9e685b9abcde33517dd8ce0d6f62475609f7f05d
[linux-3.10.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@redhat.com>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29
30 #include <asm/uaccess.h>
31 #include <asm/cacheflush.h>
32 #include <asm/tlb.h>
33 #include <asm/mmu_context.h>
34
35 #ifndef arch_mmap_check
36 #define arch_mmap_check(addr, len, flags)       (0)
37 #endif
38
39 static void unmap_region(struct mm_struct *mm,
40                 struct vm_area_struct *vma, struct vm_area_struct *prev,
41                 unsigned long start, unsigned long end);
42
43 /*
44  * WARNING: the debugging will use recursive algorithms so never enable this
45  * unless you know what you are doing.
46  */
47 #undef DEBUG_MM_RB
48
49 /* description of effects of mapping type and prot in current implementation.
50  * this is due to the limited x86 page protection hardware.  The expected
51  * behavior is in parens:
52  *
53  * map_type     prot
54  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
55  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
56  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
57  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
58  *              
59  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
60  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
61  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
62  *
63  */
64 pgprot_t protection_map[16] = {
65         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
66         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
67 };
68
69 pgprot_t vm_get_page_prot(unsigned long vm_flags)
70 {
71         return protection_map[vm_flags &
72                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
73 }
74 EXPORT_SYMBOL(vm_get_page_prot);
75
76 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
77 int sysctl_overcommit_ratio = 50;       /* default is 50% */
78 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
79 atomic_t vm_committed_space = ATOMIC_INIT(0);
80
81 /*
82  * Check that a process has enough memory to allocate a new virtual
83  * mapping. 0 means there is enough memory for the allocation to
84  * succeed and -ENOMEM implies there is not.
85  *
86  * We currently support three overcommit policies, which are set via the
87  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
88  *
89  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
90  * Additional code 2002 Jul 20 by Robert Love.
91  *
92  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
93  *
94  * Note this is a helper function intended to be used by LSMs which
95  * wish to use this logic.
96  */
97 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
98 {
99         unsigned long free, allowed;
100
101         vm_acct_memory(pages);
102
103         /*
104          * Sometimes we want to use more memory than we have
105          */
106         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
107                 return 0;
108
109         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
110                 unsigned long n;
111
112                 free = global_page_state(NR_FILE_PAGES);
113                 free += nr_swap_pages;
114
115                 /*
116                  * Any slabs which are created with the
117                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
118                  * which are reclaimable, under pressure.  The dentry
119                  * cache and most inode caches should fall into this
120                  */
121                 free += global_page_state(NR_SLAB_RECLAIMABLE);
122
123                 /*
124                  * Leave the last 3% for root
125                  */
126                 if (!cap_sys_admin)
127                         free -= free / 32;
128
129                 if (free > pages)
130                         return 0;
131
132                 /*
133                  * nr_free_pages() is very expensive on large systems,
134                  * only call if we're about to fail.
135                  */
136                 n = nr_free_pages();
137
138                 /*
139                  * Leave reserved pages. The pages are not for anonymous pages.
140                  */
141                 if (n <= totalreserve_pages)
142                         goto error;
143                 else
144                         n -= totalreserve_pages;
145
146                 /*
147                  * Leave the last 3% for root
148                  */
149                 if (!cap_sys_admin)
150                         n -= n / 32;
151                 free += n;
152
153                 if (free > pages)
154                         return 0;
155
156                 goto error;
157         }
158
159         allowed = (totalram_pages - hugetlb_total_pages())
160                 * sysctl_overcommit_ratio / 100;
161         /*
162          * Leave the last 3% for root
163          */
164         if (!cap_sys_admin)
165                 allowed -= allowed / 32;
166         allowed += total_swap_pages;
167
168         /* Don't let a single process grow too big:
169            leave 3% of the size of this process for other processes */
170         allowed -= mm->total_vm / 32;
171
172         /*
173          * cast `allowed' as a signed long because vm_committed_space
174          * sometimes has a negative value
175          */
176         if (atomic_read(&vm_committed_space) < (long)allowed)
177                 return 0;
178 error:
179         vm_unacct_memory(pages);
180
181         return -ENOMEM;
182 }
183
184 EXPORT_SYMBOL(__vm_enough_memory);
185
186 /*
187  * Requires inode->i_mapping->i_mmap_lock
188  */
189 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
190                 struct file *file, struct address_space *mapping)
191 {
192         if (vma->vm_flags & VM_DENYWRITE)
193                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
194         if (vma->vm_flags & VM_SHARED)
195                 mapping->i_mmap_writable--;
196
197         flush_dcache_mmap_lock(mapping);
198         if (unlikely(vma->vm_flags & VM_NONLINEAR))
199                 list_del_init(&vma->shared.vm_set.list);
200         else
201                 vma_prio_tree_remove(vma, &mapping->i_mmap);
202         flush_dcache_mmap_unlock(mapping);
203 }
204
205 /*
206  * Unlink a file-based vm structure from its prio_tree, to hide
207  * vma from rmap and vmtruncate before freeing its page tables.
208  */
209 void unlink_file_vma(struct vm_area_struct *vma)
210 {
211         struct file *file = vma->vm_file;
212
213         if (file) {
214                 struct address_space *mapping = file->f_mapping;
215                 spin_lock(&mapping->i_mmap_lock);
216                 __remove_shared_vm_struct(vma, file, mapping);
217                 spin_unlock(&mapping->i_mmap_lock);
218         }
219 }
220
221 /*
222  * Close a vm structure and free it, returning the next.
223  */
224 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
225 {
226         struct vm_area_struct *next = vma->vm_next;
227
228         might_sleep();
229         if (vma->vm_ops && vma->vm_ops->close)
230                 vma->vm_ops->close(vma);
231         if (vma->vm_file)
232                 fput(vma->vm_file);
233         mpol_free(vma_policy(vma));
234         kmem_cache_free(vm_area_cachep, vma);
235         return next;
236 }
237
238 asmlinkage unsigned long sys_brk(unsigned long brk)
239 {
240         unsigned long rlim, retval;
241         unsigned long newbrk, oldbrk;
242         struct mm_struct *mm = current->mm;
243
244         down_write(&mm->mmap_sem);
245
246         if (brk < mm->end_code)
247                 goto out;
248
249         /*
250          * Check against rlimit here. If this check is done later after the test
251          * of oldbrk with newbrk then it can escape the test and let the data
252          * segment grow beyond its set limit the in case where the limit is
253          * not page aligned -Ram Gupta
254          */
255         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
256         if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
257                 goto out;
258
259         newbrk = PAGE_ALIGN(brk);
260         oldbrk = PAGE_ALIGN(mm->brk);
261         if (oldbrk == newbrk)
262                 goto set_brk;
263
264         /* Always allow shrinking brk. */
265         if (brk <= mm->brk) {
266                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
267                         goto set_brk;
268                 goto out;
269         }
270
271         /* Check against existing mmap mappings. */
272         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
273                 goto out;
274
275         /* Ok, looks good - let it rip. */
276         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
277                 goto out;
278 set_brk:
279         mm->brk = brk;
280 out:
281         retval = mm->brk;
282         up_write(&mm->mmap_sem);
283         return retval;
284 }
285
286 #ifdef DEBUG_MM_RB
287 static int browse_rb(struct rb_root *root)
288 {
289         int i = 0, j;
290         struct rb_node *nd, *pn = NULL;
291         unsigned long prev = 0, pend = 0;
292
293         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
294                 struct vm_area_struct *vma;
295                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
296                 if (vma->vm_start < prev)
297                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
298                 if (vma->vm_start < pend)
299                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
300                 if (vma->vm_start > vma->vm_end)
301                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
302                 i++;
303                 pn = nd;
304                 prev = vma->vm_start;
305                 pend = vma->vm_end;
306         }
307         j = 0;
308         for (nd = pn; nd; nd = rb_prev(nd)) {
309                 j++;
310         }
311         if (i != j)
312                 printk("backwards %d, forwards %d\n", j, i), i = 0;
313         return i;
314 }
315
316 void validate_mm(struct mm_struct *mm)
317 {
318         int bug = 0;
319         int i = 0;
320         struct vm_area_struct *tmp = mm->mmap;
321         while (tmp) {
322                 tmp = tmp->vm_next;
323                 i++;
324         }
325         if (i != mm->map_count)
326                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
327         i = browse_rb(&mm->mm_rb);
328         if (i != mm->map_count)
329                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
