[PATCH] Small cleanup in quota.h
[linux-2.6] / 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/mm.h>
11 #include <linux/shm.h>
12 #include <linux/mman.h>
13 #include <linux/pagemap.h>
14 #include <linux/swap.h>
15 #include <linux/syscalls.h>
16 #include <linux/capability.h>
17 #include <linux/init.h>
18 #include <linux/file.h>
19 #include <linux/fs.h>
20 #include <linux/personality.h>
21 #include <linux/security.h>
22 #include <linux/hugetlb.h>
23 #include <linux/profile.h>
24 #include <linux/module.h>
25 #include <linux/mount.h>
26 #include <linux/mempolicy.h>
27 #include <linux/rmap.h>
28
29 #include <asm/uaccess.h>
30 #include <asm/cacheflush.h>
31 #include <asm/tlb.h>
32
33 static void unmap_region(struct mm_struct *mm,
34                 struct vm_area_struct *vma, struct vm_area_struct *prev,
35                 unsigned long start, unsigned long end);
36
37 /*
38  * WARNING: the debugging will use recursive algorithms so never enable this
39  * unless you know what you are doing.
40  */
41 #undef DEBUG_MM_RB
42
43 /* description of effects of mapping type and prot in current implementation.
44  * this is due to the limited x86 page protection hardware.  The expected
45  * behavior is in parens:
46  *
47  * map_type     prot
48  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
49  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
50  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
51  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
52  *              
53  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
54  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
55  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
56  *
57  */
58 pgprot_t protection_map[16] = {
59         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
60         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
61 };
62
63 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
64 int sysctl_overcommit_ratio = 50;       /* default is 50% */
65 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
66 atomic_t vm_committed_space = ATOMIC_INIT(0);
67
68 /*
69  * Check that a process has enough memory to allocate a new virtual
70  * mapping. 0 means there is enough memory for the allocation to
71  * succeed and -ENOMEM implies there is not.
72  *
73  * We currently support three overcommit policies, which are set via the
74  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
75  *
76  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
77  * Additional code 2002 Jul 20 by Robert Love.
78  *
79  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
80  *
81  * Note this is a helper function intended to be used by LSMs which
82  * wish to use this logic.
83  */
84 int __vm_enough_memory(long pages, int cap_sys_admin)
85 {
86         unsigned long free, allowed;
87
88         vm_acct_memory(pages);
89
90         /*
91          * Sometimes we want to use more memory than we have
92          */
93         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
94                 return 0;
95
96         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
97                 unsigned long n;
98
99                 free = get_page_cache_size();
100                 free += nr_swap_pages;
101
102                 /*
103                  * Any slabs which are created with the
104                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
105                  * which are reclaimable, under pressure.  The dentry
106                  * cache and most inode caches should fall into this
107                  */
108                 free += atomic_read(&slab_reclaim_pages);
109
110                 /*
111                  * Leave the last 3% for root
112                  */
113                 if (!cap_sys_admin)
114                         free -= free / 32;
115
116                 if (free > pages)
117                         return 0;
118
119                 /*
120                  * nr_free_pages() is very expensive on large systems,
121                  * only call if we're about to fail.
122                  */
123                 n = nr_free_pages();
124                 if (!cap_sys_admin)
125                         n -= n / 32;
126                 free += n;
127
128                 if (free > pages)
129                         return 0;
130                 vm_unacct_memory(pages);
131                 return -ENOMEM;
132         }
133
134         allowed = (totalram_pages - hugetlb_total_pages())
135                 * sysctl_overcommit_ratio / 100;
136         /*
137          * Leave the last 3% for root
138          */
139         if (!cap_sys_admin)
140                 allowed -= allowed / 32;
141         allowed += total_swap_pages;
142
143         /* Don't let a single process grow too big:
144            leave 3% of the size of this process for other processes */
145         allowed -= current->mm->total_vm / 32;
146
147         /*
148          * cast `allowed' as a signed long because vm_committed_space
149          * sometimes has a negative value
150          */
151         if (atomic_read(&vm_committed_space) < (long)allowed)
152                 return 0;
153
154         vm_unacct_memory(pages);
155
156         return -ENOMEM;
157 }
158
159 EXPORT_SYMBOL(__vm_enough_memory);
160
161 /*
162  * Requires inode->i_mapping->i_mmap_lock
163  */
164 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
165                 struct file *file, struct address_space *mapping)
166 {
167         if (vma->vm_flags & VM_DENYWRITE)
168                 atomic_inc(&file->f_dentry->d_inode->i_writecount);
169         if (vma->vm_flags & VM_SHARED)
170                 mapping->i_mmap_writable--;
171
172         flush_dcache_mmap_lock(mapping);
173         if (unlikely(vma->vm_flags & VM_NONLINEAR))
174                 list_del_init(&vma->shared.vm_set.list);
175         else
176                 vma_prio_tree_remove(vma, &mapping->i_mmap);
177         flush_dcache_mmap_unlock(mapping);
178 }
179
180 /*
181  * Unlink a file-based vm structure from its prio_tree, to hide
182  * vma from rmap and vmtruncate before freeing its page tables.
183  */
184 void unlink_file_vma(struct vm_area_struct *vma)
185 {
186         struct file *file = vma->vm_file;
187
188         if (file) {
189                 struct address_space *mapping = file->f_mapping;
190                 spin_lock(&mapping->i_mmap_lock);
191                 __remove_shared_vm_struct(vma, file, mapping);
192                 spin_unlock(&mapping->i_mmap_lock);
193         }
194 }
195
196 /*
197  * Close a vm structure and free it, returning the next.
198  */
199 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
200 {
201         struct vm_area_struct *next = vma->vm_next;
202
203         might_sleep();
204         if (vma->vm_ops && vma->vm_ops->close)
205                 vma->vm_ops->close(vma);
206         if (vma->vm_file)
207                 fput(vma->vm_file);
208         mpol_free(vma_policy(vma));
209         kmem_cache_free(vm_area_cachep, vma);
210         return next;
211 }
212
213 asmlinkage unsigned long sys_brk(unsigned long brk)
214 {
215         unsigned long rlim, retval;
216         unsigned long newbrk, oldbrk;
217         struct mm_struct *mm = current->mm;
218
219         down_write(&mm->mmap_sem);
220
221         if (brk < mm->end_code)
222                 goto out;
223         newbrk = PAGE_ALIGN(brk);
224         oldbrk = PAGE_ALIGN(mm->brk);
225         if (oldbrk == newbrk)
226                 goto set_brk;
227
228         /* Always allow shrinking brk. */
229         if (brk <= mm->brk) {
230                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
231                         goto set_brk;
232                 goto out;
233         }
234
235         /* Check against rlimit.. */
236         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
237         if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
238                 goto out;
239
240         /* Check against existing mmap mappings. */
241         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
242                 goto out;
243
244         /* Ok, looks good - let it rip. */
245         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
246                 goto out;
247 set_brk:
248         mm->brk = brk;
249 out:
250         retval = mm->brk;
251         up_write(&mm->mmap_sem);
252         return retval;
253 }
254
255 #ifdef DEBUG_MM_RB
256 static int browse_rb(struct rb_root *root)
257 {
258         int i = 0, j;
259         struct rb_node *nd, *pn = NULL;
260         unsigned long prev = 0, pend = 0;
261
262         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
263                 struct vm_area_struct *vma;
264                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
265                 if (vma->vm_start < prev)
266                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
267                 if (vma->vm_start < pend)
268                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
269                 if (vma->vm_start > vma->vm_end)
270                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
271                 i++;
272                 pn = nd;
273         }
274         j = 0;
275         for (nd = pn; nd; nd = rb_prev(nd)) {
276                 j++;
277         }
278         if (i != j)
279                 printk("backwards %d, forwards %d\n", j, i), i = 0;
280         return i;
281 }
282
283 void validate_mm(struct mm_struct *mm)
284 {
285         int bug = 0;
286         int i = 0;
287         struct vm_area_struct *tmp = mm->mmap;
288         while (tmp) {
289                 tmp = tmp->vm_next;
290                 i++;
291         }
