mm: have order > 0 compaction start near a pageblock with free pages
[linux-3.10.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 #include <linux/uprobes.h>
34
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
37 #include <asm/tlb.h>
38 #include <asm/mmu_context.h>
39
40 #include "internal.h"
41
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags)       (0)
44 #endif
45
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len)              (addr)
48 #endif
49
50 static void unmap_region(struct mm_struct *mm,
51                 struct vm_area_struct *vma, struct vm_area_struct *prev,
52                 unsigned long start, unsigned long end);
53
54 /*
55  * WARNING: the debugging will use recursive algorithms so never enable this
56  * unless you know what you are doing.
57  */
58 #undef DEBUG_MM_RB
59
60 /* description of effects of mapping type and prot in current implementation.
61  * this is due to the limited x86 page protection hardware.  The expected
62  * behavior is in parens:
63  *
64  * map_type     prot
65  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
66  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
67  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
68  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
69  *              
70  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
71  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
72  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
73  *
74  */
75 pgprot_t protection_map[16] = {
76         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
77         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
78 };
79
80 pgprot_t vm_get_page_prot(unsigned long vm_flags)
81 {
82         return __pgprot(pgprot_val(protection_map[vm_flags &
83                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
84                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
85 }
86 EXPORT_SYMBOL(vm_get_page_prot);
87
88 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
89 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
90 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
91 /*
92  * Make sure vm_committed_as in one cacheline and not cacheline shared with
93  * other variables. It can be updated by several CPUs frequently.
94  */
95 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
96
97 /*
98  * Check that a process has enough memory to allocate a new virtual
99  * mapping. 0 means there is enough memory for the allocation to
100  * succeed and -ENOMEM implies there is not.
101  *
102  * We currently support three overcommit policies, which are set via the
103  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
104  *
105  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
106  * Additional code 2002 Jul 20 by Robert Love.
107  *
108  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
109  *
110  * Note this is a helper function intended to be used by LSMs which
111  * wish to use this logic.
112  */
113 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
114 {
115         unsigned long free, allowed;
116
117         vm_acct_memory(pages);
118
119         /*
120          * Sometimes we want to use more memory than we have
121          */
122         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
123                 return 0;
124
125         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
126                 free = global_page_state(NR_FREE_PAGES);
127                 free += global_page_state(NR_FILE_PAGES);
128
129                 /*
130                  * shmem pages shouldn't be counted as free in this
131                  * case, they can't be purged, only swapped out, and
132                  * that won't affect the overall amount of available
133                  * memory in the system.
134                  */
135                 free -= global_page_state(NR_SHMEM);
136
137                 free += nr_swap_pages;
138
139                 /*
140                  * Any slabs which are created with the
141                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
142                  * which are reclaimable, under pressure.  The dentry
143                  * cache and most inode caches should fall into this
144                  */
145                 free += global_page_state(NR_SLAB_RECLAIMABLE);
146
147                 /*
148                  * Leave reserved pages. The pages are not for anonymous pages.
149                  */
150                 if (free <= totalreserve_pages)
151                         goto error;
152                 else
153                         free -= totalreserve_pages;
154
155                 /*
156                  * Leave the last 3% for root
157                  */
158                 if (!cap_sys_admin)
159                         free -= free / 32;
160
161                 if (free > pages)
162                         return 0;
163
164                 goto error;
165         }
166
167         allowed = (totalram_pages - hugetlb_total_pages())
168                 * sysctl_overcommit_ratio / 100;
169         /*
170          * Leave the last 3% for root
171          */
172         if (!cap_sys_admin)
173                 allowed -= allowed / 32;
174         allowed += total_swap_pages;
175
176         /* Don't let a single process grow too big:
177            leave 3% of the size of this process for other processes */
178         if (mm)
179                 allowed -= mm->total_vm / 32;
180
181         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
182                 return 0;
183 error:
184         vm_unacct_memory(pages);
185
186         return -ENOMEM;
187 }
188
189 /*
190  * Requires inode->i_mapping->i_mmap_mutex
191  */
192 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193                 struct file *file, struct address_space *mapping)
194 {
195         if (vma->vm_flags & VM_DENYWRITE)
196                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
197         if (vma->vm_flags & VM_SHARED)
198                 mapping->i_mmap_writable--;
199
200         flush_dcache_mmap_lock(mapping);
201         if (unlikely(vma->vm_flags & VM_NONLINEAR))
202                 list_del_init(&vma->shared.vm_set.list);
203         else
204                 vma_prio_tree_remove(vma, &mapping->i_mmap);
205         flush_dcache_mmap_unlock(mapping);
206 }
207
208 /*
209  * Unlink a file-based vm structure from its prio_tree, to hide
210  * vma from rmap and vmtruncate before freeing its page tables.
211  */
212 void unlink_file_vma(struct vm_area_struct *vma)
213 {
214         struct file *file = vma->vm_file;
215
216         if (file) {
217                 struct address_space *mapping = file->f_mapping;
218                 mutex_lock(&mapping->i_mmap_mutex);
219                 __remove_shared_vm_struct(vma, file, mapping);
220                 mutex_unlock(&mapping->i_mmap_mutex);
221         }
222 }
223
224 /*
225  * Close a vm structure and free it, returning the next.
226  */
227 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
228 {
229         struct vm_area_struct *next = vma->vm_next;
230
231         might_sleep();
232         if (vma->vm_ops && vma->vm_ops->close)
233                 vma->vm_ops->close(vma);
234         if (vma->vm_file) {
235                 fput(vma->vm_file);
236                 if (vma->vm_flags & VM_EXECUTABLE)
237                         removed_exe_file_vma(vma->vm_mm);
238         }
239         mpol_put(vma_policy(vma));
240         kmem_cache_free(vm_area_cachep, vma);
241         return next;
242 }
243
244 static unsigned long do_brk(unsigned long addr, unsigned long len);
245
246 SYSCALL_DEFINE1(brk, unsigned long, brk)
247 {
248         unsigned long rlim, retval;
249         unsigned long newbrk, oldbrk;
250         struct mm_struct *mm = current->mm;
251         unsigned long min_brk;
252
253         down_write(&mm->mmap_sem);
254
255 #ifdef CONFIG_COMPAT_BRK
256         /*
257          * CONFIG_COMPAT_BRK can still be overridden by setting
258          * randomize_va_space to 2, which will still cause mm->start_brk
259          * to be arbitrarily shifted
260          */
261         if (current->brk_randomized)
262                 min_brk = mm->start_brk;
263         else
264                 min_brk = mm->end_data;
265 #else
266         min_brk = mm->start_brk;
267 #endif
268         if (brk < min_brk)
269                 goto out;
270
271         /*
272          * Check against rlimit here. If this check is done later after the test
273          * of oldbrk with newbrk then it can escape the test and let the data
274          * segment grow beyond its set limit the in case where the limit is
275          * not page aligned -Ram Gupta
276          */
277         rlim = rlimit(RLIMIT_DATA);
278         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
279                         (mm->end_data - mm->start_data) > rlim)
280                 goto out;
281
282         newbrk = PAGE_ALIGN(brk);
283         oldbrk = PAGE_ALIGN(mm->brk);
284         if (oldbrk == newbrk)
285                 goto set_brk;
286
287         /* Always allow shrinking brk. */
288         if (brk <= mm->brk) {
289                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
290                         goto set_brk;
291                 goto out;
292         }
293
294         /* Check against existing mmap mappings. */
295         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
296                 goto out;
297
298         /* Ok, looks good - let it rip. */
299         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
300                 goto out;
301 set_brk:
302         mm->brk = brk;
303 out:
304         retval = mm->brk;
305         up_write(&mm->mmap_sem);
306         return retval;
307 }
308
309 #ifdef DEBUG_MM_RB
310 static int browse_rb(struct rb_root *root)
311 {
312         int i = 0, j;
313         struct rb_node *nd, *pn = NULL;
314         unsigned long prev = 0, pend = 0;
315
316         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
317                 struct vm_area_struct *vma;
318                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
319                 if (vma->vm_start < prev)
320                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
321                 if (vma->vm_start < pend)
322                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
323                 if (vma->vm_start > vma->vm_end)
324                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
325                 i++;
326                 pn = nd;
327                 prev = vma->vm_start;
328                 pend = vma->vm_end;
329         }
330         j = 0;
331         for (nd = pn; nd; nd = rb_prev(nd)) {
332                 j++;
333         }
334         if (i != j)
335                 printk("backwards %d, forwards %d\n", j, i), i = 0;
336         return i;
337 }
338
339 void validate_mm(struct mm_struct *mm)
340 {
341         int bug = 0;
342         int i = 0;
343         struct vm_area_struct *tmp = mm->mmap;
344         while (tmp) {
345                 tmp = tmp->vm_next;
346                 i++;
347         }
348         if (i != mm->map_count)
349                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
350         i = browse_rb(&mm->mm_rb);
351         if (i != mm->map_count)
352                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
353         BUG_ON(bug);
354 }
355 #else
356 #define validate_mm(mm) do { } while (0)
357 #endif
358
359 static struct vm_area_struct *
360 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
361                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
362                 struct rb_node ** rb_parent)
363 {
364         struct vm_area_struct * vma;
365         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
366
367         __rb_link = &mm->mm_rb.rb_node;
368         rb_prev = __rb_parent = NULL;
369         vma = NULL;
370
371         while (*__rb_link) {
372                 struct vm_area_struct *vma_tmp;
373
374                 __rb_parent = *__rb_link;
375                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
376
377                 if (vma_tmp->vm_end > addr) {
378                         vma = vma_tmp;
379                         if (vma_tmp->vm_start <= addr)
380                                 break;
381                         __rb_link = &__rb_parent->rb_left;
382                 } else {
383                         rb_prev = __rb_parent;
384                         __rb_link = &__rb_parent->rb_right;
385                 }
386         }
387
388         *pprev = NULL;
389         if (rb_prev)
390                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
391         *rb_link = __rb_link;
392         *rb_parent = __rb_parent;
393         return vma;
394 }
395
396 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
397                 struct rb_node **rb_link, struct rb_node *rb_parent)
398 {
399         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
400         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
401 }
402
403 static void __vma_link_file(struct vm_area_struct *vma)
404 {
405         struct file *file;
406
407         file = vma->vm_file;
408         if (file) {
409                 struct address_space *mapping = file->f_mapping;
410
411                 if (vma->vm_flags & VM_DENYWRITE)
412                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
413                 if (vma->vm_flags & VM_SHARED)
414                         mapping->i_mmap_writable++;
415
416                 flush_dcache_mmap_lock(mapping);
417                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
418                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
419                 else
420                         vma_prio_tree_insert(vma, &mapping->i_mmap);
421                 flush_dcache_mmap_unlock(mapping);
422         }
423 }
424
425 static void
426 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
427         struct vm_area_struct *prev, struct rb_node **rb_link,
428         struct rb_node *rb_parent)
429 {
430         __vma_link_list(mm, vma, prev, rb_parent);
431         __vma_link_rb(mm, vma, rb_link, rb_parent);
432 }
433
434 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435                         struct vm_area_struct *prev, struct rb_node **rb_link,
436                         struct rb_node *rb_parent)
437 {
438         struct address_space *mapping = NULL;
439
440         if (vma->vm_file)
441                 mapping = vma->vm_file->f_mapping;
442
443         if (mapping)
444                 mutex_lock(&mapping->i_mmap_mutex);
445
446         __vma_link(mm, vma, prev, rb_link, rb_parent);
447         __vma_link_file(vma);
448
449         if (mapping)
450                 mutex_unlock(&mapping->i_mmap_mutex);
451
452         mm->map_count++;
453         validate_mm(mm);
454 }
455
456 /*
457  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
458  * mm's list and rbtree.  It has already been inserted into the prio_tree.