330         BUG_ON(bug);
331 }
332 #else
333 #define validate_mm(mm) do { } while (0)
334 #endif
335
336 static struct vm_area_struct *
337 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
338                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
339                 struct rb_node ** rb_parent)
340 {
341         struct vm_area_struct * vma;
342         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
343
344         __rb_link = &mm->mm_rb.rb_node;
345         rb_prev = __rb_parent = NULL;
346         vma = NULL;
347
348         while (*__rb_link) {
349                 struct vm_area_struct *vma_tmp;
350
351                 __rb_parent = *__rb_link;
352                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
353
354                 if (vma_tmp->vm_end > addr) {
355                         vma = vma_tmp;
356                         if (vma_tmp->vm_start <= addr)
357                                 return vma;
358                         __rb_link = &__rb_parent->rb_left;
359                 } else {
360                         rb_prev = __rb_parent;
361                         __rb_link = &__rb_parent->rb_right;
362                 }
363         }
364
365         *pprev = NULL;
366         if (rb_prev)
367                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
368         *rb_link = __rb_link;
369         *rb_parent = __rb_parent;
370         return vma;
371 }
372
373 static inline void
374 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
375                 struct vm_area_struct *prev, struct rb_node *rb_parent)
376 {
377         if (prev) {
378                 vma->vm_next = prev->vm_next;
379                 prev->vm_next = vma;
380         } else {
381                 mm->mmap = vma;
382                 if (rb_parent)
383                         vma->vm_next = rb_entry(rb_parent,
384                                         struct vm_area_struct, vm_rb);
385                 else
386                         vma->vm_next = NULL;
387         }
388 }
389
390 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
391                 struct rb_node **rb_link, struct rb_node *rb_parent)
392 {
393         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
394         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
395 }
396
397 static inline void __vma_link_file(struct vm_area_struct *vma)
398 {
399         struct file * file;
400
401         file = vma->vm_file;
402         if (file) {
403                 struct address_space *mapping = file->f_mapping;
404
405                 if (vma->vm_flags & VM_DENYWRITE)
406                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
407                 if (vma->vm_flags & VM_SHARED)
408                         mapping->i_mmap_writable++;
409
410                 flush_dcache_mmap_lock(mapping);
411                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
412                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
413                 else
414                         vma_prio_tree_insert(vma, &mapping->i_mmap);
415                 flush_dcache_mmap_unlock(mapping);
416         }
417 }
418
419 static void
420 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
421         struct vm_area_struct *prev, struct rb_node **rb_link,
422         struct rb_node *rb_parent)
423 {
424         __vma_link_list(mm, vma, prev, rb_parent);
425         __vma_link_rb(mm, vma, rb_link, rb_parent);
426         __anon_vma_link(vma);
427 }
428
429 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
430                         struct vm_area_struct *prev, struct rb_node **rb_link,
431                         struct rb_node *rb_parent)
432 {
433         struct address_space *mapping = NULL;
434
435         if (vma->vm_file)
436                 mapping = vma->vm_file->f_mapping;
437
438         if (mapping) {
439                 spin_lock(&mapping->i_mmap_lock);
440                 vma->vm_truncate_count = mapping->truncate_count;
441         }
442         anon_vma_lock(vma);
443
444         __vma_link(mm, vma, prev, rb_link, rb_parent);
445         __vma_link_file(vma);
446
447         anon_vma_unlock(vma);
448         if (mapping)
449                 spin_unlock(&mapping->i_mmap_lock);
450
451         mm->map_count++;
452         validate_mm(mm);
453 }
454
455 /*
456  * Helper for vma_adjust in the split_vma insert case:
457  * insert vm structure into list and rbtree and anon_vma,
458  * but it has already been inserted into prio_tree earlier.
459  */
460 static void
461 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
462 {
463         struct vm_area_struct * __vma, * prev;
464         struct rb_node ** rb_link, * rb_parent;
465
466         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
467         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
468         __vma_link(mm, vma, prev, rb_link, rb_parent);
469         mm->map_count++;
470 }
471
472 static inline void
473 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
474                 struct vm_area_struct *prev)
475 {
476         prev->vm_next = vma->vm_next;
477         rb_erase(&vma->vm_rb, &mm->mm_rb);
478         if (mm->mmap_cache == vma)
479                 mm->mmap_cache = prev;
480 }
481
482 /*
483  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
484  * is already present in an i_mmap tree without adjusting the tree.
485  * The following helper function should be used when such adjustments
486  * are necessary.  The "insert" vma (if any) is to be inserted
487  * before we drop the necessary locks.
488  */
489 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
490         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
491 {
492         struct mm_struct *mm = vma->vm_mm;
493         struct vm_area_struct *next = vma->vm_next;
494         struct vm_area_struct *importer = NULL;
495         struct address_space *mapping = NULL;
496         struct prio_tree_root *root = NULL;
497         struct file *file = vma->vm_file;
498         struct anon_vma *anon_vma = NULL;
499         long adjust_next = 0;
500         int remove_next = 0;
501
502         if (next && !insert) {
503                 if (end >= next->vm_end) {
504                         /*
505                          * vma expands, overlapping all the next, and
506                          * perhaps the one after too (mprotect case 6).
507                          */
508 again:                  remove_next = 1 + (end > next->vm_end);
509                         end = next->vm_end;
510                         anon_vma = next->anon_vma;
511                         importer = vma;
512                 } else if (end > next->vm_start) {
513                         /*
514                          * vma expands, overlapping part of the next:
515                          * mprotect case 5 shifting the boundary up.
516                          */
517                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
518                         anon_vma = next->anon_vma;
519                         importer = vma;
520                 } else if (end < vma->vm_end) {
521                         /*
522                          * vma shrinks, and !insert tells it's not
523                          * split_vma inserting another: so it must be
524                          * mprotect case 4 shifting the boundary down.
525                          */
526                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
527                         anon_vma = next->anon_vma;
528                         importer = next;
529                 }
530         }
531
532         if (file) {
533                 mapping = file->f_mapping;
534                 if (!(vma->vm_flags & VM_NONLINEAR))
535                         root = &mapping->i_mmap;
536                 spin_lock(&mapping->i_mmap_lock);
537                 if (importer &&
538                     vma->vm_truncate_count != next->vm_truncate_count) {
539                         /*
540                          * unmap_mapping_range might be in progress:
541                          * ensure that the expanding vma is rescanned.
542                          */
543                         importer->vm_truncate_count = 0;
544                 }
545                 if (insert) {
546                         insert->vm_truncate_count = vma->vm_truncate_count;
547                         /*
548                          * Put into prio_tree now, so instantiated pages
549                          * are visible to arm/parisc __flush_dcache_page
550                          * throughout; but we cannot insert into address
551                          * space until vma start or end is updated.
552                          */
553                         __vma_link_file(insert);
554                 }
555         }
556
557         /*
558          * When changing only vma->vm_end, we don't really need
559          * anon_vma lock: but is that case worth optimizing out?
560          */
561         if (vma->anon_vma)
562                 anon_vma = vma->anon_vma;
563         if (anon_vma) {
564                 spin_lock(&anon_vma->lock);
565                 /*
566                  * Easily overlooked: when mprotect shifts the boundary,
567                  * make sure the expanding vma has anon_vma set if the
568                  * shrinking vma had, to cover any anon pages imported.
569                  */
570                 if (importer && !importer->anon_vma) {
571                         importer->anon_vma = anon_vma;
572                         __anon_vma_link(importer);
573                 }
574         }
575
576         if (root) {
577                 flush_dcache_mmap_lock(mapping);
578                 vma_prio_tree_remove(vma, root);
579                 if (adjust_next)
580                         vma_prio_tree_remove(next, root);
581         }
582
583         vma->vm_start = start;
584         vma->vm_end = end;
585         vma->vm_pgoff = pgoff;
586         if (adjust_next) {
587                 next->vm_start += adjust_next << PAGE_SHIFT;
588                 next->vm_pgoff += adjust_next;
589         }
590
591         if (root) {
592                 if (adjust_next)
593                         vma_prio_tree_insert(next, root);
594                 vma_prio_tree_insert(vma, root);
595                 flush_dcache_mmap_unlock(mapping);
596         }
597
598         if (remove_next) {
599                 /*
600                  * vma_merge has merged next into vma, and needs
601                  * us to remove next before dropping the locks.