292         if (i != mm->map_count)
293                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
294         i = browse_rb(&mm->mm_rb);
295         if (i != mm->map_count)
296                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
297         if (bug)
298                 BUG();
299 }
300 #else
301 #define validate_mm(mm) do { } while (0)
302 #endif
303
304 static struct vm_area_struct *
305 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
306                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
307                 struct rb_node ** rb_parent)
308 {
309         struct vm_area_struct * vma;
310         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
311
312         __rb_link = &mm->mm_rb.rb_node;
313         rb_prev = __rb_parent = NULL;
314         vma = NULL;
315
316         while (*__rb_link) {
317                 struct vm_area_struct *vma_tmp;
318
319                 __rb_parent = *__rb_link;
320                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
321
322                 if (vma_tmp->vm_end > addr) {
323                         vma = vma_tmp;
324                         if (vma_tmp->vm_start <= addr)
325                                 return vma;
326                         __rb_link = &__rb_parent->rb_left;
327                 } else {
328                         rb_prev = __rb_parent;
329                         __rb_link = &__rb_parent->rb_right;
330                 }
331         }
332
333         *pprev = NULL;
334         if (rb_prev)
335                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
336         *rb_link = __rb_link;
337         *rb_parent = __rb_parent;
338         return vma;
339 }
340
341 static inline void
342 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
343                 struct vm_area_struct *prev, struct rb_node *rb_parent)
344 {
345         if (prev) {
346                 vma->vm_next = prev->vm_next;
347                 prev->vm_next = vma;
348         } else {
349                 mm->mmap = vma;
350                 if (rb_parent)
351                         vma->vm_next = rb_entry(rb_parent,
352                                         struct vm_area_struct, vm_rb);
353                 else
354                         vma->vm_next = NULL;
355         }
356 }
357
358 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
359                 struct rb_node **rb_link, struct rb_node *rb_parent)
360 {
361         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
362         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
363 }
364
365 static inline void __vma_link_file(struct vm_area_struct *vma)
366 {
367         struct file * file;
368
369         file = vma->vm_file;
370         if (file) {
371                 struct address_space *mapping = file->f_mapping;
372
373                 if (vma->vm_flags & VM_DENYWRITE)
374                         atomic_dec(&file->f_dentry->d_inode->i_writecount);
375                 if (vma->vm_flags & VM_SHARED)
376                         mapping->i_mmap_writable++;
377
378                 flush_dcache_mmap_lock(mapping);
379                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
380                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
381                 else
382                         vma_prio_tree_insert(vma, &mapping->i_mmap);
383                 flush_dcache_mmap_unlock(mapping);
384         }
385 }
386
387 static void
388 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
389         struct vm_area_struct *prev, struct rb_node **rb_link,
390         struct rb_node *rb_parent)
391 {
392         __vma_link_list(mm, vma, prev, rb_parent);
393         __vma_link_rb(mm, vma, rb_link, rb_parent);
394         __anon_vma_link(vma);
395 }
396
397 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
398                         struct vm_area_struct *prev, struct rb_node **rb_link,
399                         struct rb_node *rb_parent)
400 {
401         struct address_space *mapping = NULL;
402
403         if (vma->vm_file)
404                 mapping = vma->vm_file->f_mapping;
405
406         if (mapping) {
407                 spin_lock(&mapping->i_mmap_lock);
408                 vma->vm_truncate_count = mapping->truncate_count;
409         }
410         anon_vma_lock(vma);
411
412         __vma_link(mm, vma, prev, rb_link, rb_parent);
413         __vma_link_file(vma);
414
415         anon_vma_unlock(vma);
416         if (mapping)
417                 spin_unlock(&mapping->i_mmap_lock);
418
419         mm->map_count++;
420         validate_mm(mm);
421 }
422
423 /*
424  * Helper for vma_adjust in the split_vma insert case:
425  * insert vm structure into list and rbtree and anon_vma,
426  * but it has already been inserted into prio_tree earlier.
427  */
428 static void
429 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
430 {
431         struct vm_area_struct * __vma, * prev;
432         struct rb_node ** rb_link, * rb_parent;
433
434         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
435         if (__vma && __vma->vm_start < vma->vm_end)
436                 BUG();
437         __vma_link(mm, vma, prev, rb_link, rb_parent);
438         mm->map_count++;
439 }
440
441 static inline void
442 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
443                 struct vm_area_struct *prev)
444 {
445         prev->vm_next = vma->vm_next;
446         rb_erase(&vma->vm_rb, &mm->mm_rb);
447         if (mm->mmap_cache == vma)
448                 mm->mmap_cache = prev;
449 }
450
451 /*
452  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
453  * is already present in an i_mmap tree without adjusting the tree.
454  * The following helper function should be used when such adjustments
455  * are necessary.  The "insert" vma (if any) is to be inserted
456  * before we drop the necessary locks.
457  */
458 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
459         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
460 {
461         struct mm_struct *mm = vma->vm_mm;
462         struct vm_area_struct *next = vma->vm_next;
463         struct vm_area_struct *importer = NULL;
464         struct address_space *mapping = NULL;
465         struct prio_tree_root *root = NULL;
466         struct file *file = vma->vm_file;
467         struct anon_vma *anon_vma = NULL;
468         long adjust_next = 0;
469         int remove_next = 0;
470
471         if (next && !insert) {
472                 if (end >= next->vm_end) {
473                         /*
474                          * vma expands, overlapping all the next, and
475                          * perhaps the one after too (mprotect case 6).
476                          */
477 again:                  remove_next = 1 + (end > next->vm_end);
478                         end = next->vm_end;
479                         anon_vma = next->anon_vma;
480                         importer = vma;
481                 } else if (end > next->vm_start) {
482                         /*
483                          * vma expands, overlapping part of the next:
484                          * mprotect case 5 shifting the boundary up.
485                          */
486                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
487                         anon_vma = next->anon_vma;
488                         importer = vma;
489                 } else if (end < vma->vm_end) {
490                         /*
491                          * vma shrinks, and !insert tells it's not
492                          * split_vma inserting another: so it must be
493                          * mprotect case 4 shifting the boundary down.
494                          */
495                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
496                         anon_vma = next->anon_vma;
497                         importer = next;
498                 }
499         }
500
501         if (file) {
502                 mapping = file->f_mapping;
503                 if (!(vma->vm_flags & VM_NONLINEAR))
504                         root = &mapping->i_mmap;
505                 spin_lock(&mapping->i_mmap_lock);
506                 if (importer &&
507                     vma->vm_truncate_count != next->vm_truncate_count) {
508                         /*
509                          * unmap_mapping_range might be in progress:
510                          * ensure that the expanding vma is rescanned.
511                          */
512                         importer->vm_truncate_count = 0;
513                 }
514                 if (insert) {
515                         insert->vm_truncate_count = vma->vm_truncate_count;
516                         /*
517                          * Put into prio_tree now, so instantiated pages
518                          * are visible to arm/parisc __flush_dcache_page
519                          * throughout; but we cannot insert into address
520                          * space until vma start or end is updated.
521                          */
522                         __vma_link_file(insert);
523                 }
524         }
525
526         /*
527          * When changing only vma->vm_end, we don't really need
528          * anon_vma lock: but is that case worth optimizing out?
529          */
530         if (vma->anon_vma)
531                 anon_vma = vma->anon_vma;
532         if (anon_vma) {
533                 spin_lock(&anon_vma->lock);
534                 /*
535                  * Easily overlooked: when mprotect shifts the boundary,
536                  * make sure the expanding vma has anon_vma set if the
537                  * shrinking vma had, to cover any anon pages imported.
