VM/Security: add security hook to do_brk
[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/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29
30 #include <asm/uaccess.h>
31 #include <asm/cacheflush.h>
32 #include <asm/tlb.h>
33 #include <asm/mmu_context.h>
34
35 #ifndef arch_mmap_check
36 #define arch_mmap_check(addr, len, flags)       (0)
37 #endif
38
39 static void unmap_region(struct mm_struct *mm,
40                 struct vm_area_struct *vma, struct vm_area_struct *prev,
41                 unsigned long start, unsigned long end);
42
43 /*
44  * WARNING: the debugging will use recursive algorithms so never enable this
45  * unless you know what you are doing.
46  */
47 #undef DEBUG_MM_RB
48
49 /* description of effects of mapping type and prot in current implementation.
50  * this is due to the limited x86 page protection hardware.  The expected
51  * behavior is in parens:
52  *
53  * map_type     prot
54  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
55  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
56  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
57  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
58  *              
59  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
60  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
61  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
62  *
63  */
64 pgprot_t protection_map[16] = {
65         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
66         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
67 };
68
69 pgprot_t vm_get_page_prot(unsigned long vm_flags)
70 {
71         return protection_map[vm_flags &
72                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
73 }
74 EXPORT_SYMBOL(vm_get_page_prot);
75
76 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
77 int sysctl_overcommit_ratio = 50;       /* default is 50% */
78 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
79 atomic_t vm_committed_space = ATOMIC_INIT(0);
80
81 /*
82  * Check that a process has enough memory to allocate a new virtual
83  * mapping. 0 means there is enough memory for the allocation to
84  * succeed and -ENOMEM implies there is not.
85  *
86  * We currently support three overcommit policies, which are set via the
87  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
88  *
89  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
90  * Additional code 2002 Jul 20 by Robert Love.
91  *
92  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
93  *
94  * Note this is a helper function intended to be used by LSMs which
95  * wish to use this logic.
96  */
97 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
98 {
99         unsigned long free, allowed;
100
101         vm_acct_memory(pages);
102
103         /*
104          * Sometimes we want to use more memory than we have
105          */
106         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
107                 return 0;
108
109         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
110                 unsigned long n;
111
112                 free = global_page_state(NR_FILE_PAGES);
113                 free += nr_swap_pages;
114
115                 /*
116                  * Any slabs which are created with the
117                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
118                  * which are reclaimable, under pressure.  The dentry
119                  * cache and most inode caches should fall into this
120                  */
121                 free += global_page_state(NR_SLAB_RECLAIMABLE);
122
123                 /*
124                  * Leave the last 3% for root
125                  */
126                 if (!cap_sys_admin)
127                         free -= free / 32;
128
129                 if (free > pages)
130                         return 0;
131
132                 /*
133                  * nr_free_pages() is very expensive on large systems,
134                  * only call if we're about to fail.
135                  */
136                 n = nr_free_pages();
137
138                 /*
139                  * Leave reserved pages. The pages are not for anonymous pages.
140                  */
141                 if (n <= totalreserve_pages)
142                         goto error;
143                 else
144                         n -= totalreserve_pages;
145
146                 /*
147                  * Leave the last 3% for root
148                  */
149                 if (!cap_sys_admin)
150                         n -= n / 32;
151                 free += n;
152
153                 if (free > pages)
154                         return 0;
155
156                 goto error;
157         }
158
159         allowed = (totalram_pages - hugetlb_total_pages())
160                 * sysctl_overcommit_ratio / 100;
161         /*
162          * Leave the last 3% for root
163          */
164         if (!cap_sys_admin)
165                 allowed -= allowed / 32;
166         allowed += total_swap_pages;
167
168         /* Don't let a single process grow too big:
169            leave 3% of the size of this process for other processes */
170         allowed -= mm->total_vm / 32;
171
172         /*
173          * cast `allowed' as a signed long because vm_committed_space
174          * sometimes has a negative value
175          */
176         if (atomic_read(&vm_committed_space) < (long)allowed)
177                 return 0;
178 error:
179         vm_unacct_memory(pages);
