VM: add "vm_munmap()" helper function
[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 static unsigned long do_brk(unsigned long addr, unsigned long len);
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         /*
256          * CONFIG_COMPAT_BRK can still be overridden by setting
257          * randomize_va_space to 2, which will still cause mm->start_brk
258          * to be arbitrarily shifted
259          */
260         if (current->brk_randomized)
261                 min_brk = mm->start_brk;
262         else
263                 min_brk = mm->end_data;
264 #else
265         min_brk = mm->start_brk;
266 #endif
267         if (brk < min_brk)
268                 goto out;
269
270         /*
271          * Check against rlimit here. If this check is done later after the test
272          * of oldbrk with newbrk then it can escape the test and let the data
273          * segment grow beyond its set limit the in case where the limit is
274          * not page aligned -Ram Gupta
275          */
276         rlim = rlimit(RLIMIT_DATA);
277         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
278                         (mm->end_data - mm->start_data) > rlim)
279                 goto out;
280
281         newbrk = PAGE_ALIGN(brk);
282         oldbrk = PAGE_ALIGN(mm->brk);
283         if (oldbrk == newbrk)
284                 goto set_brk;
285
286         /* Always allow shrinking brk. */
287         if (brk <= mm->brk) {
288                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
289                         goto set_brk;
290                 goto out;
291         }
292
293         /* Check against existing mmap mappings. */
294         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
295                 goto out;
296
297         /* Ok, looks good - let it rip. */
298         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
299                 goto out;
300 set_brk:
301         mm->brk = brk;
302 out:
303         retval = mm->brk;
304         up_write(&mm->mmap_sem);
305         return retval;
306 }
307
308 #ifdef DEBUG_MM_RB
309 static int browse_rb(struct rb_root *root)
310 {
311         int i = 0, j;
312         struct rb_node *nd, *pn = NULL;
313         unsigned long prev = 0, pend = 0;
314
315         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
316                 struct vm_area_struct *vma;
317                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
318                 if (vma->vm_start < prev)
319                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
320                 if (vma->vm_start < pend)
321                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
322                 if (vma->vm_start > vma->vm_end)
323                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
324                 i++;
325                 pn = nd;
326                 prev = vma->vm_start;
327                 pend = vma->vm_end;
328         }
329         j = 0;
330         for (nd = pn; nd; nd = rb_prev(nd)) {
331                 j++;
332         }
333         if (i != j)
334                 printk("backwards %d, forwards %d\n", j, i), i = 0;
335         return i;
336 }
337
338 void validate_mm(struct mm_struct *mm)
339 {
340         int bug = 0;
341         int i = 0;
342         struct vm_area_struct *tmp = mm->mmap;
343         while (tmp) {
344                 tmp = tmp->vm_next;
345                 i++;
346         }
347         if (i != mm->map_count)
348                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
349         i = browse_rb(&mm->mm_rb);
350         if (i != mm->map_count)
351                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
352         BUG_ON(bug);
353 }
354 #else
355 #define validate_mm(mm) do { } while (0)
356 #endif
357
358 static struct vm_area_struct *
359 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
360                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
361                 struct rb_node ** rb_parent)
362 {
363         struct vm_area_struct * vma;
364         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
365
366         __rb_link = &mm->mm_rb.rb_node;
367         rb_prev = __rb_parent = NULL;
368         vma = NULL;
369
370         while (*__rb_link) {
371                 struct vm_area_struct *vma_tmp;
372
373                 __rb_parent = *__rb_link;
374                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
375
376                 if (vma_tmp->vm_end > addr) {
377                         vma = vma_tmp;
378                         if (vma_tmp->vm_start <= addr)
379                                 break;
380                         __rb_link = &__rb_parent->rb_left;
381                 } else {
382                         rb_prev = __rb_parent;
383                         __rb_link = &__rb_parent->rb_right;
384                 }
385         }
386
387         *pprev = NULL;
388         if (rb_prev)
389                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
390         *rb_link = __rb_link;
391         *rb_parent = __rb_parent;
392         return vma;
393 }
394
395 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
396                 struct rb_node **rb_link, struct rb_node *rb_parent)
397 {
398         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
399         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
400 }
401
402 static void __vma_link_file(struct vm_area_struct *vma)
403 {
404         struct file *file;
405
406         file = vma->vm_file;
407         if (file) {
408                 struct address_space *mapping = file->f_mapping;
409
410                 if (vma->vm_flags & VM_DENYWRITE)
411                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
412                 if (vma->vm_flags & VM_SHARED)
413                         mapping->i_mmap_writable++;
414
415                 flush_dcache_mmap_lock(mapping);
416                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
417                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
418                 else
419                         vma_prio_tree_insert(vma, &mapping->i_mmap);
420                 flush_dcache_mmap_unlock(mapping);
421         }
422 }
423
424 static void
425 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
426         struct vm_area_struct *prev, struct rb_node **rb_link,
427         struct rb_node *rb_parent)
428 {
429         __vma_link_list(mm, vma, prev, rb_parent);
430         __vma_link_rb(mm, vma, rb_link, rb_parent);
431 }
432
433 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
434                         struct vm_area_struct *prev, struct rb_node **rb_link,
435                         struct rb_node *rb_parent)
436 {
437         struct address_space *mapping = NULL;
438
439         if (vma->vm_file)
440                 mapping = vma->vm_file->f_mapping;
441
442         if (mapping)
443                 mutex_lock(&mapping->i_mmap_mutex);
444
445         __vma_link(mm, vma, prev, rb_link, rb_parent);
446         __vma_link_file(vma);
447
448         if (mapping)
449                 mutex_unlock(&mapping->i_mmap_mutex);
450
451         mm->map_count++;
452         validate_mm(mm);
453 }
454
455 /*
456  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
457  * mm's list and rbtree.  It has already been inserted into the prio_tree.
458  */
459 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
460 {
461         struct vm_area_struct *__vma, *prev;
462         struct rb_node **rb_link, *rb_parent;
463
464         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
465         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
466         __vma_link(mm, vma, prev, rb_link, rb_parent);
467         mm->map_count++;
468 }
469
470 static inline void
471 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
472                 struct vm_area_struct *prev)
473 {
474         struct vm_area_struct *next = vma->vm_next;
475
476         prev->vm_next = next;
477         if (next)
478                 next->vm_prev = prev;
479         rb_erase(&vma->vm_rb, &mm->mm_rb);
480         if (mm->mmap_cache == vma)
481                 mm->mmap_cache = prev;
482 }
483
484 /*
485  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
486  * is already present in an i_mmap tree without adjusting the tree.
487  * The following helper function should be used when such adjustments
488  * are necessary.  The "insert" vma (if any) is to be inserted
489  * before we drop the necessary locks.
490  */
491 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
492         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
493 {
494         struct mm_struct *mm = vma->vm_mm;
495         struct vm_area_struct *next = vma->vm_next;
496         struct vm_area_struct *importer = NULL;
497         struct address_space *mapping = NULL;
498         struct prio_tree_root *root = NULL;
499         struct anon_vma *anon_vma = NULL;
500         struct file *file = vma->vm_file;
501         long adjust_next = 0;
502         int remove_next = 0;
503
504         if (next && !insert) {
505                 struct vm_area_struct *exporter = NULL;
506
507                 if (end >= next->vm_end) {
508                         /*
509                          * vma expands, overlapping all the next, and
510                          * perhaps the one after too (mprotect case 6).
511                          */
512 again:                  remove_next = 1 + (end > next->vm_end);
513                         end = next->vm_end;
514                         exporter = next;
515                         importer = vma;
516                 } else if (end > next->vm_start) {
517                         /*
518                          * vma expands, overlapping part of the next:
519                          * mprotect case 5 shifting the boundary up.
520                          */
521                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
522                         exporter = next;
523                         importer = vma;
524                 } else if (end < vma->vm_end) {
525                         /*
526                          * vma shrinks, and !insert tells it's not
527                          * split_vma inserting another: so it must be
528                          * mprotect case 4 shifting the boundary down.
529                          */
530                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
531                         exporter = vma;
532                         importer = next;
533                 }
534
535                 /*
536                  * Easily overlooked: when mprotect shifts the boundary,
537                  * make sure the expanding vma has anon_vma set if the
538                  * shrinking vma had, to cover any anon pages imported.
539                  */
540                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
541                         if (anon_vma_clone(importer, exporter))
542                                 return -ENOMEM;
543                         importer->anon_vma = exporter->anon_vma;
544                 }
545         }
546
547         if (file) {
548                 mapping = file->f_mapping;
549                 if (!(vma->vm_flags & VM_NONLINEAR))
550                         root = &mapping->i_mmap;
551                 mutex_lock(&mapping->i_mmap_mutex);
552                 if (insert) {
553                         /*
554                          * Put into prio_tree now, so instantiated pages
555                          * are visible to arm/parisc __flush_dcache_page
556                          * throughout; but we cannot insert into address
557                          * space until vma start or end is updated.
558                          */
559                         __vma_link_file(insert);
560                 }
561         }
562
563         vma_adjust_trans_huge(vma, start, end, adjust_next);
564
565         /*
566          * When changing only vma->vm_end, we don't really need anon_vma
567          * lock. This is a fairly rare case by itself, but the anon_vma
568          * lock may be shared between many sibling processes.  Skipping
569          * the lock for brk adjustments makes a difference sometimes.
