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