459  */
460 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
461 {
462         struct vm_area_struct *__vma, *prev;
463         struct rb_node **rb_link, *rb_parent;
464
465         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467         __vma_link(mm, vma, prev, rb_link, rb_parent);
468         mm->map_count++;
469 }
470
471 static inline void
472 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473                 struct vm_area_struct *prev)
474 {
475         struct vm_area_struct *next = vma->vm_next;
476
477         prev->vm_next = next;
478         if (next)
479                 next->vm_prev = prev;
480         rb_erase(&vma->vm_rb, &mm->mm_rb);
481         if (mm->mmap_cache == vma)
482                 mm->mmap_cache = prev;
483 }
484
485 /*
486  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487  * is already present in an i_mmap tree without adjusting the tree.
488  * The following helper function should be used when such adjustments
489  * are necessary.  The "insert" vma (if any) is to be inserted
490  * before we drop the necessary locks.
491  */
492 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
493         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
494 {
495         struct mm_struct *mm = vma->vm_mm;
496         struct vm_area_struct *next = vma->vm_next;
497         struct vm_area_struct *importer = NULL;
498         struct address_space *mapping = NULL;
499         struct prio_tree_root *root = NULL;
500         struct anon_vma *anon_vma = NULL;
501         struct file *file = vma->vm_file;
502         long adjust_next = 0;
503         int remove_next = 0;
504
505         if (next && !insert) {
506                 struct vm_area_struct *exporter = NULL;
507
508                 if (end >= next->vm_end) {
509                         /*
510                          * vma expands, overlapping all the next, and
511                          * perhaps the one after too (mprotect case 6).
512                          */
513 again:                  remove_next = 1 + (end > next->vm_end);
514                         end = next->vm_end;
515                         exporter = next;
516                         importer = vma;
517                 } else if (end > next->vm_start) {
518                         /*
519                          * vma expands, overlapping part of the next:
520                          * mprotect case 5 shifting the boundary up.
521                          */
522                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
523                         exporter = next;
524                         importer = vma;
525                 } else if (end < vma->vm_end) {
526                         /*
527                          * vma shrinks, and !insert tells it's not
528                          * split_vma inserting another: so it must be
529                          * mprotect case 4 shifting the boundary down.
530                          */
531                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
532                         exporter = vma;
533                         importer = next;
534                 }
535
536                 /*
537                  * Easily overlooked: when mprotect shifts the boundary,
538                  * make sure the expanding vma has anon_vma set if the
539                  * shrinking vma had, to cover any anon pages imported.
540                  */
541                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
542                         if (anon_vma_clone(importer, exporter))
543                                 return -ENOMEM;
544                         importer->anon_vma = exporter->anon_vma;
545                 }
546         }
547
548         if (file) {
549                 mapping = file->f_mapping;
550                 if (!(vma->vm_flags & VM_NONLINEAR)) {
551                         root = &mapping->i_mmap;
552                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
553
554                         if (adjust_next)
555                                 uprobe_munmap(next, next->vm_start,
556                                                         next->vm_end);
557                 }
558
559                 mutex_lock(&mapping->i_mmap_mutex);
560                 if (insert) {
561                         /*
562                          * Put into prio_tree now, so instantiated pages
563                          * are visible to arm/parisc __flush_dcache_page
564                          * throughout; but we cannot insert into address
565                          * space until vma start or end is updated.
566                          */
567                         __vma_link_file(insert);
568                 }
569         }
570
571         vma_adjust_trans_huge(vma, start, end, adjust_next);
572
573         /*
574          * When changing only vma->vm_end, we don't really need anon_vma
575          * lock. This is a fairly rare case by itself, but the anon_vma
576          * lock may be shared between many sibling processes.  Skipping
577          * the lock for brk adjustments makes a difference sometimes.
578          */
579         if (vma->anon_vma && (importer || start != vma->vm_start)) {
580                 anon_vma = vma->anon_vma;
581                 anon_vma_lock(anon_vma);
582         }
583
584         if (root) {
585                 flush_dcache_mmap_lock(mapping);
586                 vma_prio_tree_remove(vma, root);
587                 if (adjust_next)
588                         vma_prio_tree_remove(next, root);
589         }
590
591         vma->vm_start = start;
592         vma->vm_end = end;
593         vma->vm_pgoff = pgoff;
594         if (adjust_next) {
595                 next->vm_start += adjust_next << PAGE_SHIFT;
596                 next->vm_pgoff += adjust_next;
597         }
598
599         if (root) {
600                 if (adjust_next)
601                         vma_prio_tree_insert(next, root);
602                 vma_prio_tree_insert(vma, root);
603                 flush_dcache_mmap_unlock(mapping);
604         }
605
606         if (remove_next) {
607                 /*
608                  * vma_merge has merged next into vma, and needs
609                  * us to remove next before dropping the locks.
610                  */
611                 __vma_unlink(mm, next, vma);
612                 if (file)
613                         __remove_shared_vm_struct(next, file, mapping);
614         } else if (insert) {
615                 /*
616                  * split_vma has split insert from vma, and needs
617                  * us to insert it before dropping the locks
618                  * (it may either follow vma or precede it).
619                  */
620                 __insert_vm_struct(mm, insert);
621         }
622
623         if (anon_vma)
624                 anon_vma_unlock(anon_vma);
625         if (mapping)
626                 mutex_unlock(&mapping->i_mmap_mutex);
627
628         if (root) {
629                 uprobe_mmap(vma);
630
631                 if (adjust_next)
632                         uprobe_mmap(next);
633         }
634
635         if (remove_next) {
636                 if (file) {
637                         uprobe_munmap(next, next->vm_start, next->vm_end);
638                         fput(file);
639                         if (next->vm_flags & VM_EXECUTABLE)
640                                 removed_exe_file_vma(mm);
641                 }
642                 if (next->anon_vma)
643                         anon_vma_merge(vma, next);
644                 mm->map_count--;
645                 mpol_put(vma_policy(next));
646                 kmem_cache_free(vm_area_cachep, next);
647                 /*
648                  * In mprotect's case 6 (see comments on vma_merge),
649                  * we must remove another next too. It would clutter
650                  * up the code too much to do both in one go.
651                  */
652                 if (remove_next == 2) {
653                         next = vma->vm_next;
654                         goto again;
655                 }
656         }
657         if (insert && file)
658                 uprobe_mmap(insert);
659
660         validate_mm(mm);
661
662         return 0;
663 }
664
665 /*
666  * If the vma has a ->close operation then the driver probably needs to release
667  * per-vma resources, so we don't attempt to merge those.
668  */
669 static inline int is_mergeable_vma(struct vm_area_struct *vma,
670                         struct file *file, unsigned long vm_flags)
671 {
672         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
673         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
674                 return 0;
675         if (vma->vm_file != file)
676                 return 0;
677         if (vma->vm_ops && vma->vm_ops->close)
678                 return 0;
679         return 1;
680 }
681
682 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
683                                         struct anon_vma *anon_vma2,
684                                         struct vm_area_struct *vma)
685 {
686         /*
687          * The list_is_singular() test is to avoid merging VMA cloned from
688          * parents. This can improve scalability caused by anon_vma lock.
689          */
690         if ((!anon_vma1 || !anon_vma2) && (!vma ||
691                 list_is_singular(&vma->anon_vma_chain)))
692                 return 1;
693         return anon_vma1 == anon_vma2;
694 }
695
696 /*
697  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
698  * in front of (at a lower virtual address and file offset than) the vma.
699  *
700  * We cannot merge two vmas if they have differently assigned (non-NULL)
701  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
702  *
703  * We don't check here for the merged mmap wrapping around the end of pagecache
704  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
705  * wrap, nor mmaps which cover the final page at index -1UL.
706  */
707 static int
708 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
709         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
710 {
711         if (is_mergeable_vma(vma, file, vm_flags) &&
712             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
713                 if (vma->vm_pgoff == vm_pgoff)
714                         return 1;
715         }
716         return 0;
717 }
718
719 /*
720  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
721  * beyond (at a higher virtual address and file offset than) the vma.
722  *
723  * We cannot merge two vmas if they have differently assigned (non-NULL)
724  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
725  */
726 static int
727 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
728         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
729 {
730         if (is_mergeable_vma(vma, file, vm_flags) &&
731             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
732                 pgoff_t vm_pglen;
733                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
734                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
735                         return 1;
736         }
737         return 0;
738 }
739
740 /*
741  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
742  * whether that can be merged with its predecessor or its successor.
743  * Or both (it neatly fills a hole).