570          */
571         if (vma->anon_vma && (importer || start != vma->vm_start)) {
572                 anon_vma = vma->anon_vma;
573                 anon_vma_lock(anon_vma);
574         }
575
576         if (root) {
577                 flush_dcache_mmap_lock(mapping);
578                 vma_prio_tree_remove(vma, root);
579                 if (adjust_next)
580                         vma_prio_tree_remove(next, root);
581         }
582
583         vma->vm_start = start;
584         vma->vm_end = end;
585         vma->vm_pgoff = pgoff;
586         if (adjust_next) {
587                 next->vm_start += adjust_next << PAGE_SHIFT;
588                 next->vm_pgoff += adjust_next;
589         }
590
591         if (root) {
592                 if (adjust_next)
593                         vma_prio_tree_insert(next, root);
594                 vma_prio_tree_insert(vma, root);
595                 flush_dcache_mmap_unlock(mapping);
596         }
597
598         if (remove_next) {
599                 /*
600                  * vma_merge has merged next into vma, and needs
601                  * us to remove next before dropping the locks.
602                  */
603                 __vma_unlink(mm, next, vma);
604                 if (file)
605                         __remove_shared_vm_struct(next, file, mapping);
606         } else if (insert) {
607                 /*
608                  * split_vma has split insert from vma, and needs
609                  * us to insert it before dropping the locks
610                  * (it may either follow vma or precede it).
611                  */
612                 __insert_vm_struct(mm, insert);
613         }
614
615         if (anon_vma)
616                 anon_vma_unlock(anon_vma);
617         if (mapping)
618                 mutex_unlock(&mapping->i_mmap_mutex);
619
620         if (remove_next) {
621                 if (file) {
622                         fput(file);
623                         if (next->vm_flags & VM_EXECUTABLE)
624                                 removed_exe_file_vma(mm);
625                 }
626                 if (next->anon_vma)
627                         anon_vma_merge(vma, next);
628                 mm->map_count--;
629                 mpol_put(vma_policy(next));
630                 kmem_cache_free(vm_area_cachep, next);
631                 /*
632                  * In mprotect's case 6 (see comments on vma_merge),
633                  * we must remove another next too. It would clutter
634                  * up the code too much to do both in one go.
635                  */
636                 if (remove_next == 2) {
637                         next = vma->vm_next;
638                         goto again;
639                 }
640         }
641
642         validate_mm(mm);
643
644         return 0;
645 }
646
647 /*
648  * If the vma has a ->close operation then the driver probably needs to release
649  * per-vma resources, so we don't attempt to merge those.
650  */
651 static inline int is_mergeable_vma(struct vm_area_struct *vma,
652                         struct file *file, unsigned long vm_flags)
653 {
654         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
655         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
656                 return 0;
657         if (vma->vm_file != file)
658                 return 0;
659         if (vma->vm_ops && vma->vm_ops->close)
660                 return 0;
661         return 1;
662 }
663
664 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
665                                         struct anon_vma *anon_vma2,
666                                         struct vm_area_struct *vma)
667 {
668         /*
669          * The list_is_singular() test is to avoid merging VMA cloned from
670          * parents. This can improve scalability caused by anon_vma lock.
671          */
672         if ((!anon_vma1 || !anon_vma2) && (!vma ||
673                 list_is_singular(&vma->anon_vma_chain)))
674                 return 1;
675         return anon_vma1 == anon_vma2;
676 }
677
678 /*
679  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
680  * in front of (at a lower virtual address and file offset than) the vma.
681  *
682  * We cannot merge two vmas if they have differently assigned (non-NULL)
683  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
684  *
685  * We don't check here for the merged mmap wrapping around the end of pagecache
686  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
687  * wrap, nor mmaps which cover the final page at index -1UL.
688  */
689 static int
690 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
691         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
692 {
693         if (is_mergeable_vma(vma, file, vm_flags) &&
694             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
695                 if (vma->vm_pgoff == vm_pgoff)
696                         return 1;
697         }
698         return 0;
699 }
700
701 /*
702  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
703  * beyond (at a higher virtual address and file offset than) the vma.
704  *
705  * We cannot merge two vmas if they have differently assigned (non-NULL)
706  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
707  */
708 static int
709 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
710         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
711 {
712         if (is_mergeable_vma(vma, file, vm_flags) &&
713             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
714                 pgoff_t vm_pglen;
715                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
716                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
717                         return 1;
718         }
719         return 0;
720 }
721
722 /*
723  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
724  * whether that can be merged with its predecessor or its successor.
725  * Or both (it neatly fills a hole).
726  *
727  * In most cases - when called for mmap, brk or mremap - [addr,end) is
728  * certain not to be mapped by the time vma_merge is called; but when
729  * called for mprotect, it is certain to be already mapped (either at
730  * an offset within prev, or at the start of next), and the flags of
731  * this area are about to be changed to vm_flags - and the no-change
732  * case has already been eliminated.
733  *
734  * The following mprotect cases have to be considered, where AAAA is
735  * the area passed down from mprotect_fixup, never extending beyond one
736  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
737  *
738  *     AAAA             AAAA                AAAA          AAAA
739  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
740  *    cannot merge    might become    might become    might become
741  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
742  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
743  *    mremap move:                                    PPPPNNNNNNNN 8
744  *        AAAA
745  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
746  *    might become    case 1 below    case 2 below    case 3 below
747  *
748  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
749  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
750  */
751 struct vm_area_struct *vma_merge(struct mm_struct *mm,
752                         struct vm_area_struct *prev, unsigned long addr,
753                         unsigned long end, unsigned long vm_flags,
754                         struct anon_vma *anon_vma, struct file *file,
755                         pgoff_t pgoff, struct mempolicy *policy)
756 {
757         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
758         struct vm_area_struct *area, *next;
759         int err;
760
761         /*
762          * We later require that vma->vm_flags == vm_flags,
763          * so this tests vma->vm_flags & VM_SPECIAL, too.
764          */
765         if (vm_flags & VM_SPECIAL)
766                 return NULL;
767
768         if (prev)
769                 next = prev->vm_next;
770         else
771                 next = mm->mmap;
772         area = next;
773         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
774                 next = next->vm_next;
775
776         /*
777          * Can it merge with the predecessor?
778          */
779         if (prev && prev->vm_end == addr &&
780                         mpol_equal(vma_policy(prev), policy) &&
781                         can_vma_merge_after(prev, vm_flags,
782                                                 anon_vma, file, pgoff)) {
783                 /*
784                  * OK, it can.  Can we now merge in the successor as well?
785                  */
786                 if (next && end == next->vm_start &&
787                                 mpol_equal(policy, vma_policy(next)) &&
788                                 can_vma_merge_before(next, vm_flags,
789                                         anon_vma, file, pgoff+pglen) &&
790                                 is_mergeable_anon_vma(prev->anon_vma,
791                                                       next->anon_vma, NULL)) {
792                                                         /* cases 1, 6 */
793                         err = vma_adjust(prev, prev->vm_start,
794                                 next->vm_end, prev->vm_pgoff, NULL);
795                 } else                                  /* cases 2, 5, 7 */
796                         err = vma_adjust(prev, prev->vm_start,
797                                 end, prev->vm_pgoff, NULL);
798                 if (err)
799                         return NULL;
800                 khugepaged_enter_vma_merge(prev);
801                 return prev;
802         }
803
804         /*
805          * Can this new request be merged in front of next?
806          */
807         if (next && end == next->vm_start &&
808                         mpol_equal(policy, vma_policy(next)) &&
809                         can_vma_merge_before(next, vm_flags,
810                                         anon_vma, file, pgoff+pglen)) {
811                 if (prev && addr < prev->vm_end)        /* case 4 */
812                         err = vma_adjust(prev, prev->vm_start,
813                                 addr, prev->vm_pgoff, NULL);
814                 else                                    /* cases 3, 8 */
815                         err = vma_adjust(area, addr, next->vm_end,
816                                 next->vm_pgoff - pglen, NULL);
817                 if (err)
818                         return NULL;
819                 khugepaged_enter_vma_merge(area);
820                 return area;
821         }
822
823         return NULL;
824 }
825
826 /*
827  * Rough compatbility check to quickly see if it's even worth looking
828  * at sharing an anon_vma.
829  *
830  * They need to have the same vm_file, and the flags can only differ
831  * in things that mprotect may change.
832  *
833  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
834  * we can merge the two vma's. For example, we refuse to merge a vma if
835  * there is a vm_ops->close() function, because that indicates that the
836  * driver is doing some kind of reference counting. But that doesn't
837  * really matter for the anon_vma sharing case.
838  */
839 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
840 {
841         return a->vm_end == b->vm_start &&
842                 mpol_equal(vma_policy(a), vma_policy(b)) &&
843                 a->vm_file == b->vm_file &&
844                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
845                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
846 }
847
848 /*
849  * Do some basic sanity checking to see if we can re-use the anon_vma
850  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
851  * the same as 'old', the other will be the new one that is trying
852  * to share the anon_vma.
853  *
854  * NOTE! This runs with mm_sem held for reading, so it is possible that
855  * the anon_vma of 'old' is concurrently in the process of being set up
856  * by another page fault trying to merge _that_. But that's ok: if it
857  * is being set up, that automatically means that it will be a singleton
858  * acceptable for merging, so we can do all of this optimistically. But
859  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
860  *
861  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
862  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
863  * is to return an anon_vma that is "complex" due to having gone through
864  * a fork).