538                  */
539                 if (importer && !importer->anon_vma) {
540                         importer->anon_vma = anon_vma;
541                         __anon_vma_link(importer);
542                 }
543         }
544
545         if (root) {
546                 flush_dcache_mmap_lock(mapping);
547                 vma_prio_tree_remove(vma, root);
548                 if (adjust_next)
549                         vma_prio_tree_remove(next, root);
550         }
551
552         vma->vm_start = start;
553         vma->vm_end = end;
554         vma->vm_pgoff = pgoff;
555         if (adjust_next) {
556                 next->vm_start += adjust_next << PAGE_SHIFT;
557                 next->vm_pgoff += adjust_next;
558         }
559
560         if (root) {
561                 if (adjust_next)
562                         vma_prio_tree_insert(next, root);
563                 vma_prio_tree_insert(vma, root);
564                 flush_dcache_mmap_unlock(mapping);
565         }
566
567         if (remove_next) {
568                 /*
569                  * vma_merge has merged next into vma, and needs
570                  * us to remove next before dropping the locks.
571                  */
572                 __vma_unlink(mm, next, vma);
573                 if (file)
574                         __remove_shared_vm_struct(next, file, mapping);
575                 if (next->anon_vma)
576                         __anon_vma_merge(vma, next);
577         } else if (insert) {
578                 /*
579                  * split_vma has split insert from vma, and needs
580                  * us to insert it before dropping the locks
581                  * (it may either follow vma or precede it).
582                  */
583                 __insert_vm_struct(mm, insert);
584         }
585
586         if (anon_vma)
587                 spin_unlock(&anon_vma->lock);
588         if (mapping)
589                 spin_unlock(&mapping->i_mmap_lock);
590
591         if (remove_next) {
592                 if (file)
593                         fput(file);
594                 mm->map_count--;
595                 mpol_free(vma_policy(next));
596                 kmem_cache_free(vm_area_cachep, next);
597                 /*
598                  * In mprotect's case 6 (see comments on vma_merge),
599                  * we must remove another next too. It would clutter
600                  * up the code too much to do both in one go.
601                  */
602                 if (remove_next == 2) {
603                         next = vma->vm_next;
604                         goto again;
605                 }
606         }
607
608         validate_mm(mm);
609 }
610
611 /*
612  * If the vma has a ->close operation then the driver probably needs to release
613  * per-vma resources, so we don't attempt to merge those.
614  */
615 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
616
617 static inline int is_mergeable_vma(struct vm_area_struct *vma,
618                         struct file *file, unsigned long vm_flags)
619 {
620         if (vma->vm_flags != vm_flags)
621                 return 0;
622         if (vma->vm_file != file)
623                 return 0;
624         if (vma->vm_ops && vma->vm_ops->close)
625                 return 0;
626         return 1;
627 }
628
629 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
630                                         struct anon_vma *anon_vma2)
631 {
632         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
633 }
634
635 /*
636  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
637  * in front of (at a lower virtual address and file offset than) the vma.
638  *
639  * We cannot merge two vmas if they have differently assigned (non-NULL)
640  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
641  *
642  * We don't check here for the merged mmap wrapping around the end of pagecache
643  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
644  * wrap, nor mmaps which cover the final page at index -1UL.
645  */
646 static int
647 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
648         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
649 {
650         if (is_mergeable_vma(vma, file, vm_flags) &&
651             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
652                 if (vma->vm_pgoff == vm_pgoff)
653                         return 1;
654         }
655         return 0;
656 }
657
658 /*
659  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
660  * beyond (at a higher virtual address and file offset than) the vma.
661  *
662  * We cannot merge two vmas if they have differently assigned (non-NULL)
663  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
664  */
665 static int
666 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
667         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
668 {
669         if (is_mergeable_vma(vma, file, vm_flags) &&
670             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
671                 pgoff_t vm_pglen;
672                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
673                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
674                         return 1;
675         }
676         return 0;
677 }
678
679 /*
680  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
681  * whether that can be merged with its predecessor or its successor.
682  * Or both (it neatly fills a hole).
683  *
684  * In most cases - when called for mmap, brk or mremap - [addr,end) is
685  * certain not to be mapped by the time vma_merge is called; but when
686  * called for mprotect, it is certain to be already mapped (either at
687  * an offset within prev, or at the start of next), and the flags of
688  * this area are about to be changed to vm_flags - and the no-change
689  * case has already been eliminated.
690  *
691  * The following mprotect cases have to be considered, where AAAA is
692  * the area passed down from mprotect_fixup, never extending beyond one
693  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
694  *
695  *     AAAA             AAAA                AAAA          AAAA
696  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
697  *    cannot merge    might become    might become    might become
698  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
699  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
700  *    mremap move:                                    PPPPNNNNNNNN 8
701  *        AAAA
702  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
703  *    might become    case 1 below    case 2 below    case 3 below
704  *
705  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
706  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
707  */
708 struct vm_area_struct *vma_merge(struct mm_struct *mm,
709                         struct vm_area_struct *prev, unsigned long addr,
710                         unsigned long end, unsigned long vm_flags,
711                         struct anon_vma *anon_vma, struct file *file,
712                         pgoff_t pgoff, struct mempolicy *policy)
713 {
714         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
715         struct vm_area_struct *area, *next;
716
717         /*
718          * We later require that vma->vm_flags == vm_flags,
719          * so this tests vma->vm_flags & VM_SPECIAL, too.
720          */
721         if (vm_flags & VM_SPECIAL)
722                 return NULL;
723
724         if (prev)
725                 next = prev->vm_next;
726         else
727                 next = mm->mmap;
728         area = next;
729         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
730                 next = next->vm_next;
731
732         /*
733          * Can it merge with the predecessor?
734          */
735         if (prev && prev->vm_end == addr &&
736                         mpol_equal(vma_policy(prev), policy) &&
737                         can_vma_merge_after(prev, vm_flags,
738                                                 anon_vma, file, pgoff)) {
739                 /*
740                  * OK, it can.  Can we now merge in the successor as well?
741                  */
742                 if (next && end == next->vm_start &&
743                                 mpol_equal(policy, vma_policy(next)) &&
744                                 can_vma_merge_before(next, vm_flags,
745                                         anon_vma, file, pgoff+pglen) &&
746                                 is_mergeable_anon_vma(prev->anon_vma,
747                                                       next->anon_vma)) {
748                                                         /* cases 1, 6 */
749                         vma_adjust(prev, prev->vm_start,
750                                 next->vm_end, prev->vm_pgoff, NULL);
751                 } else                                  /* cases 2, 5, 7 */
752                         vma_adjust(prev, prev->vm_start,
753                                 end, prev->vm_pgoff, NULL);
754                 return prev;
755         }
756
757         /*
758          * Can this new request be merged in front of next?
759          */
760         if (next && end == next->vm_start &&
761                         mpol_equal(policy, vma_policy(next)) &&
762                         can_vma_merge_before(next, vm_flags,
763                                         anon_vma, file, pgoff+pglen)) {
764                 if (prev && addr < prev->vm_end)        /* case 4 */
765                         vma_adjust(prev, prev->vm_start,
766                                 addr, prev->vm_pgoff, NULL);
767                 else                                    /* cases 3, 8 */
768                         vma_adjust(area, addr, next->vm_end,
769                                 next->vm_pgoff - pglen, NULL);
770                 return area;
771         }
772
773         return NULL;
774 }
775
776 /*
777  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
778  * neighbouring vmas for a suitable anon_vma, before it goes off
779  * to allocate a new anon_vma.  It checks because a repetitive
780  * sequence of mprotects and faults may otherwise lead to distinct
781  * anon_vmas being allocated, preventing vma merge in subsequent
782  * mprotect.
783  */
784 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
785 {
786         struct vm_area_struct *near;
787         unsigned long vm_flags;
788
789         near = vma->vm_next;
790         if (!near)
791                 goto try_prev;
792
793         /*
794          * Since only mprotect tries to remerge vmas, match flags
795          * which might be mprotected into each other later on.
796          * Neither mlock nor madvise tries to remerge at present,
797          * so leave their flags as obstructing a merge.