744  *
745  * In most cases - when called for mmap, brk or mremap - [addr,end) is
746  * certain not to be mapped by the time vma_merge is called; but when
747  * called for mprotect, it is certain to be already mapped (either at
748  * an offset within prev, or at the start of next), and the flags of
749  * this area are about to be changed to vm_flags - and the no-change
750  * case has already been eliminated.
751  *
752  * The following mprotect cases have to be considered, where AAAA is
753  * the area passed down from mprotect_fixup, never extending beyond one
754  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
755  *
756  *     AAAA             AAAA                AAAA          AAAA
757  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
758  *    cannot merge    might become    might become    might become
759  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
760  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
761  *    mremap move:                                    PPPPNNNNNNNN 8
762  *        AAAA
763  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
764  *    might become    case 1 below    case 2 below    case 3 below
765  *
766  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
767  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
768  */
769 struct vm_area_struct *vma_merge(struct mm_struct *mm,
770                         struct vm_area_struct *prev, unsigned long addr,
771                         unsigned long end, unsigned long vm_flags,
772                         struct anon_vma *anon_vma, struct file *file,
773                         pgoff_t pgoff, struct mempolicy *policy)
774 {
775         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
776         struct vm_area_struct *area, *next;
777         int err;
778
779         /*
780          * We later require that vma->vm_flags == vm_flags,
781          * so this tests vma->vm_flags & VM_SPECIAL, too.
782          */
783         if (vm_flags & VM_SPECIAL)
784                 return NULL;
785
786         if (prev)
787                 next = prev->vm_next;
788         else
789                 next = mm->mmap;
790         area = next;
791         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
792                 next = next->vm_next;
793
794         /*
795          * Can it merge with the predecessor?
796          */
797         if (prev && prev->vm_end == addr &&
798                         mpol_equal(vma_policy(prev), policy) &&
799                         can_vma_merge_after(prev, vm_flags,
800                                                 anon_vma, file, pgoff)) {
801                 /*
802                  * OK, it can.  Can we now merge in the successor as well?
803                  */
804                 if (next && end == next->vm_start &&
805                                 mpol_equal(policy, vma_policy(next)) &&
806                                 can_vma_merge_before(next, vm_flags,
807                                         anon_vma, file, pgoff+pglen) &&
808                                 is_mergeable_anon_vma(prev->anon_vma,
809                                                       next->anon_vma, NULL)) {
810                                                         /* cases 1, 6 */
811                         err = vma_adjust(prev, prev->vm_start,
812                                 next->vm_end, prev->vm_pgoff, NULL);
813                 } else                                  /* cases 2, 5, 7 */
814                         err = vma_adjust(prev, prev->vm_start,
815                                 end, prev->vm_pgoff, NULL);
816                 if (err)
817                         return NULL;
818                 khugepaged_enter_vma_merge(prev);
819                 return prev;
820         }
821
822         /*
823          * Can this new request be merged in front of next?
824          */
825         if (next && end == next->vm_start &&
826                         mpol_equal(policy, vma_policy(next)) &&
827                         can_vma_merge_before(next, vm_flags,
828                                         anon_vma, file, pgoff+pglen)) {
829                 if (prev && addr < prev->vm_end)        /* case 4 */
830                         err = vma_adjust(prev, prev->vm_start,
831                                 addr, prev->vm_pgoff, NULL);
832                 else                                    /* cases 3, 8 */
833                         err = vma_adjust(area, addr, next->vm_end,
834                                 next->vm_pgoff - pglen, NULL);
835                 if (err)
836                         return NULL;
837                 khugepaged_enter_vma_merge(area);
838                 return area;
839         }
840
841         return NULL;
842 }
843
844 /*
845  * Rough compatbility check to quickly see if it's even worth looking
846  * at sharing an anon_vma.
847  *
848  * They need to have the same vm_file, and the flags can only differ
849  * in things that mprotect may change.
850  *
851  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
852  * we can merge the two vma's. For example, we refuse to merge a vma if
853  * there is a vm_ops->close() function, because that indicates that the
854  * driver is doing some kind of reference counting. But that doesn't
855  * really matter for the anon_vma sharing case.
856  */
857 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
858 {
859         return a->vm_end == b->vm_start &&
860                 mpol_equal(vma_policy(a), vma_policy(b)) &&
861                 a->vm_file == b->vm_file &&
862                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
863                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
864 }
865
866 /*
867  * Do some basic sanity checking to see if we can re-use the anon_vma
868  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
869  * the same as 'old', the other will be the new one that is trying
870  * to share the anon_vma.
871  *
872  * NOTE! This runs with mm_sem held for reading, so it is possible that
873  * the anon_vma of 'old' is concurrently in the process of being set up
874  * by another page fault trying to merge _that_. But that's ok: if it
875  * is being set up, that automatically means that it will be a singleton
876  * acceptable for merging, so we can do all of this optimistically. But
877  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
878  *
879  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
880  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
881  * is to return an anon_vma that is "complex" due to having gone through
882  * a fork).
883  *
884  * We also make sure that the two vma's are compatible (adjacent,
885  * and with the same memory policies). That's all stable, even with just
886  * a read lock on the mm_sem.
887  */
888 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
889 {
890         if (anon_vma_compatible(a, b)) {
891                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
892
893                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
894                         return anon_vma;
895         }
896         return NULL;
897 }
898
899 /*
900  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
901  * neighbouring vmas for a suitable anon_vma, before it goes off
902  * to allocate a new anon_vma.  It checks because a repetitive
903  * sequence of mprotects and faults may otherwise lead to distinct
904  * anon_vmas being allocated, preventing vma merge in subsequent
905  * mprotect.
906  */
907 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
908 {
909         struct anon_vma *anon_vma;
910         struct vm_area_struct *near;
911
912         near = vma->vm_next;
913         if (!near)
914                 goto try_prev;
915
916         anon_vma = reusable_anon_vma(near, vma, near);
917         if (anon_vma)
918                 return anon_vma;
919 try_prev:
920         near = vma->vm_prev;
921         if (!near)
922                 goto none;
923
924         anon_vma = reusable_anon_vma(near, near, vma);
925         if (anon_vma)
926                 return anon_vma;
927 none:
928         /*
929          * There's no absolute need to look only at touching neighbours:
930          * we could search further afield for "compatible" anon_vmas.
931          * But it would probably just be a waste of time searching,
932          * or lead to too many vmas hanging off the same anon_vma.
933          * We're trying to allow mprotect remerging later on,
934          * not trying to minimize memory used for anon_vmas.
935          */
936         return NULL;
937 }
938
939 #ifdef CONFIG_PROC_FS
940 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
941                                                 struct file *file, long pages)
942 {
943         const unsigned long stack_flags
944                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
945
946         mm->total_vm += pages;
947
948         if (file) {
949                 mm->shared_vm += pages;
950                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
951                         mm->exec_vm += pages;
952         } else if (flags & stack_flags)
953                 mm->stack_vm += pages;
954         if (flags & (VM_RESERVED|VM_IO))
955                 mm->reserved_vm += pages;
956 }
957 #endif /* CONFIG_PROC_FS */
958
959 /*
960  * If a hint addr is less than mmap_min_addr change hint to be as
961  * low as possible but still greater than mmap_min_addr
962  */
963 static inline unsigned long round_hint_to_min(unsigned long hint)
964 {
965         hint &= PAGE_MASK;
966         if (((void *)hint != NULL) &&
967             (hint < mmap_min_addr))
968                 return PAGE_ALIGN(mmap_min_addr);
969         return hint;
970 }
971
972 /*
973  * The caller must hold down_write(&current->mm->mmap_sem).
974  */
975
976 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
977                         unsigned long len, unsigned long prot,
978                         unsigned long flags, unsigned long pgoff)
979 {
980         struct mm_struct * mm = current->mm;
981         struct inode *inode;
982         vm_flags_t vm_flags;
983
984         /*
985          * Does the application expect PROT_READ to imply PROT_EXEC?
986          *
987          * (the exception is when the underlying filesystem is noexec
988          *  mounted, in which case we dont add PROT_EXEC.)
989          */
990         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
991                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
992                         prot |= PROT_EXEC;
993
994         if (!len)
995                 return -EINVAL;
996
997         if (!(flags & MAP_FIXED))
998                 addr = round_hint_to_min(addr);
999
1000         /* Careful about overflows.. */
1001         len = PAGE_ALIGN(len);
1002         if (!len)
1003                 return -ENOMEM;
1004
1005         /* offset overflow? */
1006         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1007                return -EOVERFLOW;
1008
1009         /* Too many mappings? */
1010         if (mm->map_count > sysctl_max_map_count)
1011                 return -ENOMEM;
1012
1013         /* Obtain the address to map to. we verify (or select) it and ensure
1014          * that it represents a valid section of the address space.
1015          */
1016         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1017         if (addr & ~PAGE_MASK)
1018                 return addr;
1019
1020         /* Do simple checking here so the lower-level routines won't have
1021          * to. we assume access permissions have been handled by the open
1022          * of the memory object, so we don't do any here.
1023          */
1024         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1025                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1026
1027         if (flags & MAP_LOCKED)
1028                 if (!can_do_mlock())
1029                         return -EPERM;
1030
1031         /* mlock MCL_FUTURE? */
1032         if (vm_flags & VM_LOCKED) {
1033                 unsigned long locked, lock_limit;
1034                 locked = len >> PAGE_SHIFT;
1035                 locked += mm->locked_vm;
1036                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1037                 lock_limit >>= PAGE_SHIFT;
1038                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1039                         return -EAGAIN;
1040         }
1041
1042         inode = file ? file->f_path.dentry->d_inode : NULL;
1043
1044         if (file) {
1045                 switch (flags & MAP_TYPE) {
1046                 case MAP_SHARED:
1047                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1048                                 return -EACCES;
1049
1050                         /*
1051                          * Make sure we don't allow writing to an append-only
1052                          * file..
1053                          */
1054                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1055                                 return -EACCES;
1056
1057                         /*
1058                          * Make sure there are no mandatory locks on the file.
