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