602                  */
603                 __vma_unlink(mm, next, vma);
604                 if (file)
605                         __remove_shared_vm_struct(next, file, mapping);
606                 if (next->anon_vma)
607                         __anon_vma_merge(vma, next);
608         } else if (insert) {
609                 /*
610                  * split_vma has split insert from vma, and needs
611                  * us to insert it before dropping the locks
612                  * (it may either follow vma or precede it).
613                  */
614                 __insert_vm_struct(mm, insert);
615         }
616
617         if (anon_vma)
618                 spin_unlock(&anon_vma->lock);
619         if (mapping)
620                 spin_unlock(&mapping->i_mmap_lock);
621
622         if (remove_next) {
623                 if (file)
624                         fput(file);
625                 mm->map_count--;
626                 mpol_free(vma_policy(next));
627                 kmem_cache_free(vm_area_cachep, next);
628                 /*
629                  * In mprotect's case 6 (see comments on vma_merge),
630                  * we must remove another next too. It would clutter
631                  * up the code too much to do both in one go.
632                  */
633                 if (remove_next == 2) {
634                         next = vma->vm_next;
635                         goto again;
636                 }
637         }
638
639         validate_mm(mm);
640 }
641
642 /*
643  * If the vma has a ->close operation then the driver probably needs to release
644  * per-vma resources, so we don't attempt to merge those.
645  */
646 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
647
648 static inline int is_mergeable_vma(struct vm_area_struct *vma,
649                         struct file *file, unsigned long vm_flags)
650 {
651         if (vma->vm_flags != vm_flags)
652                 return 0;
653         if (vma->vm_file != file)
654                 return 0;
655         if (vma->vm_ops && vma->vm_ops->close)
656                 return 0;
657         return 1;
658 }
659
660 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
661                                         struct anon_vma *anon_vma2)
662 {
663         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
664 }
665
666 /*
667  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
668  * in front of (at a lower virtual address and file offset than) the vma.
669  *
670  * We cannot merge two vmas if they have differently assigned (non-NULL)
671  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
672  *
673  * We don't check here for the merged mmap wrapping around the end of pagecache
674  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
675  * wrap, nor mmaps which cover the final page at index -1UL.
676  */
677 static int
678 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
679         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
680 {
681         if (is_mergeable_vma(vma, file, vm_flags) &&
682             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
683                 if (vma->vm_pgoff == vm_pgoff)
684                         return 1;
685         }
686         return 0;
687 }
688
689 /*
690  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
691  * beyond (at a higher virtual address and file offset than) the vma.
692  *
693  * We cannot merge two vmas if they have differently assigned (non-NULL)
694  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
695  */
696 static int
697 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
698         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
699 {
700         if (is_mergeable_vma(vma, file, vm_flags) &&
701             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
702                 pgoff_t vm_pglen;
703                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
704                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
705                         return 1;
706         }
707         return 0;
708 }
709
710 /*
711  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
712  * whether that can be merged with its predecessor or its successor.
713  * Or both (it neatly fills a hole).
714  *
715  * In most cases - when called for mmap, brk or mremap - [addr,end) is
716  * certain not to be mapped by the time vma_merge is called; but when
717  * called for mprotect, it is certain to be already mapped (either at
718  * an offset within prev, or at the start of next), and the flags of
719  * this area are about to be changed to vm_flags - and the no-change
720  * case has already been eliminated.
721  *
722  * The following mprotect cases have to be considered, where AAAA is
723  * the area passed down from mprotect_fixup, never extending beyond one
724  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
725  *
726  *     AAAA             AAAA                AAAA          AAAA
727  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
728  *    cannot merge    might become    might become    might become
729  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
730  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
731  *    mremap move:                                    PPPPNNNNNNNN 8
732  *        AAAA
733  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
734  *    might become    case 1 below    case 2 below    case 3 below
735  *
736  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
737  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
738  */
739 struct vm_area_struct *vma_merge(struct mm_struct *mm,
740                         struct vm_area_struct *prev, unsigned long addr,
741                         unsigned long end, unsigned long vm_flags,
742                         struct anon_vma *anon_vma, struct file *file,
743                         pgoff_t pgoff, struct mempolicy *policy)
744 {
745         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
746         struct vm_area_struct *area, *next;
747
748         /*
749          * We later require that vma->vm_flags == vm_flags,
750          * so this tests vma->vm_flags & VM_SPECIAL, too.
751          */
752         if (vm_flags & VM_SPECIAL)
753                 return NULL;
754
755         if (prev)
756                 next = prev->vm_next;
757         else
758                 next = mm->mmap;
759         area = next;
760         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
761                 next = next->vm_next;
762
763         /*
764          * Can it merge with the predecessor?
765          */
766         if (prev && prev->vm_end == addr &&
767                         mpol_equal(vma_policy(prev), policy) &&
768                         can_vma_merge_after(prev, vm_flags,
769                                                 anon_vma, file, pgoff)) {
770                 /*
771                  * OK, it can.  Can we now merge in the successor as well?
772                  */
773                 if (next && end == next->vm_start &&
774                                 mpol_equal(policy, vma_policy(next)) &&
775                                 can_vma_merge_before(next, vm_flags,
776                                         anon_vma, file, pgoff+pglen) &&
777                                 is_mergeable_anon_vma(prev->anon_vma,
778                                                       next->anon_vma)) {
779                                                         /* cases 1, 6 */
780                         vma_adjust(prev, prev->vm_start,
781                                 next->vm_end, prev->vm_pgoff, NULL);
782                 } else                                  /* cases 2, 5, 7 */
783                         vma_adjust(prev, prev->vm_start,
784                                 end, prev->vm_pgoff, NULL);
785                 return prev;
786         }
787
788         /*
789          * Can this new request be merged in front of next?
790          */
791         if (next && end == next->vm_start &&
792                         mpol_equal(policy, vma_policy(next)) &&
793                         can_vma_merge_before(next, vm_flags,
794                                         anon_vma, file, pgoff+pglen)) {
795                 if (prev && addr < prev->vm_end)        /* case 4 */
796                         vma_adjust(prev, prev->vm_start,
797                                 addr, prev->vm_pgoff, NULL);
798                 else                                    /* cases 3, 8 */
799                         vma_adjust(area, addr, next->vm_end,
800                                 next->vm_pgoff - pglen, NULL);
801                 return area;
802         }
803
804         return NULL;
805 }
806
807 /*
808  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
809  * neighbouring vmas for a suitable anon_vma, before it goes off
810  * to allocate a new anon_vma.  It checks because a repetitive
811  * sequence of mprotects and faults may otherwise lead to distinct
812  * anon_vmas being allocated, preventing vma merge in subsequent
813  * mprotect.
814  */
815 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
816 {
817         struct vm_area_struct *near;
818         unsigned long vm_flags;
819
820         near = vma->vm_next;
821         if (!near)
822                 goto try_prev;
823
824         /*
825          * Since only mprotect tries to remerge vmas, match flags
826          * which might be mprotected into each other later on.
827          * Neither mlock nor madvise tries to remerge at present,
828          * so leave their flags as obstructing a merge.
829          */
830         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
831         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
832
833         if (near->anon_vma && vma->vm_end == near->vm_start &&
834                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
835                         can_vma_merge_before(near, vm_flags,
836                                 NULL, vma->vm_file, vma->vm_pgoff +
837                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
838                 return near->anon_vma;
839 try_prev:
840         /*
841          * It is potentially slow to have to call find_vma_prev here.
842          * But it's only on the first write fault on the vma, not
843          * every time, and we could devise a way to avoid it later
844          * (e.g. stash info in next's anon_vma_node when assigning
845          * an anon_vma, or when trying vma_merge).  Another time.
846          */
847         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
848         if (!near)
849                 goto none;
850
851         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
852         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
853
854         if (near->anon_vma && near->vm_end == vma->vm_start &&
855                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
856                         can_vma_merge_after(near, vm_flags,
857                                 NULL, vma->vm_file, vma->vm_pgoff))
858                 return near->anon_vma;
859 none:
860         /*
861          * There's no absolute need to look only at touching neighbours:
862          * we could search further afield for "compatible" anon_vmas.
863          * But it would probably just be a waste of time searching,
864          * or lead to too many vmas hanging off the same anon_vma.
865          * We're trying to allow mprotect remerging later on,
866          * not trying to minimize memory used for anon_vmas.