865  *
866  * We also make sure that the two vma's are compatible (adjacent,
867  * and with the same memory policies). That's all stable, even with just
868  * a read lock on the mm_sem.
869  */
870 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
871 {
872         if (anon_vma_compatible(a, b)) {
873                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
874
875                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
876                         return anon_vma;
877         }
878         return NULL;
879 }
880
881 /*
882  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
883  * neighbouring vmas for a suitable anon_vma, before it goes off
884  * to allocate a new anon_vma.  It checks because a repetitive
885  * sequence of mprotects and faults may otherwise lead to distinct
886  * anon_vmas being allocated, preventing vma merge in subsequent
887  * mprotect.
888  */
889 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
890 {
891         struct anon_vma *anon_vma;
892         struct vm_area_struct *near;
893
894         near = vma->vm_next;
895         if (!near)
896                 goto try_prev;
897
898         anon_vma = reusable_anon_vma(near, vma, near);
899         if (anon_vma)
900                 return anon_vma;
901 try_prev:
902         near = vma->vm_prev;
903         if (!near)
904                 goto none;
905
906         anon_vma = reusable_anon_vma(near, near, vma);
907         if (anon_vma)
908                 return anon_vma;
909 none:
910         /*
911          * There's no absolute need to look only at touching neighbours:
912          * we could search further afield for "compatible" anon_vmas.
913          * But it would probably just be a waste of time searching,
914          * or lead to too many vmas hanging off the same anon_vma.
915          * We're trying to allow mprotect remerging later on,
916          * not trying to minimize memory used for anon_vmas.
917          */
918         return NULL;
919 }
920
921 #ifdef CONFIG_PROC_FS
922 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
923                                                 struct file *file, long pages)
924 {
925         const unsigned long stack_flags
926                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
927
928         if (file) {
929                 mm->shared_vm += pages;
930                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
931                         mm->exec_vm += pages;
932         } else if (flags & stack_flags)
933                 mm->stack_vm += pages;
934         if (flags & (VM_RESERVED|VM_IO))
935                 mm->reserved_vm += pages;
936 }
937 #endif /* CONFIG_PROC_FS */
938
939 /*
940  * If a hint addr is less than mmap_min_addr change hint to be as
941  * low as possible but still greater than mmap_min_addr
942  */
943 static inline unsigned long round_hint_to_min(unsigned long hint)
944 {
945         hint &= PAGE_MASK;
946         if (((void *)hint != NULL) &&
947             (hint < mmap_min_addr))
948                 return PAGE_ALIGN(mmap_min_addr);
949         return hint;
950 }
951
952 /*
953  * The caller must hold down_write(&current->mm->mmap_sem).
954  */
955
956 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
957                         unsigned long len, unsigned long prot,
958                         unsigned long flags, unsigned long pgoff)
959 {
960         struct mm_struct * mm = current->mm;
961         struct inode *inode;
962         vm_flags_t vm_flags;
963         int error;
964         unsigned long reqprot = prot;
965
966         /*
967          * Does the application expect PROT_READ to imply PROT_EXEC?
968          *
969          * (the exception is when the underlying filesystem is noexec
970          *  mounted, in which case we dont add PROT_EXEC.)
971          */
972         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
973                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
974                         prot |= PROT_EXEC;
975
976         if (!len)
977                 return -EINVAL;
978
979         if (!(flags & MAP_FIXED))
980                 addr = round_hint_to_min(addr);
981
982         /* Careful about overflows.. */
983         len = PAGE_ALIGN(len);
984         if (!len)
985                 return -ENOMEM;
986
987         /* offset overflow? */
988         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
989                return -EOVERFLOW;
990
991         /* Too many mappings? */
992         if (mm->map_count > sysctl_max_map_count)
993                 return -ENOMEM;
994
995         /* Obtain the address to map to. we verify (or select) it and ensure
996          * that it represents a valid section of the address space.
997          */
998         addr = get_unmapped_area(file, addr, len, pgoff, flags);
999         if (addr & ~PAGE_MASK)
1000                 return addr;
1001
1002         /* Do simple checking here so the lower-level routines won't have
1003          * to. we assume access permissions have been handled by the open
1004          * of the memory object, so we don't do any here.
1005          */
1006         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1007                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1008
1009         if (flags & MAP_LOCKED)
1010                 if (!can_do_mlock())
1011                         return -EPERM;
1012
1013         /* mlock MCL_FUTURE? */
1014         if (vm_flags & VM_LOCKED) {
1015                 unsigned long locked, lock_limit;
1016                 locked = len >> PAGE_SHIFT;
1017                 locked += mm->locked_vm;
1018                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1019                 lock_limit >>= PAGE_SHIFT;
1020                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1021                         return -EAGAIN;
1022         }
1023
1024         inode = file ? file->f_path.dentry->d_inode : NULL;
1025
1026         if (file) {
1027                 switch (flags & MAP_TYPE) {
1028                 case MAP_SHARED:
1029                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1030                                 return -EACCES;
1031
1032                         /*
1033                          * Make sure we don't allow writing to an append-only
1034                          * file..
1035                          */
1036                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1037                                 return -EACCES;
1038
1039                         /*
1040                          * Make sure there are no mandatory locks on the file.
1041                          */
1042                         if (locks_verify_locked(inode))
1043                                 return -EAGAIN;
1044
1045                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1046                         if (!(file->f_mode & FMODE_WRITE))
1047                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1048
1049                         /* fall through */
1050                 case MAP_PRIVATE:
1051                         if (!(file->f_mode & FMODE_READ))
1052                                 return -EACCES;
1053                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1054                                 if (vm_flags & VM_EXEC)
1055                                         return -EPERM;
1056                                 vm_flags &= ~VM_MAYEXEC;
1057                         }
1058
1059                         if (!file->f_op || !file->f_op->mmap)
1060                                 return -ENODEV;
1061                         break;
1062
1063                 default:
1064                         return -EINVAL;
1065                 }
1066         } else {
1067                 switch (flags & MAP_TYPE) {
1068                 case MAP_SHARED:
1069                         /*
1070                          * Ignore pgoff.
1071                          */
1072                         pgoff = 0;
1073                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1074                         break;
1075                 case MAP_PRIVATE:
1076                         /*
1077                          * Set pgoff according to addr for anon_vma.
1078                          */
1079                         pgoff = addr >> PAGE_SHIFT;
1080                         break;
1081                 default:
1082                         return -EINVAL;
1083                 }
1084         }
1085
1086         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1087         if (error)
1088                 return error;
1089
1090         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1091 }
1092 EXPORT_SYMBOL(do_mmap_pgoff);
1093
1094 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1095                 unsigned long, prot, unsigned long, flags,
1096                 unsigned long, fd, unsigned long, pgoff)
1097 {
1098         struct file *file = NULL;
1099         unsigned long retval = -EBADF;
1100
1101         if (!(flags & MAP_ANONYMOUS)) {
1102                 audit_mmap_fd(fd, flags);
1103                 if (unlikely(flags & MAP_HUGETLB))
1104                         return -EINVAL;
1105                 file = fget(fd);
1106                 if (!file)
1107                         goto out;
1108         } else if (flags & MAP_HUGETLB) {
1109                 struct user_struct *user = NULL;
1110                 /*
1111                  * VM_NORESERVE is used because the reservations will be
1112                  * taken when vm_ops->mmap() is called
1113                  * A dummy user value is used because we are not locking
1114                  * memory so no accounting is necessary
1115                  */
1116                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1117                                                 VM_NORESERVE, &user,
1118                                                 HUGETLB_ANONHUGE_INODE);
1119                 if (IS_ERR(file))
1120                         return PTR_ERR(file);
1121         }
1122
1123         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1124
1125         down_write(&current->mm->mmap_sem);
1126         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1127         up_write(&current->mm->mmap_sem);
1128
1129         if (file)
1130                 fput(file);
1131 out:
1132         return retval;
1133 }
1134
1135 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1136 struct mmap_arg_struct {
1137         unsigned long addr;
1138         unsigned long len;
1139         unsigned long prot;
1140         unsigned long flags;
1141         unsigned long fd;
1142         unsigned long offset;
1143 };
1144
1145 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1146 {
1147         struct mmap_arg_struct a;
1148
1149         if (copy_from_user(&a, arg, sizeof(a)))
1150                 return -EFAULT;
1151         if (a.offset & ~PAGE_MASK)
1152                 return -EINVAL;
1153
1154         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1155                               a.offset >> PAGE_SHIFT);
1156 }
1157 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1158
1159 /*
1160  * Some shared mappigns will want the pages marked read-only
1161  * to track write events. If so, we'll downgrade vm_page_prot
1162  * to the private version (using protection_map[] without the
1163  * VM_SHARED bit).
1164  */
1165 int vma_wants_writenotify(struct vm_area_struct *vma)
1166 {
1167         vm_flags_t vm_flags = vma->vm_flags;
1168
1169         /* If it was private or non-writable, the write bit is already clear */
1170         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1171                 return 0;
1172
1173         /* The backer wishes to know when pages are first written to? */
1174         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1175                 return 1;
1176
1177         /* The open routine did something to the protections already? */
1178         if (pgprot_val(vma->vm_page_prot) !=
1179             pgprot_val(vm_get_page_prot(vm_flags)))
1180                 return 0;
1181
1182         /* Specialty mapping? */
1183         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1184                 return 0;
1185
1186         /* Can the mapping track the dirty pages? */
1187         return vma->vm_file && vma->vm_file->f_mapping &&
1188                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1189 }
1190
1191 /*
1192  * We account for memory if it's a private writeable mapping,
1193  * not hugepages and VM_NORESERVE wasn't set.