1059                          */
1060                         if (locks_verify_locked(inode))
1061                                 return -EAGAIN;
1062
1063                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1064                         if (!(file->f_mode & FMODE_WRITE))
1065                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1066
1067                         /* fall through */
1068                 case MAP_PRIVATE:
1069                         if (!(file->f_mode & FMODE_READ))
1070                                 return -EACCES;
1071                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1072                                 if (vm_flags & VM_EXEC)
1073                                         return -EPERM;
1074                                 vm_flags &= ~VM_MAYEXEC;
1075                         }
1076
1077                         if (!file->f_op || !file->f_op->mmap)
1078                                 return -ENODEV;
1079                         break;
1080
1081                 default:
1082                         return -EINVAL;
1083                 }
1084         } else {
1085                 switch (flags & MAP_TYPE) {
1086                 case MAP_SHARED:
1087                         /*
1088                          * Ignore pgoff.
1089                          */
1090                         pgoff = 0;
1091                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1092                         break;
1093                 case MAP_PRIVATE:
1094                         /*
1095                          * Set pgoff according to addr for anon_vma.
1096                          */
1097                         pgoff = addr >> PAGE_SHIFT;
1098                         break;
1099                 default:
1100                         return -EINVAL;
1101                 }
1102         }
1103
1104         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1105 }
1106
1107 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1108                 unsigned long, prot, unsigned long, flags,
1109                 unsigned long, fd, unsigned long, pgoff)
1110 {
1111         struct file *file = NULL;
1112         unsigned long retval = -EBADF;
1113
1114         if (!(flags & MAP_ANONYMOUS)) {
1115                 audit_mmap_fd(fd, flags);
1116                 if (unlikely(flags & MAP_HUGETLB))
1117                         return -EINVAL;
1118                 file = fget(fd);
1119                 if (!file)
1120                         goto out;
1121         } else if (flags & MAP_HUGETLB) {
1122                 struct user_struct *user = NULL;
1123                 /*
1124                  * VM_NORESERVE is used because the reservations will be
1125                  * taken when vm_ops->mmap() is called
1126                  * A dummy user value is used because we are not locking
1127                  * memory so no accounting is necessary
1128                  */
1129                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1130                                                 VM_NORESERVE, &user,
1131                                                 HUGETLB_ANONHUGE_INODE);
1132                 if (IS_ERR(file))
1133                         return PTR_ERR(file);
1134         }
1135
1136         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1137
1138         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1139         if (file)
1140                 fput(file);
1141 out:
1142         return retval;
1143 }
1144
1145 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1146 struct mmap_arg_struct {
1147         unsigned long addr;
1148         unsigned long len;
1149         unsigned long prot;
1150         unsigned long flags;
1151         unsigned long fd;
1152         unsigned long offset;
1153 };
1154
1155 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1156 {
1157         struct mmap_arg_struct a;
1158
1159         if (copy_from_user(&a, arg, sizeof(a)))
1160                 return -EFAULT;
1161         if (a.offset & ~PAGE_MASK)
1162                 return -EINVAL;
1163
1164         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1165                               a.offset >> PAGE_SHIFT);
1166 }
1167 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1168
1169 /*
1170  * Some shared mappigns will want the pages marked read-only
1171  * to track write events. If so, we'll downgrade vm_page_prot
1172  * to the private version (using protection_map[] without the
1173  * VM_SHARED bit).
1174  */
1175 int vma_wants_writenotify(struct vm_area_struct *vma)
1176 {
1177         vm_flags_t vm_flags = vma->vm_flags;
1178
1179         /* If it was private or non-writable, the write bit is already clear */
1180         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1181                 return 0;
1182
1183         /* The backer wishes to know when pages are first written to? */
1184         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1185                 return 1;
1186
1187         /* The open routine did something to the protections already? */
1188         if (pgprot_val(vma->vm_page_prot) !=
1189             pgprot_val(vm_get_page_prot(vm_flags)))
1190                 return 0;
1191
1192         /* Specialty mapping? */
1193         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1194                 return 0;
1195
1196         /* Can the mapping track the dirty pages? */
1197         return vma->vm_file && vma->vm_file->f_mapping &&
1198                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1199 }
1200
1201 /*
1202  * We account for memory if it's a private writeable mapping,
1203  * not hugepages and VM_NORESERVE wasn't set.
1204  */
1205 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1206 {
1207         /*
1208          * hugetlb has its own accounting separate from the core VM
1209          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1210          */
1211         if (file && is_file_hugepages(file))
1212                 return 0;
1213
1214         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1215 }
1216
1217 unsigned long mmap_region(struct file *file, unsigned long addr,
1218                           unsigned long len, unsigned long flags,
1219                           vm_flags_t vm_flags, unsigned long pgoff)
1220 {
1221         struct mm_struct *mm = current->mm;
1222         struct vm_area_struct *vma, *prev;
1223         int correct_wcount = 0;
1224         int error;
1225         struct rb_node **rb_link, *rb_parent;
1226         unsigned long charged = 0;
1227         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1228
1229         /* Clear old maps */
1230         error = -ENOMEM;
1231 munmap_back:
1232         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1233         if (vma && vma->vm_start < addr + len) {
1234                 if (do_munmap(mm, addr, len))
1235                         return -ENOMEM;
1236                 goto munmap_back;
1237         }
1238
1239         /* Check against address space limit. */
1240         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1241                 return -ENOMEM;
1242
1243         /*
1244          * Set 'VM_NORESERVE' if we should not account for the
1245          * memory use of this mapping.
1246          */
1247         if ((flags & MAP_NORESERVE)) {
1248                 /* We honor MAP_NORESERVE if allowed to overcommit */
1249                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1250                         vm_flags |= VM_NORESERVE;
1251
1252                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1253                 if (file && is_file_hugepages(file))
1254                         vm_flags |= VM_NORESERVE;
1255         }
1256
1257         /*
1258          * Private writable mapping: check memory availability
1259          */
1260         if (accountable_mapping(file, vm_flags)) {
1261                 charged = len >> PAGE_SHIFT;
1262                 if (security_vm_enough_memory_mm(mm, charged))
1263                         return -ENOMEM;
1264                 vm_flags |= VM_ACCOUNT;
1265         }
1266
1267         /*
1268          * Can we just expand an old mapping?
1269          */
1270         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1271         if (vma)
1272                 goto out;
1273
1274         /*
1275          * Determine the object being mapped and call the appropriate
1276          * specific mapper. the address has already been validated, but
1277          * not unmapped, but the maps are removed from the list.
1278          */
1279         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1280         if (!vma) {
1281                 error = -ENOMEM;
1282                 goto unacct_error;
1283         }
1284
1285         vma->vm_mm = mm;
1286         vma->vm_start = addr;
1287         vma->vm_end = addr + len;
1288         vma->vm_flags = vm_flags;
1289         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1290         vma->vm_pgoff = pgoff;
1291         INIT_LIST_HEAD(&vma->anon_vma_chain);
1292
1293         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1294
1295         if (file) {
1296                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1297                         goto free_vma;
1298                 if (vm_flags & VM_DENYWRITE) {
1299                         error = deny_write_access(file);
1300                         if (error)
1301                                 goto free_vma;
1302                         correct_wcount = 1;
1303                 }
1304                 vma->vm_file = file;
1305                 get_file(file);
1306                 error = file->f_op->mmap(file, vma);
1307                 if (error)
1308                         goto unmap_and_free_vma;
1309                 if (vm_flags & VM_EXECUTABLE)
1310                         added_exe_file_vma(mm);
1311
1312                 /* Can addr have changed??
1313                  *
1314                  * Answer: Yes, several device drivers can do it in their
1315                  *         f_op->mmap method. -DaveM
1316                  */
1317                 addr = vma->vm_start;
1318                 pgoff = vma->vm_pgoff;
1319                 vm_flags = vma->vm_flags;
1320         } else if (vm_flags & VM_SHARED) {
1321                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1322                         goto free_vma;
1323                 error = shmem_zero_setup(vma);
1324                 if (error)
1325                         goto free_vma;
1326         }
1327
1328         if (vma_wants_writenotify(vma)) {
1329                 pgprot_t pprot = vma->vm_page_prot;
1330
1331                 /* Can vma->vm_page_prot have changed??
1332                  *
1333                  * Answer: Yes, drivers may have changed it in their
1334                  *         f_op->mmap method.
1335                  *
1336                  * Ensures that vmas marked as uncached stay that way.
1337                  */
1338                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1339                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1340                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1341         }
1342
1343         vma_link(mm, vma, prev, rb_link, rb_parent);
1344         file = vma->vm_file;
1345
1346         /* Once vma denies write, undo our temporary denial count */
1347         if (correct_wcount)
1348                 atomic_inc(&inode->i_writecount);
1349 out:
1350         perf_event_mmap(vma);
1351
1352         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1353         if (vm_flags & VM_LOCKED) {
1354                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1355                         mm->locked_vm += (len >> PAGE_SHIFT);
1356         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1357                 make_pages_present(addr, addr + len);
1358
1359         if (file && uprobe_mmap(vma))
1360                 /* matching probes but cannot insert */
1361                 goto unmap_and_free_vma;
1362
1363         return addr;
1364
1365 unmap_and_free_vma:
1366         if (correct_wcount)
1367                 atomic_inc(&inode->i_writecount);
1368         vma->vm_file = NULL;
1369         fput(file);
1370
1371         /* Undo any partial mapping done by a device driver. */
1372         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1373         charged = 0;
1374 free_vma:
1375         kmem_cache_free(vm_area_cachep, vma);
1376 unacct_error:
1377         if (charged)
1378                 vm_unacct_memory(charged);
1379         return error;
1380 }
1381
1382 /* Get an address range which is currently unmapped.
1383  * For shmat() with addr=0.
1384  *
1385  * Ugly calling convention alert:
1386  * Return value with the low bits set means error value,
1387  * ie
1388  *      if (ret & ~PAGE_MASK)
1389  *              error = ret;
1390  *
1391  * This function "knows" that -ENOMEM has the bits set.