180
181         return -ENOMEM;
182 }
183
184 /*
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                 prev = vma->vm_start;
303                 pend = vma->vm_end;
304         }
305         j = 0;
306         for (nd = pn; nd; nd = rb_prev(nd)) {
307                 j++;
308         }
309         if (i != j)
310                 printk("backwards %d, forwards %d\n", j, i), i = 0;
311         return i;
312 }
313
314 void validate_mm(struct mm_struct *mm)
315 {
316         int bug = 0;
317         int i = 0;
318         struct vm_area_struct *tmp = mm->mmap;
319         while (tmp) {
320                 tmp = tmp->vm_next;
321                 i++;
322         }
323         if (i != mm->map_count)
324                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
325         i = browse_rb(&mm->mm_rb);
326         if (i != mm->map_count)
327                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
328         BUG_ON(bug);
329 }
330 #else
331 #define validate_mm(mm) do { } while (0)
332 #endif
333
334 static struct vm_area_struct *
335 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
336                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
337                 struct rb_node ** rb_parent)
338 {
339         struct vm_area_struct * vma;
340         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
341
342         __rb_link = &mm->mm_rb.rb_node;
343         rb_prev = __rb_parent = NULL;
344         vma = NULL;
345
346         while (*__rb_link) {
347                 struct vm_area_struct *vma_tmp;
348
349                 __rb_parent = *__rb_link;
350                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
351
352                 if (vma_tmp->vm_end > addr) {
353                         vma = vma_tmp;
354                         if (vma_tmp->vm_start <= addr)
355                                 return vma;
356                         __rb_link = &__rb_parent->rb_left;
357                 } else {
358                         rb_prev = __rb_parent;
359                         __rb_link = &__rb_parent->rb_right;
360                 }
361         }
362
363         *pprev = NULL;
364         if (rb_prev)
365                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
366         *rb_link = __rb_link;
367         *rb_parent = __rb_parent;
368         return vma;
369 }
370
371 static inline void
372 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
373                 struct vm_area_struct *prev, struct rb_node *rb_parent)
374 {
375         if (prev) {
376                 vma->vm_next = prev->vm_next;
377                 prev->vm_next = vma;
378         } else {
379                 mm->mmap = vma;
380                 if (rb_parent)
381                         vma->vm_next = rb_entry(rb_parent,
382                                         struct vm_area_struct, vm_rb);
383                 else
384                         vma->vm_next = NULL;
385         }
386 }
387
388 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
389                 struct rb_node **rb_link, struct rb_node *rb_parent)
390 {
391         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
392         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
393 }
394
395 static inline void __vma_link_file(struct vm_area_struct *vma)
396 {
397         struct file * file;
398
399         file = vma->vm_file;
400         if (file) {
401                 struct address_space *mapping = file->f_mapping;
402
403                 if (vma->vm_flags & VM_DENYWRITE)
404                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
405                 if (vma->vm_flags & VM_SHARED)
406                         mapping->i_mmap_writable++;
407
408                 flush_dcache_mmap_lock(mapping);
409                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
410                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
411                 else
412                         vma_prio_tree_insert(vma, &mapping->i_mmap);
413                 flush_dcache_mmap_unlock(mapping);
414         }
415 }
416
417 static void
418 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
419         struct vm_area_struct *prev, struct rb_node **rb_link,
420         struct rb_node *rb_parent)
421 {
422         __vma_link_list(mm, vma, prev, rb_parent);
423         __vma_link_rb(mm, vma, rb_link, rb_parent);
424         __anon_vma_link(vma);
425 }
426
427 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
428                         struct vm_area_struct *prev, struct rb_node **rb_link,
429                         struct rb_node *rb_parent)
430 {
431         struct address_space *mapping = NULL;
432
433         if (vma->vm_file)
434                 mapping = vma->vm_file->f_mapping;
435
436         if (mapping) {
437                 spin_lock(&mapping->i_mmap_lock);
438                 vma->vm_truncate_count = mapping->truncate_count;
439         }
440         anon_vma_lock(vma);
441
442         __vma_link(mm, vma, prev, rb_link, rb_parent);
443         __vma_link_file(vma);
444
445         anon_vma_unlock(vma);
446         if (mapping)
447                 spin_unlock(&mapping->i_mmap_lock);
448
449         mm->map_count++;
450         validate_mm(mm);
451 }
452
453 /*
454  * Helper for vma_adjust in the split_vma insert case:
455  * insert vm structure into list and rbtree and anon_vma,
456  * but it has already been inserted into prio_tree earlier.