867          */
868         return NULL;
869 }
870
871 #ifdef CONFIG_PROC_FS
872 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
873                                                 struct file *file, long pages)
874 {
875         const unsigned long stack_flags
876                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
877
878         if (file) {
879                 mm->shared_vm += pages;
880                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
881                         mm->exec_vm += pages;
882         } else if (flags & stack_flags)
883                 mm->stack_vm += pages;
884         if (flags & (VM_RESERVED|VM_IO))
885                 mm->reserved_vm += pages;
886 }
887 #endif /* CONFIG_PROC_FS */
888
889 /*
890  * The caller must hold down_write(current->mm->mmap_sem).
891  */
892
893 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
894                         unsigned long len, unsigned long prot,
895                         unsigned long flags, unsigned long pgoff)
896 {
897         struct mm_struct * mm = current->mm;
898         struct inode *inode;
899         unsigned int vm_flags;
900         int error;
901         int accountable = 1;
902         unsigned long reqprot = prot;
903
904         /*
905          * Does the application expect PROT_READ to imply PROT_EXEC?
906          *
907          * (the exception is when the underlying filesystem is noexec
908          *  mounted, in which case we dont add PROT_EXEC.)
909          */
910         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
911                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
912                         prot |= PROT_EXEC;
913
914         if (!len)
915                 return -EINVAL;
916
917         error = arch_mmap_check(addr, len, flags);
918         if (error)
919                 return error;
920
921         /* Careful about overflows.. */
922         len = PAGE_ALIGN(len);
923         if (!len || len > TASK_SIZE)
924                 return -ENOMEM;
925
926         /* offset overflow? */
927         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
928                return -EOVERFLOW;
929
930         /* Too many mappings? */
931         if (mm->map_count > sysctl_max_map_count)
932                 return -ENOMEM;
933
934         /* Obtain the address to map to. we verify (or select) it and ensure
935          * that it represents a valid section of the address space.
936          */
937         addr = get_unmapped_area(file, addr, len, pgoff, flags);
938         if (addr & ~PAGE_MASK)
939                 return addr;
940
941         /* Do simple checking here so the lower-level routines won't have
942          * to. we assume access permissions have been handled by the open
943          * of the memory object, so we don't do any here.
944          */
945         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
946                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
947
948         if (flags & MAP_LOCKED) {
949                 if (!can_do_mlock())
950                         return -EPERM;
951                 vm_flags |= VM_LOCKED;
952         }
953         /* mlock MCL_FUTURE? */
954         if (vm_flags & VM_LOCKED) {
955                 unsigned long locked, lock_limit;
956                 locked = len >> PAGE_SHIFT;
957                 locked += mm->locked_vm;
958                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
959                 lock_limit >>= PAGE_SHIFT;
960                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
961                         return -EAGAIN;
962         }
963
964         inode = file ? file->f_path.dentry->d_inode : NULL;
965
966         if (file) {
967                 switch (flags & MAP_TYPE) {
968                 case MAP_SHARED:
969                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
970                                 return -EACCES;
971
972                         /*
973                          * Make sure we don't allow writing to an append-only
974                          * file..
975                          */
976                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
977                                 return -EACCES;
978
979                         /*
980                          * Make sure there are no mandatory locks on the file.
981                          */
982                         if (locks_verify_locked(inode))
983                                 return -EAGAIN;
984
985                         vm_flags |= VM_SHARED | VM_MAYSHARE;
986                         if (!(file->f_mode & FMODE_WRITE))
987                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
988
989                         /* fall through */
990                 case MAP_PRIVATE:
991                         if (!(file->f_mode & FMODE_READ))
992                                 return -EACCES;
993                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
994                                 if (vm_flags & VM_EXEC)
995                                         return -EPERM;
996                                 vm_flags &= ~VM_MAYEXEC;
997                         }
998                         if (is_file_hugepages(file))
999                                 accountable = 0;
1000
1001                         if (!file->f_op || !file->f_op->mmap)
1002                                 return -ENODEV;
1003                         break;
1004
1005                 default:
1006                         return -EINVAL;
1007                 }
1008         } else {
1009                 switch (flags & MAP_TYPE) {
1010                 case MAP_SHARED:
1011                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1012                         break;
1013                 case MAP_PRIVATE:
1014                         /*
1015                          * Set pgoff according to addr for anon_vma.
1016                          */
1017                         pgoff = addr >> PAGE_SHIFT;
1018                         break;
1019                 default:
1020                         return -EINVAL;
1021                 }
1022         }
1023
1024         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1025         if (error)
1026                 return error;
1027
1028         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1029                            accountable);
1030 }
1031 EXPORT_SYMBOL(do_mmap_pgoff);
1032
1033 /*
1034  * Some shared mappigns will want the pages marked read-only
1035  * to track write events. If so, we'll downgrade vm_page_prot
1036  * to the private version (using protection_map[] without the
1037  * VM_SHARED bit).
1038  */
1039 int vma_wants_writenotify(struct vm_area_struct *vma)
1040 {
1041         unsigned int vm_flags = vma->vm_flags;
1042
1043         /* If it was private or non-writable, the write bit is already clear */
1044         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1045                 return 0;
1046
1047         /* The backer wishes to know when pages are first written to? */
1048         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1049                 return 1;
1050
1051         /* The open routine did something to the protections already? */
1052         if (pgprot_val(vma->vm_page_prot) !=
1053             pgprot_val(protection_map[vm_flags &
1054                     (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]))
1055                 return 0;
1056
1057         /* Specialty mapping? */
1058         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1059                 return 0;
1060
1061         /* Can the mapping track the dirty pages? */
1062         return vma->vm_file && vma->vm_file->f_mapping &&
1063                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1064 }
1065
1066
1067 unsigned long mmap_region(struct file *file, unsigned long addr,
1068                           unsigned long len, unsigned long flags,
1069                           unsigned int vm_flags, unsigned long pgoff,
1070                           int accountable)
1071 {
1072         struct mm_struct *mm = current->mm;
1073         struct vm_area_struct *vma, *prev;
1074         int correct_wcount = 0;
1075         int error;
1076         struct rb_node **rb_link, *rb_parent;
1077         unsigned long charged = 0;
1078         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1079
1080         /* Clear old maps */
1081         error = -ENOMEM;
1082 munmap_back:
1083         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1084         if (vma && vma->vm_start < addr + len) {
1085                 if (do_munmap(mm, addr, len))
1086                         return -ENOMEM;
1087                 goto munmap_back;
1088         }
1089
1090         /* Check against address space limit. */
1091         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1092                 return -ENOMEM;
1093
1094         if (accountable && (!(flags & MAP_NORESERVE) ||
1095                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1096                 if (vm_flags & VM_SHARED) {
1097                         /* Check memory availability in shmem_file_setup? */
1098                         vm_flags |= VM_ACCOUNT;
1099                 } else if (vm_flags & VM_WRITE) {
1100                         /*
1101                          * Private writable mapping: check memory availability
1102                          */
1103                         charged = len >> PAGE_SHIFT;
1104                         if (security_vm_enough_memory(charged))
1105                                 return -ENOMEM;
1106                         vm_flags |= VM_ACCOUNT;
1107                 }
1108         }
1109
1110         /*
1111          * Can we just expand an old private anonymous mapping?
1112          * The VM_SHARED test is necessary because shmem_zero_setup
1113          * will create the file object for a shared anonymous map below.
1114          */
1115         if (!file && !(vm_flags & VM_SHARED) &&
1116             vma_merge(mm, prev, addr, addr + len, vm_flags,
1117                                         NULL, NULL, pgoff, NULL))
1118                 goto out;
1119
1120         /*
1121          * Determine the object being mapped and call the appropriate
1122          * specific mapper. the address has already been validated, but
1123          * not unmapped, but the maps are removed from the list.
1124          */
1125         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1126         if (!vma) {
1127                 error = -ENOMEM;
1128                 goto unacct_error;
1129         }
1130
1131         vma->vm_mm = mm;
1132         vma->vm_start = addr;
1133         vma->vm_end = addr + len;
1134         vma->vm_flags = vm_flags;
1135         vma->vm_page_prot = protection_map[vm_flags &
1136                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1137         vma->vm_pgoff = pgoff;
1138
1139         if (file) {
1140                 error = -EINVAL;
1141                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1142                         goto free_vma;
1143                 if (vm_flags & VM_DENYWRITE) {
1144                         error = deny_write_access(file);
1145                         if (error)
1146                                 goto free_vma;
1147                         correct_wcount = 1;
1148                 }
1149                 vma->vm_file = file;
1150                 get_file(file);
1151                 error = file->f_op->mmap(file, vma);
1152                 if (error)
1153                         goto unmap_and_free_vma;
1154         } else if (vm_flags & VM_SHARED) {
1155                 error = shmem_zero_setup(vma);
1156                 if (error)
1157                         goto free_vma;
1158         }
1159
1160         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1161          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1162          * that memory reservation must be checked; but that reservation
1163          * belongs to shared memory object, not to vma: so now clear it.