1194  */
1195 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1196 {
1197         /*
1198          * hugetlb has its own accounting separate from the core VM
1199          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1200          */
1201         if (file && is_file_hugepages(file))
1202                 return 0;
1203
1204         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1205 }
1206
1207 unsigned long mmap_region(struct file *file, unsigned long addr,
1208                           unsigned long len, unsigned long flags,
1209                           vm_flags_t vm_flags, unsigned long pgoff)
1210 {
1211         struct mm_struct *mm = current->mm;
1212         struct vm_area_struct *vma, *prev;
1213         int correct_wcount = 0;
1214         int error;
1215         struct rb_node **rb_link, *rb_parent;
1216         unsigned long charged = 0;
1217         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1218
1219         /* Clear old maps */
1220         error = -ENOMEM;
1221 munmap_back:
1222         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1223         if (vma && vma->vm_start < addr + len) {
1224                 if (do_munmap(mm, addr, len))
1225                         return -ENOMEM;
1226                 goto munmap_back;
1227         }
1228
1229         /* Check against address space limit. */
1230         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1231                 return -ENOMEM;
1232
1233         /*
1234          * Set 'VM_NORESERVE' if we should not account for the
1235          * memory use of this mapping.
1236          */
1237         if ((flags & MAP_NORESERVE)) {
1238                 /* We honor MAP_NORESERVE if allowed to overcommit */
1239                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1240                         vm_flags |= VM_NORESERVE;
1241
1242                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1243                 if (file && is_file_hugepages(file))
1244                         vm_flags |= VM_NORESERVE;
1245         }
1246
1247         /*
1248          * Private writable mapping: check memory availability
1249          */
1250         if (accountable_mapping(file, vm_flags)) {
1251                 charged = len >> PAGE_SHIFT;
1252                 if (security_vm_enough_memory_mm(mm, charged))
1253                         return -ENOMEM;
1254                 vm_flags |= VM_ACCOUNT;
1255         }
1256
1257         /*
1258          * Can we just expand an old mapping?
1259          */
1260         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1261         if (vma)
1262                 goto out;
1263
1264         /*
1265          * Determine the object being mapped and call the appropriate
1266          * specific mapper. the address has already been validated, but
1267          * not unmapped, but the maps are removed from the list.
1268          */
1269         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1270         if (!vma) {
1271                 error = -ENOMEM;
1272                 goto unacct_error;
1273         }
1274
1275         vma->vm_mm = mm;
1276         vma->vm_start = addr;
1277         vma->vm_end = addr + len;
1278         vma->vm_flags = vm_flags;
1279         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1280         vma->vm_pgoff = pgoff;
1281         INIT_LIST_HEAD(&vma->anon_vma_chain);
1282
1283         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1284
1285         if (file) {
1286                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1287                         goto free_vma;
1288                 if (vm_flags & VM_DENYWRITE) {
1289                         error = deny_write_access(file);
1290                         if (error)
1291                                 goto free_vma;
1292                         correct_wcount = 1;
1293                 }
1294                 vma->vm_file = file;
1295                 get_file(file);
1296                 error = file->f_op->mmap(file, vma);
1297                 if (error)
1298                         goto unmap_and_free_vma;
1299                 if (vm_flags & VM_EXECUTABLE)
1300                         added_exe_file_vma(mm);
1301
1302                 /* Can addr have changed??
1303                  *
1304                  * Answer: Yes, several device drivers can do it in their
1305                  *         f_op->mmap method. -DaveM
1306                  */
1307                 addr = vma->vm_start;
1308                 pgoff = vma->vm_pgoff;
1309                 vm_flags = vma->vm_flags;
1310         } else if (vm_flags & VM_SHARED) {
1311                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1312                         goto free_vma;
1313                 error = shmem_zero_setup(vma);
1314                 if (error)
1315                         goto free_vma;
1316         }
1317
1318         if (vma_wants_writenotify(vma)) {
1319                 pgprot_t pprot = vma->vm_page_prot;
1320
1321                 /* Can vma->vm_page_prot have changed??
1322                  *
1323                  * Answer: Yes, drivers may have changed it in their
1324                  *         f_op->mmap method.
1325                  *
1326                  * Ensures that vmas marked as uncached stay that way.
1327                  */
1328                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1329                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1330                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1331         }
1332
1333         vma_link(mm, vma, prev, rb_link, rb_parent);
1334         file = vma->vm_file;
1335
1336         /* Once vma denies write, undo our temporary denial count */
1337         if (correct_wcount)
1338                 atomic_inc(&inode->i_writecount);
1339 out:
1340         perf_event_mmap(vma);
1341
1342         mm->total_vm += len >> PAGE_SHIFT;
1343         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1344         if (vm_flags & VM_LOCKED) {
1345                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1346                         mm->locked_vm += (len >> PAGE_SHIFT);
1347         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1348                 make_pages_present(addr, addr + len);
1349         return addr;
1350
1351 unmap_and_free_vma:
1352         if (correct_wcount)
1353                 atomic_inc(&inode->i_writecount);
1354         vma->vm_file = NULL;
1355         fput(file);
1356
1357         /* Undo any partial mapping done by a device driver. */
1358         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1359         charged = 0;
1360 free_vma:
1361         kmem_cache_free(vm_area_cachep, vma);
1362 unacct_error:
1363         if (charged)
1364                 vm_unacct_memory(charged);
1365         return error;
1366 }
1367
1368 /* Get an address range which is currently unmapped.
1369  * For shmat() with addr=0.
1370  *
1371  * Ugly calling convention alert:
1372  * Return value with the low bits set means error value,
1373  * ie
1374  *      if (ret & ~PAGE_MASK)
1375  *              error = ret;
1376  *
1377  * This function "knows" that -ENOMEM has the bits set.
1378  */
1379 #ifndef HAVE_ARCH_UNMAPPED_AREA
1380 unsigned long
1381 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1382                 unsigned long len, unsigned long pgoff, unsigned long flags)
1383 {
1384         struct mm_struct *mm = current->mm;
1385         struct vm_area_struct *vma;
1386         unsigned long start_addr;
1387
1388         if (len > TASK_SIZE)
1389                 return -ENOMEM;
1390
1391         if (flags & MAP_FIXED)
1392                 return addr;
1393
1394         if (addr) {
1395                 addr = PAGE_ALIGN(addr);
1396                 vma = find_vma(mm, addr);
1397                 if (TASK_SIZE - len >= addr &&
1398                     (!vma || addr + len <= vma->vm_start))
1399                         return addr;
1400         }
1401         if (len > mm->cached_hole_size) {
1402                 start_addr = addr = mm->free_area_cache;
1403         } else {
1404                 start_addr = addr = TASK_UNMAPPED_BASE;
1405                 mm->cached_hole_size = 0;
1406         }
1407
1408 full_search:
1409         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1410                 /* At this point:  (!vma || addr < vma->vm_end). */
1411                 if (TASK_SIZE - len < addr) {
1412                         /*
1413                          * Start a new search - just in case we missed
1414                          * some holes.
1415                          */
1416                         if (start_addr != TASK_UNMAPPED_BASE) {
1417                                 addr = TASK_UNMAPPED_BASE;
1418                                 start_addr = addr;
1419                                 mm->cached_hole_size = 0;
1420                                 goto full_search;
1421                         }
1422                         return -ENOMEM;
1423                 }
1424                 if (!vma || addr + len <= vma->vm_start) {
1425                         /*
1426                          * Remember the place where we stopped the search:
1427                          */
1428                         mm->free_area_cache = addr + len;
1429                         return addr;
1430                 }
1431                 if (addr + mm->cached_hole_size < vma->vm_start)
1432                         mm->cached_hole_size = vma->vm_start - addr;
1433                 addr = vma->vm_end;
1434         }
1435 }
1436 #endif  
1437
1438 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1439 {
1440         /*
1441          * Is this a new hole at the lowest possible address?
1442          */
1443         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1444                 mm->free_area_cache = addr;
1445 }
1446
1447 /*
1448  * This mmap-allocator allocates new areas top-down from below the
1449  * stack's low limit (the base):
1450  */
1451 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1452 unsigned long
1453 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1454                           const unsigned long len, const unsigned long pgoff,
1455                           const unsigned long flags)
1456 {
1457         struct vm_area_struct *vma;
1458         struct mm_struct *mm = current->mm;
1459         unsigned long addr = addr0, start_addr;
1460
1461         /* requested length too big for entire address space */
1462         if (len > TASK_SIZE)
1463                 return -ENOMEM;
1464
1465         if (flags & MAP_FIXED)
1466                 return addr;
1467
1468         /* requesting a specific address */
1469         if (addr) {
1470                 addr = PAGE_ALIGN(addr);
1471                 vma = find_vma(mm, addr);
1472                 if (TASK_SIZE - len >= addr &&
1473                                 (!vma || addr + len <= vma->vm_start))
1474                         return addr;
1475         }
1476
1477         /* check if free_area_cache is useful for us */
1478         if (len <= mm->cached_hole_size) {
1479                 mm->cached_hole_size = 0;
1480                 mm->free_area_cache = mm->mmap_base;
1481         }
1482
1483 try_again:
1484         /* either no address requested or can't fit in requested address hole */
1485         start_addr = addr = mm->free_area_cache;
1486
1487         if (addr < len)
1488                 goto fail;
1489
1490         addr -= len;
1491         do {
1492                 /*
1493                  * Lookup failure means no vma is above this address,
1494                  * else if new region fits below vma->vm_start,
1495                  * return with success:
1496                  */
1497                 vma = find_vma(mm, addr);
1498                 if (!vma || addr+len <= vma->vm_start)
1499                         /* remember the address as a hint for next time */
1500                         return (mm->free_area_cache = addr);
1501
1502                 /* remember the largest hole we saw so far */
1503                 if (addr + mm->cached_hole_size < vma->vm_start)
1504                         mm->cached_hole_size = vma->vm_start - addr;
1505
1506                 /* try just below the current vma->vm_start */
1507                 addr = vma->vm_start-len;
1508         } while (len < vma->vm_start);
1509
1510 fail:
1511         /*
1512          * if hint left us with no space for the requested
1513          * mapping then try again:
1514          *
1515          * Note: this is different with the case of bottomup
1516          * which does the fully line-search, but we use find_vma
1517          * here that causes some holes skipped.
