Driver core: more fallout from class_device changes for pcmcia
[linux-2.6.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/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 #ifndef arch_mmap_check
34 #define arch_mmap_check(addr, len, flags)       (0)
35 #endif
36
37 static void unmap_region(struct mm_struct *mm,
38                 struct vm_area_struct *vma, struct vm_area_struct *prev,
39                 unsigned long start, unsigned long end);
40
41 /*
42  * WARNING: the debugging will use recursive algorithms so never enable this
43  * unless you know what you are doing.
44  */
45 #undef DEBUG_MM_RB
46
47 /* description of effects of mapping type and prot in current implementation.
48  * this is due to the limited x86 page protection hardware.  The expected
49  * behavior is in parens:
50  *
51  * map_type     prot
52  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
53  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
54  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
55  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
56  *              
57  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
58  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
59  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
60  *
61  */
62 pgprot_t protection_map[16] = {
63         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
64         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
65 };
66
67 pgprot_t vm_get_page_prot(unsigned long vm_flags)
68 {
69         return protection_map[vm_flags &
70                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
71 }
72 EXPORT_SYMBOL(vm_get_page_prot);
73
74 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
75 int sysctl_overcommit_ratio = 50;       /* default is 50% */
76 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
77 atomic_t vm_committed_space = ATOMIC_INIT(0);
78
79 /*
80  * Check that a process has enough memory to allocate a new virtual
81  * mapping. 0 means there is enough memory for the allocation to
82  * succeed and -ENOMEM implies there is not.
83  *
84  * We currently support three overcommit policies, which are set via the
85  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
86  *
87  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
88  * Additional code 2002 Jul 20 by Robert Love.
89  *
90  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
91  *
92  * Note this is a helper function intended to be used by LSMs which
93  * wish to use this logic.
94  */
95 int __vm_enough_memory(long pages, int cap_sys_admin)
96 {
97         unsigned long free, allowed;
98
99         vm_acct_memory(pages);
100
101         /*
102          * Sometimes we want to use more memory than we have
103          */
104         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
105                 return 0;
106
107         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
108                 unsigned long n;
109
110                 free = global_page_state(NR_FILE_PAGES);
111                 free += nr_swap_pages;
112
113                 /*
114                  * Any slabs which are created with the
115                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
116                  * which are reclaimable, under pressure.  The dentry
117                  * cache and most inode caches should fall into this
118                  */
119                 free += global_page_state(NR_SLAB_RECLAIMABLE);
120
121                 /*
122                  * Leave the last 3% for root
123                  */
124                 if (!cap_sys_admin)
125                         free -= free / 32;
126
127                 if (free > pages)
128                         return 0;
129
130                 /*
131                  * nr_free_pages() is very expensive on large systems,
132                  * only call if we're about to fail.
133                  */
134                 n = nr_free_pages();
135
136                 /*
137                  * Leave reserved pages. The pages are not for anonymous pages.
138                  */
139                 if (n <= totalreserve_pages)
140                         goto error;
141                 else
142                         n -= totalreserve_pages;
143
144                 /*
145                  * Leave the last 3% for root
146                  */
147                 if (!cap_sys_admin)
148                         n -= n / 32;
149                 free += n;
150
151                 if (free > pages)
152                         return 0;
153
154                 goto error;
155         }
156
157         allowed = (totalram_pages - hugetlb_total_pages())
158                 * sysctl_overcommit_ratio / 100;
159         /*
160          * Leave the last 3% for root
161          */
162         if (!cap_sys_admin)
163                 allowed -= allowed / 32;
164         allowed += total_swap_pages;
165
166         /* Don't let a single process grow too big:
167            leave 3% of the size of this process for other processes */
168         allowed -= current->mm->total_vm / 32;
169
170         /*
171          * cast `allowed' as a signed long because vm_committed_space
172          * sometimes has a negative value
173          */
174         if (atomic_read(&vm_committed_space) < (long)allowed)
175                 return 0;
176 error:
177         vm_unacct_memory(pages);
178
179         return -ENOMEM;
180 }
181
182 EXPORT_SYMBOL(__vm_enough_memory);
183
184 /*
185  * Requires inode->i_mapping->i_mmap_lock
186  */
187 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
188                 struct file *file, struct address_space *mapping)
189 {
190         if (vma->vm_flags & VM_DENYWRITE)
191                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
192         if (vma->vm_flags & VM_SHARED)
193                 mapping->i_mmap_writable--;
194
195         flush_dcache_mmap_lock(mapping);
196         if (unlikely(vma->vm_flags & VM_NONLINEAR))
197                 list_del_init(&vma->shared.vm_set.list);
198         else
199                 vma_prio_tree_remove(vma, &mapping->i_mmap);
200         flush_dcache_mmap_unlock(mapping);
201 }
202
203 /*
204  * Unlink a file-based vm structure from its prio_tree, to hide
205  * vma from rmap and vmtruncate before freeing its page tables.
206  */
207 void unlink_file_vma(struct vm_area_struct *vma)
208 {
209         struct file *file = vma->vm_file;
210
211         if (file) {
212                 struct address_space *mapping = file->f_mapping;
213                 spin_lock(&mapping->i_mmap_lock);
214                 __remove_shared_vm_struct(vma, file, mapping);
215                 spin_unlock(&mapping->i_mmap_lock);
216         }
217 }
218
219 /*
220  * Close a vm structure and free it, returning the next.
221  */
222 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
223 {
224         struct vm_area_struct *next = vma->vm_next;
225
226         might_sleep();
227         if (vma->vm_ops && vma->vm_ops->close)
228                 vma->vm_ops->close(vma);
229         if (vma->vm_file)
230                 fput(vma->vm_file);
231         mpol_free(vma_policy(vma));
232         kmem_cache_free(vm_area_cachep, vma);
233         return next;
234 }
235
236 asmlinkage unsigned long sys_brk(unsigned long brk)
237 {
238         unsigned long rlim, retval;
239         unsigned long newbrk, oldbrk;
240         struct mm_struct *mm = current->mm;
241
242         down_write(&mm->mmap_sem);
243
244         if (brk < mm->end_code)
245                 goto out;
246
247         /*
248          * Check against rlimit here. If this check is done later after the test
249          * of oldbrk with newbrk then it can escape the test and let the data
250          * segment grow beyond its set limit the in case where the limit is
251          * not page aligned -Ram Gupta
252          */
253         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
254         if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
255                 goto out;
256
257         newbrk = PAGE_ALIGN(brk);
258         oldbrk = PAGE_ALIGN(mm->brk);
259         if (oldbrk == newbrk)
260                 goto set_brk;
261
262         /* Always allow shrinking brk. */
263         if (brk <= mm->brk) {
264                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
265                         goto set_brk;
266                 goto out;
267         }
268
269         /* Check against existing mmap mappings. */
270         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
271                 goto out;
272
273         /* Ok, looks good - let it rip. */
274         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
275                 goto out;
276 set_brk:
277         mm->brk = brk;
278 out:
279         retval = mm->brk;
280         up_write(&mm->mmap_sem);
281         return retval;
282 }
283
284 #ifdef DEBUG_MM_RB
285 static int browse_rb(struct rb_root *root)
286 {
287         int i = 0, j;
288         struct rb_node *nd, *pn = NULL;
289         unsigned long prev = 0, pend = 0;
290
291         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
292                 struct vm_area_struct *vma;
293                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
294                 if (vma->vm_start < prev)
295                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
296                 if (vma->vm_start < pend)
297                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
298                 if (vma->vm_start > vma->vm_end)
299                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
300                 i++;
301                 pn = nd;
302         }
303         j = 0;
304         for (nd = pn; nd; nd = rb_prev(nd)) {
305                 j++;
306         }
307         if (i != j)
308                 printk("backwards %d, forwards %d\n", j, i), i = 0;
309         return i;
310 }
311
312 void validate_mm(struct mm_struct *mm)
313 {
314         int bug = 0;
315         int i = 0;
316         struct vm_area_struct *tmp = mm->mmap;
317         while (tmp) {
318                 tmp = tmp->vm_next;
319                 i++;
320         }
321         if (i != mm->map_count)
322                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
323         i = browse_rb(&mm->mm_rb);
324         if (i != mm->map_count)
325                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
326         BUG_ON(bug);
327 }
328 #else
329 #define validate_mm(mm) do { } while (0)
