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