1164          */
1165         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1166                 vma->vm_flags &= ~VM_ACCOUNT;
1167
1168         /* Can addr have changed??
1169          *
1170          * Answer: Yes, several device drivers can do it in their
1171          *         f_op->mmap method. -DaveM
1172          */
1173         addr = vma->vm_start;
1174         pgoff = vma->vm_pgoff;
1175         vm_flags = vma->vm_flags;
1176
1177         if (vma_wants_writenotify(vma))
1178                 vma->vm_page_prot =
1179                         protection_map[vm_flags & (VM_READ|VM_WRITE|VM_EXEC)];
1180
1181         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1182                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1183                 file = vma->vm_file;
1184                 vma_link(mm, vma, prev, rb_link, rb_parent);
1185                 if (correct_wcount)
1186                         atomic_inc(&inode->i_writecount);
1187         } else {
1188                 if (file) {
1189                         if (correct_wcount)
1190                                 atomic_inc(&inode->i_writecount);
1191                         fput(file);
1192                 }
1193                 mpol_free(vma_policy(vma));
1194                 kmem_cache_free(vm_area_cachep, vma);
1195         }
1196 out:    
1197         mm->total_vm += len >> PAGE_SHIFT;
1198         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1199         if (vm_flags & VM_LOCKED) {
1200                 mm->locked_vm += len >> PAGE_SHIFT;
1201                 make_pages_present(addr, addr + len);
1202         }
1203         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1204                 make_pages_present(addr, addr + len);
1205         return addr;
1206
1207 unmap_and_free_vma:
1208         if (correct_wcount)
1209                 atomic_inc(&inode->i_writecount);
1210         vma->vm_file = NULL;
1211         fput(file);
1212
1213         /* Undo any partial mapping done by a device driver. */
1214         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1215         charged = 0;
1216 free_vma:
1217         kmem_cache_free(vm_area_cachep, vma);
1218 unacct_error:
1219         if (charged)
1220                 vm_unacct_memory(charged);
1221         return error;
1222 }
1223
1224 /* Get an address range which is currently unmapped.
1225  * For shmat() with addr=0.
1226  *
1227  * Ugly calling convention alert:
1228  * Return value with the low bits set means error value,
1229  * ie
1230  *      if (ret & ~PAGE_MASK)
1231  *              error = ret;
1232  *
1233  * This function "knows" that -ENOMEM has the bits set.
1234  */
1235 #ifndef HAVE_ARCH_UNMAPPED_AREA
1236 unsigned long
1237 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1238                 unsigned long len, unsigned long pgoff, unsigned long flags)
1239 {
1240         struct mm_struct *mm = current->mm;
1241         struct vm_area_struct *vma;
1242         unsigned long start_addr;
1243
1244         if (len > TASK_SIZE)
1245                 return -ENOMEM;
1246
1247         if (flags & MAP_FIXED)
1248                 return addr;
1249
1250         if (addr) {
1251                 addr = PAGE_ALIGN(addr);
1252                 vma = find_vma(mm, addr);
1253                 if (TASK_SIZE - len >= addr &&
1254                     (!vma || addr + len <= vma->vm_start))
1255                         return addr;
1256         }
1257         if (len > mm->cached_hole_size) {
1258                 start_addr = addr = mm->free_area_cache;
1259         } else {
1260                 start_addr = addr = TASK_UNMAPPED_BASE;
1261                 mm->cached_hole_size = 0;
1262         }
1263
1264 full_search:
1265         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1266                 /* At this point:  (!vma || addr < vma->vm_end). */
1267                 if (TASK_SIZE - len < addr) {
1268                         /*
1269                          * Start a new search - just in case we missed
1270                          * some holes.
1271                          */
1272                         if (start_addr != TASK_UNMAPPED_BASE) {
1273                                 addr = TASK_UNMAPPED_BASE;
1274                                 start_addr = addr;
1275                                 mm->cached_hole_size = 0;
1276                                 goto full_search;
1277                         }
1278                         return -ENOMEM;
1279                 }
1280                 if (!vma || addr + len <= vma->vm_start) {
1281                         /*
1282                          * Remember the place where we stopped the search:
1283                          */
1284                         mm->free_area_cache = addr + len;
1285                         return addr;
1286                 }
1287                 if (addr + mm->cached_hole_size < vma->vm_start)
1288                         mm->cached_hole_size = vma->vm_start - addr;
1289                 addr = vma->vm_end;
1290         }
1291 }
1292 #endif  
1293
1294 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1295 {
1296         /*
1297          * Is this a new hole at the lowest possible address?
1298          */
1299         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1300                 mm->free_area_cache = addr;
1301                 mm->cached_hole_size = ~0UL;
1302         }
1303 }
1304
1305 /*
1306  * This mmap-allocator allocates new areas top-down from below the
1307  * stack's low limit (the base):
1308  */
1309 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1310 unsigned long
1311 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1312                           const unsigned long len, const unsigned long pgoff,
1313                           const unsigned long flags)
1314 {
1315         struct vm_area_struct *vma;
1316         struct mm_struct *mm = current->mm;
1317         unsigned long addr = addr0;
1318
1319         /* requested length too big for entire address space */
1320         if (len > TASK_SIZE)
1321                 return -ENOMEM;
1322
1323         if (flags & MAP_FIXED)
1324                 return addr;
1325
1326         /* requesting a specific address */
1327         if (addr) {
1328                 addr = PAGE_ALIGN(addr);
1329                 vma = find_vma(mm, addr);
1330                 if (TASK_SIZE - len >= addr &&
1331                                 (!vma || addr + len <= vma->vm_start))
1332                         return addr;
1333         }
1334
1335         /* check if free_area_cache is useful for us */
1336         if (len <= mm->cached_hole_size) {
1337                 mm->cached_hole_size = 0;
1338                 mm->free_area_cache = mm->mmap_base;
1339         }
1340
1341         /* either no address requested or can't fit in requested address hole */
1342         addr = mm->free_area_cache;
1343
1344         /* make sure it can fit in the remaining address space */
1345         if (addr > len) {
1346                 vma = find_vma(mm, addr-len);
1347                 if (!vma || addr <= vma->vm_start)
1348                         /* remember the address as a hint for next time */
1349                         return (mm->free_area_cache = addr-len);
1350         }
1351
1352         if (mm->mmap_base < len)
1353                 goto bottomup;
1354
1355         addr = mm->mmap_base-len;
1356
1357         do {
1358                 /*
1359                  * Lookup failure means no vma is above this address,
1360                  * else if new region fits below vma->vm_start,
1361                  * return with success:
1362                  */
1363                 vma = find_vma(mm, addr);
1364                 if (!vma || addr+len <= vma->vm_start)
1365                         /* remember the address as a hint for next time */
1366                         return (mm->free_area_cache = addr);
1367
1368                 /* remember the largest hole we saw so far */
1369                 if (addr + mm->cached_hole_size < vma->vm_start)
1370                         mm->cached_hole_size = vma->vm_start - addr;
1371
1372                 /* try just below the current vma->vm_start */
1373                 addr = vma->vm_start-len;
1374         } while (len < vma->vm_start);
1375
1376 bottomup:
1377         /*
1378          * A failed mmap() very likely causes application failure,
1379          * so fall back to the bottom-up function here. This scenario
1380          * can happen with large stack limits and large mmap()
1381          * allocations.
1382          */
1383         mm->cached_hole_size = ~0UL;
1384         mm->free_area_cache = TASK_UNMAPPED_BASE;
1385         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1386         /*
1387          * Restore the topdown base:
1388          */
1389         mm->free_area_cache = mm->mmap_base;
1390         mm->cached_hole_size = ~0UL;
1391
1392         return addr;
1393 }
1394 #endif
1395
1396 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1397 {
1398         /*
1399          * Is this a new hole at the highest possible address?