457  */
458 static void
459 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
460 {
461         struct vm_area_struct * __vma, * prev;
462         struct rb_node ** rb_link, * rb_parent;
463
464         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
465         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
466         __vma_link(mm, vma, prev, rb_link, rb_parent);
467         mm->map_count++;
468 }
469
470 static inline void
471 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
472                 struct vm_area_struct *prev)
473 {
474         prev->vm_next = vma->vm_next;
475         rb_erase(&vma->vm_rb, &mm->mm_rb);
476         if (mm->mmap_cache == vma)
477                 mm->mmap_cache = prev;
478 }
479
480 /*
481  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
482  * is already present in an i_mmap tree without adjusting the tree.
483  * The following helper function should be used when such adjustments
484  * are necessary.  The "insert" vma (if any) is to be inserted
485  * before we drop the necessary locks.
486  */
487 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
488         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
489 {
490         struct mm_struct *mm = vma->vm_mm;
491         struct vm_area_struct *next = vma->vm_next;
492         struct vm_area_struct *importer = NULL;
493         struct address_space *mapping = NULL;
494         struct prio_tree_root *root = NULL;
495         struct file *file = vma->vm_file;
496         struct anon_vma *anon_vma = NULL;
497         long adjust_next = 0;
498         int remove_next = 0;
499
500         if (next && !insert) {
501                 if (end >= next->vm_end) {
502                         /*
503                          * vma expands, overlapping all the next, and
504                          * perhaps the one after too (mprotect case 6).
505                          */
506 again:                  remove_next = 1 + (end > next->vm_end);
507                         end = next->vm_end;
508                         anon_vma = next->anon_vma;
509                         importer = vma;
510                 } else if (end > next->vm_start) {
511                         /*
512                          * vma expands, overlapping part of the next:
513                          * mprotect case 5 shifting the boundary up.
514                          */
515                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
516                         anon_vma = next->anon_vma;
517                         importer = vma;
518                 } else if (end < vma->vm_end) {
519                         /*
520                          * vma shrinks, and !insert tells it's not
521                          * split_vma inserting another: so it must be
522                          * mprotect case 4 shifting the boundary down.
523                          */
524                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
525                         anon_vma = next->anon_vma;
526                         importer = next;
527                 }
528         }
529
530         if (file) {
531                 mapping = file->f_mapping;
532                 if (!(vma->vm_flags & VM_NONLINEAR))
533                         root = &mapping->i_mmap;
534                 spin_lock(&mapping->i_mmap_lock);
535                 if (importer &&
536                     vma->vm_truncate_count != next->vm_truncate_count) {
537                         /*
538                          * unmap_mapping_range might be in progress:
539                          * ensure that the expanding vma is rescanned.
540                          */
541                         importer->vm_truncate_count = 0;
542                 }
543                 if (insert) {
544                         insert->vm_truncate_count = vma->vm_truncate_count;
545                         /*
546                          * Put into prio_tree now, so instantiated pages
547                          * are visible to arm/parisc __flush_dcache_page
548                          * throughout; but we cannot insert into address
549                          * space until vma start or end is updated.
550                          */
551                         __vma_link_file(insert);
552                 }
553         }
554
555         /*
556          * When changing only vma->vm_end, we don't really need
557          * anon_vma lock: but is that case worth optimizing out?
558          */
559         if (vma->anon_vma)
560                 anon_vma = vma->anon_vma;
561         if (anon_vma) {
562                 spin_lock(&anon_vma->lock);
563                 /*
564                  * Easily overlooked: when mprotect shifts the boundary,
565                  * make sure the expanding vma has anon_vma set if the
566                  * shrinking vma had, to cover any anon pages imported.
567                  */
568                 if (importer && !importer->anon_vma) {
569                         importer->anon_vma = anon_vma;
570                         __anon_vma_link(importer);
571                 }
572         }
573
574         if (root) {
575                 flush_dcache_mmap_lock(mapping);
576                 vma_prio_tree_remove(vma, root);
577                 if (adjust_next)
578                         vma_prio_tree_remove(next, root);
579         }
580
581         vma->vm_start = start;
582         vma->vm_end = end;
583         vma->vm_pgoff = pgoff;
584         if (adjust_next) {
585                 next->vm_start += adjust_next << PAGE_SHIFT;
586                 next->vm_pgoff += adjust_next;
587         }
588
589         if (root) {
590                 if (adjust_next)
591                         vma_prio_tree_insert(next, root);
592                 vma_prio_tree_insert(vma, root);
593                 flush_dcache_mmap_unlock(mapping);
594         }
595
596         if (remove_next) {
597                 /*
598                  * vma_merge has merged next into vma, and needs
599                  * us to remove next before dropping the locks.