330 #endif
331
332 static struct vm_area_struct *
333 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
334                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
335                 struct rb_node ** rb_parent)
336 {
337         struct vm_area_struct * vma;
338         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
339
340         __rb_link = &mm->mm_rb.rb_node;
341         rb_prev = __rb_parent = NULL;
342         vma = NULL;
343
344         while (*__rb_link) {
345                 struct vm_area_struct *vma_tmp;
346
347                 __rb_parent = *__rb_link;
348                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
349
350                 if (vma_tmp->vm_end > addr) {
351                         vma = vma_tmp;
352                         if (vma_tmp->vm_start <= addr)
353                                 return vma;
354                         __rb_link = &__rb_parent->rb_left;
355                 } else {
356                         rb_prev = __rb_parent;
357                         __rb_link = &__rb_parent->rb_right;
358                 }
359         }
360
361         *pprev = NULL;
362         if (rb_prev)
363                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
364         *rb_link = __rb_link;
365         *rb_parent = __rb_parent;
366         return vma;
367 }
368
369 static inline void
370 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
371                 struct vm_area_struct *prev, struct rb_node *rb_parent)
372 {
373         if (prev) {
374                 vma->vm_next = prev->vm_next;
375                 prev->vm_next = vma;
376         } else {
377                 mm->mmap = vma;
378                 if (rb_parent)
379                         vma->vm_next = rb_entry(rb_parent,
380                                         struct vm_area_struct, vm_rb);
381                 else
382                         vma->vm_next = NULL;
383         }
384 }
385
386 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
387                 struct rb_node **rb_link, struct rb_node *rb_parent)
388 {
389         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
390         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
391 }
392
393 static inline void __vma_link_file(struct vm_area_struct *vma)
394 {
395         struct file * file;
396
397         file = vma->vm_file;
398         if (file) {
399                 struct address_space *mapping = file->f_mapping;
400
401                 if (vma->vm_flags & VM_DENYWRITE)
402                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
403                 if (vma->vm_flags & VM_SHARED)
404                         mapping->i_mmap_writable++;
405
406                 flush_dcache_mmap_lock(mapping);
407                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
408                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
409                 else
410                         vma_prio_tree_insert(vma, &mapping->i_mmap);
411                 flush_dcache_mmap_unlock(mapping);
412         }
413 }
414
415 static void
416 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
417         struct vm_area_struct *prev, struct rb_node **rb_link,
418         struct rb_node *rb_parent)
419 {
420         __vma_link_list(mm, vma, prev, rb_parent);
421         __vma_link_rb(mm, vma, rb_link, rb_parent);
422         __anon_vma_link(vma);
423 }
424
425 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
426                         struct vm_area_struct *prev, struct rb_node **rb_link,
427                         struct rb_node *rb_parent)
428 {
429         struct address_space *mapping = NULL;
430
431         if (vma->vm_file)
432                 mapping = vma->vm_file->f_mapping;
433
434         if (mapping) {
435                 spin_lock(&mapping->i_mmap_lock);
436                 vma->vm_truncate_count = mapping->truncate_count;
437         }
438         anon_vma_lock(vma);
439
440         __vma_link(mm, vma, prev, rb_link, rb_parent);
441         __vma_link_file(vma);
442
443         anon_vma_unlock(vma);
444         if (mapping)
445                 spin_unlock(&mapping->i_mmap_lock);
446
447         mm->map_count++;
448         validate_mm(mm);
449 }
450
451 /*
452  * Helper for vma_adjust in the split_vma insert case:
453  * insert vm structure into list and rbtree and anon_vma,
454  * but it has already been inserted into prio_tree earlier.
455  */
456 static void
457 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
458 {
459         struct vm_area_struct * __vma, * prev;
460         struct rb_node ** rb_link, * rb_parent;
461
462         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
463         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
464         __vma_link(mm, vma, prev, rb_link, rb_parent);
465         mm->map_count++;
466 }
467
468 static inline void
469 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
470                 struct vm_area_struct *prev)
471 {
472         prev->vm_next = vma->vm_next;
473         rb_erase(&vma->vm_rb, &mm->mm_rb);
474         if (mm->mmap_cache == vma)
475                 mm->mmap_cache = prev;
476 }
477
478 /*
479  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
480  * is already present in an i_mmap tree without adjusting the tree.
481  * The following helper function should be used when such adjustments
482  * are necessary.  The "insert" vma (if any) is to be inserted
483  * before we drop the necessary locks.
484  */
485 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
486         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
487 {
488         struct mm_struct *mm = vma->vm_mm;
489         struct vm_area_struct *next = vma->vm_next;
490         struct vm_area_struct *importer = NULL;
491         struct address_space *mapping = NULL;
492         struct prio_tree_root *root = NULL;
493         struct file *file = vma->vm_file;
494         struct anon_vma *anon_vma = NULL;
495         long adjust_next = 0;
496         int remove_next = 0;
497
498         if (next && !insert) {
499                 if (end >= next->vm_end) {
500                         /*
501                          * vma expands, overlapping all the next, and
502                          * perhaps the one after too (mprotect case 6).
503                          */
504 again:                  remove_next = 1 + (end > next->vm_end);
505                         end = next->vm_end;
506                         anon_vma = next->anon_vma;
507                         importer = vma;
508                 } else if (end > next->vm_start) {
509                         /*
510                          * vma expands, overlapping part of the next:
511                          * mprotect case 5 shifting the boundary up.
512                          */
513                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
514                         anon_vma = next->anon_vma;
515                         importer = vma;
516                 } else if (end < vma->vm_end) {
517                         /*
518                          * vma shrinks, and !insert tells it's not
519                          * split_vma inserting another: so it must be
520                          * mprotect case 4 shifting the boundary down.
521                          */
522                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
523                         anon_vma = next->anon_vma;
524                         importer = next;
525                 }
526         }
527
528         if (file) {
529                 mapping = file->f_mapping;
530                 if (!(vma->vm_flags & VM_NONLINEAR))
531                         root = &mapping->i_mmap;
532                 spin_lock(&mapping->i_mmap_lock);
533                 if (importer &&
534                     vma->vm_truncate_count != next->vm_truncate_count) {
535                         /*
536                          * unmap_mapping_range might be in progress:
537                          * ensure that the expanding vma is rescanned.
538                          */
539                         importer->vm_truncate_count = 0;
540                 }
541                 if (insert) {
542                         insert->vm_truncate_count = vma->vm_truncate_count;
543                         /*
544                          * Put into prio_tree now, so instantiated pages
545                          * are visible to arm/parisc __flush_dcache_page
546                          * throughout; but we cannot insert into address
547                          * space until vma start or end is updated.
548                          */
549                         __vma_link_file(insert);
550                 }
551         }
552
553         /*
554          * When changing only vma->vm_end, we don't really need
555          * anon_vma lock: but is that case worth optimizing out?
556          */
557         if (vma->anon_vma)
558                 anon_vma = vma->anon_vma;
559         if (anon_vma) {
560                 spin_lock(&anon_vma->lock);
561                 /*
562                  * Easily overlooked: when mprotect shifts the boundary,
563                  * make sure the expanding vma has anon_vma set if the
564                  * shrinking vma had, to cover any anon pages imported.
565                  */
566                 if (importer && !importer->anon_vma) {
567                         importer->anon_vma = anon_vma;
568                         __anon_vma_link(importer);
569                 }
570         }
571
572         if (root) {
573                 flush_dcache_mmap_lock(mapping);
574                 vma_prio_tree_remove(vma, root);
575                 if (adjust_next)
576                         vma_prio_tree_remove(next, root);
577         }
578
579         vma->vm_start = start;
580         vma->vm_end = end;
581         vma->vm_pgoff = pgoff;
582         if (adjust_next) {
583                 next->vm_start += adjust_next << PAGE_SHIFT;
584                 next->vm_pgoff += adjust_next;
585         }
586
587         if (root) {
588                 if (adjust_next)
589                         vma_prio_tree_insert(next, root);
590                 vma_prio_tree_insert(vma, root);
591                 flush_dcache_mmap_unlock(mapping);
592         }
593
594         if (remove_next) {
595                 /*
596                  * vma_merge has merged next into vma, and needs
597                  * us to remove next before dropping the locks.