798          */
799         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
800         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
801
802         if (near->anon_vma && vma->vm_end == near->vm_start &&
803                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
804                         can_vma_merge_before(near, vm_flags,
805                                 NULL, vma->vm_file, vma->vm_pgoff +
806                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
807                 return near->anon_vma;
808 try_prev:
809         /*
810          * It is potentially slow to have to call find_vma_prev here.
811          * But it's only on the first write fault on the vma, not
812          * every time, and we could devise a way to avoid it later
813          * (e.g. stash info in next's anon_vma_node when assigning
814          * an anon_vma, or when trying vma_merge).  Another time.
815          */
816         if (find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma)
817                 BUG();
818         if (!near)
819                 goto none;
820
821         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
822         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
823
824         if (near->anon_vma && near->vm_end == vma->vm_start &&
825                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
826                         can_vma_merge_after(near, vm_flags,
827                                 NULL, vma->vm_file, vma->vm_pgoff))
828                 return near->anon_vma;
829 none:
830         /*
831          * There's no absolute need to look only at touching neighbours:
832          * we could search further afield for "compatible" anon_vmas.
833          * But it would probably just be a waste of time searching,
834          * or lead to too many vmas hanging off the same anon_vma.
835          * We're trying to allow mprotect remerging later on,
836          * not trying to minimize memory used for anon_vmas.
837          */
838         return NULL;
839 }
840
841 #ifdef CONFIG_PROC_FS
842 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
843                                                 struct file *file, long pages)
844 {
845         const unsigned long stack_flags
846                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
847
848         if (file) {
849                 mm->shared_vm += pages;
850                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
851                         mm->exec_vm += pages;
852         } else if (flags & stack_flags)
853                 mm->stack_vm += pages;
854         if (flags & (VM_RESERVED|VM_IO))
855                 mm->reserved_vm += pages;
856 }
857 #endif /* CONFIG_PROC_FS */
858
859 /*
860  * The caller must hold down_write(current->mm->mmap_sem).
861  */
862
863 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
864                         unsigned long len, unsigned long prot,
865                         unsigned long flags, unsigned long pgoff)
866 {
867         struct mm_struct * mm = current->mm;
868         struct vm_area_struct * vma, * prev;
869         struct inode *inode;
870         unsigned int vm_flags;
871         int correct_wcount = 0;
872         int error;
873         struct rb_node ** rb_link, * rb_parent;
874         int accountable = 1;
875         unsigned long charged = 0, reqprot = prot;
876
877         if (file) {
878                 if (is_file_hugepages(file))
879                         accountable = 0;
880
881                 if (!file->f_op || !file->f_op->mmap)
882                         return -ENODEV;
883
884                 if ((prot & PROT_EXEC) &&
885                     (file->f_vfsmnt->mnt_flags & MNT_NOEXEC))
886                         return -EPERM;
887         }
888         /*
889          * Does the application expect PROT_READ to imply PROT_EXEC?
890          *
891          * (the exception is when the underlying filesystem is noexec
892          *  mounted, in which case we dont add PROT_EXEC.)
893          */
894         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
895                 if (!(file && (file->f_vfsmnt->mnt_flags & MNT_NOEXEC)))
896                         prot |= PROT_EXEC;
897
898         if (!len)
899                 return -EINVAL;
900
901         /* Careful about overflows.. */
902         len = PAGE_ALIGN(len);
903         if (!len || len > TASK_SIZE)
904                 return -ENOMEM;
905
906         /* offset overflow? */
907         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
908                return -EOVERFLOW;
909
910         /* Too many mappings? */
911         if (mm->map_count > sysctl_max_map_count)
912                 return -ENOMEM;
913
914         /* Obtain the address to map to. we verify (or select) it and ensure
915          * that it represents a valid section of the address space.
916          */
917         addr = get_unmapped_area(file, addr, len, pgoff, flags);
918         if (addr & ~PAGE_MASK)
919                 return addr;
920
921         /* Do simple checking here so the lower-level routines won't have
922          * to. we assume access permissions have been handled by the open
923          * of the memory object, so we don't do any here.
924          */
925         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
926                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
927
928         if (flags & MAP_LOCKED) {
929                 if (!can_do_mlock())
930                         return -EPERM;
931                 vm_flags |= VM_LOCKED;
932         }
933         /* mlock MCL_FUTURE? */
934         if (vm_flags & VM_LOCKED) {
935                 unsigned long locked, lock_limit;
936                 locked = len >> PAGE_SHIFT;
937                 locked += mm->locked_vm;
938                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
939                 lock_limit >>= PAGE_SHIFT;
940                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
941                         return -EAGAIN;
942         }
943
944         inode = file ? file->f_dentry->d_inode : NULL;
945
946         if (file) {
947                 switch (flags & MAP_TYPE) {
948                 case MAP_SHARED:
949                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
950                                 return -EACCES;
951
952                         /*
953                          * Make sure we don't allow writing to an append-only
954                          * file..
955                          */
956                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
957                                 return -EACCES;
958
959                         /*
960                          * Make sure there are no mandatory locks on the file.
961                          */
962                         if (locks_verify_locked(inode))
963                                 return -EAGAIN;
964
965                         vm_flags |= VM_SHARED | VM_MAYSHARE;
966                         if (!(file->f_mode & FMODE_WRITE))
967                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
968
969                         /* fall through */
970                 case MAP_PRIVATE:
971                         if (!(file->f_mode & FMODE_READ))
972                                 return -EACCES;
973                         break;
974
975                 default:
976                         return -EINVAL;
977                 }
978         } else {
979                 switch (flags & MAP_TYPE) {
980                 case MAP_SHARED:
981                         vm_flags |= VM_SHARED | VM_MAYSHARE;
982                         break;
983                 case MAP_PRIVATE:
984                         /*
985                          * Set pgoff according to addr for anon_vma.
986                          */
987                         pgoff = addr >> PAGE_SHIFT;
988                         break;
989                 default:
990                         return -EINVAL;
991                 }
992         }
993
994         error = security_file_mmap(file, reqprot, prot, flags);
995         if (error)
996                 return error;
997                 
998         /* Clear old maps */
999         error = -ENOMEM;
1000 munmap_back:
1001         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1002         if (vma && vma->vm_start < addr + len) {
1003                 if (do_munmap(mm, addr, len))
1004                         return -ENOMEM;
1005                 goto munmap_back;
1006         }
1007
1008         /* Check against address space limit. */
1009         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1010                 return -ENOMEM;
1011
1012         if (accountable && (!(flags & MAP_NORESERVE) ||
1013                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1014                 if (vm_flags & VM_SHARED) {
1015                         /* Check memory availability in shmem_file_setup? */
1016                         vm_flags |= VM_ACCOUNT;
1017                 } else if (vm_flags & VM_WRITE) {
1018                         /*
1019                          * Private writable mapping: check memory availability
1020                          */
1021                         charged = len >> PAGE_SHIFT;
1022                         if (security_vm_enough_memory(charged))
1023                                 return -ENOMEM;
1024                         vm_flags |= VM_ACCOUNT;
1025                 }
1026         }
1027
1028         /*
1029          * Can we just expand an old private anonymous mapping?
1030          * The VM_SHARED test is necessary because shmem_zero_setup
1031          * will create the file object for a shared anonymous map below.
1032          */
1033         if (!file && !(vm_flags & VM_SHARED) &&
1034             vma_merge(mm, prev, addr, addr + len, vm_flags,
1035                                         NULL, NULL, pgoff, NULL))
1036                 goto out;
1037
1038         /*
1039          * Determine the object being mapped and call the appropriate
1040          * specific mapper. the address has already been validated, but
1041          * not unmapped, but the maps are removed from the list.