1400          */
1401         if (addr > mm->free_area_cache)
1402                 mm->free_area_cache = addr;
1403
1404         /* dont allow allocations above current base */
1405         if (mm->free_area_cache > mm->mmap_base)
1406                 mm->free_area_cache = mm->mmap_base;
1407 }
1408
1409 unsigned long
1410 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1411                 unsigned long pgoff, unsigned long flags)
1412 {
1413         unsigned long (*get_area)(struct file *, unsigned long,
1414                                   unsigned long, unsigned long, unsigned long);
1415
1416         get_area = current->mm->get_unmapped_area;
1417         if (file && file->f_op && file->f_op->get_unmapped_area)
1418                 get_area = file->f_op->get_unmapped_area;
1419         addr = get_area(file, addr, len, pgoff, flags);
1420         if (IS_ERR_VALUE(addr))
1421                 return addr;
1422
1423         if (addr > TASK_SIZE - len)
1424                 return -ENOMEM;
1425         if (addr & ~PAGE_MASK)
1426                 return -EINVAL;
1427
1428         return addr;
1429 }
1430
1431 EXPORT_SYMBOL(get_unmapped_area);
1432
1433 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1434 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1435 {
1436         struct vm_area_struct *vma = NULL;
1437
1438         if (mm) {
1439                 /* Check the cache first. */
1440                 /* (Cache hit rate is typically around 35%.) */
1441                 vma = mm->mmap_cache;
1442                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1443                         struct rb_node * rb_node;
1444
1445                         rb_node = mm->mm_rb.rb_node;
1446                         vma = NULL;
1447
1448                         while (rb_node) {
1449                                 struct vm_area_struct * vma_tmp;
1450
1451                                 vma_tmp = rb_entry(rb_node,
1452                                                 struct vm_area_struct, vm_rb);
1453
1454                                 if (vma_tmp->vm_end > addr) {
1455                                         vma = vma_tmp;
1456                                         if (vma_tmp->vm_start <= addr)
1457                                                 break;
1458                                         rb_node = rb_node->rb_left;
1459                                 } else
1460                                         rb_node = rb_node->rb_right;
1461                         }
1462                         if (vma)
1463                                 mm->mmap_cache = vma;
1464                 }
1465         }
1466         return vma;
1467 }
1468
1469 EXPORT_SYMBOL(find_vma);
1470
1471 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1472 struct vm_area_struct *
1473 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1474                         struct vm_area_struct **pprev)
1475 {
1476         struct vm_area_struct *vma = NULL, *prev = NULL;
1477         struct rb_node * rb_node;
1478         if (!mm)
1479                 goto out;
1480
1481         /* Guard against addr being lower than the first VMA */
1482         vma = mm->mmap;
1483
1484         /* Go through the RB tree quickly. */
1485         rb_node = mm->mm_rb.rb_node;
1486
1487         while (rb_node) {
1488                 struct vm_area_struct *vma_tmp;
1489                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1490
1491                 if (addr < vma_tmp->vm_end) {
1492                         rb_node = rb_node->rb_left;
1493                 } else {
1494                         prev = vma_tmp;
1495                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1496                                 break;
1497                         rb_node = rb_node->rb_right;
1498                 }
1499         }
1500
1501 out:
1502         *pprev = prev;
1503         return prev ? prev->vm_next : vma;
1504 }
1505
1506 /*
1507  * Verify that the stack growth is acceptable and
1508  * update accounting. This is shared with both the
1509  * grow-up and grow-down cases.
1510  */
1511 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1512 {
1513         struct mm_struct *mm = vma->vm_mm;
1514         struct rlimit *rlim = current->signal->rlim;
1515         unsigned long new_start;
1516
1517         /* address space limit tests */
1518         if (!may_expand_vm(mm, grow))
1519                 return -ENOMEM;
1520
1521         /* Stack limit test */
1522         if (size > rlim[RLIMIT_STACK].rlim_cur)
1523                 return -ENOMEM;
1524
1525         /* mlock limit tests */
1526         if (vma->vm_flags & VM_LOCKED) {
1527                 unsigned long locked;
1528                 unsigned long limit;
1529                 locked = mm->locked_vm + grow;
1530                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1531                 if (locked > limit && !capable(CAP_IPC_LOCK))
1532                         return -ENOMEM;
1533         }
1534
1535         /* Check to ensure the stack will not grow into a hugetlb-only region */
1536         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1537                         vma->vm_end - size;
1538         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1539                 return -EFAULT;
1540
1541         /*
1542          * Overcommit..  This must be the final test, as it will
1543          * update security statistics.
1544          */
1545         if (security_vm_enough_memory(grow))
1546                 return -ENOMEM;
1547
1548         /* Ok, everything looks good - let it rip */
1549         mm->total_vm += grow;
1550         if (vma->vm_flags & VM_LOCKED)
1551                 mm->locked_vm += grow;
1552         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1553         return 0;
1554 }
1555
1556 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1557 /*
1558  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1559  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1560  */
1561 #ifndef CONFIG_IA64
1562 static inline
1563 #endif
1564 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1565 {
1566         int error;
1567
1568         if (!(vma->vm_flags & VM_GROWSUP))
1569                 return -EFAULT;
1570
1571         /*
1572          * We must make sure the anon_vma is allocated
1573          * so that the anon_vma locking is not a noop.
1574          */
1575         if (unlikely(anon_vma_prepare(vma)))
1576                 return -ENOMEM;
1577         anon_vma_lock(vma);
1578
1579         /*
1580          * vma->vm_start/vm_end cannot change under us because the caller
1581          * is required to hold the mmap_sem in read mode.  We need the
1582          * anon_vma lock to serialize against concurrent expand_stacks.
1583          * Also guard against wrapping around to address 0.
1584          */
1585         if (address < PAGE_ALIGN(address+4))
1586                 address = PAGE_ALIGN(address+4);
1587         else {
1588                 anon_vma_unlock(vma);
1589                 return -ENOMEM;
1590         }
1591         error = 0;
1592
1593         /* Somebody else might have raced and expanded it already */
1594         if (address > vma->vm_end) {
1595                 unsigned long size, grow;
1596
1597                 size = address - vma->vm_start;
1598                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1599
1600                 error = acct_stack_growth(vma, size, grow);
1601                 if (!error)
1602                         vma->vm_end = address;
1603         }
1604         anon_vma_unlock(vma);
1605         return error;
1606 }
1607 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1608
1609 /*
1610  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1611  */
1612 static inline int expand_downwards(struct vm_area_struct *vma,
1613                                    unsigned long address)
1614 {
1615         int error;
1616
1617         /*
1618          * We must make sure the anon_vma is allocated
1619          * so that the anon_vma locking is not a noop.
1620          */
1621         if (unlikely(anon_vma_prepare(vma)))
1622                 return -ENOMEM;
1623         anon_vma_lock(vma);
1624
1625         /*
1626          * vma->vm_start/vm_end cannot change under us because the caller
1627          * is required to hold the mmap_sem in read mode.  We need the
1628          * anon_vma lock to serialize against concurrent expand_stacks.
1629          */
1630         address &= PAGE_MASK;
1631         error = 0;
1632
1633         /* Somebody else might have raced and expanded it already */
1634         if (address < vma->vm_start) {
1635                 unsigned long size, grow;
1636
1637                 size = vma->vm_end - address;
1638                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1639
1640                 error = acct_stack_growth(vma, size, grow);
1641                 if (!error) {
1642                         vma->vm_start = address;
1643                         vma->vm_pgoff -= grow;
1644                 }
1645         }
1646         anon_vma_unlock(vma);
1647         return error;
1648 }
1649
1650 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1651 {
1652         return expand_downwards(vma, address);
1653 }
1654
1655 #ifdef CONFIG_STACK_GROWSUP
1656 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1657 {
1658         return expand_upwards(vma, address);
1659 }
1660
1661 struct vm_area_struct *
1662 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1663 {
1664         struct vm_area_struct *vma, *prev;
1665
1666         addr &= PAGE_MASK;
1667         vma = find_vma_prev(mm, addr, &prev);
1668         if (vma && (vma->vm_start <= addr))
1669                 return vma;
1670         if (!prev || expand_stack(prev, addr))
1671                 return NULL;
1672         if (prev->vm_flags & VM_LOCKED)
1673                 make_pages_present(addr, prev->vm_end);
1674         return prev;
1675 }
1676 #else
1677 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1678 {
1679         return expand_downwards(vma, address);
1680 }
1681
1682 struct vm_area_struct *
1683 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1684 {
1685         struct vm_area_struct * vma;
1686         unsigned long start;
1687
1688         addr &= PAGE_MASK;
1689         vma = find_vma(mm,addr);
1690         if (!vma)
1691                 return NULL;
1692         if (vma->vm_start <= addr)
1693                 return vma;
1694         if (!(vma->vm_flags & VM_GROWSDOWN))
1695                 return NULL;
1696         start = vma->vm_start;
1697         if (expand_stack(vma, addr))
1698                 return NULL;
1699         if (vma->vm_flags & VM_LOCKED)
1700                 make_pages_present(addr, start);
1701         return vma;
1702 }
1703 #endif
1704
1705 /*
1706  * Ok - we have the memory areas we should free on the vma list,
1707  * so release them, and do the vma updates.