598                  */
599                 __vma_unlink(mm, next, vma);
600                 if (file)
601                         __remove_shared_vm_struct(next, file, mapping);
602                 if (next->anon_vma)
603                         __anon_vma_merge(vma, next);
604         } else if (insert) {
605                 /*
606                  * split_vma has split insert from vma, and needs
607                  * us to insert it before dropping the locks
608                  * (it may either follow vma or precede it).
609                  */
610                 __insert_vm_struct(mm, insert);
611         }
612
613         if (anon_vma)
614                 spin_unlock(&anon_vma->lock);
615         if (mapping)
616                 spin_unlock(&mapping->i_mmap_lock);
617
618         if (remove_next) {
619                 if (file)
620                         fput(file);
621                 mm->map_count--;
622                 mpol_free(vma_policy(next));
623                 kmem_cache_free(vm_area_cachep, next);
624                 /*
625                  * In mprotect's case 6 (see comments on vma_merge),
626                  * we must remove another next too. It would clutter
627                  * up the code too much to do both in one go.
628                  */
629                 if (remove_next == 2) {
630                         next = vma->vm_next;
631                         goto again;
632                 }
633         }
634
635         validate_mm(mm);
636 }
637
638 /*
639  * If the vma has a ->close operation then the driver probably needs to release
640  * per-vma resources, so we don't attempt to merge those.
641  */
642 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
643
644 static inline int is_mergeable_vma(struct vm_area_struct *vma,
645                         struct file *file, unsigned long vm_flags)
646 {
647         if (vma->vm_flags != vm_flags)
648                 return 0;
649         if (vma->vm_file != file)
650                 return 0;
651         if (vma->vm_ops && vma->vm_ops->close)
652                 return 0;
653         return 1;
654 }
655
656 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
657                                         struct anon_vma *anon_vma2)
658 {
659         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
660 }
661
662 /*
663  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
664  * in front of (at a lower virtual address and file offset than) the vma.
665  *
666  * We cannot merge two vmas if they have differently assigned (non-NULL)
667  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
668  *
669  * We don't check here for the merged mmap wrapping around the end of pagecache
670  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
671  * wrap, nor mmaps which cover the final page at index -1UL.
672  */
673 static int
674 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
675         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
676 {
677         if (is_mergeable_vma(vma, file, vm_flags) &&
678             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
679                 if (vma->vm_pgoff == vm_pgoff)
680                         return 1;
681         }
682         return 0;
683 }
684
685 /*
686  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
687  * beyond (at a higher virtual address and file offset than) the vma.
688  *
689  * We cannot merge two vmas if they have differently assigned (non-NULL)
690  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
691  */
692 static int
693 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
694         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
695 {
696         if (is_mergeable_vma(vma, file, vm_flags) &&
697             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
698                 pgoff_t vm_pglen;
699                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
700                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
701                         return 1;
702         }
703         return 0;
704 }
705
706 /*
707  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
708  * whether that can be merged with its predecessor or its successor.
709  * Or both (it neatly fills a hole).
710  *
711  * In most cases - when called for mmap, brk or mremap - [addr,end) is
712  * certain not to be mapped by the time vma_merge is called; but when
713  * called for mprotect, it is certain to be already mapped (either at
714  * an offset within prev, or at the start of next), and the flags of
715  * this area are about to be changed to vm_flags - and the no-change
716  * case has already been eliminated.
717  *
718  * The following mprotect cases have to be considered, where AAAA is
719  * the area passed down from mprotect_fixup, never extending beyond one
720  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
721  *
722  *     AAAA             AAAA                AAAA          AAAA
723  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
724  *    cannot merge    might become    might become    might become
725  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
726  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
727  *    mremap move:                                    PPPPNNNNNNNN 8
728  *        AAAA
729  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
730  *    might become    case 1 below    case 2 below    case 3 below
731  *
732  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
733  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
734  */
735 struct vm_area_struct *vma_merge(struct mm_struct *mm,
736                         struct vm_area_struct *prev, unsigned long addr,
737                         unsigned long end, unsigned long vm_flags,
738                         struct anon_vma *anon_vma, struct file *file,
739                         pgoff_t pgoff, struct mempolicy *policy)
740 {
741         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
742         struct vm_area_struct *area, *next;
743
744         /*
745          * We later require that vma->vm_flags == vm_flags,
746          * so this tests vma->vm_flags & VM_SPECIAL, too.
747          */
748         if (vm_flags & VM_SPECIAL)
749                 return NULL;
750
751         if (prev)
752                 next = prev->vm_next;
753         else
754                 next = mm->mmap;
755         area = next;
756         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
757                 next = next->vm_next;
758
759         /*
760          * Can it merge with the predecessor?
761          */
762         if (prev && prev->vm_end == addr &&
763                         mpol_equal(vma_policy(prev), policy) &&
764                         can_vma_merge_after(prev, vm_flags,
765                                                 anon_vma, file, pgoff)) {
766                 /*
767                  * OK, it can.  Can we now merge in the successor as well?
768                  */
769                 if (next && end == next->vm_start &&
770                                 mpol_equal(policy, vma_policy(next)) &&
771                                 can_vma_merge_before(next, vm_flags,
772                                         anon_vma, file, pgoff+pglen) &&
773                                 is_mergeable_anon_vma(prev->anon_vma,
774                                                       next->anon_vma)) {
775                                                         /* cases 1, 6 */
776                         vma_adjust(prev, prev->vm_start,
777                                 next->vm_end, prev->vm_pgoff, NULL);
778                 } else                                  /* cases 2, 5, 7 */
779                         vma_adjust(prev, prev->vm_start,
780                                 end, prev->vm_pgoff, NULL);
781                 return prev;
782         }
783
784         /*
785          * Can this new request be merged in front of next?
786          */
787         if (next && end == next->vm_start &&
788                         mpol_equal(policy, vma_policy(next)) &&
789                         can_vma_merge_before(next, vm_flags,
790                                         anon_vma, file, pgoff+pglen)) {
791                 if (prev && addr < prev->vm_end)        /* case 4 */
792                         vma_adjust(prev, prev->vm_start,
793                                 addr, prev->vm_pgoff, NULL);
794                 else                                    /* cases 3, 8 */
795                         vma_adjust(area, addr, next->vm_end,
796                                 next->vm_pgoff - pglen, NULL);
797                 return area;
798         }
799
800         return NULL;
801 }
802
803 /*
804  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
805  * neighbouring vmas for a suitable anon_vma, before it goes off
806  * to allocate a new anon_vma.  It checks because a repetitive
807  * sequence of mprotects and faults may otherwise lead to distinct
808  * anon_vmas being allocated, preventing vma merge in subsequent
809  * mprotect.
810  */
811 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
812 {
813         struct vm_area_struct *near;
814         unsigned long vm_flags;
815
816         near = vma->vm_next;
817         if (!near)
818                 goto try_prev;
819
820         /*
821          * Since only mprotect tries to remerge vmas, match flags
822          * which might be mprotected into each other later on.
823          * Neither mlock nor madvise tries to remerge at present,
824          * so leave their flags as obstructing a merge.
825          */
826         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
827         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
828
829         if (near->anon_vma && vma->vm_end == near->vm_start &&
830                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
831                         can_vma_merge_before(near, vm_flags,
832                                 NULL, vma->vm_file, vma->vm_pgoff +
833                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
834                 return near->anon_vma;
835 try_prev:
836         /*
837          * It is potentially slow to have to call find_vma_prev here.
838          * But it's only on the first write fault on the vma, not
839          * every time, and we could devise a way to avoid it later
840          * (e.g. stash info in next's anon_vma_node when assigning
841          * an anon_vma, or when trying vma_merge).  Another time.
842          */
843         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
844         if (!near)
845                 goto none;
846
847         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
848         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
849
850         if (near->anon_vma && near->vm_end == vma->vm_start &&
851                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
852                         can_vma_merge_after(near, vm_flags,
853                                 NULL, vma->vm_file, vma->vm_pgoff))
854                 return near->anon_vma;
855 none:
856         /*
857          * There's no absolute need to look only at touching neighbours:
858          * we could search further afield for "compatible" anon_vmas.
859          * But it would probably just be a waste of time searching,
860          * or lead to too many vmas hanging off the same anon_vma.
861          * We're trying to allow mprotect remerging later on,
862          * not trying to minimize memory used for anon_vmas.