1042          */
1043         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1044         if (!vma) {
1045                 error = -ENOMEM;
1046                 goto unacct_error;
1047         }
1048
1049         vma->vm_mm = mm;
1050         vma->vm_start = addr;
1051         vma->vm_end = addr + len;
1052         vma->vm_flags = vm_flags;
1053         vma->vm_page_prot = protection_map[vm_flags & 0x0f];
1054         vma->vm_pgoff = pgoff;
1055
1056         if (file) {
1057                 error = -EINVAL;
1058                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1059                         goto free_vma;
1060                 if (vm_flags & VM_DENYWRITE) {
1061                         error = deny_write_access(file);
1062                         if (error)
1063                                 goto free_vma;
1064                         correct_wcount = 1;
1065                 }
1066                 vma->vm_file = file;
1067                 get_file(file);
1068                 error = file->f_op->mmap(file, vma);
1069                 if (error)
1070                         goto unmap_and_free_vma;
1071         } else if (vm_flags & VM_SHARED) {
1072                 error = shmem_zero_setup(vma);
1073                 if (error)
1074                         goto free_vma;
1075         }
1076
1077         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1078          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1079          * that memory reservation must be checked; but that reservation
1080          * belongs to shared memory object, not to vma: so now clear it.
1081          */
1082         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1083                 vma->vm_flags &= ~VM_ACCOUNT;
1084
1085         /* Can addr have changed??
1086          *
1087          * Answer: Yes, several device drivers can do it in their
1088          *         f_op->mmap method. -DaveM
1089          */
1090         addr = vma->vm_start;
1091         pgoff = vma->vm_pgoff;
1092         vm_flags = vma->vm_flags;
1093
1094         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1095                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1096                 file = vma->vm_file;
1097                 vma_link(mm, vma, prev, rb_link, rb_parent);
1098                 if (correct_wcount)
1099                         atomic_inc(&inode->i_writecount);
1100         } else {
1101                 if (file) {
1102                         if (correct_wcount)
1103                                 atomic_inc(&inode->i_writecount);
1104                         fput(file);
1105                 }
1106                 mpol_free(vma_policy(vma));
1107                 kmem_cache_free(vm_area_cachep, vma);
1108         }
1109 out:    
1110         mm->total_vm += len >> PAGE_SHIFT;
1111         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1112         if (vm_flags & VM_LOCKED) {
1113                 mm->locked_vm += len >> PAGE_SHIFT;
1114                 make_pages_present(addr, addr + len);
1115         }
1116         if (flags & MAP_POPULATE) {
1117                 up_write(&mm->mmap_sem);
1118                 sys_remap_file_pages(addr, len, 0,
1119                                         pgoff, flags & MAP_NONBLOCK);
1120                 down_write(&mm->mmap_sem);
1121         }
1122         return addr;
1123
1124 unmap_and_free_vma:
1125         if (correct_wcount)
1126                 atomic_inc(&inode->i_writecount);
1127         vma->vm_file = NULL;
1128         fput(file);
1129
1130         /* Undo any partial mapping done by a device driver. */
1131         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1132         charged = 0;
1133 free_vma:
1134         kmem_cache_free(vm_area_cachep, vma);
1135 unacct_error:
1136         if (charged)
1137                 vm_unacct_memory(charged);
1138         return error;
1139 }
1140
1141 EXPORT_SYMBOL(do_mmap_pgoff);
1142
1143 /* Get an address range which is currently unmapped.
1144  * For shmat() with addr=0.
1145  *
1146  * Ugly calling convention alert:
1147  * Return value with the low bits set means error value,
1148  * ie
1149  *      if (ret & ~PAGE_MASK)
1150  *              error = ret;
1151  *
1152  * This function "knows" that -ENOMEM has the bits set.
1153  */
1154 #ifndef HAVE_ARCH_UNMAPPED_AREA
1155 unsigned long
1156 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1157                 unsigned long len, unsigned long pgoff, unsigned long flags)
1158 {
1159         struct mm_struct *mm = current->mm;
1160         struct vm_area_struct *vma;
1161         unsigned long start_addr;
1162
1163         if (len > TASK_SIZE)
1164                 return -ENOMEM;
1165
1166         if (addr) {
1167                 addr = PAGE_ALIGN(addr);
1168                 vma = find_vma(mm, addr);
1169                 if (TASK_SIZE - len >= addr &&
1170                     (!vma || addr + len <= vma->vm_start))
1171                         return addr;
1172         }
1173         if (len > mm->cached_hole_size) {
1174                 start_addr = addr = mm->free_area_cache;
1175         } else {
1176                 start_addr = addr = TASK_UNMAPPED_BASE;
1177                 mm->cached_hole_size = 0;
1178         }
1179
1180 full_search:
1181         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1182                 /* At this point:  (!vma || addr < vma->vm_end). */
1183                 if (TASK_SIZE - len < addr) {
1184                         /*
1185                          * Start a new search - just in case we missed
1186                          * some holes.
1187                          */
1188                         if (start_addr != TASK_UNMAPPED_BASE) {
1189                                 addr = TASK_UNMAPPED_BASE;
1190                                 start_addr = addr;
1191                                 mm->cached_hole_size = 0;
1192                                 goto full_search;
1193                         }
1194                         return -ENOMEM;
1195                 }
1196                 if (!vma || addr + len <= vma->vm_start) {
1197                         /*
1198                          * Remember the place where we stopped the search:
1199                          */
1200                         mm->free_area_cache = addr + len;
1201                         return addr;
1202                 }
1203                 if (addr + mm->cached_hole_size < vma->vm_start)
1204                         mm->cached_hole_size = vma->vm_start - addr;
1205                 addr = vma->vm_end;
1206         }
1207 }
1208 #endif  
1209
1210 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1211 {
1212         /*
1213          * Is this a new hole at the lowest possible address?
1214          */
1215         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1216                 mm->free_area_cache = addr;
1217                 mm->cached_hole_size = ~0UL;
1218         }
1219 }
1220
1221 /*
1222  * This mmap-allocator allocates new areas top-down from below the
1223  * stack's low limit (the base):
1224  */
1225 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1226 unsigned long
1227 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1228                           const unsigned long len, const unsigned long pgoff,
1229                           const unsigned long flags)
1230 {
1231         struct vm_area_struct *vma;
1232         struct mm_struct *mm = current->mm;
1233         unsigned long addr = addr0;
1234
1235         /* requested length too big for entire address space */
1236         if (len > TASK_SIZE)
1237                 return -ENOMEM;
1238
1239         /* requesting a specific address */
1240         if (addr) {
1241                 addr = PAGE_ALIGN(addr);
1242                 vma = find_vma(mm, addr);
1243                 if (TASK_SIZE - len >= addr &&
1244                                 (!vma || addr + len <= vma->vm_start))
1245                         return addr;
1246         }
1247
1248         /* check if free_area_cache is useful for us */
1249         if (len <= mm->cached_hole_size) {
1250                 mm->cached_hole_size = 0;
1251                 mm->free_area_cache = mm->mmap_base;
1252         }
1253
1254         /* either no address requested or can't fit in requested address hole */
1255         addr = mm->free_area_cache;
1256
1257         /* make sure it can fit in the remaining address space */
1258         if (addr > len) {
1259                 vma = find_vma(mm, addr-len);
1260                 if (!vma || addr <= vma->vm_start)
1261                         /* remember the address as a hint for next time */
1262                         return (mm->free_area_cache = addr-len);
1263         }
1264
1265         if (mm->mmap_base < len)
1266                 goto bottomup;
1267
1268         addr = mm->mmap_base-len;
1269
1270         do {
1271                 /*
1272                  * Lookup failure means no vma is above this address,
1273                  * else if new region fits below vma->vm_start,
1274                  * return with success:
1275                  */
1276                 vma = find_vma(mm, addr);
1277                 if (!vma || addr+len <= vma->vm_start)
1278                         /* remember the address as a hint for next time */
1279                         return (mm->free_area_cache = addr);
1280
1281                 /* remember the largest hole we saw so far */
1282                 if (addr + mm->cached_hole_size < vma->vm_start)
1283                         mm->cached_hole_size = vma->vm_start - addr;
1284
1285                 /* try just below the current vma->vm_start */
1286                 addr = vma->vm_start-len;
1287         } while (len < vma->vm_start);
1288
1289 bottomup:
1290         /*
1291          * A failed mmap() very likely causes application failure,
1292          * so fall back to the bottom-up function here. This scenario
1293          * can happen with large stack limits and large mmap()
1294          * allocations.
