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