1518          */
1519         if (start_addr != mm->mmap_base) {
1520                 mm->free_area_cache = mm->mmap_base;
1521                 mm->cached_hole_size = 0;
1522                 goto try_again;
1523         }
1524
1525         /*
1526          * A failed mmap() very likely causes application failure,
1527          * so fall back to the bottom-up function here. This scenario
1528          * can happen with large stack limits and large mmap()
1529          * allocations.
1530          */
1531         mm->cached_hole_size = ~0UL;
1532         mm->free_area_cache = TASK_UNMAPPED_BASE;
1533         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1534         /*
1535          * Restore the topdown base:
1536          */
1537         mm->free_area_cache = mm->mmap_base;
1538         mm->cached_hole_size = ~0UL;
1539
1540         return addr;
1541 }
1542 #endif
1543
1544 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1545 {
1546         /*
1547          * Is this a new hole at the highest possible address?
1548          */
1549         if (addr > mm->free_area_cache)
1550                 mm->free_area_cache = addr;
1551
1552         /* dont allow allocations above current base */
1553         if (mm->free_area_cache > mm->mmap_base)
1554                 mm->free_area_cache = mm->mmap_base;
1555 }
1556
1557 unsigned long
1558 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1559                 unsigned long pgoff, unsigned long flags)
1560 {
1561         unsigned long (*get_area)(struct file *, unsigned long,
1562                                   unsigned long, unsigned long, unsigned long);
1563
1564         unsigned long error = arch_mmap_check(addr, len, flags);
1565         if (error)
1566                 return error;
1567
1568         /* Careful about overflows.. */
1569         if (len > TASK_SIZE)
1570                 return -ENOMEM;
1571
1572         get_area = current->mm->get_unmapped_area;
1573         if (file && file->f_op && file->f_op->get_unmapped_area)
1574                 get_area = file->f_op->get_unmapped_area;
1575         addr = get_area(file, addr, len, pgoff, flags);
1576         if (IS_ERR_VALUE(addr))
1577                 return addr;
1578
1579         if (addr > TASK_SIZE - len)
1580                 return -ENOMEM;
1581         if (addr & ~PAGE_MASK)
1582                 return -EINVAL;
1583
1584         return arch_rebalance_pgtables(addr, len);
1585 }
1586
1587 EXPORT_SYMBOL(get_unmapped_area);
1588
1589 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1590 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1591 {
1592         struct vm_area_struct *vma = NULL;
1593
1594         if (mm) {
1595                 /* Check the cache first. */
1596                 /* (Cache hit rate is typically around 35%.) */
1597                 vma = mm->mmap_cache;
1598                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1599                         struct rb_node * rb_node;
1600
1601                         rb_node = mm->mm_rb.rb_node;
1602                         vma = NULL;
1603
1604                         while (rb_node) {
1605                                 struct vm_area_struct * vma_tmp;
1606
1607                                 vma_tmp = rb_entry(rb_node,
1608                                                 struct vm_area_struct, vm_rb);
1609
1610                                 if (vma_tmp->vm_end > addr) {
1611                                         vma = vma_tmp;
1612                                         if (vma_tmp->vm_start <= addr)
1613                                                 break;
1614                                         rb_node = rb_node->rb_left;
1615                                 } else
1616                                         rb_node = rb_node->rb_right;
1617                         }
1618                         if (vma)
1619                                 mm->mmap_cache = vma;
1620                 }
1621         }
1622         return vma;
1623 }
1624
1625 EXPORT_SYMBOL(find_vma);
1626
1627 /*
1628  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1629  */
1630 struct vm_area_struct *
1631 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1632                         struct vm_area_struct **pprev)
1633 {
1634         struct vm_area_struct *vma;
1635
1636         vma = find_vma(mm, addr);
1637         if (vma) {
1638                 *pprev = vma->vm_prev;
1639         } else {
1640                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1641                 *pprev = NULL;
1642                 while (rb_node) {
1643                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1644                         rb_node = rb_node->rb_right;
1645                 }
1646         }
1647         return vma;
1648 }
1649
1650 /*
1651  * Verify that the stack growth is acceptable and
1652  * update accounting. This is shared with both the
1653  * grow-up and grow-down cases.
1654  */
1655 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1656 {
1657         struct mm_struct *mm = vma->vm_mm;
1658         struct rlimit *rlim = current->signal->rlim;
1659         unsigned long new_start;
1660
1661         /* address space limit tests */
1662         if (!may_expand_vm(mm, grow))
1663                 return -ENOMEM;
1664
1665         /* Stack limit test */
1666         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1667                 return -ENOMEM;
1668
1669         /* mlock limit tests */
1670         if (vma->vm_flags & VM_LOCKED) {
1671                 unsigned long locked;
1672                 unsigned long limit;
1673                 locked = mm->locked_vm + grow;
1674                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1675                 limit >>= PAGE_SHIFT;
1676                 if (locked > limit && !capable(CAP_IPC_LOCK))
1677                         return -ENOMEM;
1678         }
1679
1680         /* Check to ensure the stack will not grow into a hugetlb-only region */
1681         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1682                         vma->vm_end - size;
1683         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1684                 return -EFAULT;
1685
1686         /*
1687          * Overcommit..  This must be the final test, as it will
1688          * update security statistics.
1689          */
1690         if (security_vm_enough_memory_mm(mm, grow))
1691                 return -ENOMEM;
1692
1693         /* Ok, everything looks good - let it rip */
1694         mm->total_vm += grow;
1695         if (vma->vm_flags & VM_LOCKED)
1696                 mm->locked_vm += grow;
1697         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1698         return 0;
1699 }
1700
1701 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1702 /*
1703  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1704  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1705  */
1706 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1707 {
1708         int error;
1709
1710         if (!(vma->vm_flags & VM_GROWSUP))
1711                 return -EFAULT;
1712
1713         /*
1714          * We must make sure the anon_vma is allocated
1715          * so that the anon_vma locking is not a noop.
1716          */
1717         if (unlikely(anon_vma_prepare(vma)))
1718                 return -ENOMEM;
1719         vma_lock_anon_vma(vma);
1720
1721         /*
1722          * vma->vm_start/vm_end cannot change under us because the caller
1723          * is required to hold the mmap_sem in read mode.  We need the
1724          * anon_vma lock to serialize against concurrent expand_stacks.
1725          * Also guard against wrapping around to address 0.
1726          */
1727         if (address < PAGE_ALIGN(address+4))
1728                 address = PAGE_ALIGN(address+4);
1729         else {
1730                 vma_unlock_anon_vma(vma);
1731                 return -ENOMEM;
1732         }
1733         error = 0;
1734
1735         /* Somebody else might have raced and expanded it already */
1736         if (address > vma->vm_end) {
1737                 unsigned long size, grow;
1738
1739                 size = address - vma->vm_start;
1740                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1741
1742                 error = -ENOMEM;
1743                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1744                         error = acct_stack_growth(vma, size, grow);
1745                         if (!error) {
1746                                 vma->vm_end = address;
1747                                 perf_event_mmap(vma);
1748                         }
1749                 }
1750         }
1751         vma_unlock_anon_vma(vma);
1752         khugepaged_enter_vma_merge(vma);
1753         return error;
1754 }
1755 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1756
1757 /*
1758  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1759  */
1760 int expand_downwards(struct vm_area_struct *vma,
1761                                    unsigned long address)
1762 {
1763         int error;
1764
1765         /*
1766          * We must make sure the anon_vma is allocated
1767          * so that the anon_vma locking is not a noop.
1768          */
1769         if (unlikely(anon_vma_prepare(vma)))
1770                 return -ENOMEM;
1771
1772         address &= PAGE_MASK;
1773         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1774         if (error)
1775                 return error;
1776
1777         vma_lock_anon_vma(vma);
1778
1779         /*
1780          * vma->vm_start/vm_end cannot change under us because the caller
1781          * is required to hold the mmap_sem in read mode.  We need the
1782          * anon_vma lock to serialize against concurrent expand_stacks.