1295          */
1296         mm->cached_hole_size = ~0UL;
1297         mm->free_area_cache = TASK_UNMAPPED_BASE;
1298         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1299         /*
1300          * Restore the topdown base:
1301          */
1302         mm->free_area_cache = mm->mmap_base;
1303         mm->cached_hole_size = ~0UL;
1304
1305         return addr;
1306 }
1307 #endif
1308
1309 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1310 {
1311         /*
1312          * Is this a new hole at the highest possible address?
1313          */
1314         if (addr > mm->free_area_cache)
1315                 mm->free_area_cache = addr;
1316
1317         /* dont allow allocations above current base */
1318         if (mm->free_area_cache > mm->mmap_base)
1319                 mm->free_area_cache = mm->mmap_base;
1320 }
1321
1322 unsigned long
1323 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1324                 unsigned long pgoff, unsigned long flags)
1325 {
1326         unsigned long ret;
1327
1328         if (!(flags & MAP_FIXED)) {
1329                 unsigned long (*get_area)(struct file *, unsigned long, unsigned long, unsigned long, unsigned long);
1330
1331                 get_area = current->mm->get_unmapped_area;
1332                 if (file && file->f_op && file->f_op->get_unmapped_area)
1333                         get_area = file->f_op->get_unmapped_area;
1334                 addr = get_area(file, addr, len, pgoff, flags);
1335                 if (IS_ERR_VALUE(addr))
1336                         return addr;
1337         }
1338
1339         if (addr > TASK_SIZE - len)
1340                 return -ENOMEM;
1341         if (addr & ~PAGE_MASK)
1342                 return -EINVAL;
1343         if (file && is_file_hugepages(file))  {
1344                 /*
1345                  * Check if the given range is hugepage aligned, and
1346                  * can be made suitable for hugepages.
1347                  */
1348                 ret = prepare_hugepage_range(addr, len);
1349         } else {
1350                 /*
1351                  * Ensure that a normal request is not falling in a
1352                  * reserved hugepage range.  For some archs like IA-64,
1353                  * there is a separate region for hugepages.
1354                  */
1355                 ret = is_hugepage_only_range(current->mm, addr, len);
1356         }
1357         if (ret)
1358                 return -EINVAL;
1359         return addr;
1360 }
1361
1362 EXPORT_SYMBOL(get_unmapped_area);
1363
1364 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1365 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1366 {
1367         struct vm_area_struct *vma = NULL;
1368
1369         if (mm) {
1370                 /* Check the cache first. */
1371                 /* (Cache hit rate is typically around 35%.) */
1372                 vma = mm->mmap_cache;
1373                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1374                         struct rb_node * rb_node;
1375
1376                         rb_node = mm->mm_rb.rb_node;
1377                         vma = NULL;
1378
1379                         while (rb_node) {
1380                                 struct vm_area_struct * vma_tmp;
1381
1382                                 vma_tmp = rb_entry(rb_node,
1383                                                 struct vm_area_struct, vm_rb);
1384
1385                                 if (vma_tmp->vm_end > addr) {
1386                                         vma = vma_tmp;
1387                                         if (vma_tmp->vm_start <= addr)
1388                                                 break;
1389                                         rb_node = rb_node->rb_left;
1390                                 } else
1391                                         rb_node = rb_node->rb_right;
1392                         }
1393                         if (vma)
1394                                 mm->mmap_cache = vma;
1395                 }
1396         }
1397         return vma;
1398 }
1399
1400 EXPORT_SYMBOL(find_vma);
1401
1402 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1403 struct vm_area_struct *
1404 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1405                         struct vm_area_struct **pprev)
1406 {
1407         struct vm_area_struct *vma = NULL, *prev = NULL;
1408         struct rb_node * rb_node;
1409         if (!mm)
1410                 goto out;
1411
1412         /* Guard against addr being lower than the first VMA */
1413         vma = mm->mmap;
1414
1415         /* Go through the RB tree quickly. */
1416         rb_node = mm->mm_rb.rb_node;
1417
1418         while (rb_node) {
1419                 struct vm_area_struct *vma_tmp;
1420                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1421
1422                 if (addr < vma_tmp->vm_end) {
1423                         rb_node = rb_node->rb_left;
1424                 } else {
1425                         prev = vma_tmp;
1426                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1427                                 break;
1428                         rb_node = rb_node->rb_right;
1429                 }
1430         }
1431
1432 out:
1433         *pprev = prev;
1434         return prev ? prev->vm_next : vma;
1435 }
1436
1437 /*
1438  * Verify that the stack growth is acceptable and
1439  * update accounting. This is shared with both the
1440  * grow-up and grow-down cases.
1441  */
1442 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1443 {
1444         struct mm_struct *mm = vma->vm_mm;
1445         struct rlimit *rlim = current->signal->rlim;
1446
1447         /* address space limit tests */
1448         if (!may_expand_vm(mm, grow))
1449                 return -ENOMEM;
1450
1451         /* Stack limit test */
1452         if (size > rlim[RLIMIT_STACK].rlim_cur)
1453                 return -ENOMEM;
1454
1455         /* mlock limit tests */
1456         if (vma->vm_flags & VM_LOCKED) {
1457                 unsigned long locked;
1458                 unsigned long limit;
1459                 locked = mm->locked_vm + grow;
1460                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1461                 if (locked > limit && !capable(CAP_IPC_LOCK))
1462                         return -ENOMEM;
1463         }
1464
1465         /*
1466          * Overcommit..  This must be the final test, as it will
1467          * update security statistics.
1468          */
1469         if (security_vm_enough_memory(grow))
1470                 return -ENOMEM;
1471
1472         /* Ok, everything looks good - let it rip */
1473         mm->total_vm += grow;
1474         if (vma->vm_flags & VM_LOCKED)
1475                 mm->locked_vm += grow;
1476         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1477         return 0;
1478 }
1479
1480 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1481 /*
1482  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1483  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1484  */
1485 #ifndef CONFIG_IA64
1486 static inline
1487 #endif
1488 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1489 {
1490         int error;
1491
1492         if (!(vma->vm_flags & VM_GROWSUP))
1493                 return -EFAULT;
1494
1495         /*
1496          * We must make sure the anon_vma is allocated
1497          * so that the anon_vma locking is not a noop.
1498          */
1499         if (unlikely(anon_vma_prepare(vma)))
1500                 return -ENOMEM;
1501         anon_vma_lock(vma);
1502
1503         /*
1504          * vma->vm_start/vm_end cannot change under us because the caller
1505          * is required to hold the mmap_sem in read mode.  We need the
1506          * anon_vma lock to serialize against concurrent expand_stacks.
1507          */
1508         address += 4 + PAGE_SIZE - 1;
1509         address &= PAGE_MASK;
1510         error = 0;
1511
1512         /* Somebody else might have raced and expanded it already */
1513         if (address > vma->vm_end) {
1514                 unsigned long size, grow;
1515
1516                 size = address - vma->vm_start;
1517                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1518
1519                 error = acct_stack_growth(vma, size, grow);
1520                 if (!error)
1521                         vma->vm_end = address;
1522         }
1523         anon_vma_unlock(vma);
1524         return error;
1525 }
1526 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1527
1528 #ifdef CONFIG_STACK_GROWSUP
1529 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1530 {
1531         return expand_upwards(vma, address);
1532 }
1533
1534 struct vm_area_struct *
1535 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1536 {
1537         struct vm_area_struct *vma, *prev;
1538
1539         addr &= PAGE_MASK;
1540         vma = find_vma_prev(mm, addr, &prev);
1541         if (vma && (vma->vm_start <= addr))
1542                 return vma;
1543         if (!prev || expand_stack(prev, addr))
1544                 return NULL;
1545         if (prev->vm_flags & VM_LOCKED) {
1546                 make_pages_present(addr, prev->vm_end);
1547         }
1548         return prev;
1549 }
1550 #else
1551 /*
1552  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1553  */
1554 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1555 {
1556         int error;
1557
1558         /*
1559          * We must make sure the anon_vma is allocated
1560          * so that the anon_vma locking is not a noop.