1392  */
1393 #ifndef HAVE_ARCH_UNMAPPED_AREA
1394 unsigned long
1395 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1396                 unsigned long len, unsigned long pgoff, unsigned long flags)
1397 {
1398         struct mm_struct *mm = current->mm;
1399         struct vm_area_struct *vma;
1400         unsigned long start_addr;
1401
1402         if (len > TASK_SIZE)
1403                 return -ENOMEM;
1404
1405         if (flags & MAP_FIXED)
1406                 return addr;
1407
1408         if (addr) {
1409                 addr = PAGE_ALIGN(addr);
1410                 vma = find_vma(mm, addr);
1411                 if (TASK_SIZE - len >= addr &&
1412                     (!vma || addr + len <= vma->vm_start))
1413                         return addr;
1414         }
1415         if (len > mm->cached_hole_size) {
1416                 start_addr = addr = mm->free_area_cache;
1417         } else {
1418                 start_addr = addr = TASK_UNMAPPED_BASE;
1419                 mm->cached_hole_size = 0;
1420         }
1421
1422 full_search:
1423         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1424                 /* At this point:  (!vma || addr < vma->vm_end). */
1425                 if (TASK_SIZE - len < addr) {
1426                         /*
1427                          * Start a new search - just in case we missed
1428                          * some holes.
1429                          */
1430                         if (start_addr != TASK_UNMAPPED_BASE) {
1431                                 addr = TASK_UNMAPPED_BASE;
1432                                 start_addr = addr;
1433                                 mm->cached_hole_size = 0;
1434                                 goto full_search;
1435                         }
1436                         return -ENOMEM;
1437                 }
1438                 if (!vma || addr + len <= vma->vm_start) {
1439                         /*
1440                          * Remember the place where we stopped the search:
1441                          */
1442                         mm->free_area_cache = addr + len;
1443                         return addr;
1444                 }
1445                 if (addr + mm->cached_hole_size < vma->vm_start)
1446                         mm->cached_hole_size = vma->vm_start - addr;
1447                 addr = vma->vm_end;
1448         }
1449 }
1450 #endif  
1451
1452 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1453 {
1454         /*
1455          * Is this a new hole at the lowest possible address?
1456          */
1457         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1458                 mm->free_area_cache = addr;
1459 }
1460
1461 /*
1462  * This mmap-allocator allocates new areas top-down from below the
1463  * stack's low limit (the base):
1464  */
1465 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1466 unsigned long
1467 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1468                           const unsigned long len, const unsigned long pgoff,
1469                           const unsigned long flags)
1470 {
1471         struct vm_area_struct *vma;
1472         struct mm_struct *mm = current->mm;
1473         unsigned long addr = addr0, start_addr;
1474
1475         /* requested length too big for entire address space */
1476         if (len > TASK_SIZE)
1477                 return -ENOMEM;
1478
1479         if (flags & MAP_FIXED)
1480                 return addr;
1481
1482         /* requesting a specific address */
1483         if (addr) {
1484                 addr = PAGE_ALIGN(addr);
1485                 vma = find_vma(mm, addr);
1486                 if (TASK_SIZE - len >= addr &&
1487                                 (!vma || addr + len <= vma->vm_start))
1488                         return addr;
1489         }
1490
1491         /* check if free_area_cache is useful for us */
1492         if (len <= mm->cached_hole_size) {
1493                 mm->cached_hole_size = 0;
1494                 mm->free_area_cache = mm->mmap_base;
1495         }
1496
1497 try_again:
1498         /* either no address requested or can't fit in requested address hole */
1499         start_addr = addr = mm->free_area_cache;
1500
1501         if (addr < len)
1502                 goto fail;
1503
1504         addr -= len;
1505         do {
1506                 /*
1507                  * Lookup failure means no vma is above this address,
1508                  * else if new region fits below vma->vm_start,
1509                  * return with success:
1510                  */
1511                 vma = find_vma(mm, addr);
1512                 if (!vma || addr+len <= vma->vm_start)
1513                         /* remember the address as a hint for next time */
1514                         return (mm->free_area_cache = addr);
1515
1516                 /* remember the largest hole we saw so far */
1517                 if (addr + mm->cached_hole_size < vma->vm_start)
1518                         mm->cached_hole_size = vma->vm_start - addr;
1519
1520                 /* try just below the current vma->vm_start */
1521                 addr = vma->vm_start-len;
1522         } while (len < vma->vm_start);
1523
1524 fail:
1525         /*
1526          * if hint left us with no space for the requested
1527          * mapping then try again:
1528          *
1529          * Note: this is different with the case of bottomup
1530          * which does the fully line-search, but we use find_vma
1531          * here that causes some holes skipped.
1532          */
1533         if (start_addr != mm->mmap_base) {
1534                 mm->free_area_cache = mm->mmap_base;
1535                 mm->cached_hole_size = 0;
1536                 goto try_again;
1537         }
1538
1539         /*
1540          * A failed mmap() very likely causes application failure,
1541          * so fall back to the bottom-up function here. This scenario
1542          * can happen with large stack limits and large mmap()
1543          * allocations.
1544          */
1545         mm->cached_hole_size = ~0UL;
1546         mm->free_area_cache = TASK_UNMAPPED_BASE;
1547         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1548         /*
1549          * Restore the topdown base:
1550          */
1551         mm->free_area_cache = mm->mmap_base;
1552         mm->cached_hole_size = ~0UL;
1553
1554         return addr;
1555 }
1556 #endif
1557
1558 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1559 {
1560         /*
1561          * Is this a new hole at the highest possible address?
1562          */
1563         if (addr > mm->free_area_cache)
1564                 mm->free_area_cache = addr;
1565
1566         /* dont allow allocations above current base */
1567         if (mm->free_area_cache > mm->mmap_base)
1568                 mm->free_area_cache = mm->mmap_base;
1569 }
1570
1571 unsigned long
1572 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1573                 unsigned long pgoff, unsigned long flags)
1574 {
1575         unsigned long (*get_area)(struct file *, unsigned long,
1576                                   unsigned long, unsigned long, unsigned long);
1577
1578         unsigned long error = arch_mmap_check(addr, len, flags);
1579         if (error)
1580                 return error;
1581
1582         /* Careful about overflows.. */
1583         if (len > TASK_SIZE)
1584                 return -ENOMEM;
1585
1586         get_area = current->mm->get_unmapped_area;
1587         if (file && file->f_op && file->f_op->get_unmapped_area)
1588                 get_area = file->f_op->get_unmapped_area;
1589         addr = get_area(file, addr, len, pgoff, flags);
1590         if (IS_ERR_VALUE(addr))
1591                 return addr;
1592
1593         if (addr > TASK_SIZE - len)
1594                 return -ENOMEM;
1595         if (addr & ~PAGE_MASK)
1596                 return -EINVAL;
1597
1598         addr = arch_rebalance_pgtables(addr, len);
1599         error = security_mmap_addr(addr);
1600         return error ? error : addr;
1601 }
1602
1603 EXPORT_SYMBOL(get_unmapped_area);
1604
1605 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1606 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1607 {
1608         struct vm_area_struct *vma = NULL;
1609
1610         if (WARN_ON_ONCE(!mm))          /* Remove this in linux-3.6 */
1611                 return NULL;
1612
1613         /* Check the cache first. */
1614         /* (Cache hit rate is typically around 35%.) */
1615         vma = mm->mmap_cache;
1616         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1617                 struct rb_node *rb_node;
1618
1619                 rb_node = mm->mm_rb.rb_node;
1620                 vma = NULL;
1621
1622                 while (rb_node) {
1623                         struct vm_area_struct *vma_tmp;
1624
1625                         vma_tmp = rb_entry(rb_node,
1626                                            struct vm_area_struct, vm_rb);
1627
1628                         if (vma_tmp->vm_end > addr) {
1629                                 vma = vma_tmp;
1630                                 if (vma_tmp->vm_start <= addr)
1631                                         break;
1632                                 rb_node = rb_node->rb_left;
1633                         } else
1634                                 rb_node = rb_node->rb_right;
1635                 }
1636                 if (vma)
1637                         mm->mmap_cache = vma;
1638         }
1639         return vma;
1640 }
1641
1642 EXPORT_SYMBOL(find_vma);
1643
1644 /*
1645  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1646  */
1647 struct vm_area_struct *
1648 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1649                         struct vm_area_struct **pprev)
1650 {
1651         struct vm_area_struct *vma;
1652
1653         vma = find_vma(mm, addr);
1654         if (vma) {
1655                 *pprev = vma->vm_prev;
1656         } else {
1657                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1658                 *pprev = NULL;
1659                 while (rb_node) {
1660                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1661                         rb_node = rb_node->rb_right;
1662                 }
1663         }
1664         return vma;
1665 }
1666
1667 /*
1668  * Verify that the stack growth is acceptable and
1669  * update accounting. This is shared with both the
1670  * grow-up and grow-down cases.
1671  */
1672 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1673 {
1674         struct mm_struct *mm = vma->vm_mm;
1675         struct rlimit *rlim = current->signal->rlim;
1676         unsigned long new_start;
1677
1678         /* address space limit tests */
1679         if (!may_expand_vm(mm, grow))
1680                 return -ENOMEM;
1681
1682         /* Stack limit test */
1683         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1684                 return -ENOMEM;
1685
1686         /* mlock limit tests */
1687         if (vma->vm_flags & VM_LOCKED) {
1688                 unsigned long locked;
1689                 unsigned long limit;
1690                 locked = mm->locked_vm + grow;
1691                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1692                 limit >>= PAGE_SHIFT;
1693                 if (locked > limit && !capable(CAP_IPC_LOCK))
1694                         return -ENOMEM;
1695         }
1696
1697         /* Check to ensure the stack will not grow into a hugetlb-only region */
1698         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1699                         vma->vm_end - size;
1700         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1701                 return -EFAULT;
1702
1703         /*
1704          * Overcommit..  This must be the final test, as it will
1705          * update security statistics.
1706          */
1707         if (security_vm_enough_memory_mm(mm, grow))
1708                 return -ENOMEM;
1709
1710         /* Ok, everything looks good - let it rip */
1711         if (vma->vm_flags & VM_LOCKED)
1712                 mm->locked_vm += grow;
1713         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1714         return 0;
1715 }
1716
1717 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1718 /*
1719  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1720  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1721  */
1722 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1723 {
1724         int error;
1725
1726         if (!(vma->vm_flags & VM_GROWSUP))
1727                 return -EFAULT;
1728
1729         /*
1730          * We must make sure the anon_vma is allocated
1731          * so that the anon_vma locking is not a noop.