1708  *
1709  * Called with the mm semaphore held.
1710  */
1711 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1712 {
1713         /* Update high watermark before we lower total_vm */
1714         update_hiwater_vm(mm);
1715         do {
1716                 long nrpages = vma_pages(vma);
1717
1718                 mm->total_vm -= nrpages;
1719                 if (vma->vm_flags & VM_LOCKED)
1720                         mm->locked_vm -= nrpages;
1721                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1722                 vma = remove_vma(vma);
1723         } while (vma);
1724         validate_mm(mm);
1725 }
1726
1727 /*
1728  * Get rid of page table information in the indicated region.
1729  *
1730  * Called with the mm semaphore held.
1731  */
1732 static void unmap_region(struct mm_struct *mm,
1733                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1734                 unsigned long start, unsigned long end)
1735 {
1736         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1737         struct mmu_gather *tlb;
1738         unsigned long nr_accounted = 0;
1739
1740         lru_add_drain();
1741         tlb = tlb_gather_mmu(mm, 0);
1742         update_hiwater_rss(mm);
1743         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1744         vm_unacct_memory(nr_accounted);
1745         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1746                                  next? next->vm_start: 0);
1747         tlb_finish_mmu(tlb, start, end);
1748 }
1749
1750 /*
1751  * Create a list of vma's touched by the unmap, removing them from the mm's
1752  * vma list as we go..
1753  */
1754 static void
1755 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1756         struct vm_area_struct *prev, unsigned long end)
1757 {
1758         struct vm_area_struct **insertion_point;
1759         struct vm_area_struct *tail_vma = NULL;
1760         unsigned long addr;
1761
1762         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1763         do {
1764                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1765                 mm->map_count--;
1766                 tail_vma = vma;
1767                 vma = vma->vm_next;
1768         } while (vma && vma->vm_start < end);
1769         *insertion_point = vma;
1770         tail_vma->vm_next = NULL;
1771         if (mm->unmap_area == arch_unmap_area)
1772                 addr = prev ? prev->vm_end : mm->mmap_base;
1773         else
1774                 addr = vma ?  vma->vm_start : mm->mmap_base;
1775         mm->unmap_area(mm, addr);
1776         mm->mmap_cache = NULL;          /* Kill the cache. */
1777 }
1778
1779 /*
1780  * Split a vma into two pieces at address 'addr', a new vma is allocated
1781  * either for the first part or the tail.
1782  */
1783 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1784               unsigned long addr, int new_below)
1785 {
1786         struct mempolicy *pol;
1787         struct vm_area_struct *new;
1788
1789         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1790                 return -EINVAL;
1791
1792         if (mm->map_count >= sysctl_max_map_count)
1793                 return -ENOMEM;
1794
1795         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1796         if (!new)
1797                 return -ENOMEM;
1798
1799         /* most fields are the same, copy all, and then fixup */
1800         *new = *vma;
1801
1802         if (new_below)
1803                 new->vm_end = addr;
1804         else {
1805                 new->vm_start = addr;
1806                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1807         }
1808
1809         pol = mpol_copy(vma_policy(vma));
1810         if (IS_ERR(pol)) {
1811                 kmem_cache_free(vm_area_cachep, new);
1812                 return PTR_ERR(pol);
1813         }
1814         vma_set_policy(new, pol);
1815
1816         if (new->vm_file)
1817                 get_file(new->vm_file);
1818
1819         if (new->vm_ops && new->vm_ops->open)
1820                 new->vm_ops->open(new);
1821
1822         if (new_below)
1823                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1824                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1825         else
1826                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1827
1828         return 0;
1829 }
1830
1831 /* Munmap is split into 2 main parts -- this part which finds
1832  * what needs doing, and the areas themselves, which do the
1833  * work.  This now handles partial unmappings.
1834  * Jeremy Fitzhardinge <jeremy@goop.org>
1835  */
1836 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1837 {
1838         unsigned long end;
1839         struct vm_area_struct *vma, *prev, *last;
1840
1841         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1842                 return -EINVAL;
1843
1844         if ((len = PAGE_ALIGN(len)) == 0)
1845                 return -EINVAL;
1846
1847         /* Find the first overlapping VMA */
1848         vma = find_vma_prev(mm, start, &prev);
1849         if (!vma)
1850                 return 0;
1851         /* we have  start < vma->vm_end  */
1852
1853         /* if it doesn't overlap, we have nothing.. */
1854         end = start + len;
1855         if (vma->vm_start >= end)
1856                 return 0;
1857
1858         /*
1859          * If we need to split any vma, do it now to save pain later.
1860          *
1861          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1862          * unmapped vm_area_struct will remain in use: so lower split_vma
1863          * places tmp vma above, and higher split_vma places tmp vma below.
1864          */
1865         if (start > vma->vm_start) {
1866                 int error = split_vma(mm, vma, start, 0);
1867                 if (error)
1868                         return error;
1869                 prev = vma;
1870         }
1871
1872         /* Does it split the last one? */
1873         last = find_vma(mm, end);
1874         if (last && end > last->vm_start) {
1875                 int error = split_vma(mm, last, end, 1);
1876                 if (error)
1877                         return error;
1878         }
1879         vma = prev? prev->vm_next: mm->mmap;
1880
1881         /*
1882          * Remove the vma's, and unmap the actual pages
1883          */
1884         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1885         unmap_region(mm, vma, prev, start, end);
1886
1887         /* Fix up all other VM information */
1888         remove_vma_list(mm, vma);
1889
1890         return 0;
1891 }
1892
1893 EXPORT_SYMBOL(do_munmap);
1894
1895 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1896 {
1897         int ret;
1898         struct mm_struct *mm = current->mm;
1899
1900         profile_munmap(addr);
1901
1902         down_write(&mm->mmap_sem);
1903         ret = do_munmap(mm, addr, len);
1904         up_write(&mm->mmap_sem);
1905         return ret;
1906 }
1907
1908 static inline void verify_mm_writelocked(struct mm_struct *mm)
1909 {
1910 #ifdef CONFIG_DEBUG_VM
1911         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1912                 WARN_ON(1);
1913                 up_read(&mm->mmap_sem);
1914         }
1915 #endif
1916 }
1917
1918 /*
1919  *  this is really a simplified "do_mmap".  it only handles
1920  *  anonymous maps.  eventually we may be able to do some
1921  *  brk-specific accounting here.
1922  */
1923 unsigned long do_brk(unsigned long addr, unsigned long len)
1924 {
1925         struct mm_struct * mm = current->mm;
1926         struct vm_area_struct * vma, * prev;
1927         unsigned long flags;
1928         struct rb_node ** rb_link, * rb_parent;
1929         pgoff_t pgoff = addr >> PAGE_SHIFT;
1930         int error;
1931
1932         len = PAGE_ALIGN(len);
1933         if (!len)
1934                 return addr;
1935
1936         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1937                 return -EINVAL;
1938
1939         if (is_hugepage_only_range(mm, addr, len))
1940                 return -EINVAL;
1941
1942         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1943
1944         error = arch_mmap_check(addr, len, flags);
1945         if (error)
1946                 return error;
1947
1948         /*
1949          * mlock MCL_FUTURE?
1950          */
1951         if (mm->def_flags & VM_LOCKED) {
1952                 unsigned long locked, lock_limit;
1953                 locked = len >> PAGE_SHIFT;
1954                 locked += mm->locked_vm;
1955                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1956                 lock_limit >>= PAGE_SHIFT;
1957                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1958                         return -EAGAIN;
1959         }
1960
1961         /*
1962          * mm->mmap_sem is required to protect against another thread
1963          * changing the mappings in case we sleep.