1561          */
1562         if (unlikely(anon_vma_prepare(vma)))
1563                 return -ENOMEM;
1564         anon_vma_lock(vma);
1565
1566         /*
1567          * vma->vm_start/vm_end cannot change under us because the caller
1568          * is required to hold the mmap_sem in read mode.  We need the
1569          * anon_vma lock to serialize against concurrent expand_stacks.
1570          */
1571         address &= PAGE_MASK;
1572         error = 0;
1573
1574         /* Somebody else might have raced and expanded it already */
1575         if (address < vma->vm_start) {
1576                 unsigned long size, grow;
1577
1578                 size = vma->vm_end - address;
1579                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1580
1581                 error = acct_stack_growth(vma, size, grow);
1582                 if (!error) {
1583                         vma->vm_start = address;
1584                         vma->vm_pgoff -= grow;
1585                 }
1586         }
1587         anon_vma_unlock(vma);
1588         return error;
1589 }
1590
1591 struct vm_area_struct *
1592 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1593 {
1594         struct vm_area_struct * vma;
1595         unsigned long start;
1596
1597         addr &= PAGE_MASK;
1598         vma = find_vma(mm,addr);
1599         if (!vma)
1600                 return NULL;
1601         if (vma->vm_start <= addr)
1602                 return vma;
1603         if (!(vma->vm_flags & VM_GROWSDOWN))
1604                 return NULL;
1605         start = vma->vm_start;
1606         if (expand_stack(vma, addr))
1607                 return NULL;
1608         if (vma->vm_flags & VM_LOCKED) {
1609                 make_pages_present(addr, start);
1610         }
1611         return vma;
1612 }
1613 #endif
1614
1615 /*
1616  * Ok - we have the memory areas we should free on the vma list,
1617  * so release them, and do the vma updates.
1618  *
1619  * Called with the mm semaphore held.
1620  */
1621 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1622 {
1623         /* Update high watermark before we lower total_vm */
1624         update_hiwater_vm(mm);
1625         do {
1626                 long nrpages = vma_pages(vma);
1627
1628                 mm->total_vm -= nrpages;
1629                 if (vma->vm_flags & VM_LOCKED)
1630                         mm->locked_vm -= nrpages;
1631                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1632                 vma = remove_vma(vma);
1633         } while (vma);
1634         validate_mm(mm);
1635 }
1636
1637 /*
1638  * Get rid of page table information in the indicated region.
1639  *
1640  * Called with the mm semaphore held.
1641  */
1642 static void unmap_region(struct mm_struct *mm,
1643                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1644                 unsigned long start, unsigned long end)
1645 {
1646         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1647         struct mmu_gather *tlb;
1648         unsigned long nr_accounted = 0;
1649
1650         lru_add_drain();
1651         tlb = tlb_gather_mmu(mm, 0);
1652         update_hiwater_rss(mm);
1653         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1654         vm_unacct_memory(nr_accounted);
1655         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1656                                  next? next->vm_start: 0);
1657         tlb_finish_mmu(tlb, start, end);
1658 }
1659
1660 /*
1661  * Create a list of vma's touched by the unmap, removing them from the mm's
1662  * vma list as we go..
1663  */
1664 static void
1665 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1666         struct vm_area_struct *prev, unsigned long end)
1667 {
1668         struct vm_area_struct **insertion_point;
1669         struct vm_area_struct *tail_vma = NULL;
1670         unsigned long addr;
1671
1672         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1673         do {
1674                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1675                 mm->map_count--;
1676                 tail_vma = vma;
1677                 vma = vma->vm_next;
1678         } while (vma && vma->vm_start < end);
1679         *insertion_point = vma;
1680         tail_vma->vm_next = NULL;
1681         if (mm->unmap_area == arch_unmap_area)
1682                 addr = prev ? prev->vm_end : mm->mmap_base;
1683         else
1684                 addr = vma ?  vma->vm_start : mm->mmap_base;
1685         mm->unmap_area(mm, addr);
1686         mm->mmap_cache = NULL;          /* Kill the cache. */
1687 }
1688
1689 /*
1690  * Split a vma into two pieces at address 'addr', a new vma is allocated
1691  * either for the first part or the the tail.
1692  */
1693 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1694               unsigned long addr, int new_below)
1695 {
1696         struct mempolicy *pol;
1697         struct vm_area_struct *new;
1698
1699         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1700                 return -EINVAL;
1701
1702         if (mm->map_count >= sysctl_max_map_count)
1703                 return -ENOMEM;
1704
1705         new = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1706         if (!new)
1707                 return -ENOMEM;
1708
1709         /* most fields are the same, copy all, and then fixup */
1710         *new = *vma;
1711
1712         if (new_below)
1713                 new->vm_end = addr;
1714         else {
1715                 new->vm_start = addr;
1716                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1717         }
1718
1719         pol = mpol_copy(vma_policy(vma));
1720         if (IS_ERR(pol)) {
1721                 kmem_cache_free(vm_area_cachep, new);
1722                 return PTR_ERR(pol);
1723         }
1724         vma_set_policy(new, pol);
1725
1726         if (new->vm_file)
1727                 get_file(new->vm_file);
1728
1729         if (new->vm_ops && new->vm_ops->open)
1730                 new->vm_ops->open(new);
1731
1732         if (new_below)
1733                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1734                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1735         else
1736                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1737
1738         return 0;
1739 }
1740
1741 /* Munmap is split into 2 main parts -- this part which finds
1742  * what needs doing, and the areas themselves, which do the
1743  * work.  This now handles partial unmappings.
1744  * Jeremy Fitzhardinge <jeremy@goop.org>
1745  */
1746 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1747 {
1748         unsigned long end;
1749         struct vm_area_struct *vma, *prev, *last;
1750
1751         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1752                 return -EINVAL;
1753
1754         if ((len = PAGE_ALIGN(len)) == 0)
1755                 return -EINVAL;
1756
1757         /* Find the first overlapping VMA */
1758         vma = find_vma_prev(mm, start, &prev);
1759         if (!vma)
1760                 return 0;
1761         /* we have  start < vma->vm_end  */
1762
1763         /* if it doesn't overlap, we have nothing.. */
1764         end = start + len;
1765         if (vma->vm_start >= end)
1766                 return 0;
1767
1768         /*
1769          * If we need to split any vma, do it now to save pain later.
1770          *
1771          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1772          * unmapped vm_area_struct will remain in use: so lower split_vma
1773          * places tmp vma above, and higher split_vma places tmp vma below.
1774          */
1775         if (start > vma->vm_start) {
1776                 int error = split_vma(mm, vma, start, 0);
1777                 if (error)
1778                         return error;
1779                 prev = vma;
1780         }
1781
1782         /* Does it split the last one? */
1783         last = find_vma(mm, end);
1784         if (last && end > last->vm_start) {
1785                 int error = split_vma(mm, last, end, 1);
1786                 if (error)
1787                         return error;
1788         }
1789         vma = prev? prev->vm_next: mm->mmap;
1790
1791         /*
1792          * Remove the vma's, and unmap the actual pages
1793          */
1794         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1795         unmap_region(mm, vma, prev, start, end);
1796
1797         /* Fix up all other VM information */
1798         remove_vma_list(mm, vma);
1799
1800         return 0;
1801 }
1802
1803 EXPORT_SYMBOL(do_munmap);
1804
1805 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1806 {
1807         int ret;
1808         struct mm_struct *mm = current->mm;
1809
1810         profile_munmap(addr);
1811
1812         down_write(&mm->mmap_sem);
1813         ret = do_munmap(mm, addr, len);
1814         up_write(&mm->mmap_sem);
1815         return ret;
1816 }
1817
1818 static inline void verify_mm_writelocked(struct mm_struct *mm)
1819 {
1820 #ifdef CONFIG_DEBUG_VM
1821         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1822                 WARN_ON(1);
1823                 up_read(&mm->mmap_sem);
1824         }
1825 #endif
1826 }
1827
1828 /*
1829  *  this is really a simplified "do_mmap".  it only handles
1830  *  anonymous maps.  eventually we may be able to do some
1831  *  brk-specific accounting here.