600                  */
601                 __vma_unlink(mm, next, vma);
602                 if (file)
603                         __remove_shared_vm_struct(next, file, mapping);
604                 if (next->anon_vma)
605                         __anon_vma_merge(vma, next);
606         } else if (insert) {
607                 /*
608                  * split_vma has split insert from vma, and needs
609                  * us to insert it before dropping the locks
610                  * (it may either follow vma or precede it).
611                  */
612                 __insert_vm_struct(mm, insert);
613         }
614
615         if (anon_vma)
616                 spin_unlock(&anon_vma->lock);
617         if (mapping)
618                 spin_unlock(&mapping->i_mmap_lock);
619
620         if (remove_next) {
621                 if (file)
622                         fput(file);
623                 mm->map_count--;
624                 mpol_free(vma_policy(next));
625                 kmem_cache_free(vm_area_cachep, next);
626                 /*
627                  * In mprotect's case 6 (see comments on vma_merge),
628                  * we must remove another next too. It would clutter
629                  * up the code too much to do both in one go.
630                  */
631                 if (remove_next == 2) {
632                         next = vma->vm_next;
633                         goto again;
634                 }
635         }
636
637         validate_mm(mm);
638 }
639
640 /*
641  * If the vma has a ->close operation then the driver probably needs to release
642  * per-vma resources, so we don't attempt to merge those.
643  */
644 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
645
646 static inline int is_mergeable_vma(struct vm_area_struct *vma,
647                         struct file *file, unsigned long vm_flags)
648 {
649         if (vma->vm_flags != vm_flags)
650                 return 0;
651         if (vma->vm_file != file)
652                 return 0;
653         if (vma->vm_ops && vma->vm_ops->close)
654                 return 0;
655         return 1;
656 }
657
658 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
659                                         struct anon_vma *anon_vma2)
660 {
661         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
662 }
663
664 /*
665  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
666  * in front of (at a lower virtual address and file offset than) the vma.
667  *
668  * We cannot merge two vmas if they have differently assigned (non-NULL)
669  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
670  *
671  * We don't check here for the merged mmap wrapping around the end of pagecache
672  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
673  * wrap, nor mmaps which cover the final page at index -1UL.
674  */
675 static int
676 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
677         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
678 {
679         if (is_mergeable_vma(vma, file, vm_flags) &&
680             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
681                 if (vma->vm_pgoff == vm_pgoff)
682                         return 1;
683         }
684         return 0;
685 }
686
687 /*
688  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
689  * beyond (at a higher virtual address and file offset than) the vma.
690  *
691  * We cannot merge two vmas if they have differently assigned (non-NULL)
692  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
693  */
694 static int
695 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
696         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
697 {
698         if (is_mergeable_vma(vma, file, vm_flags) &&
699             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
700                 pgoff_t vm_pglen;
701                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
702                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
703                         return 1;
704         }
705         return 0;
706 }
707
708 /*
709  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
710  * whether that can be merged with its predecessor or its successor.
711  * Or both (it neatly fills a hole).
712  *
713  * In most cases - when called for mmap, brk or mremap - [addr,end) is
714  * certain not to be mapped by the time vma_merge is called; but when
715  * called for mprotect, it is certain to be already mapped (either at
716  * an offset within prev, or at the start of next), and the flags of
717  * this area are about to be changed to vm_flags - and the no-change
718  * case has already been eliminated.
719  *
720  * The following mprotect cases have to be considered, where AAAA is
721  * the area passed down from mprotect_fixup, never extending beyond one
722  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
723  *
724  *     AAAA             AAAA                AAAA          AAAA
725  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
726  *    cannot merge    might become    might become    might become
727  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
728  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
729  *    mremap move:                                    PPPPNNNNNNNN 8
730  *        AAAA
731  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
732  *    might become    case 1 below    case 2 below    case 3 below
733  *
734  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
735  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
736  */
737 struct vm_area_struct *vma_merge(struct mm_struct *mm,
738                         struct vm_area_struct *prev, unsigned long addr,
739                         unsigned long end, unsigned long vm_flags,
740                         struct anon_vma *anon_vma, struct file *file,
741                         pgoff_t pgoff, struct mempolicy *policy)
742 {
743         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
744         struct vm_area_struct *area, *next;
745
746         /*
747          * We later require that vma->vm_flags == vm_flags,
748          * so this tests vma->vm_flags & VM_SPECIAL, too.