1732          */
1733         if (unlikely(anon_vma_prepare(vma)))
1734                 return -ENOMEM;
1735         vma_lock_anon_vma(vma);
1736
1737         /*
1738          * vma->vm_start/vm_end cannot change under us because the caller
1739          * is required to hold the mmap_sem in read mode.  We need the
1740          * anon_vma lock to serialize against concurrent expand_stacks.
1741          * Also guard against wrapping around to address 0.
1742          */
1743         if (address < PAGE_ALIGN(address+4))
1744                 address = PAGE_ALIGN(address+4);
1745         else {
1746                 vma_unlock_anon_vma(vma);
1747                 return -ENOMEM;
1748         }
1749         error = 0;
1750
1751         /* Somebody else might have raced and expanded it already */
1752         if (address > vma->vm_end) {
1753                 unsigned long size, grow;
1754
1755                 size = address - vma->vm_start;
1756                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1757
1758                 error = -ENOMEM;
1759                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1760                         error = acct_stack_growth(vma, size, grow);
1761                         if (!error) {
1762                                 vma->vm_end = address;
1763                                 perf_event_mmap(vma);
1764                         }
1765                 }
1766         }
1767         vma_unlock_anon_vma(vma);
1768         khugepaged_enter_vma_merge(vma);
1769         return error;
1770 }
1771 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1772
1773 /*
1774  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1775  */
1776 int expand_downwards(struct vm_area_struct *vma,
1777                                    unsigned long address)
1778 {
1779         int error;
1780
1781         /*
1782          * We must make sure the anon_vma is allocated
1783          * so that the anon_vma locking is not a noop.
1784          */
1785         if (unlikely(anon_vma_prepare(vma)))
1786                 return -ENOMEM;
1787
1788         address &= PAGE_MASK;
1789         error = security_mmap_addr(address);
1790         if (error)
1791                 return error;
1792
1793         vma_lock_anon_vma(vma);
1794
1795         /*
1796          * vma->vm_start/vm_end cannot change under us because the caller
1797          * is required to hold the mmap_sem in read mode.  We need the
1798          * anon_vma lock to serialize against concurrent expand_stacks.
1799          */
1800
1801         /* Somebody else might have raced and expanded it already */
1802         if (address < vma->vm_start) {
1803                 unsigned long size, grow;
1804
1805                 size = vma->vm_end - address;
1806                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1807
1808                 error = -ENOMEM;
1809                 if (grow <= vma->vm_pgoff) {
1810                         error = acct_stack_growth(vma, size, grow);
1811                         if (!error) {
1812                                 vma->vm_start = address;
1813                                 vma->vm_pgoff -= grow;
1814                                 perf_event_mmap(vma);
1815                         }
1816                 }
1817         }
1818         vma_unlock_anon_vma(vma);
1819         khugepaged_enter_vma_merge(vma);
1820         return error;
1821 }
1822
1823 #ifdef CONFIG_STACK_GROWSUP
1824 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1825 {
1826         return expand_upwards(vma, address);
1827 }
1828
1829 struct vm_area_struct *
1830 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1831 {
1832         struct vm_area_struct *vma, *prev;
1833
1834         addr &= PAGE_MASK;
1835         vma = find_vma_prev(mm, addr, &prev);
1836         if (vma && (vma->vm_start <= addr))
1837                 return vma;
1838         if (!prev || expand_stack(prev, addr))
1839                 return NULL;
1840         if (prev->vm_flags & VM_LOCKED) {
1841                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1842         }
1843         return prev;
1844 }
1845 #else
1846 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1847 {
1848         return expand_downwards(vma, address);
1849 }
1850
1851 struct vm_area_struct *
1852 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1853 {
1854         struct vm_area_struct * vma;
1855         unsigned long start;
1856
1857         addr &= PAGE_MASK;
1858         vma = find_vma(mm,addr);
1859         if (!vma)
1860                 return NULL;
1861         if (vma->vm_start <= addr)
1862                 return vma;
1863         if (!(vma->vm_flags & VM_GROWSDOWN))
1864                 return NULL;
1865         start = vma->vm_start;
1866         if (expand_stack(vma, addr))
1867                 return NULL;
1868         if (vma->vm_flags & VM_LOCKED) {
1869                 mlock_vma_pages_range(vma, addr, start);
1870         }
1871         return vma;
1872 }
1873 #endif
1874
1875 /*
1876  * Ok - we have the memory areas we should free on the vma list,
1877  * so release them, and do the vma updates.
1878  *
1879  * Called with the mm semaphore held.
1880  */
1881 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1882 {
1883         unsigned long nr_accounted = 0;
1884
1885         /* Update high watermark before we lower total_vm */
1886         update_hiwater_vm(mm);
1887         do {
1888                 long nrpages = vma_pages(vma);
1889
1890                 if (vma->vm_flags & VM_ACCOUNT)
1891                         nr_accounted += nrpages;
1892                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1893                 vma = remove_vma(vma);
1894         } while (vma);
1895         vm_unacct_memory(nr_accounted);
1896         validate_mm(mm);
1897 }
1898
1899 /*
1900  * Get rid of page table information in the indicated region.
1901  *
1902  * Called with the mm semaphore held.
1903  */
1904 static void unmap_region(struct mm_struct *mm,
1905                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1906                 unsigned long start, unsigned long end)
1907 {
1908         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1909         struct mmu_gather tlb;
1910
1911         lru_add_drain();
1912         tlb_gather_mmu(&tlb, mm, 0);
1913         update_hiwater_rss(mm);
1914         unmap_vmas(&tlb, vma, start, end);
1915         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1916                                  next ? next->vm_start : 0);
1917         tlb_finish_mmu(&tlb, start, end);
1918 }
1919
1920 /*
1921  * Create a list of vma's touched by the unmap, removing them from the mm's
1922  * vma list as we go..
1923  */
1924 static void
1925 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1926         struct vm_area_struct *prev, unsigned long end)
1927 {
1928         struct vm_area_struct **insertion_point;
1929         struct vm_area_struct *tail_vma = NULL;
1930         unsigned long addr;
1931
1932         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1933         vma->vm_prev = NULL;
1934         do {
1935                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1936                 mm->map_count--;
1937                 tail_vma = vma;
1938                 vma = vma->vm_next;
1939         } while (vma && vma->vm_start < end);
1940         *insertion_point = vma;
1941         if (vma)
1942                 vma->vm_prev = prev;
1943         tail_vma->vm_next = NULL;
1944         if (mm->unmap_area == arch_unmap_area)
1945                 addr = prev ? prev->vm_end : mm->mmap_base;
1946         else
1947                 addr = vma ?  vma->vm_start : mm->mmap_base;
1948         mm->unmap_area(mm, addr);
1949         mm->mmap_cache = NULL;          /* Kill the cache. */
1950 }
1951
1952 /*
1953  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1954  * munmap path where it doesn't make sense to fail.
1955  */
1956 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1957               unsigned long addr, int new_below)
1958 {
1959         struct mempolicy *pol;
1960         struct vm_area_struct *new;
1961         int err = -ENOMEM;
1962
1963         if (is_vm_hugetlb_page(vma) && (addr &
1964                                         ~(huge_page_mask(hstate_vma(vma)))))
1965                 return -EINVAL;
1966
1967         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1968         if (!new)
1969                 goto out_err;
1970
1971         /* most fields are the same, copy all, and then fixup */
1972         *new = *vma;
1973
1974         INIT_LIST_HEAD(&new->anon_vma_chain);
1975
1976         if (new_below)
1977                 new->vm_end = addr;
1978         else {
1979                 new->vm_start = addr;
1980                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1981         }
1982
1983         pol = mpol_dup(vma_policy(vma));
1984         if (IS_ERR(pol)) {
1985                 err = PTR_ERR(pol);
1986                 goto out_free_vma;
1987         }
1988         vma_set_policy(new, pol);
1989
1990         if (anon_vma_clone(new, vma))
1991                 goto out_free_mpol;
1992
1993         if (new->vm_file) {
1994                 get_file(new->vm_file);
1995                 if (vma->vm_flags & VM_EXECUTABLE)
1996                         added_exe_file_vma(mm);
1997         }
1998
1999         if (new->vm_ops && new->vm_ops->open)
2000                 new->vm_ops->open(new);
2001
2002         if (new_below)
2003                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2004                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2005         else
2006                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2007
2008         /* Success. */
2009         if (!err)
2010                 return 0;
2011
2012         /* Clean everything up if vma_adjust failed. */
2013         if (new->vm_ops && new->vm_ops->close)
2014                 new->vm_ops->close(new);
2015         if (new->vm_file) {
2016                 if (vma->vm_flags & VM_EXECUTABLE)
2017                         removed_exe_file_vma(mm);
2018                 fput(new->vm_file);
2019         }
2020         unlink_anon_vmas(new);
2021  out_free_mpol:
2022         mpol_put(pol);
2023  out_free_vma:
2024         kmem_cache_free(vm_area_cachep, new);
2025  out_err:
2026         return err;
2027 }
2028
2029 /*
2030  * Split a vma into two pieces at address 'addr', a new vma is allocated
2031  * either for the first part or the tail.
2032  */
2033 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2034               unsigned long addr, int new_below)
2035 {
2036         if (mm->map_count >= sysctl_max_map_count)
2037                 return -ENOMEM;
2038
2039         return __split_vma(mm, vma, addr, new_below);
2040 }
2041
2042 /* Munmap is split into 2 main parts -- this part which finds
2043  * what needs doing, and the areas themselves, which do the
2044  * work.  This now handles partial unmappings.
2045  * Jeremy Fitzhardinge <jeremy@goop.org>
2046  */
2047 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2048 {
2049         unsigned long end;
2050         struct vm_area_struct *vma, *prev, *last;
2051
2052         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2053                 return -EINVAL;
2054
2055         if ((len = PAGE_ALIGN(len)) == 0)
2056                 return -EINVAL;
2057
2058         /* Find the first overlapping VMA */
2059         vma = find_vma(mm, start);
2060         if (!vma)
2061                 return 0;
2062         prev = vma->vm_prev;
2063         /* we have  start < vma->vm_end  */
2064
2065         /* if it doesn't overlap, we have nothing.. */
2066         end = start + len;
2067         if (vma->vm_start >= end)
2068                 return 0;
2069
2070         /*
2071          * If we need to split any vma, do it now to save pain later.
