_vma->vm_end = addr + len;
2381                         new_vma->vm_pgoff = pgoff;
2382                         if (new_vma->vm_file) {
2383                                 get_file(new_vma->vm_file);
2384                                 if (vma->vm_flags & VM_EXECUTABLE)
2385                                         added_exe_file_vma(mm);
2386                         }
2387                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2388                                 new_vma->vm_ops->open(new_vma);
2389                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2390                 }
2391         }
2392         return new_vma;
2393
2394  out_free_mempol:
2395         mpol_put(pol);
2396  out_free_vma:
2397         kmem_cache_free(vm_area_cachep, new_vma);
2398         return NULL;
2399 }
2400
2401 /*
2402  * Return true if the calling process may expand its vm space by the passed
2403  * number of pages
2404  */
2405 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2406 {
2407         unsigned long cur = mm->total_vm;       /* pages */
2408         unsigned long lim;
2409
2410         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2411
2412         if (cur + npages > lim)
2413                 return 0;
2414         return 1;
2415 }
2416
2417
2418 static int special_mapping_fault(struct vm_area_struct *vma,
2419                                 struct vm_fault *vmf)
2420 {
2421         pgoff_t pgoff;
2422         struct page **pages;
2423
2424         /*
2425          * special mappings have no vm_file, and in that case, the mm
2426          * uses vm_pgoff internally. So we have to subtract it from here.
2427          * We are allowed to do this because we are the mm; do not copy
2428          * this code into drivers!
2429          */
2430         pgoff = vmf->pgoff - vma->vm_pgoff;
2431
2432         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2433                 pgoff--;
2434
2435         if (*pages) {
2436                 struct page *page = *pages;
2437                 get_page(page);
2438                 vmf->page = page;
2439                 return 0;
2440         }
2441
2442         return VM_FAULT_SIGBUS;
2443 }
2444
2445 /*
2446  * Having a close hook prevents vma merging regardless of flags.
2447  */
2448 static void special_mapping_close(struct vm_area_struct *vma)
2449 {
2450 }
2451
2452 static const struct vm_operations_struct special_mapping_vmops = {
2453         .close = special_mapping_close,
2454         .fault = special_mapping_fault,
2455 };
2456
2457 /*
2458  * Called with mm->mmap_sem held for writing.
2459  * Insert a new vma covering the given region, with the given flags.
2460  * Its pages are supplied by the given array of struct page *.
2461  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2462  * The region past the last page supplied will always produce SIGBUS.
2463  * The array pointer and the pages it points to are assumed to stay alive
2464  * for as long as this mapping might exist.
2465  */
2466 int install_special_mapping(struct mm_struct *mm,
2467                             unsigned long addr, unsigned long len,
2468                             unsigned long vm_flags, struct page **pages)
2469 {
2470         int ret;
2471         struct vm_area_struct *vma;
2472
2473         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2474         if (unlikely(vma == NULL))
2475                 return -ENOMEM;
2476
2477         INIT_LIST_HEAD(&vma->anon_vma_chain);
2478         vma->vm_mm = mm;
2479         vma->vm_start = addr;
2480         vma->vm_end = addr + len;
2481
2482         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2483         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2484
2485         vma->vm_ops = &special_mapping_vmops;
2486         vma->vm_private_data = pages;
2487
2488         ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2489         if (ret)
2490                 goto out;
2491
2492         ret = insert_vm_struct(mm, vma);
2493         if (ret)
2494                 goto out;
2495
2496         mm->total_vm += len >> PAGE_SHIFT;
2497
2498         perf_event_mmap(vma);
2499
2500         return 0;
2501
2502 out:
2503         kmem_cache_free(vm_area_cachep, vma);
2504         return ret;
2505 }
2506
2507 static DEFINE_MUTEX(mm_all_locks_mutex);
2508
2509 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2510 {
2511         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2512                 /*
2513                  * The LSB of head.next can't change from under us
2514                  * because we hold the mm_all_locks_mutex.
2515                  */
2516                 spin_lock_nest_lock(&anon_vma->root->lock, &mm->mmap_sem);
2517                 /*
2518                  * We can safely modify head.next after taking the
2519                  * anon_vma->root->lock. If some other vma in this mm shares
2520                  * the same anon_vma we won't take it again.
2521                  *
2522                  * No need of atomic instructions here, head.next
2523                  * can't change from under us thanks to the
2524                  * anon_vma->root->lock.
2525                  */
2526                 if (__test_and_set_bit(0, (unsigned long *)
2527                                        &anon_vma->root->head.next))
2528                         BUG();
2529         }
2530 }
2531
2532 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2533 {
2534         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2535                 /*
2536                  * AS_MM_ALL_LOCKS can't change from under us because
2537                  * we hold the mm_all_locks_mutex.