1832  */
1833 unsigned long do_brk(unsigned long addr, unsigned long len)
1834 {
1835         struct mm_struct * mm = current->mm;
1836         struct vm_area_struct * vma, * prev;
1837         unsigned long flags;
1838         struct rb_node ** rb_link, * rb_parent;
1839         pgoff_t pgoff = addr >> PAGE_SHIFT;
1840
1841         len = PAGE_ALIGN(len);
1842         if (!len)
1843                 return addr;
1844
1845         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1846                 return -EINVAL;
1847
1848         /*
1849          * mlock MCL_FUTURE?
1850          */
1851         if (mm->def_flags & VM_LOCKED) {
1852                 unsigned long locked, lock_limit;
1853                 locked = len >> PAGE_SHIFT;
1854                 locked += mm->locked_vm;
1855                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1856                 lock_limit >>= PAGE_SHIFT;
1857                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1858                         return -EAGAIN;
1859         }
1860
1861         /*
1862          * mm->mmap_sem is required to protect against another thread
1863          * changing the mappings in case we sleep.
1864          */
1865         verify_mm_writelocked(mm);
1866
1867         /*
1868          * Clear old maps.  this also does some error checking for us
1869          */
1870  munmap_back:
1871         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1872         if (vma && vma->vm_start < addr + len) {
1873                 if (do_munmap(mm, addr, len))
1874                         return -ENOMEM;
1875                 goto munmap_back;
1876         }
1877
1878         /* Check against address space limits *after* clearing old maps... */
1879         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1880                 return -ENOMEM;
1881
1882         if (mm->map_count > sysctl_max_map_count)
1883                 return -ENOMEM;
1884
1885         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1886                 return -ENOMEM;
1887
1888         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1889
1890         /* Can we just expand an old private anonymous mapping? */
1891         if (vma_merge(mm, prev, addr, addr + len, flags,
1892                                         NULL, NULL, pgoff, NULL))
1893                 goto out;
1894
1895         /*
1896          * create a vma struct for an anonymous mapping
1897          */
1898         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1899         if (!vma) {
1900                 vm_unacct_memory(len >> PAGE_SHIFT);
1901                 return -ENOMEM;
1902         }
1903
1904         vma->vm_mm = mm;
1905         vma->vm_start = addr;
1906         vma->vm_end = addr + len;
1907         vma->vm_pgoff = pgoff;
1908         vma->vm_flags = flags;
1909         vma->vm_page_prot = protection_map[flags & 0x0f];
1910         vma_link(mm, vma, prev, rb_link, rb_parent);
1911 out:
1912         mm->total_vm += len >> PAGE_SHIFT;
1913         if (flags & VM_LOCKED) {
1914                 mm->locked_vm += len >> PAGE_SHIFT;
1915                 make_pages_present(addr, addr + len);
1916         }
1917         return addr;
1918 }
1919
1920 EXPORT_SYMBOL(do_brk);
1921
1922 /* Release all mmaps. */
1923 void exit_mmap(struct mm_struct *mm)
1924 {
1925         struct mmu_gather *tlb;
1926         struct vm_area_struct *vma = mm->mmap;
1927         unsigned long nr_accounted = 0;
1928         unsigned long end;
1929
1930         lru_add_drain();
1931         flush_cache_mm(mm);
1932         tlb = tlb_gather_mmu(mm, 1);
1933         /* Don't update_hiwater_rss(mm) here, do_exit already did */
1934         /* Use -1 here to ensure all VMAs in the mm are unmapped */
1935         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
1936         vm_unacct_memory(nr_accounted);
1937         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
1938         tlb_finish_mmu(tlb, 0, end);
1939
1940         /*
1941          * Walk the list again, actually closing and freeing it,
1942          * with preemption enabled, without holding any MM locks.
1943          */
1944         while (vma)
1945                 vma = remove_vma(vma);
1946
1947         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
1948 }
1949
1950 /* Insert vm structure into process list sorted by address
1951  * and into the inode's i_mmap tree.  If vm_file is non-NULL
1952  * then i_mmap_lock is taken here.
1953  */
1954 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
1955 {
1956         struct vm_area_struct * __vma, * prev;
1957         struct rb_node ** rb_link, * rb_parent;
1958
1959         /*
1960          * The vm_pgoff of a purely anonymous vma should be irrelevant
1961          * until its first write fault, when page's anon_vma and index
1962          * are set.  But now set the vm_pgoff it will almost certainly
1963          * end up with (unless mremap moves it elsewhere before that
1964          * first wfault), so /proc/pid/maps tells a consistent story.
1965          *
1966          * By setting it to reflect the virtual start address of the
1967          * vma, merges and splits can happen in a seamless way, just
1968          * using the existing file pgoff checks and manipulations.
1969          * Similarly in do_mmap_pgoff and in do_brk.
1970          */
1971         if (!vma->vm_file) {
1972                 BUG_ON(vma->anon_vma);
1973                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
1974         }
1975         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
1976         if (__vma && __vma->vm_start < vma->vm_end)
1977                 return -ENOMEM;
1978         if ((vma->vm_flags & VM_ACCOUNT) &&
1979              security_vm_enough_memory(vma_pages(vma)))
1980                 return -ENOMEM;
1981         vma_link(mm, vma, prev, rb_link, rb_parent);
1982         return 0;
1983 }
1984
1985 /*
1986  * Copy the vma structure to a new location in the same mm,
1987  * prior to moving page table entries, to effect an mremap move.
1988  */
1989 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
1990         unsigned long addr, unsigned long len, pgoff_t pgoff)
1991 {
1992         struct vm_area_struct *vma = *vmap;
1993         unsigned long vma_start = vma->vm_start;
1994         struct mm_struct *mm = vma->vm_mm;
1995         struct vm_area_struct *new_vma, *prev;
1996         struct rb_node **rb_link, *rb_parent;
1997         struct mempolicy *pol;
1998
1999         /*
2000          * If anonymous vma has not yet been faulted, update new pgoff
2001          * to match new location, to increase its chance of merging.
2002          */
2003         if (!vma->vm_file && !vma->anon_vma)
2004                 pgoff = addr >> PAGE_SHIFT;
2005
2006         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2007         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2008                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2009         if (new_vma) {
2010                 /*
2011                  * Source vma may have been merged into new_vma
2012                  */
2013                 if (vma_start >= new_vma->vm_start &&
2014                     vma_start < new_vma->vm_end)
2015                         *vmap = new_vma;
2016         } else {
2017                 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2018                 if (new_vma) {
2019                         *new_vma = *vma;
2020                         pol = mpol_copy(vma_policy(vma));
2021                         if (IS_ERR(pol)) {
2022                                 kmem_cache_free(vm_area_cachep, new_vma);
2023                                 return NULL;
2024                         }
2025                         vma_set_policy(new_vma, pol);
2026                         new_vma->vm_start = addr;
2027                         new_vma->vm_end = addr + len;
2028                         new_vma->vm_pgoff = pgoff;
2029                         if (new_vma->vm_file)
2030                                 get_file(new_vma->vm_file);
2031                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2032                                 new_vma->vm_ops->open(new_vma);
2033                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2034                 }
2035         }
2036         return new_vma;
2037 }
2038
2039 /*
2040  * Return true if the calling process may expand its vm space by the passed
2041  * number of pages
2042  */
2043 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2044 {
2045         unsigned long cur = mm->total_vm;       /* pages */
2046         unsigned long lim;
2047
2048         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2049
2050         if (cur + npages > lim)
2051                 return 0;
2052         return 1;
2053 }