2072          *
2073          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2074          * unmapped vm_area_struct will remain in use: so lower split_vma
2075          * places tmp vma above, and higher split_vma places tmp vma below.
2076          */
2077         if (start > vma->vm_start) {
2078                 int error;
2079
2080                 /*
2081                  * Make sure that map_count on return from munmap() will
2082                  * not exceed its limit; but let map_count go just above
2083                  * its limit temporarily, to help free resources as expected.
2084                  */
2085                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2086                         return -ENOMEM;
2087
2088                 error = __split_vma(mm, vma, start, 0);
2089                 if (error)
2090                         return error;
2091                 prev = vma;
2092         }
2093
2094         /* Does it split the last one? */
2095         last = find_vma(mm, end);
2096         if (last && end > last->vm_start) {
2097                 int error = __split_vma(mm, last, end, 1);
2098                 if (error)
2099                         return error;
2100         }
2101         vma = prev? prev->vm_next: mm->mmap;
2102
2103         /*
2104          * unlock any mlock()ed ranges before detaching vmas
2105          */
2106         if (mm->locked_vm) {
2107                 struct vm_area_struct *tmp = vma;
2108                 while (tmp && tmp->vm_start < end) {
2109                         if (tmp->vm_flags & VM_LOCKED) {
2110                                 mm->locked_vm -= vma_pages(tmp);
2111                                 munlock_vma_pages_all(tmp);
2112                         }
2113                         tmp = tmp->vm_next;
2114                 }
2115         }
2116
2117         /*
2118          * Remove the vma's, and unmap the actual pages
2119          */
2120         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2121         unmap_region(mm, vma, prev, start, end);
2122
2123         /* Fix up all other VM information */
2124         remove_vma_list(mm, vma);
2125
2126         return 0;
2127 }
2128
2129 int vm_munmap(unsigned long start, size_t len)
2130 {
2131         int ret;
2132         struct mm_struct *mm = current->mm;
2133
2134         down_write(&mm->mmap_sem);
2135         ret = do_munmap(mm, start, len);
2136         up_write(&mm->mmap_sem);
2137         return ret;
2138 }
2139 EXPORT_SYMBOL(vm_munmap);
2140
2141 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2142 {
2143         profile_munmap(addr);
2144         return vm_munmap(addr, len);
2145 }
2146
2147 static inline void verify_mm_writelocked(struct mm_struct *mm)
2148 {
2149 #ifdef CONFIG_DEBUG_VM
2150         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2151                 WARN_ON(1);
2152                 up_read(&mm->mmap_sem);
2153         }
2154 #endif
2155 }
2156
2157 /*
2158  *  this is really a simplified "do_mmap".  it only handles
2159  *  anonymous maps.  eventually we may be able to do some
2160  *  brk-specific accounting here.
2161  */
2162 static unsigned long do_brk(unsigned long addr, unsigned long len)
2163 {
2164         struct mm_struct * mm = current->mm;
2165         struct vm_area_struct * vma, * prev;
2166         unsigned long flags;
2167         struct rb_node ** rb_link, * rb_parent;
2168         pgoff_t pgoff = addr >> PAGE_SHIFT;
2169         int error;
2170
2171         len = PAGE_ALIGN(len);
2172         if (!len)
2173                 return addr;
2174
2175         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2176
2177         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2178         if (error & ~PAGE_MASK)
2179                 return error;
2180
2181         /*
2182          * mlock MCL_FUTURE?
2183          */
2184         if (mm->def_flags & VM_LOCKED) {
2185                 unsigned long locked, lock_limit;
2186                 locked = len >> PAGE_SHIFT;
2187                 locked += mm->locked_vm;
2188                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2189                 lock_limit >>= PAGE_SHIFT;
2190                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2191                         return -EAGAIN;
2192         }
2193
2194         /*
2195          * mm->mmap_sem is required to protect against another thread
2196          * changing the mappings in case we sleep.
2197          */
2198         verify_mm_writelocked(mm);
2199
2200         /*
2201          * Clear old maps.  this also does some error checking for us
2202          */
2203  munmap_back:
2204         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2205         if (vma && vma->vm_start < addr + len) {
2206                 if (do_munmap(mm, addr, len))
2207                         return -ENOMEM;
2208                 goto munmap_back;
2209         }
2210
2211         /* Check against address space limits *after* clearing old maps... */
2212         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2213                 return -ENOMEM;
2214
2215         if (mm->map_count > sysctl_max_map_count)
2216                 return -ENOMEM;
2217
2218         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2219                 return -ENOMEM;
2220
2221         /* Can we just expand an old private anonymous mapping? */
2222         vma = vma_merge(mm, prev, addr, addr + len, flags,
2223                                         NULL, NULL, pgoff, NULL);
2224         if (vma)
2225                 goto out;
2226
2227         /*
2228          * create a vma struct for an anonymous mapping
2229          */
2230         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2231         if (!vma) {
2232                 vm_unacct_memory(len >> PAGE_SHIFT);
2233                 return -ENOMEM;
2234         }
2235
2236         INIT_LIST_HEAD(&vma->anon_vma_chain);
2237         vma->vm_mm = mm;
2238         vma->vm_start = addr;
2239         vma->vm_end = addr + len;
2240         vma->vm_pgoff = pgoff;
2241         vma->vm_flags = flags;
2242         vma->vm_page_prot = vm_get_page_prot(flags);
2243         vma_link(mm, vma, prev, rb_link, rb_parent);
2244 out:
2245         perf_event_mmap(vma);
2246         mm->total_vm += len >> PAGE_SHIFT;
2247         if (flags & VM_LOCKED) {
2248                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2249                         mm->locked_vm += (len >> PAGE_SHIFT);
2250         }
2251         return addr;
2252 }
2253
2254 unsigned long vm_brk(unsigned long addr, unsigned long len)
2255 {
2256         struct mm_struct *mm = current->mm;
2257         unsigned long ret;
2258
2259         down_write(&mm->mmap_sem);
2260         ret = do_brk(addr, len);
2261         up_write(&mm->mmap_sem);
2262         return ret;
2263 }
2264 EXPORT_SYMBOL(vm_brk);
2265
2266 /* Release all mmaps. */
2267 void exit_mmap(struct mm_struct *mm)
2268 {
2269         struct mmu_gather tlb;
2270         struct vm_area_struct *vma;
2271         unsigned long nr_accounted = 0;
2272
2273         /* mm's last user has gone, and its about to be pulled down */
2274         mmu_notifier_release(mm);
2275
2276         if (mm->locked_vm) {
2277                 vma = mm->mmap;
2278                 while (vma) {
2279                         if (vma->vm_flags & VM_LOCKED)
2280                                 munlock_vma_pages_all(vma);
2281                         vma = vma->vm_next;
2282                 }
2283         }
2284
2285         arch_exit_mmap(mm);
2286
2287         vma = mm->mmap;
2288         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2289                 return;
2290
2291         lru_add_drain();
2292         flush_cache_mm(mm);
2293         tlb_gather_mmu(&tlb, mm, 1);
2294         /* update_hiwater_rss(mm) here? but nobody should be looking */
2295         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2296         unmap_vmas(&tlb, vma, 0, -1);
2297
2298         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2299         tlb_finish_mmu(&tlb, 0, -1);
2300
2301         /*
2302          * Walk the list again, actually closing and freeing it,
2303          * with preemption enabled, without holding any MM locks.
2304          */
2305         while (vma) {
2306                 if (vma->vm_flags & VM_ACCOUNT)
2307                         nr_accounted += vma_pages(vma);
2308                 vma = remove_vma(vma);
2309         }
2310         vm_unacct_memory(nr_accounted);
2311
2312         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2313 }
2314
2315 /* Insert vm structure into process list sorted by address
2316  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2317  * then i_mmap_mutex is taken here.
2318  */
2319 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2320 {
2321         struct vm_area_struct * __vma, * prev;
2322         struct rb_node ** rb_link, * rb_parent;
2323
2324         /*
2325          * The vm_pgoff of a purely anonymous vma should be irrelevant
2326          * until its first write fault, when page's anon_vma and index
2327          * are set.  But now set the vm_pgoff it will almost certainly
2328          * end up with (unless mremap moves it elsewhere before that
2329          * first wfault), so /proc/pid/maps tells a consistent story.
2330          *
2331          * By setting it to reflect the virtual start address of the
2332          * vma, merges and splits can happen in a seamless way, just
2333          * using the existing file pgoff checks and manipulations.
2334          * Similarly in do_mmap_pgoff and in do_brk.
2335          */
2336         if (!vma->vm_file) {
2337                 BUG_ON(vma->anon_vma);
2338                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2339         }
2340         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2341         if (__vma && __vma->vm_start < vma->vm_end)
2342                 return -ENOMEM;
2343         if ((vma->vm_flags & VM_ACCOUNT) &&
2344              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2345                 return -ENOMEM;
2346
2347         vma_link(mm, vma, prev, rb_link, rb_parent);
2348         return 0;
2349 }
2350
2351 /*
2352  * Copy the vma structure to a new location in the same mm,
2353  * prior to moving page table entries, to effect an mremap move.
2354  */
2355 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2356         unsigned long addr, unsigned long len, pgoff_t pgoff)
2357 {
2358         struct vm_area_struct *vma = *vmap;
2359         unsigned long vma_start = vma->vm_start;
2360         struct mm_struct *mm = vma->vm_mm;
2361         struct vm_area_struct *new_vma, *prev;
2362         struct rb_node **rb_link, *rb_parent;
2363         struct mempolicy *pol;
2364         bool faulted_in_anon_vma = true;
2365
2366         /*
2367          * If anonymous vma has not yet been faulted, update new pgoff
2368          * to match new location, to increase its chance of merging.