1783          */
1784
1785         /* Somebody else might have raced and expanded it already */
1786         if (address < vma->vm_start) {
1787                 unsigned long size, grow;
1788
1789                 size = vma->vm_end - address;
1790                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1791
1792                 error = -ENOMEM;
1793                 if (grow <= vma->vm_pgoff) {
1794                         error = acct_stack_growth(vma, size, grow);
1795                         if (!error) {
1796                                 vma->vm_start = address;
1797                                 vma->vm_pgoff -= grow;
1798                                 perf_event_mmap(vma);
1799                         }
1800                 }
1801         }
1802         vma_unlock_anon_vma(vma);
1803         khugepaged_enter_vma_merge(vma);
1804         return error;
1805 }
1806
1807 #ifdef CONFIG_STACK_GROWSUP
1808 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1809 {
1810         return expand_upwards(vma, address);
1811 }
1812
1813 struct vm_area_struct *
1814 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1815 {
1816         struct vm_area_struct *vma, *prev;
1817
1818         addr &= PAGE_MASK;
1819         vma = find_vma_prev(mm, addr, &prev);
1820         if (vma && (vma->vm_start <= addr))
1821                 return vma;
1822         if (!prev || expand_stack(prev, addr))
1823                 return NULL;
1824         if (prev->vm_flags & VM_LOCKED) {
1825                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1826         }
1827         return prev;
1828 }
1829 #else
1830 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1831 {
1832         return expand_downwards(vma, address);
1833 }
1834
1835 struct vm_area_struct *
1836 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1837 {
1838         struct vm_area_struct * vma;
1839         unsigned long start;
1840
1841         addr &= PAGE_MASK;
1842         vma = find_vma(mm,addr);
1843         if (!vma)
1844                 return NULL;
1845         if (vma->vm_start <= addr)
1846                 return vma;
1847         if (!(vma->vm_flags & VM_GROWSDOWN))
1848                 return NULL;
1849         start = vma->vm_start;
1850         if (expand_stack(vma, addr))
1851                 return NULL;
1852         if (vma->vm_flags & VM_LOCKED) {
1853                 mlock_vma_pages_range(vma, addr, start);
1854         }
1855         return vma;
1856 }
1857 #endif
1858
1859 /*
1860  * Ok - we have the memory areas we should free on the vma list,
1861  * so release them, and do the vma updates.
1862  *
1863  * Called with the mm semaphore held.
1864  */
1865 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1866 {
1867         /* Update high watermark before we lower total_vm */
1868         update_hiwater_vm(mm);
1869         do {
1870                 long nrpages = vma_pages(vma);
1871
1872                 mm->total_vm -= nrpages;
1873                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1874                 vma = remove_vma(vma);
1875         } while (vma);
1876         validate_mm(mm);
1877 }
1878
1879 /*
1880  * Get rid of page table information in the indicated region.
1881  *
1882  * Called with the mm semaphore held.
1883  */
1884 static void unmap_region(struct mm_struct *mm,
1885                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1886                 unsigned long start, unsigned long end)
1887 {
1888         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1889         struct mmu_gather tlb;
1890         unsigned long nr_accounted = 0;
1891
1892         lru_add_drain();
1893         tlb_gather_mmu(&tlb, mm, 0);
1894         update_hiwater_rss(mm);
1895         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1896         vm_unacct_memory(nr_accounted);
1897         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1898                                  next ? next->vm_start : 0);
1899         tlb_finish_mmu(&tlb, start, end);
1900 }
1901
1902 /*
1903  * Create a list of vma's touched by the unmap, removing them from the mm's
1904  * vma list as we go..
1905  */
1906 static void
1907 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1908         struct vm_area_struct *prev, unsigned long end)
1909 {
1910         struct vm_area_struct **insertion_point;
1911         struct vm_area_struct *tail_vma = NULL;
1912         unsigned long addr;
1913
1914         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1915         vma->vm_prev = NULL;
1916         do {
1917                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1918                 mm->map_count--;
1919                 tail_vma = vma;
1920                 vma = vma->vm_next;
1921         } while (vma && vma->vm_start < end);
1922         *insertion_point = vma;
1923         if (vma)
1924                 vma->vm_prev = prev;
1925         tail_vma->vm_next = NULL;
1926         if (mm->unmap_area == arch_unmap_area)
1927                 addr = prev ? prev->vm_end : mm->mmap_base;
1928         else
1929                 addr = vma ?  vma->vm_start : mm->mmap_base;
1930         mm->unmap_area(mm, addr);
1931         mm->mmap_cache = NULL;          /* Kill the cache. */
1932 }
1933
1934 /*
1935  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1936  * munmap path where it doesn't make sense to fail.
1937  */
1938 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1939               unsigned long addr, int new_below)
1940 {
1941         struct mempolicy *pol;
1942         struct vm_area_struct *new;
1943         int err = -ENOMEM;
1944
1945         if (is_vm_hugetlb_page(vma) && (addr &
1946                                         ~(huge_page_mask(hstate_vma(vma)))))
1947                 return -EINVAL;
1948
1949         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1950         if (!new)
1951                 goto out_err;
1952
1953         /* most fields are the same, copy all, and then fixup */
1954         *new = *vma;
1955
1956         INIT_LIST_HEAD(&new->anon_vma_chain);
1957
1958         if (new_below)
1959                 new->vm_end = addr;
1960         else {
1961                 new->vm_start = addr;
1962                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1963         }
1964
1965         pol = mpol_dup(vma_policy(vma));
1966         if (IS_ERR(pol)) {
1967                 err = PTR_ERR(pol);
1968                 goto out_free_vma;
1969         }
1970         vma_set_policy(new, pol);
1971
1972         if (anon_vma_clone(new, vma))
1973                 goto out_free_mpol;
1974
1975         if (new->vm_file) {
1976                 get_file(new->vm_file);
1977                 if (vma->vm_flags & VM_EXECUTABLE)
1978                         added_exe_file_vma(mm);
1979         }
1980
1981         if (new->vm_ops && new->vm_ops->open)
1982                 new->vm_ops->open(new);
1983
1984         if (new_below)
1985                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1986                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1987         else
1988                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1989
1990         /* Success. */
1991         if (!err)
1992                 return 0;
1993
1994         /* Clean everything up if vma_adjust failed. */
1995         if (new->vm_ops && new->vm_ops->close)
1996                 new->vm_ops->close(new);
1997         if (new->vm_file) {
1998                 if (vma->vm_flags & VM_EXECUTABLE)
1999                         removed_exe_file_vma(mm);
2000                 fput(new->vm_file);
2001         }
2002         unlink_anon_vmas(new);
2003  out_free_mpol:
2004         mpol_put(pol);
2005  out_free_vma:
2006         kmem_cache_free(vm_area_cachep, new);
2007  out_err:
2008         return err;
2009 }
2010
2011 /*
2012  * Split a vma into two pieces at address 'addr', a new vma is allocated
2013  * either for the first part or the tail.
2014  */
2015 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2016               unsigned long addr, int new_below)
2017 {
2018         if (mm->map_count >= sysctl_max_map_count)
2019                 return -ENOMEM;
2020
2021         return __split_vma(mm, vma, addr, new_below);
2022 }
2023
2024 /* Munmap is split into 2 main parts -- this part which finds
2025  * what needs doing, and the areas themselves, which do the
2026  * work.  This now handles partial unmappings.
2027  * Jeremy Fitzhardinge <jeremy@goop.org>
2028  */
2029 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2030 {
2031         unsigned long end;
2032         struct vm_area_struct *vma, *prev, *last;
2033
2034         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2035                 return -EINVAL;
2036
2037         if ((len = PAGE_ALIGN(len)) == 0)
2038                 return -EINVAL;
2039
2040         /* Find the first overlapping VMA */
2041         vma = find_vma(mm, start);
2042         if (!vma)
2043                 return 0;
2044         prev = vma->vm_prev;
2045         /* we have  start < vma->vm_end  */
2046
2047         /* if it doesn't overlap, we have nothing.. */
2048         end = start + len;
2049         if (vma->vm_start >= end)
2050                 return 0;
2051
2052         /*
2053          * If we need to split any vma, do it now to save pain later.
2054          *
2055          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2056          * unmapped vm_area_struct will remain in use: so lower split_vma
2057          * places tmp vma above, and higher split_vma places tmp vma below.
2058          */
2059         if (start > vma->vm_start) {
2060                 int error;
2061
2062                 /*
2063                  * Make sure that map_count on return from munmap() will
2064                  * not exceed its limit; but let map_count go just above
2065                  * its limit temporarily, to help free resources as expected.