863          */
864         return NULL;
865 }
866
867 #ifdef CONFIG_PROC_FS
868 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
869                                                 struct file *file, long pages)
870 {
871         const unsigned long stack_flags
872                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
873
874         if (file) {
875                 mm->shared_vm += pages;
876                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
877                         mm->exec_vm += pages;
878         } else if (flags & stack_flags)
879                 mm->stack_vm += pages;
880         if (flags & (VM_RESERVED|VM_IO))
881                 mm->reserved_vm += pages;
882 }
883 #endif /* CONFIG_PROC_FS */
884
885 /*
886  * The caller must hold down_write(current->mm->mmap_sem).
887  */
888
889 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
890                         unsigned long len, unsigned long prot,
891                         unsigned long flags, unsigned long pgoff)
892 {
893         struct mm_struct * mm = current->mm;
894         struct vm_area_struct * vma, * prev;
895         struct inode *inode;
896         unsigned int vm_flags;
897         int correct_wcount = 0;
898         int error;
899         struct rb_node ** rb_link, * rb_parent;
900         int accountable = 1;
901         unsigned long charged = 0, reqprot = prot;
902
903         /*
904          * Does the application expect PROT_READ to imply PROT_EXEC?
905          *
906          * (the exception is when the underlying filesystem is noexec
907          *  mounted, in which case we dont add PROT_EXEC.)
908          */
909         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
910                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
911                         prot |= PROT_EXEC;
912
913         if (!len)
914                 return -EINVAL;
915
916         error = arch_mmap_check(addr, len, flags);
917         if (error)
918                 return error;
919
920         /* Careful about overflows.. */
921         len = PAGE_ALIGN(len);
922         if (!len || len > TASK_SIZE)
923                 return -ENOMEM;
924
925         /* offset overflow? */
926         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
927                return -EOVERFLOW;
928
929         /* Too many mappings? */
930         if (mm->map_count > sysctl_max_map_count)
931                 return -ENOMEM;
932
933         /* Obtain the address to map to. we verify (or select) it and ensure
934          * that it represents a valid section of the address space.
935          */
936         addr = get_unmapped_area(file, addr, len, pgoff, flags);
937         if (addr & ~PAGE_MASK)
938                 return addr;
939
940         /* Do simple checking here so the lower-level routines won't have
941          * to. we assume access permissions have been handled by the open
942          * of the memory object, so we don't do any here.
943          */
944         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
945                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
946
947         if (flags & MAP_LOCKED) {
948                 if (!can_do_mlock())
949                         return -EPERM;
950                 vm_flags |= VM_LOCKED;
951         }
952         /* mlock MCL_FUTURE? */
953         if (vm_flags & VM_LOCKED) {
954                 unsigned long locked, lock_limit;
955                 locked = len >> PAGE_SHIFT;
956                 locked += mm->locked_vm;
957                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
958                 lock_limit >>= PAGE_SHIFT;
959                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
960                         return -EAGAIN;
961         }
962
963         inode = file ? file->f_path.dentry->d_inode : NULL;
964
965         if (file) {
966                 switch (flags & MAP_TYPE) {
967                 case MAP_SHARED:
968                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
969                                 return -EACCES;
970
971                         /*
972                          * Make sure we don't allow writing to an append-only
973                          * file..
974                          */
975                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
976                                 return -EACCES;
977
978                         /*
979                          * Make sure there are no mandatory locks on the file.
980                          */
981                         if (locks_verify_locked(inode))
982                                 return -EAGAIN;
983
984                         vm_flags |= VM_SHARED | VM_MAYSHARE;
985                         if (!(file->f_mode & FMODE_WRITE))
986                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
987
988                         /* fall through */
989                 case MAP_PRIVATE:
990                         if (!(file->f_mode & FMODE_READ))
991                                 return -EACCES;
992                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
993                                 if (vm_flags & VM_EXEC)
994                                         return -EPERM;
995                                 vm_flags &= ~VM_MAYEXEC;
996                         }
997                         if (is_file_hugepages(file))
998                                 accountable = 0;
999
1000                         if (!file->f_op || !file->f_op->mmap)
1001                                 return -ENODEV;
1002                         break;
1003
1004                 default:
1005                         return -EINVAL;
1006                 }
1007         } else {
1008                 switch (flags & MAP_TYPE) {
1009                 case MAP_SHARED:
1010                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1011                         break;
1012                 case MAP_PRIVATE:
1013                         /*
1014                          * Set pgoff according to addr for anon_vma.
1015                          */
1016                         pgoff = addr >> PAGE_SHIFT;
1017                         break;
1018                 default:
1019                         return -EINVAL;
1020                 }
1021         }
1022
1023         error = security_file_mmap(file, reqprot, prot, flags);
1024         if (error)
1025                 return error;
1026                 
1027         /* Clear old maps */
1028         error = -ENOMEM;
1029 munmap_back:
1030         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1031         if (vma && vma->vm_start < addr + len) {
1032                 if (do_munmap(mm, addr, len))
1033                         return -ENOMEM;
1034                 goto munmap_back;
1035         }
1036
1037         /* Check against address space limit. */
1038         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1039                 return -ENOMEM;
1040
1041         if (accountable && (!(flags & MAP_NORESERVE) ||
1042                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1043                 if (vm_flags & VM_SHARED) {
1044                         /* Check memory availability in shmem_file_setup? */
1045                         vm_flags |= VM_ACCOUNT;
1046                 } else if (vm_flags & VM_WRITE) {
1047                         /*
1048                          * Private writable mapping: check memory availability
1049                          */
1050                         charged = len >> PAGE_SHIFT;
1051                         if (security_vm_enough_memory(charged))
1052                                 return -ENOMEM;
1053                         vm_flags |= VM_ACCOUNT;
1054                 }
1055         }
1056
1057         /*
1058          * Can we just expand an old private anonymous mapping?
1059          * The VM_SHARED test is necessary because shmem_zero_setup
1060          * will create the file object for a shared anonymous map below.
1061          */
1062         if (!file && !(vm_flags & VM_SHARED) &&
1063             vma_merge(mm, prev, addr, addr + len, vm_flags,
1064                                         NULL, NULL, pgoff, NULL))
1065                 goto out;
1066
1067         /*
1068          * Determine the object being mapped and call the appropriate
1069          * specific mapper. the address has already been validated, but
1070          * not unmapped, but the maps are removed from the list.
1071          */
1072         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1073         if (!vma) {
1074                 error = -ENOMEM;
1075                 goto unacct_error;
1076         }
1077
1078         vma->vm_mm = mm;
1079         vma->vm_start = addr;
1080         vma->vm_end = addr + len;
1081         vma->vm_flags = vm_flags;
1082         vma->vm_page_prot = protection_map[vm_flags &
1083                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1084         vma->vm_pgoff = pgoff;
1085
1086         if (file) {
1087                 error = -EINVAL;
1088                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1089                         goto free_vma;
1090                 if (vm_flags & VM_DENYWRITE) {
1091                         error = deny_write_access(file);
1092                         if (error)
1093                                 goto free_vma;
1094                         correct_wcount = 1;
1095                 }
1096                 vma->vm_file = file;
1097                 get_file(file);
1098                 error = file->f_op->mmap(file, vma);
1099                 if (error)
1100                         goto unmap_and_free_vma;
1101         } else if (vm_flags & VM_SHARED) {
1102                 error = shmem_zero_setup(vma);
1103                 if (error)
1104                         goto free_vma;
1105         }
1106
1107         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1108          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1109          * that memory reservation must be checked; but that reservation
1110          * belongs to shared memory object, not to vma: so now clear it.
1111          */
1112         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1113                 vma->vm_flags &= ~VM_ACCOUNT;
1114
1115         /* Can addr have changed??