749          */
750         if (vm_flags & VM_SPECIAL)
751                 return NULL;
752
753         if (prev)
754                 next = prev->vm_next;
755         else
756                 next = mm->mmap;
757         area = next;
758         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
759                 next = next->vm_next;
760
761         /*
762          * Can it merge with the predecessor?
763          */
764         if (prev && prev->vm_end == addr &&
765                         mpol_equal(vma_policy(prev), policy) &&
766                         can_vma_merge_after(prev, vm_flags,
767                                                 anon_vma, file, pgoff)) {
768                 /*
769                  * OK, it can.  Can we now merge in the successor as well?
770                  */
771                 if (next && end == next->vm_start &&
772                                 mpol_equal(policy, vma_policy(next)) &&
773                                 can_vma_merge_before(next, vm_flags,
774                                         anon_vma, file, pgoff+pglen) &&
775                                 is_mergeable_anon_vma(prev->anon_vma,
776                                                       next->anon_vma)) {
777                                                         /* cases 1, 6 */
778                         vma_adjust(prev, prev->vm_start,
779                                 next->vm_end, prev->vm_pgoff, NULL);
780                 } else                                  /* cases 2, 5, 7 */
781                         vma_adjust(prev, prev->vm_start,
782                                 end, prev->vm_pgoff, NULL);
783                 return prev;
784         }
785
786         /*
787          * Can this new request be merged in front of next?
788          */
789         if (next && end == next->vm_start &&
790                         mpol_equal(policy, vma_policy(next)) &&
791                         can_vma_merge_before(next, vm_flags,
792                                         anon_vma, file, pgoff+pglen)) {
793                 if (prev && addr < prev->vm_end)        /* case 4 */
794                         vma_adjust(prev, prev->vm_start,
795                                 addr, prev->vm_pgoff, NULL);
796                 else                                    /* cases 3, 8 */
797                         vma_adjust(area, addr, next->vm_end,
798                                 next->vm_pgoff - pglen, NULL);
799                 return area;
800         }
801
802         return NULL;
803 }
804
805 /*
806  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
807  * neighbouring vmas for a suitable anon_vma, before it goes off
808  * to allocate a new anon_vma.  It checks because a repetitive
809  * sequence of mprotects and faults may otherwise lead to distinct
810  * anon_vmas being allocated, preventing vma merge in subsequent
811  * mprotect.
812  */
813 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
814 {
815         struct vm_area_struct *near;
816         unsigned long vm_flags;
817
818         near = vma->vm_next;
819         if (!near)
820                 goto try_prev;
821
822         /*
823          * Since only mprotect tries to remerge vmas, match flags
824          * which might be mprotected into each other later on.
825          * Neither mlock nor madvise tries to remerge at present,
826          * so leave their flags as obstructing a merge.
827          */
828         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
829         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
830
831         if (near->anon_vma && vma->vm_end == near->vm_start &&
832                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
833                         can_vma_merge_before(near, vm_flags,
834                                 NULL, vma->vm_file, vma->vm_pgoff +
835                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
836                 return near->anon_vma;
837 try_prev:
838         /*
839          * It is potentially slow to have to call find_vma_prev here.
840          * But it's only on the first write fault on the vma, not
841          * every time, and we could devise a way to avoid it later
842          * (e.g. stash info in next's anon_vma_node when assigning
843          * an anon_vma, or when trying vma_merge).  Another time.
844          */
845         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
846         if (!near)
847                 goto none;
848
849         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
850         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
851
852         if (near->anon_vma && near->vm_end == vma->vm_start &&
853                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
854                         can_vma_merge_after(near, vm_flags,
855                                 NULL, vma->vm_file, vma->vm_pgoff))
856                 return near->anon_vma;
857 none:
858         /*
859          * There's no absolute need to look only at touching neighbours:
860          * we could search further afield for "compatible" anon_vmas.
861          * But it would probably just be a waste of time searching,
862          * or lead to too many vmas hanging off the same anon_vma.
863          * We're trying to allow mprotect remerging later on,
864          * not trying to minimize memory used for anon_vmas.
865          */
866         return NULL;
867 }
868
869 #ifdef CONFIG_PROC_FS
870 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
871                                                 struct file *file, long pages)
872 {
873         const unsigned long stack_flags
874                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
875
876         if (file) {
877                 mm->shared_vm += pages;
878                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
879                         mm->exec_vm += pages;
880         } else if (flags & stack_flags)
881                 mm->stack_vm += pages;
882         if (flags & (VM_RESERVED|VM_IO))
883                 mm->reserved_vm += pages;
884 }
885 #endif /* CONFIG_PROC_FS */
886
887 /*
888  * The caller must hold down_write(current->mm->mmap_sem).