2369          */
2370         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2371                 pgoff = addr >> PAGE_SHIFT;
2372                 faulted_in_anon_vma = false;
2373         }
2374
2375         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2376         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2377                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2378         if (new_vma) {
2379                 /*
2380                  * Source vma may have been merged into new_vma
2381                  */
2382                 if (unlikely(vma_start >= new_vma->vm_start &&
2383                              vma_start < new_vma->vm_end)) {
2384                         /*
2385                          * The only way we can get a vma_merge with
2386                          * self during an mremap is if the vma hasn't
2387                          * been faulted in yet and we were allowed to
2388                          * reset the dst vma->vm_pgoff to the
2389                          * destination address of the mremap to allow
2390                          * the merge to happen. mremap must change the
2391                          * vm_pgoff linearity between src and dst vmas
2392                          * (in turn preventing a vma_merge) to be
2393                          * safe. It is only safe to keep the vm_pgoff
2394                          * linear if there are no pages mapped yet.
2395                          */
2396                         VM_BUG_ON(faulted_in_anon_vma);
2397                         *vmap = new_vma;
2398                 } else
2399                         anon_vma_moveto_tail(new_vma);
2400         } else {
2401                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2402                 if (new_vma) {
2403                         *new_vma = *vma;
2404                         pol = mpol_dup(vma_policy(vma));
2405                         if (IS_ERR(pol))
2406                                 goto out_free_vma;
2407                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2408                         if (anon_vma_clone(new_vma, vma))
2409                                 goto out_free_mempol;
2410                         vma_set_policy(new_vma, pol);
2411                         new_vma->vm_start = addr;
2412                         new_vma->vm_end = addr + len;
2413                         new_vma->vm_pgoff = pgoff;
2414                         if (new_vma->vm_file) {
2415                                 get_file(new_vma->vm_file);
2416
2417                                 if (vma->vm_flags & VM_EXECUTABLE)
2418                                         added_exe_file_vma(mm);
2419                         }
2420                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2421                                 new_vma->vm_ops->open(new_vma);
2422                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2423                 }
2424         }
2425         return new_vma;
2426
2427  out_free_mempol:
2428         mpol_put(pol);
2429  out_free_vma:
2430         kmem_cache_free(vm_area_cachep, new_vma);
2431         return NULL;
2432 }
2433
2434 /*
2435  * Return true if the calling process may expand its vm space by the passed
2436  * number of pages
2437  */
2438 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2439 {
2440         unsigned long cur = mm->total_vm;       /* pages */
2441         unsigned long lim;
2442
2443         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2444
2445         if (cur + npages > lim)
2446                 return 0;
2447         return 1;
2448 }
2449
2450
2451 static int special_mapping_fault(struct vm_area_struct *vma,
2452                                 struct vm_fault *vmf)
2453 {
2454         pgoff_t pgoff;
2455         struct page **pages;
2456
2457         /*
2458          * special mappings have no vm_file, and in that case, the mm
2459          * uses vm_pgoff internally. So we have to subtract it from here.
2460          * We are allowed to do this because we are the mm; do not copy
2461          * this code into drivers!
2462          */
2463         pgoff = vmf->pgoff - vma->vm_pgoff;
2464
2465         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2466                 pgoff--;
2467
2468         if (*pages) {
2469                 struct page *page = *pages;
2470                 get_page(page);
2471                 vmf->page = page;
2472                 return 0;
2473         }
2474
2475         return VM_FAULT_SIGBUS;
2476 }
2477
2478 /*
2479  * Having a close hook prevents vma merging regardless of flags.
2480  */
2481 static void special_mapping_close(struct vm_area_struct *vma)
2482 {
2483 }
2484
2485 static const struct vm_operations_struct special_mapping_vmops = {
2486         .close = special_mapping_close,
2487         .fault = special_mapping_fault,
2488 };
2489
2490 /*
2491  * Called with mm->mmap_sem held for writing.
2492  * Insert a new vma covering the given region, with the given flags.
2493  * Its pages are supplied by the given array of struct page *.
2494  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2495  * The region past the last page supplied will always produce SIGBUS.
2496  * The array pointer and the pages it points to are assumed to stay alive
2497  * for as long as this mapping might exist.
2498  */
2499 int install_special_mapping(struct mm_struct *mm,
2500                             unsigned long addr, unsigned long len,
2501                             unsigned long vm_flags, struct page **pages)
2502 {
2503         int ret;
2504         struct vm_area_struct *vma;
2505
2506         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2507         if (unlikely(vma == NULL))
2508                 return -ENOMEM;
2509
2510         INIT_LIST_HEAD(&vma->anon_vma_chain);
2511         vma->vm_mm = mm;
2512         vma->vm_start = addr;
2513         vma->vm_end = addr + len;
2514
2515         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2516         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2517
2518         vma->vm_ops = &special_mapping_vmops;
2519         vma->vm_private_data = pages;
2520
2521         ret = insert_vm_struct(mm, vma);
2522         if (ret)
2523                 goto out;
2524
2525         mm->total_vm += len >> PAGE_SHIFT;
2526
2527         perf_event_mmap(vma);
2528
2529         return 0;
2530
2531 out:
2532         kmem_cache_free(vm_area_cachep, vma);
2533         return ret;
2534 }
2535
2536 static DEFINE_MUTEX(mm_all_locks_mutex);
2537
2538 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2539 {
2540         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2541                 /*
2542                  * The LSB of head.next can't change from under us
2543                  * because we hold the mm_all_locks_mutex.
2544                  */
2545                 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2546                 /*
2547                  * We can safely modify head.next after taking the
2548                  * anon_vma->root->mutex. If some other vma in this mm shares
2549                  * the same anon_vma we won't take it again.
2550                  *
2551                  * No need of atomic instructions here, head.next
2552                  * can't change from under us thanks to the
2553                  * anon_vma->root->mutex.
2554                  */
2555                 if (__test_and_set_bit(0, (unsigned long *)
2556                                        &anon_vma->root->head.next))
2557                         BUG();
2558         }
2559 }
2560
2561 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2562 {
2563         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2564                 /*
2565                  * AS_MM_ALL_LOCKS can't change from under us because
2566                  * we hold the mm_all_locks_mutex.
2567                  *
2568                  * Operations on ->flags have to be atomic because
2569                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2570                  * mm_all_locks_mutex, there may be other cpus
2571                  * changing other bitflags in parallel to us.
2572                  */
2573                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2574                         BUG();
2575                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2576         }
2577 }
2578
2579 /*
2580  * This operation locks against the VM for all pte/vma/mm related
2581  * operations that could ever happen on a certain mm. This includes
2582  * vmtruncate, try_to_unmap, and all page faults.
2583  *
2584  * The caller must take the mmap_sem in write mode before calling
2585  * mm_take_all_locks(). The caller isn't allowed to release the
2586  * mmap_sem until mm_drop_all_locks() returns.
2587  *
2588  * mmap_sem in write mode is required in order to block all operations
2589  * that could modify pagetables and free pages without need of
2590  * altering the vma layout (for example populate_range() with
2591  * nonlinear vmas). It's also needed in write mode to avoid new
2592  * anon_vmas to be associated with existing vmas.
2593  *
2594  * A single task can't take more than one mm_take_all_locks() in a row
2595  * or it would deadlock.
2596  *
2597  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2598  * mapping->flags avoid to take the same lock twice, if more than one
2599  * vma in this mm is backed by the same anon_vma or address_space.
2600  *
2601  * We can take all the locks in random order because the VM code
2602  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2603  * takes more than one of them in a row. Secondly we're protected
2604  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2605  *
2606  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2607  * that may have to take thousand of locks.
2608  *
2609  * mm_take_all_locks() can fail if it's interrupted by signals.
2610  */
2611 int mm_take_all_locks(struct mm_struct *mm)
2612 {
2613         struct vm_area_struct *vma;
2614         struct anon_vma_chain *avc;
2615
2616         BUG_ON(down_read_trylock(&mm->mmap_sem));
2617
2618         mutex_lock(&mm_all_locks_mutex);
2619
2620         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2621                 if (signal_pending(current))
2622                         goto out_unlock;
2623                 if (vma->vm_file && vma->vm_file->f_mapping)
2624                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2625         }
2626
2627         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2628                 if (signal_pending(current))
2629                         goto out_unlock;
2630                 if (vma->anon_vma)
2631                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2632                                 vm_lock_anon_vma(mm, avc->anon_vma);
2633         }
2634
2635         return 0;
2636
2637 out_unlock:
2638         mm_drop_all_locks(mm);
2639         return -EINTR;
2640 }
2641
2642 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2643 {
2644         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2645                 /*
2646                  * The LSB of head.next can't change to 0 from under
2647                  * us because we hold the mm_all_locks_mutex.
2648                  *
2649                  * We must however clear the bitflag before unlocking
2650                  * the vma so the users using the anon_vma->head will
2651                  * never see our bitflag.
2652                  *
2653                  * No need of atomic instructions here, head.next
2654                  * can't change from under us until we release the
2655                  * anon_vma->root->mutex.
2656                  */
2657                 if (!__test_and_clear_bit(0, (unsigned long *)
2658                                           &anon_vma->root->head.next))
2659                         BUG();
2660                 anon_vma_unlock(anon_vma);
2661         }
2662 }
2663
2664 static void vm_unlock_mapping(struct address_space *mapping)
2665 {
2666         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2667                 /*
2668                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2669                  * because we hold the mm_all_locks_mutex.
2670                  */
2671                 mutex_unlock(&mapping->i_mmap_mutex);
2672                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2673                                         &mapping->flags))
2674                         BUG();
2675         }
2676 }
2677
2678 /*
2679  * The mmap_sem cannot be released by the caller until
2680  * mm_drop_all_locks() returns.
2681  */
2682 void mm_drop_all_locks(struct mm_struct *mm)
2683 {
2684         struct vm_area_struct *vma;
2685         struct anon_vma_chain *avc;
2686
2687         BUG_ON(down_read_trylock(&mm->mmap_sem));
2688         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2689
2690         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2691                 if (vma->anon_vma)
2692                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2693                                 vm_unlock_anon_vma(avc->anon_vma);
2694                 if (vma->vm_file && vma->vm_file->f_mapping)
2695                         vm_unlock_mapping(vma->vm_file->f_mapping);
2696         }
2697
2698         mutex_unlock(&mm_all_locks_mutex);
2699 }
2700
2701 /*
2702  * initialise the VMA slab
2703  */
2704 void __init mmap_init(void)
2705 {
2706         int ret;
2707
2708         ret = percpu_counter_init(&vm_committed_as, 0);
2709         VM_BUG_ON(ret);
2710 }