2066                  */
2067                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2068                         return -ENOMEM;
2069
2070                 error = __split_vma(mm, vma, start, 0);
2071                 if (error)
2072                         return error;
2073                 prev = vma;
2074         }
2075
2076         /* Does it split the last one? */
2077         last = find_vma(mm, end);
2078         if (last && end > last->vm_start) {
2079                 int error = __split_vma(mm, last, end, 1);
2080                 if (error)
2081                         return error;
2082         }
2083         vma = prev? prev->vm_next: mm->mmap;
2084
2085         /*
2086          * unlock any mlock()ed ranges before detaching vmas
2087          */
2088         if (mm->locked_vm) {
2089                 struct vm_area_struct *tmp = vma;
2090                 while (tmp && tmp->vm_start < end) {
2091                         if (tmp->vm_flags & VM_LOCKED) {
2092                                 mm->locked_vm -= vma_pages(tmp);
2093                                 munlock_vma_pages_all(tmp);
2094                         }
2095                         tmp = tmp->vm_next;
2096                 }
2097         }
2098
2099         /*
2100          * Remove the vma's, and unmap the actual pages
2101          */
2102         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2103         unmap_region(mm, vma, prev, start, end);
2104
2105         /* Fix up all other VM information */
2106         remove_vma_list(mm, vma);
2107
2108         return 0;
2109 }
2110 EXPORT_SYMBOL(do_munmap);
2111
2112 int vm_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2113 {
2114         int ret;
2115
2116         down_write(&mm->mmap_sem);
2117         ret = do_munmap(mm, start, len);
2118         up_write(&mm->mmap_sem);
2119         return ret;
2120 }
2121 EXPORT_SYMBOL(vm_munmap);
2122
2123 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2124 {
2125         profile_munmap(addr);
2126         return vm_munmap(current->mm, addr, len);
2127 }
2128
2129 static inline void verify_mm_writelocked(struct mm_struct *mm)
2130 {
2131 #ifdef CONFIG_DEBUG_VM
2132         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2133                 WARN_ON(1);
2134                 up_read(&mm->mmap_sem);
2135         }
2136 #endif
2137 }
2138
2139 /*
2140  *  this is really a simplified "do_mmap".  it only handles
2141  *  anonymous maps.  eventually we may be able to do some
2142  *  brk-specific accounting here.
2143  */
2144 static unsigned long do_brk(unsigned long addr, unsigned long len)
2145 {
2146         struct mm_struct * mm = current->mm;
2147         struct vm_area_struct * vma, * prev;
2148         unsigned long flags;
2149         struct rb_node ** rb_link, * rb_parent;
2150         pgoff_t pgoff = addr >> PAGE_SHIFT;
2151         int error;
2152
2153         len = PAGE_ALIGN(len);
2154         if (!len)
2155                 return addr;
2156
2157         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2158         if (error)
2159                 return error;
2160
2161         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2162
2163         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2164         if (error & ~PAGE_MASK)
2165                 return error;
2166
2167         /*
2168          * mlock MCL_FUTURE?
2169          */
2170         if (mm->def_flags & VM_LOCKED) {
2171                 unsigned long locked, lock_limit;
2172                 locked = len >> PAGE_SHIFT;
2173                 locked += mm->locked_vm;
2174                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2175                 lock_limit >>= PAGE_SHIFT;
2176                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2177                         return -EAGAIN;
2178         }
2179
2180         /*
2181          * mm->mmap_sem is required to protect against another thread
2182          * changing the mappings in case we sleep.
2183          */
2184         verify_mm_writelocked(mm);
2185
2186         /*
2187          * Clear old maps.  this also does some error checking for us
2188          */
2189  munmap_back:
2190         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2191         if (vma && vma->vm_start < addr + len) {
2192                 if (do_munmap(mm, addr, len))
2193                         return -ENOMEM;
2194                 goto munmap_back;
2195         }
2196
2197         /* Check against address space limits *after* clearing old maps... */
2198         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2199                 return -ENOMEM;
2200
2201         if (mm->map_count > sysctl_max_map_count)
2202                 return -ENOMEM;
2203
2204         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2205                 return -ENOMEM;
2206
2207         /* Can we just expand an old private anonymous mapping? */
2208         vma = vma_merge(mm, prev, addr, addr + len, flags,
2209                                         NULL, NULL, pgoff, NULL);
2210         if (vma)
2211                 goto out;
2212
2213         /*
2214          * create a vma struct for an anonymous mapping
2215          */
2216         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2217         if (!vma) {
2218                 vm_unacct_memory(len >> PAGE_SHIFT);
2219                 return -ENOMEM;
2220         }
2221
2222         INIT_LIST_HEAD(&vma->anon_vma_chain);
2223         vma->vm_mm = mm;
2224         vma->vm_start = addr;
2225         vma->vm_end = addr + len;
2226         vma->vm_pgoff = pgoff;
2227         vma->vm_flags = flags;
2228         vma->vm_page_prot = vm_get_page_prot(flags);
2229         vma_link(mm, vma, prev, rb_link, rb_parent);
2230 out:
2231         perf_event_mmap(vma);
2232         mm->total_vm += len >> PAGE_SHIFT;
2233         if (flags & VM_LOCKED) {
2234                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2235                         mm->locked_vm += (len >> PAGE_SHIFT);
2236         }
2237         return addr;
2238 }
2239
2240 unsigned long vm_brk(unsigned long addr, unsigned long len)
2241 {
2242         struct mm_struct *mm = current->mm;
2243         unsigned long ret;
2244
2245         down_write(&mm->mmap_sem);
2246         ret = do_brk(addr, len);
2247         up_write(&mm->mmap_sem);
2248         return ret;
2249 }
2250 EXPORT_SYMBOL(vm_brk);
2251
2252 /* Release all mmaps. */
2253 void exit_mmap(struct mm_struct *mm)
2254 {
2255         struct mmu_gather tlb;
2256         struct vm_area_struct *vma;
2257         unsigned long nr_accounted = 0;
2258
2259         /* mm's last user has gone, and its about to be pulled down */
2260         mmu_notifier_release(mm);
2261
2262         if (mm->locked_vm) {
2263                 vma = mm->mmap;
2264                 while (vma) {
2265                         if (vma->vm_flags & VM_LOCKED)
2266                                 munlock_vma_pages_all(vma);
2267                         vma = vma->vm_next;
2268                 }
2269         }
2270
2271         arch_exit_mmap(mm);
2272
2273         vma = mm->mmap;
2274         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2275                 return;
2276
2277         lru_add_drain();
2278         flush_cache_mm(mm);
2279         tlb_gather_mmu(&tlb, mm, 1);
2280         /* update_hiwater_rss(mm) here? but nobody should be looking */
2281         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2282         unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2283         vm_unacct_memory(nr_accounted);
2284
2285         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2286         tlb_finish_mmu(&tlb, 0, -1);
2287
2288         /*
2289          * Walk the list again, actually closing and freeing it,
2290          * with preemption enabled, without holding any MM locks.
2291          */
2292         while (vma)
2293                 vma = remove_vma(vma);
2294
2295         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2296 }
2297
2298 /* Insert vm structure into process list sorted by address
2299  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2300  * then i_mmap_mutex is taken here.
2301  */
2302 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2303 {
2304         struct vm_area_struct * __vma, * prev;
2305         struct rb_node ** rb_link, * rb_parent;
2306
2307         /*
2308          * The vm_pgoff of a purely anonymous vma should be irrelevant
2309          * until its first write fault, when page's anon_vma and index
2310          * are set.  But now set the vm_pgoff it will almost certainly
2311          * end up with (unless mremap moves it elsewhere before that
2312          * first wfault), so /proc/pid/maps tells a consistent story.
2313          *
2314          * By setting it to reflect the virtual start address of the
2315          * vma, merges and splits can happen in a seamless way, just
2316          * using the existing file pgoff checks and manipulations.
2317          * Similarly in do_mmap_pgoff and in do_brk.
2318          */
2319         if (!vma->vm_file) {
2320                 BUG_ON(vma->anon_vma);
2321                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2322         }
2323         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2324         if (__vma && __vma->vm_start < vma->vm_end)
2325                 return -ENOMEM;
2326         if ((vma->vm_flags & VM_ACCOUNT) &&
2327              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2328                 return -ENOMEM;
2329         vma_link(mm, vma, prev, rb_link, rb_parent);
2330         return 0;
2331 }
2332
2333 /*
2334  * Copy the vma structure to a new location in the same mm,
2335  * prior to moving page table entries, to effect an mremap move.
2336  */
2337 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2338         unsigned long addr, unsigned long len, pgoff_t pgoff)
2339 {
2340         struct vm_area_struct *vma = *vmap;
2341         unsigned long vma_start = vma->vm_start;
2342         struct mm_struct *mm = vma->vm_mm;
2343         struct vm_area_struct *new_vma, *prev;
2344         struct rb_node **rb_link, *rb_parent;
2345         struct mempolicy *pol;
2346         bool faulted_in_anon_vma = true;
2347
2348         /*
2349          * If anonymous vma has not yet been faulted, update new pgoff
2350          * to match new location, to increase its chance of merging.
2351          */
2352         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2353                 pgoff = addr >> PAGE_SHIFT;
2354                 faulted_in_anon_vma = false;
2355         }
2356
2357         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2358         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2359                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2360         if (new_vma) {
2361                 /*
2362                  * Source vma may have been merged into new_vma
2363                  */
2364                 if (unlikely(vma_start >= new_vma->vm_start &&
2365                              vma_start < new_vma->vm_end)) {
2366                         /*
2367                          * The only way we can get a vma_merge with
2368                          * self during an mremap is if the vma hasn't
2369                          * been faulted in yet and we were allowed to
2370                          * reset the dst vma->vm_pgoff to the
2371                          * destination address of the mremap to allow
2372                          * the merge to happen. mremap must change the
2373                          * vm_pgoff linearity between src and dst vmas
2374                          * (in turn preventing a vma_merge) to be
2375                          * safe. It is only safe to keep the vm_pgoff
2376                          * linear if there are no pages mapped yet.