889  */
890
891 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
892                         unsigned long len, unsigned long prot,
893                         unsigned long flags, unsigned long pgoff)
894 {
895         struct mm_struct * mm = current->mm;
896         struct inode *inode;
897         unsigned int vm_flags;
898         int error;
899         int accountable = 1;
900         unsigned long reqprot = prot;
901
902         /*
903          * Does the application expect PROT_READ to imply PROT_EXEC?
904          *
905          * (the exception is when the underlying filesystem is noexec
906          *  mounted, in which case we dont add PROT_EXEC.)
907          */
908         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
909                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
910                         prot |= PROT_EXEC;
911
912         if (!len)
913                 return -EINVAL;
914
915         if (!(flags & MAP_FIXED))
916                 addr = round_hint_to_min(addr);
917
918         error = arch_mmap_check(addr, len, flags);
919         if (error)
920                 return error;
921
922         /* Careful about overflows.. */
923         len = PAGE_ALIGN(len);
924         if (!len || len > TASK_SIZE)
925                 return -ENOMEM;
926
927         /* offset overflow? */
928         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
929                return -EOVERFLOW;
930
931         /* Too many mappings? */
932         if (mm->map_count > sysctl_max_map_count)
933                 return -ENOMEM;
934
935         /* Obtain the address to map to. we verify (or select) it and ensure
936          * that it represents a valid section of the address space.
937          */
938         addr = get_unmapped_area(file, addr, len, pgoff, flags);
939         if (addr & ~PAGE_MASK)
940                 return addr;
941
942         /* Do simple checking here so the lower-level routines won't have
943          * to. we assume access permissions have been handled by the open
944          * of the memory object, so we don't do any here.
945          */
946         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
947                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
948
949         if (flags & MAP_LOCKED) {
950                 if (!can_do_mlock())
951                         return -EPERM;
952                 vm_flags |= VM_LOCKED;
953         }
954         /* mlock MCL_FUTURE? */
955         if (vm_flags & VM_LOCKED) {
956                 unsigned long locked, lock_limit;
957                 locked = len >> PAGE_SHIFT;
958                 locked += mm->locked_vm;
959                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
960                 lock_limit >>= PAGE_SHIFT;
961                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
962                         return -EAGAIN;
963         }
964
965         inode = file ? file->f_path.dentry->d_inode : NULL;
966
967         if (file) {
968                 switch (flags & MAP_TYPE) {
969                 case MAP_SHARED:
970                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
971                                 return -EACCES;
972
973                         /*
974                          * Make sure we don't allow writing to an append-only
975                          * file..
976                          */
977                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
978                                 return -EACCES;
979
980                         /*
981                          * Make sure there are no mandatory locks on the file.
982                          */
983                         if (locks_verify_locked(inode))
984                                 return -EAGAIN;
985
986                         vm_flags |= VM_SHARED | VM_MAYSHARE;
987                         if (!(file->f_mode & FMODE_WRITE))
988                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
989
990                         /* fall through */
991                 case MAP_PRIVATE:
992                         if (!(file->f_mode & FMODE_READ))
993                                 return -EACCES;
994                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
995                                 if (vm_flags & VM_EXEC)
996                                         return -EPERM;
997                                 vm_flags &= ~VM_MAYEXEC;
998                         }
999                         if (is_file_hugepages(file))
1000                                 accountable = 0;
1001
1002                         if (!file->f_op || !file->f_op->mmap)
1003                                 return -ENODEV;
1004                         break;
1005
1006                 default:
1007                         return -EINVAL;
1008                 }
1009         } else {
1010                 switch (flags & MAP_TYPE) {
1011                 case MAP_SHARED:
1012                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1013                         break;
1014                 case MAP_PRIVATE:
1015                         /*
1016                          * Set pgoff according to addr for anon_vma.
1017                          */
1018                         pgoff = addr >> PAGE_SHIFT;
1019                         break;
1020                 default:
1021                         return -EINVAL;
1022                 }
1023         }
1024
1025         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1026         if (error)
1027                 return error;
1028
1029         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1030                            accountable);
1031 }
1032 EXPORT_SYMBOL(do_mmap_pgoff);
1033
1034 /*
1035  * Some shared mappigns will want the pages marked read-only
1036  * to track write events. If so, we'll downgrade vm_page_prot
1037  * to the private version (using protection_map[] without the
1038  * VM_SHARED bit).