1964          */
1965         verify_mm_writelocked(mm);
1966
1967         /*
1968          * Clear old maps.  this also does some error checking for us
1969          */
1970  munmap_back:
1971         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1972         if (vma && vma->vm_start < addr + len) {
1973                 if (do_munmap(mm, addr, len))
1974                         return -ENOMEM;
1975                 goto munmap_back;
1976         }
1977
1978         /* Check against address space limits *after* clearing old maps... */
1979         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1980                 return -ENOMEM;
1981
1982         if (mm->map_count > sysctl_max_map_count)
1983                 return -ENOMEM;
1984
1985         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1986                 return -ENOMEM;
1987
1988         /* Can we just expand an old private anonymous mapping? */
1989         if (vma_merge(mm, prev, addr, addr + len, flags,
1990                                         NULL, NULL, pgoff, NULL))
1991                 goto out;
1992
1993         /*
1994          * create a vma struct for an anonymous mapping
1995          */
1996         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1997         if (!vma) {
1998                 vm_unacct_memory(len >> PAGE_SHIFT);
1999                 return -ENOMEM;
2000         }
2001
2002         vma->vm_mm = mm;
2003         vma->vm_start = addr;
2004         vma->vm_end = addr + len;
2005         vma->vm_pgoff = pgoff;
2006         vma->vm_flags = flags;
2007         vma->vm_page_prot = protection_map[flags &
2008                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
2009         vma_link(mm, vma, prev, rb_link, rb_parent);
2010 out:
2011         mm->total_vm += len >> PAGE_SHIFT;
2012         if (flags & VM_LOCKED) {
2013                 mm->locked_vm += len >> PAGE_SHIFT;
2014                 make_pages_present(addr, addr + len);
2015         }
2016         return addr;
2017 }
2018
2019 EXPORT_SYMBOL(do_brk);
2020
2021 /* Release all mmaps. */
2022 void exit_mmap(struct mm_struct *mm)
2023 {
2024         struct mmu_gather *tlb;
2025         struct vm_area_struct *vma = mm->mmap;
2026         unsigned long nr_accounted = 0;
2027         unsigned long end;
2028
2029         /* mm's last user has gone, and its about to be pulled down */
2030         arch_exit_mmap(mm);
2031
2032         lru_add_drain();
2033         flush_cache_mm(mm);
2034         tlb = tlb_gather_mmu(mm, 1);
2035         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2036         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2037         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2038         vm_unacct_memory(nr_accounted);
2039         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2040         tlb_finish_mmu(tlb, 0, end);
2041
2042         /*
2043          * Walk the list again, actually closing and freeing it,
2044          * with preemption enabled, without holding any MM locks.
2045          */
2046         while (vma)
2047                 vma = remove_vma(vma);
2048
2049         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2050 }
2051
2052 /* Insert vm structure into process list sorted by address
2053  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2054  * then i_mmap_lock is taken here.
2055  */
2056 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2057 {
2058         struct vm_area_struct * __vma, * prev;
2059         struct rb_node ** rb_link, * rb_parent;
2060
2061         /*
2062          * The vm_pgoff of a purely anonymous vma should be irrelevant
2063          * until its first write fault, when page's anon_vma and index
2064          * are set.  But now set the vm_pgoff it will almost certainly
2065          * end up with (unless mremap moves it elsewhere before that
2066          * first wfault), so /proc/pid/maps tells a consistent story.
2067          *
2068          * By setting it to reflect the virtual start address of the
2069          * vma, merges and splits can happen in a seamless way, just
2070          * using the existing file pgoff checks and manipulations.
2071          * Similarly in do_mmap_pgoff and in do_brk.
2072          */
2073         if (!vma->vm_file) {
2074                 BUG_ON(vma->anon_vma);
2075                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2076         }
2077         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2078         if (__vma && __vma->vm_start < vma->vm_end)
2079                 return -ENOMEM;
2080         if ((vma->vm_flags & VM_ACCOUNT) &&
2081              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2082                 return -ENOMEM;
2083         vma_link(mm, vma, prev, rb_link, rb_parent);
2084         return 0;
2085 }
2086
2087 /*
2088  * Copy the vma structure to a new location in the same mm,
2089  * prior to moving page table entries, to effect an mremap move.
2090  */
2091 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2092         unsigned long addr, unsigned long len, pgoff_t pgoff)
2093 {
2094         struct vm_area_struct *vma = *vmap;
2095         unsigned long vma_start = vma->vm_start;
2096         struct mm_struct *mm = vma->vm_mm;
2097         struct vm_area_struct *new_vma, *prev;
2098         struct rb_node **rb_link, *rb_parent;
2099         struct mempolicy *pol;
2100
2101         /*
2102          * If anonymous vma has not yet been faulted, update new pgoff
2103          * to match new location, to increase its chance of merging.
2104          */
2105         if (!vma->vm_file && !vma->anon_vma)
2106                 pgoff = addr >> PAGE_SHIFT;
2107
2108         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2109         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2110                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2111         if (new_vma) {
2112                 /*
2113                  * Source vma may have been merged into new_vma
2114                  */
2115                 if (vma_start >= new_vma->vm_start &&
2116                     vma_start < new_vma->vm_end)
2117                         *vmap = new_vma;
2118         } else {
2119                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2120                 if (new_vma) {
2121                         *new_vma = *vma;
2122                         pol = mpol_copy(vma_policy(vma));
2123                         if (IS_ERR(pol)) {
2124                                 kmem_cache_free(vm_area_cachep, new_vma);
2125                                 return NULL;
2126                         }
2127                         vma_set_policy(new_vma, pol);
2128                         new_vma->vm_start = addr;
2129                         new_vma->vm_end = addr + len;
2130                         new_vma->vm_pgoff = pgoff;
2131                         if (new_vma->vm_file)
2132                                 get_file(new_vma->vm_file);
2133                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2134                                 new_vma->vm_ops->open(new_vma);
2135                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2136                 }
2137         }
2138         return new_vma;
2139 }
2140
2141 /*
2142  * Return true if the calling process may expand its vm space by the passed
2143  * number of pages
2144  */
2145 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2146 {
2147         unsigned long cur = mm->total_vm;       /* pages */
2148         unsigned long lim;
2149
2150         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2151
2152         if (cur + npages > lim)
2153                 return 0;
2154         return 1;
2155 }
2156
2157
2158 static struct page *special_mapping_nopage(struct vm_area_struct *vma,
2159                                            unsigned long address, int *type)
2160 {
2161         struct page **pages;
2162
2163         BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2164
2165         address -= vma->vm_start;
2166         for (pages = vma->vm_private_data; address > 0 && *pages; ++pages)
2167                 address -= PAGE_SIZE;
2168
2169         if (*pages) {
2170                 struct page *page = *pages;
2171                 get_page(page);
2172                 return page;
2173         }
2174
2175         return NOPAGE_SIGBUS;
2176 }
2177
2178 /*
2179  * Having a close hook prevents vma merging regardless of flags.
2180  */
2181 static void special_mapping_close(struct vm_area_struct *vma)
2182 {
2183 }
2184
2185 static struct vm_operations_struct special_mapping_vmops = {
2186         .close = special_mapping_close,
2187         .nopage = special_mapping_nopage,
2188 };
2189
2190 /*
2191  * Called with mm->mmap_sem held for writing.
2192  * Insert a new vma covering the given region, with the given flags.
2193  * Its pages are supplied by the given array of struct page *.
2194  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2195  * The region past the last page supplied will always produce SIGBUS.
2196  * The array pointer and the pages it points to are assumed to stay alive
2197  * for as long as this mapping might exist.
2198  */
2199 int install_special_mapping(struct mm_struct *mm,
2200                             unsigned long addr, unsigned long len,
2201                             unsigned long vm_flags, struct page **pages)
2202 {
2203         struct vm_area_struct *vma;
2204
2205         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2206         if (unlikely(vma == NULL))
2207                 return -ENOMEM;
2208
2209         vma->vm_mm = mm;
2210         vma->vm_start = addr;
2211         vma->vm_end = addr + len;
2212
2213         vma->vm_flags = vm_flags | mm->def_flags;
2214         vma->vm_page_prot = protection_map[vma->vm_flags & 7];
2215
2216         vma->vm_ops = &special_mapping_vmops;
2217         vma->vm_private_data = pages;
2218
2219         if (unlikely(insert_vm_struct(mm, vma))) {
2220                 kmem_cache_free(vm_area_cachep, vma);
2221                 return -ENOMEM;
2222         }
2223
2224         mm->total_vm += len >> PAGE_SHIFT;
2225
2226         return 0;
2227 }