1116          *
1117          * Answer: Yes, several device drivers can do it in their
1118          *         f_op->mmap method. -DaveM
1119          */
1120         addr = vma->vm_start;
1121         pgoff = vma->vm_pgoff;
1122         vm_flags = vma->vm_flags;
1123
1124         if (vma_wants_writenotify(vma))
1125                 vma->vm_page_prot =
1126                         protection_map[vm_flags & (VM_READ|VM_WRITE|VM_EXEC)];
1127
1128         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1129                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1130                 file = vma->vm_file;
1131                 vma_link(mm, vma, prev, rb_link, rb_parent);
1132                 if (correct_wcount)
1133                         atomic_inc(&inode->i_writecount);
1134         } else {
1135                 if (file) {
1136                         if (correct_wcount)
1137                                 atomic_inc(&inode->i_writecount);
1138                         fput(file);
1139                 }
1140                 mpol_free(vma_policy(vma));
1141                 kmem_cache_free(vm_area_cachep, vma);
1142         }
1143 out:    
1144         mm->total_vm += len >> PAGE_SHIFT;
1145         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1146         if (vm_flags & VM_LOCKED) {
1147                 mm->locked_vm += len >> PAGE_SHIFT;
1148                 make_pages_present(addr, addr + len);
1149         }
1150         if (flags & MAP_POPULATE) {
1151                 up_write(&mm->mmap_sem);
1152                 sys_remap_file_pages(addr, len, 0,
1153                                         pgoff, flags & MAP_NONBLOCK);
1154                 down_write(&mm->mmap_sem);
1155         }
1156         return addr;
1157
1158 unmap_and_free_vma:
1159         if (correct_wcount)
1160                 atomic_inc(&inode->i_writecount);
1161         vma->vm_file = NULL;
1162         fput(file);
1163
1164         /* Undo any partial mapping done by a device driver. */
1165         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1166         charged = 0;
1167 free_vma:
1168         kmem_cache_free(vm_area_cachep, vma);
1169 unacct_error:
1170         if (charged)
1171                 vm_unacct_memory(charged);
1172         return error;
1173 }
1174
1175 EXPORT_SYMBOL(do_mmap_pgoff);
1176
1177 /* Get an address range which is currently unmapped.
1178  * For shmat() with addr=0.
1179  *
1180  * Ugly calling convention alert:
1181  * Return value with the low bits set means error value,
1182  * ie
1183  *      if (ret & ~PAGE_MASK)
1184  *              error = ret;
1185  *
1186  * This function "knows" that -ENOMEM has the bits set.
1187  */
1188 #ifndef HAVE_ARCH_UNMAPPED_AREA
1189 unsigned long
1190 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1191                 unsigned long len, unsigned long pgoff, unsigned long flags)
1192 {
1193         struct mm_struct *mm = current->mm;
1194         struct vm_area_struct *vma;
1195         unsigned long start_addr;
1196
1197         if (len > TASK_SIZE)
1198                 return -ENOMEM;
1199
1200         if (addr) {
1201                 addr = PAGE_ALIGN(addr);
1202                 vma = find_vma(mm, addr);
1203                 if (TASK_SIZE - len >= addr &&
1204                     (!vma || addr + len <= vma->vm_start))
1205                         return addr;
1206         }
1207         if (len > mm->cached_hole_size) {
1208                 start_addr = addr = mm->free_area_cache;
1209         } else {
1210                 start_addr = addr = TASK_UNMAPPED_BASE;
1211                 mm->cached_hole_size = 0;
1212         }
1213
1214 full_search:
1215         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1216                 /* At this point:  (!vma || addr < vma->vm_end). */
1217                 if (TASK_SIZE - len < addr) {
1218                         /*
1219                          * Start a new search - just in case we missed
1220                          * some holes.
1221                          */
1222                         if (start_addr != TASK_UNMAPPED_BASE) {
1223                                 addr = TASK_UNMAPPED_BASE;
1224                                 start_addr = addr;
1225                                 mm->cached_hole_size = 0;
1226                                 goto full_search;
1227                         }
1228                         return -ENOMEM;
1229                 }
1230                 if (!vma || addr + len <= vma->vm_start) {
1231                         /*
1232                          * Remember the place where we stopped the search:
1233                          */
1234                         mm->free_area_cache = addr + len;
1235                         return addr;
1236                 }
1237                 if (addr + mm->cached_hole_size < vma->vm_start)
1238                         mm->cached_hole_size = vma->vm_start - addr;
1239                 addr = vma->vm_end;
1240         }
1241 }
1242 #endif  
1243
1244 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1245 {
1246         /*
1247          * Is this a new hole at the lowest possible address?
1248          */
1249         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1250                 mm->free_area_cache = addr;
1251                 mm->cached_hole_size = ~0UL;
1252         }
1253 }
1254
1255 /*
1256  * This mmap-allocator allocates new areas top-down from below the
1257  * stack's low limit (the base):
1258  */
1259 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1260 unsigned long
1261 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1262                           const unsigned long len, const unsigned long pgoff,
1263                           const unsigned long flags)
1264 {
1265         struct vm_area_struct *vma;
1266         struct mm_struct *mm = current->mm;
1267         unsigned long addr = addr0;
1268
1269         /* requested length too big for entire address space */
1270         if (len > TASK_SIZE)
1271                 return -ENOMEM;
1272
1273         /* requesting a specific address */
1274         if (addr) {
1275                 addr = PAGE_ALIGN(addr);
1276                 vma = find_vma(mm, addr);
1277                 if (TASK_SIZE - len >= addr &&
1278                                 (!vma || addr + len <= vma->vm_start))
1279                         return addr;
1280         }
1281
1282         /* check if free_area_cache is useful for us */
1283         if (len <= mm->cached_hole_size) {
1284                 mm->cached_hole_size = 0;
1285                 mm->free_area_cache = mm->mmap_base;
1286         }
1287
1288         /* either no address requested or can't fit in requested address hole */
1289         addr = mm->free_area_cache;
1290
1291         /* make sure it can fit in the remaining address space */
1292         if (addr > len) {
1293                 vma = find_vma(mm, addr-len);
1294                 if (!vma || addr <= vma->vm_start)
1295                         /* remember the address as a hint for next time */
1296                         return (mm->free_area_cache = addr-len);
1297         }
1298
1299         if (mm->mmap_base < len)
1300                 goto bottomup;
1301
1302         addr = mm->mmap_base-len;
1303
1304         do {
1305                 /*
1306                  * Lookup failure means no vma is above this address,
1307                  * else if new region fits below vma->vm_start,
1308                  * return with success:
1309                  */
1310                 vma = find_vma(mm, addr);
1311                 if (!vma || addr+len <= vma->vm_start)
1312                         /* remember the address as a hint for next time */
1313                         return (mm->free_area_cache = addr);
1314
1315                 /* remember the largest hole we saw so far */
1316                 if (addr + mm->cached_hole_size < vma->vm_start)
1317                         mm->cached_hole_size = vma->vm_start - addr;
1318
1319                 /* try just below the current vma->vm_start */
1320                 addr = vma->vm_start-len;
1321         } while (len < vma->vm_start);
1322
1323 bottomup:
1324         /*
1325          * A failed mmap() very likely causes application failure,
1326          * so fall back to the bottom-up function here. This scenario
1327          * can happen with large stack limits and large mmap()
1328          * allocations.
1329          */
1330         mm->cached_hole_size = ~0UL;
1331         mm->free_area_cache = TASK_UNMAPPED_BASE;
1332         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1333         /*
1334          * Restore the topdown base:
1335          */
1336         mm->free_area_cache = mm->mmap_base;
1337         mm->cached_hole_size = ~0UL;
1338
1339         return addr;
1340 }
1341 #endif
1342
1343 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1344 {
1345         /*
1346          * Is this a new hole at the highest possible address?
1347          */
1348         if (addr > mm->free_area_cache)
1349                 mm->free_area_cache = addr;
1350
1351         /* dont allow allocations above current base */
1352         if (mm->free_area_cache > mm->mmap_base)
1353                 mm->free_area_cache = mm->mmap_base;
1354 }
1355
1356 unsigned long
1357 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1358                 unsigned long pgoff, unsigned long flags)
1359 {
1360         unsigned long ret;
1361
1362         if (!(flags & MAP_FIXED)) {
1363                 unsigned long (*get_area)(struct file *, unsigned long, unsigned long, unsigned long, unsigned long);
1364
1365                 get_area = current->mm->get_unmapped_area;
1366                 if (file && file->f_op && file->f_op->get_unmapped_area)
1367                         get_area = file->f_op->get_unmapped_area;
1368                 addr = get_area(file, addr, len, pgoff, flags);
1369                 if (IS_ERR_VALUE(addr))
1370                         return addr;
1371         }
1372
1373         if (addr > TASK_SIZE - len)
1374                 return -ENOMEM;
1375         if (addr & ~PAGE_MASK)
1376                 return -EINVAL;
1377         if (file && is_file_hugepages(file))  {
1378                 /*
1379                  * Check if the given range is hugepage aligned, and
1380                  * can be made suitable for hugepages.