1039  */
1040 int vma_wants_writenotify(struct vm_area_struct *vma)
1041 {
1042         unsigned int vm_flags = vma->vm_flags;
1043
1044         /* If it was private or non-writable, the write bit is already clear */
1045         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1046                 return 0;
1047
1048         /* The backer wishes to know when pages are first written to? */
1049         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1050                 return 1;
1051
1052         /* The open routine did something to the protections already? */
1053         if (pgprot_val(vma->vm_page_prot) !=
1054             pgprot_val(vm_get_page_prot(vm_flags)))
1055                 return 0;
1056
1057         /* Specialty mapping? */
1058         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1059                 return 0;
1060
1061         /* Can the mapping track the dirty pages? */
1062         return vma->vm_file && vma->vm_file->f_mapping &&
1063                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1064 }
1065
1066
1067 unsigned long mmap_region(struct file *file, unsigned long addr,
1068                           unsigned long len, unsigned long flags,
1069                           unsigned int vm_flags, unsigned long pgoff,
1070                           int accountable)
1071 {
1072         struct mm_struct *mm = current->mm;
1073         struct vm_area_struct *vma, *prev;
1074         int correct_wcount = 0;
1075         int error;
1076         struct rb_node **rb_link, *rb_parent;
1077         unsigned long charged = 0;
1078         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1079
1080         /* Clear old maps */
1081         error = -ENOMEM;
1082 munmap_back:
1083         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1084         if (vma && vma->vm_start < addr + len) {
1085                 if (do_munmap(mm, addr, len))
1086                         return -ENOMEM;
1087                 goto munmap_back;
1088         }
1089
1090         /* Check against address space limit. */
1091         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1092                 return -ENOMEM;
1093
1094         if (accountable && (!(flags & MAP_NORESERVE) ||
1095                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1096                 if (vm_flags & VM_SHARED) {
1097                         /* Check memory availability in shmem_file_setup? */
1098                         vm_flags |= VM_ACCOUNT;
1099                 } else if (vm_flags & VM_WRITE) {
1100                         /*
1101                          * Private writable mapping: check memory availability
1102                          */
1103                         charged = len >> PAGE_SHIFT;
1104                         if (security_vm_enough_memory(charged))
1105                                 return -ENOMEM;
1106                         vm_flags |= VM_ACCOUNT;
1107                 }
1108         }
1109
1110         /*
1111          * Can we just expand an old private anonymous mapping?
1112          * The VM_SHARED test is necessary because shmem_zero_setup
1113          * will create the file object for a shared anonymous map below.
1114          */
1115         if (!file && !(vm_flags & VM_SHARED) &&
1116             vma_merge(mm, prev, addr, addr + len, vm_flags,
1117                                         NULL, NULL, pgoff, NULL))
1118                 goto out;
1119
1120         /*
1121          * Determine the object being mapped and call the appropriate
1122          * specific mapper. the address has already been validated, but
1123          * not unmapped, but the maps are removed from the list.
1124          */
1125         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1126         if (!vma) {
1127                 error = -ENOMEM;
1128                 goto unacct_error;
1129         }
1130
1131         vma->vm_mm = mm;
1132         vma->vm_start = addr;
1133         vma->vm_end = addr + len;
1134         vma->vm_flags = vm_flags;
1135         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1136         vma->vm_pgoff = pgoff;
1137
1138         if (file) {
1139                 error = -EINVAL;
1140                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1141                         goto free_vma;
1142                 if (vm_flags & VM_DENYWRITE) {
1143                         error = deny_write_access(file);
1144                         if (error)
1145                                 goto free_vma;
1146                         correct_wcount = 1;
1147                 }
1148                 vma->vm_file = file;
1149                 get_file(file);
1150                 error = file->f_op->mmap(file, vma);
1151                 if (error)
1152                         goto unmap_and_free_vma;
1153         } else if (vm_flags & VM_SHARED) {
1154                 error = shmem_zero_setup(vma);
1155                 if (error)
1156                         goto free_vma;
1157         }
1158
1159         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1160          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1161          * that memory reservation must be checked; but that reservation
1162          * belongs to shared memory object, not to vma: so now clear it.
1163          */
1164         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1165                 vma->vm_flags &= ~VM_ACCOUNT;
1166
1167         /* Can addr have changed??