2377                          */
2378                         VM_BUG_ON(faulted_in_anon_vma);
2379                         *vmap = new_vma;
2380                 } else
2381                         anon_vma_moveto_tail(new_vma);
2382         } else {
2383                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2384                 if (new_vma) {
2385                         *new_vma = *vma;
2386                         pol = mpol_dup(vma_policy(vma));
2387                         if (IS_ERR(pol))
2388                                 goto out_free_vma;
2389                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2390                         if (anon_vma_clone(new_vma, vma))
2391                                 goto out_free_mempol;
2392                         vma_set_policy(new_vma, pol);
2393                         new_vma->vm_start = addr;
2394                         new_vma->vm_end = addr + len;
2395                         new_vma->vm_pgoff = pgoff;
2396                         if (new_vma->vm_file) {
2397                                 get_file(new_vma->vm_file);
2398                                 if (vma->vm_flags & VM_EXECUTABLE)
2399                                         added_exe_file_vma(mm);
2400                         }
2401                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2402                                 new_vma->vm_ops->open(new_vma);
2403                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2404                 }
2405         }
2406         return new_vma;
2407
2408  out_free_mempol:
2409         mpol_put(pol);
2410  out_free_vma:
2411         kmem_cache_free(vm_area_cachep, new_vma);
2412         return NULL;
2413 }
2414
2415 /*
2416  * Return true if the calling process may expand its vm space by the passed
2417  * number of pages
2418  */
2419 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2420 {
2421         unsigned long cur = mm->total_vm;       /* pages */
2422         unsigned long lim;
2423
2424         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2425
2426         if (cur + npages > lim)
2427                 return 0;
2428         return 1;
2429 }
2430
2431
2432 static int special_mapping_fault(struct vm_area_struct *vma,
2433                                 struct vm_fault *vmf)
2434 {
2435         pgoff_t pgoff;
2436         struct page **pages;
2437
2438         /*
2439          * special mappings have no vm_file, and in that case, the mm
2440          * uses vm_pgoff internally. So we have to subtract it from here.
2441          * We are allowed to do this because we are the mm; do not copy
2442          * this code into drivers!
2443          */
2444         pgoff = vmf->pgoff - vma->vm_pgoff;
2445
2446         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2447                 pgoff--;
2448
2449         if (*pages) {
2450                 struct page *page = *pages;
2451                 get_page(page);
2452                 vmf->page = page;
2453                 return 0;
2454         }
2455
2456         return VM_FAULT_SIGBUS;
2457 }
2458
2459 /*
2460  * Having a close hook prevents vma merging regardless of flags.
2461  */
2462 static void special_mapping_close(struct vm_area_struct *vma)
2463 {
2464 }
2465
2466 static const struct vm_operations_struct special_mapping_vmops = {
2467         .close = special_mapping_close,
2468         .fault = special_mapping_fault,
2469 };
2470
2471 /*
2472  * Called with mm->mmap_sem held for writing.
2473  * Insert a new vma covering the given region, with the given flags.
2474  * Its pages are supplied by the given array of struct page *.
2475  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2476  * The region past the last page supplied will always produce SIGBUS.
2477  * The array pointer and the pages it points to are assumed to stay alive
2478  * for as long as this mapping might exist.
2479  */
2480 int install_special_mapping(struct mm_struct *mm,
2481                             unsigned long addr, unsigned long len,
2482                             unsigned long vm_flags, struct page **pages)
2483 {
2484         int ret;
2485         struct vm_area_struct *vma;
2486
2487         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2488         if (unlikely(vma == NULL))
2489                 return -ENOMEM;
2490
2491         INIT_LIST_HEAD(&vma->anon_vma_chain);
2492         vma->vm_mm = mm;
2493         vma->vm_start = addr;
2494         vma->vm_end = addr + len;
2495
2496         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2497         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2498
2499         vma->vm_ops = &special_mapping_vmops;
2500         vma->vm_private_data = pages;
2501
2502         ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2503         if (ret)
2504                 goto out;
2505
2506         ret = insert_vm_struct(mm, vma);
2507         if (ret)
2508                 goto out;
2509
2510         mm->total_vm += len >> PAGE_SHIFT;
2511
2512         perf_event_mmap(vma);
2513
2514         return 0;
2515
2516 out:
2517         kmem_cache_free(vm_area_cachep, vma);
2518         return ret;
2519 }
2520
2521 static DEFINE_MUTEX(mm_all_locks_mutex);
2522
2523 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2524 {
2525         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2526                 /*
2527                  * The LSB of head.next can't change from under us
2528                  * because we hold the mm_all_locks_mutex.
2529                  */
2530                 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2531                 /*
2532                  * We can safely modify head.next after taking the
2533                  * anon_vma->root->mutex. If some other vma in this mm shares
2534                  * the same anon_vma we won't take it again.
2535                  *
2536                  * No need of atomic instructions here, head.next
2537                  * can't change from under us thanks to the
2538                  * anon_vma->root->mutex.
2539                  */
2540                 if (__test_and_set_bit(0, (unsigned long *)
2541                                        &anon_vma->root->head.next))
2542                         BUG();
2543         }
2544 }
2545
2546 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2547 {
2548         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2549                 /*
2550                  * AS_MM_ALL_LOCKS can't change from under us because
2551                  * we hold the mm_all_locks_mutex.
2552                  *
2553                  * Operations on ->flags have to be atomic because
2554                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2555                  * mm_all_locks_mutex, there may be other cpus
2556                  * changing other bitflags in parallel to us.
2557                  */
2558                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2559                         BUG();
2560                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2561         }
2562 }
2563
2564 /*
2565  * This operation locks against the VM for all pte/vma/mm related
2566  * operations that could ever happen on a certain mm. This includes
2567  * vmtruncate, try_to_unmap, and all page faults.
2568  *
2569  * The caller must take the mmap_sem in write mode before calling
2570  * mm_take_all_locks(). The caller isn't allowed to release the
2571  * mmap_sem until mm_drop_all_locks() returns.
2572  *
2573  * mmap_sem in write mode is required in order to block all operations
2574  * that could modify pagetables and free pages without need of
2575  * altering the vma layout (for example populate_range() with
2576  * nonlinear vmas). It's also needed in write mode to avoid new
2577  * anon_vmas to be associated with existing vmas.
2578  *
2579  * A single task can't take more than one mm_take_all_locks() in a row
2580  * or it would deadlock.
2581  *
2582  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2583  * mapping->flags avoid to take the same lock twice, if more than one
2584  * vma in this mm is backed by the same anon_vma or address_space.
2585  *
2586  * We can take all the locks in random order because the VM code
2587  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2588  * takes more than one of them in a row. Secondly we're protected
2589  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2590  *
2591  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2592  * that may have to take thousand of locks.
2593  *
2594  * mm_take_all_locks() can fail if it's interrupted by signals.
2595  */
2596 int mm_take_all_locks(struct mm_struct *mm)
2597 {
2598         struct vm_area_struct *vma;
2599         struct anon_vma_chain *avc;
2600
2601         BUG_ON(down_read_trylock(&mm->mmap_sem));
2602
2603         mutex_lock(&mm_all_locks_mutex);
2604
2605         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2606                 if (signal_pending(current))
2607                         goto out_unlock;
2608                 if (vma->vm_file && vma->vm_file->f_mapping)
2609                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2610         }
2611
2612         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2613                 if (signal_pending(current))
2614                         goto out_unlock;
2615                 if (vma->anon_vma)
2616                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2617                                 vm_lock_anon_vma(mm, avc->anon_vma);
2618         }
2619
2620         return 0;
2621
2622 out_unlock:
2623         mm_drop_all_locks(mm);
2624         return -EINTR;
2625 }
2626
2627 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2628 {
2629         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2630                 /*
2631                  * The LSB of head.next can't change to 0 from under
2632                  * us because we hold the mm_all_locks_mutex.
2633                  *
2634                  * We must however clear the bitflag before unlocking
2635                  * the vma so the users using the anon_vma->head will
2636                  * never see our bitflag.
2637                  *
2638                  * No need of atomic instructions here, head.next
2639                  * can't change from under us until we release the
2640                  * anon_vma->root->mutex.
2641                  */
2642                 if (!__test_and_clear_bit(0, (unsigned long *)
2643                                           &anon_vma->root->head.next))
2644                         BUG();
2645                 anon_vma_unlock(anon_vma);
2646         }
2647 }
2648
2649 static void vm_unlock_mapping(struct address_space *mapping)
2650 {
2651         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2652                 /*
2653                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2654                  * because we hold the mm_all_locks_mutex.
2655                  */
2656                 mutex_unlock(&mapping->i_mmap_mutex);
2657                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2658                                         &mapping->flags))
2659                         BUG();
2660         }
2661 }
2662
2663 /*
2664  * The mmap_sem cannot be released by the caller until
2665  * mm_drop_all_locks() returns.
2666  */
2667 void mm_drop_all_locks(struct mm_struct *mm)
2668 {
2669         struct vm_area_struct *vma;
2670         struct anon_vma_chain *avc;
2671
2672         BUG_ON(down_read_trylock(&mm->mmap_sem));
2673         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2674
2675         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2676                 if (vma->anon_vma)
2677                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2678                                 vm_unlock_anon_vma(avc->anon_vma);
2679                 if (vma->vm_file && vma->vm_file->f_mapping)
2680                         vm_unlock_mapping(vma->vm_file->f_mapping);
2681         }
2682
2683         mutex_unlock(&mm_all_locks_mutex);
2684 }
2685
2686 /*
2687  * initialise the VMA slab
2688  */
2689 void __init mmap_init(void)
2690 {
2691         int ret;
2692
2693         ret = percpu_counter_init(&vm_committed_as, 0);
2694         VM_BUG_ON(ret);
2695 }