2538                  *
2539                  * Operations on ->flags have to be atomic because
2540                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2541                  * mm_all_locks_mutex, there may be other cpus
2542                  * changing other bitflags in parallel to us.
2543                  */
2544                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2545                         BUG();
2546                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2547         }
2548 }
2549
2550 /*
2551  * This operation locks against the VM for all pte/vma/mm related
2552  * operations that could ever happen on a certain mm. This includes
2553  * vmtruncate, try_to_unmap, and all page faults.
2554  *
2555  * The caller must take the mmap_sem in write mode before calling
2556  * mm_take_all_locks(). The caller isn't allowed to release the
2557  * mmap_sem until mm_drop_all_locks() returns.
2558  *
2559  * mmap_sem in write mode is required in order to block all operations
2560  * that could modify pagetables and free pages without need of
2561  * altering the vma layout (for example populate_range() with
2562  * nonlinear vmas). It's also needed in write mode to avoid new
2563  * anon_vmas to be associated with existing vmas.
2564  *
2565  * A single task can't take more than one mm_take_all_locks() in a row
2566  * or it would deadlock.
2567  *
2568  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2569  * mapping->flags avoid to take the same lock twice, if more than one
2570  * vma in this mm is backed by the same anon_vma or address_space.
2571  *
2572  * We can take all the locks in random order because the VM code
2573  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2574  * takes more than one of them in a row. Secondly we're protected
2575  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2576  *
2577  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2578  * that may have to take thousand of locks.
2579  *
2580  * mm_take_all_locks() can fail if it's interrupted by signals.
2581  */
2582 int mm_take_all_locks(struct mm_struct *mm)
2583 {
2584         struct vm_area_struct *vma;
2585         struct anon_vma_chain *avc;
2586         int ret = -EINTR;
2587
2588         BUG_ON(down_read_trylock(&mm->mmap_sem));
2589
2590         mutex_lock(&mm_all_locks_mutex);
2591
2592         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2593                 if (signal_pending(current))
2594                         goto out_unlock;
2595                 if (vma->vm_file && vma->vm_file->f_mapping)
2596                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2597         }
2598
2599         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2600                 if (signal_pending(current))
2601                         goto out_unlock;
2602                 if (vma->anon_vma)
2603                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2604                                 vm_lock_anon_vma(mm, avc->anon_vma);
2605         }
2606
2607         ret = 0;
2608
2609 out_unlock:
2610         if (ret)
2611                 mm_drop_all_locks(mm);
2612
2613         return ret;
2614 }
2615
2616 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2617 {
2618         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2619                 /*
2620                  * The LSB of head.next can't change to 0 from under
2621                  * us because we hold the mm_all_locks_mutex.
2622                  *
2623                  * We must however clear the bitflag before unlocking
2624                  * the vma so the users using the anon_vma->head will
2625                  * never see our bitflag.
2626                  *
2627                  * No need of atomic instructions here, head.next
2628                  * can't change from under us until we release the
2629                  * anon_vma->root->lock.
2630                  */
2631                 if (!__test_and_clear_bit(0, (unsigned long *)
2632                                           &anon_vma->root->head.next))
2633                         BUG();
2634                 anon_vma_unlock(anon_vma);
2635         }
2636 }
2637
2638 static void vm_unlock_mapping(struct address_space *mapping)
2639 {
2640         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2641                 /*
2642                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2643                  * because we hold the mm_all_locks_mutex.
2644                  */
2645                 spin_unlock(&mapping->i_mmap_lock);
2646                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2647                                         &mapping->flags))
2648                         BUG();
2649         }
2650 }
2651
2652 /*
2653  * The mmap_sem cannot be released by the caller until
2654  * mm_drop_all_locks() returns.
2655  */
2656 void mm_drop_all_locks(struct mm_struct *mm)
2657 {
2658         struct vm_area_struct *vma;
2659         struct anon_vma_chain *avc;
2660
2661         BUG_ON(down_read_trylock(&mm->mmap_sem));
2662         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2663
2664         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2665                 if (vma->anon_vma)
2666                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2667                                 vm_unlock_anon_vma(avc->anon_vma);
2668                 if (vma->vm_file && vma->vm_file->f_mapping)
2669                         vm_unlock_mapping(vma->vm_file->f_mapping);
2670         }
2671
2672         mutex_unlock(&mm_all_locks_mutex);
2673 }
2674
2675 /*
2676  * initialise the VMA slab
2677  */
2678 void __init mmap_init(void)
2679 {
2680         int ret;
2681
2682         ret = percpu_counter_init(&vm_committed_as, 0);
2683         VM_BUG_ON(ret);
2684 }