1381                  */
1382                 ret = prepare_hugepage_range(addr, len, pgoff);
1383         } else {
1384                 /*
1385                  * Ensure that a normal request is not falling in a
1386                  * reserved hugepage range.  For some archs like IA-64,
1387                  * there is a separate region for hugepages.
1388                  */
1389                 ret = is_hugepage_only_range(current->mm, addr, len);
1390         }
1391         if (ret)
1392                 return -EINVAL;
1393         return addr;
1394 }
1395
1396 EXPORT_SYMBOL(get_unmapped_area);
1397
1398 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1399 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1400 {
1401         struct vm_area_struct *vma = NULL;
1402
1403         if (mm) {
1404                 /* Check the cache first. */
1405                 /* (Cache hit rate is typically around 35%.) */
1406                 vma = mm->mmap_cache;
1407                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1408                         struct rb_node * rb_node;
1409
1410                         rb_node = mm->mm_rb.rb_node;
1411                         vma = NULL;
1412
1413                         while (rb_node) {
1414                                 struct vm_area_struct * vma_tmp;
1415
1416                                 vma_tmp = rb_entry(rb_node,
1417                                                 struct vm_area_struct, vm_rb);
1418
1419                                 if (vma_tmp->vm_end > addr) {
1420                                         vma = vma_tmp;
1421                                         if (vma_tmp->vm_start <= addr)
1422                                                 break;
1423                                         rb_node = rb_node->rb_left;
1424                                 } else
1425                                         rb_node = rb_node->rb_right;
1426                         }
1427                         if (vma)
1428                                 mm->mmap_cache = vma;
1429                 }
1430         }
1431         return vma;
1432 }
1433
1434 EXPORT_SYMBOL(find_vma);
1435
1436 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1437 struct vm_area_struct *
1438 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1439                         struct vm_area_struct **pprev)
1440 {
1441         struct vm_area_struct *vma = NULL, *prev = NULL;
1442         struct rb_node * rb_node;
1443         if (!mm)
1444                 goto out;
1445
1446         /* Guard against addr being lower than the first VMA */
1447         vma = mm->mmap;
1448
1449         /* Go through the RB tree quickly. */
1450         rb_node = mm->mm_rb.rb_node;
1451
1452         while (rb_node) {
1453                 struct vm_area_struct *vma_tmp;
1454                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1455
1456                 if (addr < vma_tmp->vm_end) {
1457                         rb_node = rb_node->rb_left;
1458                 } else {
1459                         prev = vma_tmp;
1460                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1461                                 break;
1462                         rb_node = rb_node->rb_right;
1463                 }
1464         }
1465
1466 out:
1467         *pprev = prev;
1468         return prev ? prev->vm_next : vma;
1469 }
1470
1471 /*
1472  * Verify that the stack growth is acceptable and
1473  * update accounting. This is shared with both the
1474  * grow-up and grow-down cases.
1475  */
1476 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1477 {
1478         struct mm_struct *mm = vma->vm_mm;
1479         struct rlimit *rlim = current->signal->rlim;
1480         unsigned long new_start;
1481
1482         /* address space limit tests */
1483         if (!may_expand_vm(mm, grow))
1484                 return -ENOMEM;
1485
1486         /* Stack limit test */
1487         if (size > rlim[RLIMIT_STACK].rlim_cur)
1488                 return -ENOMEM;
1489
1490         /* mlock limit tests */
1491         if (vma->vm_flags & VM_LOCKED) {
1492                 unsigned long locked;
1493                 unsigned long limit;
1494                 locked = mm->locked_vm + grow;
1495                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1496                 if (locked > limit && !capable(CAP_IPC_LOCK))
1497                         return -ENOMEM;
1498         }
1499
1500         /* Check to ensure the stack will not grow into a hugetlb-only region */
1501         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1502                         vma->vm_end - size;
1503         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1504                 return -EFAULT;
1505
1506         /*
1507          * Overcommit..  This must be the final test, as it will
1508          * update security statistics.
1509          */
1510         if (security_vm_enough_memory(grow))
1511                 return -ENOMEM;
1512
1513         /* Ok, everything looks good - let it rip */
1514         mm->total_vm += grow;
1515         if (vma->vm_flags & VM_LOCKED)
1516                 mm->locked_vm += grow;
1517         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1518         return 0;
1519 }
1520
1521 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1522 /*
1523  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1524  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1525  */
1526 #ifndef CONFIG_IA64
1527 static inline
1528 #endif
1529 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1530 {
1531         int error;
1532
1533         if (!(vma->vm_flags & VM_GROWSUP))
1534                 return -EFAULT;
1535
1536         /*
1537          * We must make sure the anon_vma is allocated
1538          * so that the anon_vma locking is not a noop.
1539          */
1540         if (unlikely(anon_vma_prepare(vma)))
1541                 return -ENOMEM;
1542         anon_vma_lock(vma);
1543
1544         /*
1545          * vma->vm_start/vm_end cannot change under us because the caller
1546          * is required to hold the mmap_sem in read mode.  We need the
1547          * anon_vma lock to serialize against concurrent expand_stacks.
1548          */
1549         address += 4 + PAGE_SIZE - 1;
1550         address &= PAGE_MASK;
1551         error = 0;
1552
1553         /* Somebody else might have raced and expanded it already */
1554         if (address > vma->vm_end) {
1555                 unsigned long size, grow;
1556
1557                 size = address - vma->vm_start;
1558                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1559
1560                 error = acct_stack_growth(vma, size, grow);
1561                 if (!error)
1562                         vma->vm_end = address;
1563         }
1564         anon_vma_unlock(vma);
1565         return error;
1566 }
1567 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1568
1569 #ifdef CONFIG_STACK_GROWSUP
1570 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1571 {
1572         return expand_upwards(vma, address);
1573 }
1574
1575 struct vm_area_struct *
1576 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1577 {
1578         struct vm_area_struct *vma, *prev;
1579
1580         addr &= PAGE_MASK;
1581         vma = find_vma_prev(mm, addr, &prev);
1582         if (vma && (vma->vm_start <= addr))
1583                 return vma;
1584         if (!prev || expand_stack(prev, addr))
1585                 return NULL;
1586         if (prev->vm_flags & VM_LOCKED) {
1587                 make_pages_present(addr, prev->vm_end);
1588         }
1589         return prev;
1590 }
1591 #else
1592 /*
1593  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1594  */
1595 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1596 {
1597         int error;
1598
1599         /*
1600          * We must make sure the anon_vma is allocated
1601          * so that the anon_vma locking is not a noop.
1602          */
1603         if (unlikely(anon_vma_prepare(vma)))
1604                 return -ENOMEM;
1605         anon_vma_lock(vma);
1606
1607         /*
1608          * vma->vm_start/vm_end cannot change under us because the caller
1609          * is required to hold the mmap_sem in read mode.  We need the
1610          * anon_vma lock to serialize against concurrent expand_stacks.
1611          */
1612         address &= PAGE_MASK;
1613         error = 0;
1614
1615         /* Somebody else might have raced and expanded it already */
1616         if (address < vma->vm_start) {
1617                 unsigned long size, grow;
1618
1619                 size = vma->vm_end - address;
1620                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1621
1622                 error = acct_stack_growth(vma, size, grow);
1623                 if (!error) {
1624                         vma->vm_start = address;
1625                         vma->vm_pgoff -= grow;
1626                 }
1627         }
1628         anon_vma_unlock(vma);
1629         return error;
1630 }
1631
1632 struct vm_area_struct *
1633 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1634 {
1635         struct vm_area_struct * vma;
1636         unsigned long start;
1637
1638         addr &= PAGE_MASK;
1639         vma = find_vma(mm,addr);
1640         if (!vma)
1641                 return NULL;
1642         if (vma->vm_start <= addr)
1643                 return vma;
1644         if (!(vma->vm_flags & VM_GROWSDOWN))
1645                 return NULL;
1646         start = vma->vm_start;
1647         if (expand_stack(vma, addr))
1648                 return NULL;
1649         if (vma->vm_flags & VM_LOCKED) {
1650                 make_pages_present(addr, start);
1651         }
1652         return vma;
1653 }
1654 #endif
1655
1656 /*
1657  * Ok - we have the memory areas we should free on the vma list,
1658  * so release them, and do the vma updates.
