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