1659  *
1660  * Called with the mm semaphore held.
1661  */
1662 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1663 {
1664         /* Update high watermark before we lower total_vm */
1665         update_hiwater_vm(mm);
1666         do {
1667                 long nrpages = vma_pages(vma);
1668
1669                 mm->total_vm -= nrpages;
1670                 if (vma->vm_flags & VM_LOCKED)
1671                         mm->locked_vm -= nrpages;
1672                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1673                 vma = remove_vma(vma);
1674         } while (vma);
1675         validate_mm(mm);
1676 }
1677
1678 /*
1679  * Get rid of page table information in the indicated region.
1680  *
1681  * Called with the mm semaphore held.
1682  */
1683 static void unmap_region(struct mm_struct *mm,
1684                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1685                 unsigned long start, unsigned long end)
1686 {
1687         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1688         struct mmu_gather *tlb;
1689         unsigned long nr_accounted = 0;
1690
1691         lru_add_drain();
1692         tlb = tlb_gather_mmu(mm, 0);
1693         update_hiwater_rss(mm);
1694         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1695         vm_unacct_memory(nr_accounted);
1696         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1697                                  next? next->vm_start: 0);
1698         tlb_finish_mmu(tlb, start, end);
1699 }
1700
1701 /*
1702  * Create a list of vma's touched by the unmap, removing them from the mm's
1703  * vma list as we go..
1704  */
1705 static void
1706 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1707         struct vm_area_struct *prev, unsigned long end)
1708 {
1709         struct vm_area_struct **insertion_point;
1710         struct vm_area_struct *tail_vma = NULL;
1711         unsigned long addr;
1712
1713         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1714         do {
1715                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1716                 mm->map_count--;
1717                 tail_vma = vma;
1718                 vma = vma->vm_next;
1719         } while (vma && vma->vm_start < end);
1720         *insertion_point = vma;
1721         tail_vma->vm_next = NULL;
1722         if (mm->unmap_area == arch_unmap_area)
1723                 addr = prev ? prev->vm_end : mm->mmap_base;
1724         else
1725                 addr = vma ?  vma->vm_start : mm->mmap_base;
1726         mm->unmap_area(mm, addr);
1727         mm->mmap_cache = NULL;          /* Kill the cache. */
1728 }
1729
1730 /*
1731  * Split a vma into two pieces at address 'addr', a new vma is allocated
1732  * either for the first part or the the tail.
1733  */
1734 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1735               unsigned long addr, int new_below)
1736 {
1737         struct mempolicy *pol;
1738         struct vm_area_struct *new;
1739
1740         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1741                 return -EINVAL;
1742
1743         if (mm->map_count >= sysctl_max_map_count)
1744                 return -ENOMEM;
1745
1746         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1747         if (!new)
1748                 return -ENOMEM;
1749
1750         /* most fields are the same, copy all, and then fixup */
1751         *new = *vma;
1752
1753         if (new_below)
1754                 new->vm_end = addr;
1755         else {
1756                 new->vm_start = addr;
1757                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1758         }
1759
1760         pol = mpol_copy(vma_policy(vma));
1761         if (IS_ERR(pol)) {
1762                 kmem_cache_free(vm_area_cachep, new);
1763                 return PTR_ERR(pol);
1764         }
1765         vma_set_policy(new, pol);
1766
1767         if (new->vm_file)
1768                 get_file(new->vm_file);
1769
1770         if (new->vm_ops && new->vm_ops->open)
1771                 new->vm_ops->open(new);
1772
1773         if (new_below)
1774                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1775                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1776         else
1777                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1778
1779         return 0;
1780 }
1781
1782 /* Munmap is split into 2 main parts -- this part which finds
1783  * what needs doing, and the areas themselves, which do the
1784  * work.  This now handles partial unmappings.
1785  * Jeremy Fitzhardinge <jeremy@goop.org>
1786  */
1787 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1788 {
1789         unsigned long end;
1790         struct vm_area_struct *vma, *prev, *last;
1791
1792         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1793                 return -EINVAL;
1794
1795         if ((len = PAGE_ALIGN(len)) == 0)
1796                 return -EINVAL;
1797
1798         /* Find the first overlapping VMA */
1799         vma = find_vma_prev(mm, start, &prev);
1800         if (!vma)
1801                 return 0;
1802         /* we have  start < vma->vm_end  */
1803
1804         /* if it doesn't overlap, we have nothing.. */
1805         end = start + len;
1806         if (vma->vm_start >= end)
1807                 return 0;
1808
1809         /*
1810          * If we need to split any vma, do it now to save pain later.
1811          *
1812          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1813          * unmapped vm_area_struct will remain in use: so lower split_vma
1814          * places tmp vma above, and higher split_vma places tmp vma below.
1815          */
1816         if (start > vma->vm_start) {
1817                 int error = split_vma(mm, vma, start, 0);
1818                 if (error)
1819                         return error;
1820                 prev = vma;
1821         }
1822
1823         /* Does it split the last one? */
1824         last = find_vma(mm, end);
1825         if (last && end > last->vm_start) {
1826                 int error = split_vma(mm, last, end, 1);
1827                 if (error)
1828                         return error;
1829         }
1830         vma = prev? prev->vm_next: mm->mmap;
1831
1832         /*
1833          * Remove the vma's, and unmap the actual pages
1834          */
1835         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1836         unmap_region(mm, vma, prev, start, end);
1837
1838         /* Fix up all other VM information */
1839         remove_vma_list(mm, vma);
1840
1841         return 0;
1842 }
1843
1844 EXPORT_SYMBOL(do_munmap);
1845
1846 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1847 {
1848         int ret;
1849         struct mm_struct *mm = current->mm;
1850
1851         profile_munmap(addr);
1852
1853         down_write(&mm->mmap_sem);
1854         ret = do_munmap(mm, addr, len);
1855         up_write(&mm->mmap_sem);
1856         return ret;
1857 }
1858
1859 static inline void verify_mm_writelocked(struct mm_struct *mm)
1860 {
1861 #ifdef CONFIG_DEBUG_VM
1862         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1863                 WARN_ON(1);
1864                 up_read(&mm->mmap_sem);
1865         }
1866 #endif
1867 }
1868
1869 /*
1870  *  this is really a simplified "do_mmap".  it only handles
1871  *  anonymous maps.  eventually we may be able to do some
1872  *  brk-specific accounting here.
1873  */
1874 unsigned long do_brk(unsigned long addr, unsigned long len)
1875 {
1876         struct mm_struct * mm = current->mm;
1877         struct vm_area_struct * vma, * prev;
1878         unsigned long flags;
1879         struct rb_node ** rb_link, * rb_parent;
1880         pgoff_t pgoff = addr >> PAGE_SHIFT;
1881         int error;
1882
1883         len = PAGE_ALIGN(len);
1884         if (!len)
1885                 return addr;
1886
1887         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1888                 return -EINVAL;
1889
1890         if (is_hugepage_only_range(mm, addr, len))
1891                 return -EINVAL;
1892
1893         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1894
1895         error = arch_mmap_check(addr, len, flags);
1896         if (error)
1897                 return error;
1898
1899         /*
1900          * mlock MCL_FUTURE?
1901          */
1902         if (mm->def_flags & VM_LOCKED) {
1903                 unsigned long locked, lock_limit;
1904                 locked = len >> PAGE_SHIFT;
1905                 locked += mm->locked_vm;
1906                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1907                 lock_limit >>= PAGE_SHIFT;
1908                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1909                         return -EAGAIN;
1910         }
1911
1912         /*
1913          * mm->mmap_sem is required to protect against another thread
1914          * changing the mappings in case we sleep.
1915          */
1916         verify_mm_writelocked(mm);
1917
1918         /*
1919          * Clear old maps.  this also does some error checking for us
1920          */
1921  munmap_back:
1922         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1923         if (vma && vma->vm_start < addr + len) {
1924                 if (do_munmap(mm, addr, len))
1925                         return -ENOMEM;
1926                 goto munmap_back;
1927         }
1928
1929         /* Check against address space limits *after* clearing old maps... */
1930         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1931                 return -ENOMEM;
1932
1933         if (mm->map_count > sysctl_max_map_count)
1934                 return -ENOMEM;
1935
1936         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1937                 return -ENOMEM;
1938
1939         /* Can we just expand an old private anonymous mapping? */
1940         if (vma_merge(mm, prev, addr, addr + len, flags,
1941                                         NULL, NULL, pgoff, NULL))
1942                 goto out;
1943
1944         /*
1945          * create a vma struct for an anonymous mapping
1946          */
1947         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1948         if (!vma) {
1949                 vm_unacct_memory(len >> PAGE_SHIFT);
1950                 return -ENOMEM;
1951         }
1952
1953         vma->vm_mm = mm;
1954         vma->vm_start = addr;
1955         vma->vm_end = addr + len;
1956         vma->vm_pgoff = pgoff;
1957         vma->vm_flags = flags;
1958         vma->vm_page_prot = protection_map[flags &
1959                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
1960         vma_link(mm, vma, prev, rb_link, rb_parent);
1961 out:
1962         mm->total_vm += len >> PAGE_SHIFT;
1963         if (flags & VM_LOCKED) {
1964                 mm->locked_vm += len >> PAGE_SHIFT;
1965                 make_pages_present(addr, addr + len);
1966         }
1967         return addr;
1968 }
1969
1970 EXPORT_SYMBOL(do_brk);
1971
1972 /* Release all mmaps. */
1973 void exit_mmap(struct mm_struct *mm)
1974 {
1975         struct mmu_gather *tlb;
1976         struct vm_area_struct *vma = mm->mmap;
1977         unsigned long nr_accounted = 0;
1978         unsigned long end;
1979
1980         lru_add_drain();
1981         flush_cache_mm(mm);
1982         tlb = tlb_gather_mmu(mm, 1);
1983         /* Don't update_hiwater_rss(mm) here, do_exit already did */
1984         /* Use -1 here to ensure all VMAs in the mm are unmapped */
1985         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
1986         vm_unacct_memory(nr_accounted);
1987         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
1988         tlb_finish_mmu(tlb, 0, end);
1989
1990         /*
1991          * Walk the list again, actually closing and freeing it,
1992          * with preemption enabled, without holding any MM locks.
1993          */
1994         while (vma)
1995                 vma = remove_vma(vma);
1996
1997         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
1998 }
1999
2000 /* Insert vm structure into process list sorted by address
2001  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2002  * then i_mmap_lock is taken here.
2003  */
2004 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2005 {
2006         struct vm_area_struct * __vma, * prev;
2007         struct rb_node ** rb_link, * rb_parent;
2008
2009         /*
2010          * The vm_pgoff of a purely anonymous vma should be irrelevant
2011          * until its first write fault, when page's anon_vma and index
2012          * are set.  But now set the vm_pgoff it will almost certainly
2013          * end up with (unless mremap moves it elsewhere before that
2014          * first wfault), so /proc/pid/maps tells a consistent story.
2015          *
2016          * By setting it to reflect the virtual start address of the
2017          * vma, merges and splits can happen in a seamless way, just
2018          * using the existing file pgoff checks and manipulations.
2019          * Similarly in do_mmap_pgoff and in do_brk.
2020          */
2021         if (!vma->vm_file) {
2022                 BUG_ON(vma->anon_vma);
2023                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2024         }
2025         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2026         if (__vma && __vma->vm_start < vma->vm_end)
2027                 return -ENOMEM;
2028         if ((vma->vm_flags & VM_ACCOUNT) &&
2029              security_vm_enough_memory(vma_pages(vma)))
2030                 return -ENOMEM;
2031         vma_link(mm, vma, prev, rb_link, rb_parent);
2032         return 0;
2033 }
2034
2035 /*
2036  * Copy the vma structure to a new location in the same mm,
2037  * prior to moving page table entries, to effect an mremap move.
2038  */
2039 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2040         unsigned long addr, unsigned long len, pgoff_t pgoff)
2041 {
2042         struct vm_area_struct *vma = *vmap;
2043         unsigned long vma_start = vma->vm_start;
2044         struct mm_struct *mm = vma->vm_mm;
2045         struct vm_area_struct *new_vma, *prev;
2046         struct rb_node **rb_link, *rb_parent;
2047         struct mempolicy *pol;
2048
2049         /*
2050          * If anonymous vma has not yet been faulted, update new pgoff
2051          * to match new location, to increase its chance of merging.
2052          */
2053         if (!vma->vm_file && !vma->anon_vma)
2054                 pgoff = addr >> PAGE_SHIFT;
2055
2056         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2057         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2058                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2059         if (new_vma) {
2060                 /*
2061                  * Source vma may have been merged into new_vma
2062                  */
2063                 if (vma_start >= new_vma->vm_start &&
2064                     vma_start < new_vma->vm_end)
2065                         *vmap = new_vma;
2066         } else {
2067                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2068                 if (new_vma) {
2069                         *new_vma = *vma;
2070                         pol = mpol_copy(vma_policy(vma));
2071                         if (IS_ERR(pol)) {
2072                                 kmem_cache_free(vm_area_cachep, new_vma);
2073                                 return NULL;
2074                         }
2075                         vma_set_policy(new_vma, pol);
2076                         new_vma->vm_start = addr;
2077                         new_vma->vm_end = addr + len;
2078                         new_vma->vm_pgoff = pgoff;
2079                         if (new_vma->vm_file)
2080                                 get_file(new_vma->vm_file);
2081                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2082                                 new_vma->vm_ops->open(new_vma);
2083                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2084                 }
2085         }
2086         return new_vma;
2087 }
2088
2089 /*
2090  * Return true if the calling process may expand its vm space by the passed
2091  * number of pages
2092  */
2093 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2094 {
2095         unsigned long cur = mm->total_vm;       /* pages */
2096         unsigned long lim;
2097
2098         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2099
2100         if (cur + npages > lim)
2101                 return 0;
2102         return 1;
2103 }
2104
2105
2106 static struct page *special_mapping_nopage(struct vm_area_struct *vma,
2107                                            unsigned long address, int *type)
2108 {
2109         struct page **pages;
2110
2111         BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2112
2113         address -= vma->vm_start;
2114         for (pages = vma->vm_private_data; address > 0 && *pages; ++pages)
2115                 address -= PAGE_SIZE;
2116
2117         if (*pages) {
2118                 struct page *page = *pages;
2119                 get_page(page);
2120                 return page;
2121         }
2122
2123         return NOPAGE_SIGBUS;
2124 }
2125
2126 /*
2127  * Having a close hook prevents vma merging regardless of flags.
2128  */
2129 static void special_mapping_close(struct vm_area_struct *vma)
2130 {
2131 }
2132
2133 static struct vm_operations_struct special_mapping_vmops = {
2134         .close = special_mapping_close,
2135         .nopage = special_mapping_nopage,
2136 };
2137
2138 /*
2139  * Called with mm->mmap_sem held for writing.
2140  * Insert a new vma covering the given region, with the given flags.
2141  * Its pages are supplied by the given array of struct page *.
2142  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2143  * The region past the last page supplied will always produce SIGBUS.
2144  * The array pointer and the pages it points to are assumed to stay alive
2145  * for as long as this mapping might exist.
2146  */
2147 int install_special_mapping(struct mm_struct *mm,
2148                             unsigned long addr, unsigned long len,
2149                             unsigned long vm_flags, struct page **pages)
2150 {
2151         struct vm_area_struct *vma;
2152
2153         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2154         if (unlikely(vma == NULL))
2155                 return -ENOMEM;
2156
2157         vma->vm_mm = mm;
2158         vma->vm_start = addr;
2159         vma->vm_end = addr + len;
2160
2161         vma->vm_flags = vm_flags | mm->def_flags;
2162         vma->vm_page_prot = protection_map[vma->vm_flags & 7];
2163
2164         vma->vm_ops = &special_mapping_vmops;
2165         vma->vm_private_data = pages;
2166
2167         if (unlikely(insert_vm_struct(mm, vma))) {
2168                 kmem_cache_free(vm_area_cachep, vma);
2169                 return -ENOMEM;
2170         }
2171
2172         mm->total_vm += len >> PAGE_SHIFT;
2173
2174         return 0;
2175 }