1168          *
1169          * Answer: Yes, several device drivers can do it in their
1170          *         f_op->mmap method. -DaveM
1171          */
1172         addr = vma->vm_start;
1173         pgoff = vma->vm_pgoff;
1174         vm_flags = vma->vm_flags;
1175
1176         if (vma_wants_writenotify(vma))
1177                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1178
1179         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1180                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1181                 file = vma->vm_file;
1182                 vma_link(mm, vma, prev, rb_link, rb_parent);
1183                 if (correct_wcount)
1184                         atomic_inc(&inode->i_writecount);
1185         } else {
1186                 if (file) {
1187                         if (correct_wcount)
1188                                 atomic_inc(&inode->i_writecount);
1189                         fput(file);
1190                 }
1191                 mpol_free(vma_policy(vma));
1192                 kmem_cache_free(vm_area_cachep, vma);
1193         }
1194 out:    
1195         mm->total_vm += len >> PAGE_SHIFT;
1196         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1197         if (vm_flags & VM_LOCKED) {
1198                 mm->locked_vm += len >> PAGE_SHIFT;
1199                 make_pages_present(addr, addr + len);
1200         }
1201         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1202                 make_pages_present(addr, addr + len);
1203         return addr;
1204
1205 unmap_and_free_vma:
1206         if (correct_wcount)
1207                 atomic_inc(&inode->i_writecount);
1208         vma->vm_file = NULL;
1209         fput(file);
1210
1211         /* Undo any partial mapping done by a device driver. */
1212         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1213         charged = 0;
1214 free_vma:
1215         kmem_cache_free(vm_area_cachep, vma);
1216 unacct_error:
1217         if (charged)
1218                 vm_unacct_memory(charged);
1219         return error;
1220 }
1221
1222 /* Get an address range which is currently unmapped.
1223  * For shmat() with addr=0.
1224  *
1225  * Ugly calling convention alert:
1226  * Return value with the low bits set means error value,
1227  * ie
1228  *      if (ret & ~PAGE_MASK)
1229  *              error = ret;
1230  *
1231  * This function "knows" that -ENOMEM has the bits set.
1232  */
1233 #ifndef HAVE_ARCH_UNMAPPED_AREA
1234 unsigned long
1235 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1236                 unsigned long len, unsigned long pgoff, unsigned long flags)
1237 {
1238         struct mm_struct *mm = current->mm;
1239         struct vm_area_struct *vma;
1240         unsigned long start_addr;
1241
1242         if (len > TASK_SIZE)
1243                 return -ENOMEM;
1244
1245         if (flags & MAP_FIXED)
1246                 return addr;
1247
1248         if (addr) {
1249                 addr = PAGE_ALIGN(addr);
1250                 vma = find_vma(mm, addr);
1251                 if (TASK_SIZE - len >= addr &&
1252                     (!vma || addr + len <= vma->vm_start))
1253                         return addr;
1254         }
1255         if (len > mm->cached_hole_size) {
1256                 start_addr = addr = mm->free_area_cache;
1257         } else {
1258                 start_addr = addr = TASK_UNMAPPED_BASE;
1259                 mm->cached_hole_size = 0;
1260         }
1261
1262 full_search:
1263         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1264                 /* At this point:  (!vma || addr < vma->vm_end). */
1265                 if (TASK_SIZE - len < addr) {
1266                         /*
1267                          * Start a new search - just in case we missed
1268                          * some holes.
1269                          */
1270                         if (start_addr != TASK_UNMAPPED_BASE) {
1271                                 addr = TASK_UNMAPPED_BASE;
1272                                 start_addr = addr;
1273                                 mm->cached_hole_size = 0;
1274                                 goto full_search;
1275                         }
1276                         return -ENOMEM;
1277                 }
1278                 if (!vma || addr + len <= vma->vm_start) {
1279                         /*
1280                          * Remember the place where we stopped the search:
1281                          */
1282                         mm->free_area_cache = addr + len;
1283                         return addr;
1284                 }
1285                 if (addr + mm->cached_hole_size < vma->vm_start)
1286                         mm->cached_hole_size = vma->vm_start - addr;
1287                 addr = vma->vm_end;
1288         }
1289 }
1290 #endif  
1291
1292 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1293 {
1294         /*
1295          * Is this a new hole at the lowest possible address?
1296          */
1297         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1298                 mm->free_area_cache = addr;
1299                 mm->cached_hole_size = ~0UL;
1300         }
1301 }
1302
1303 /*
1304  * This mmap-allocator allocates new areas top-down from below the
1305  * stack's low limit (the base):
1306  */
1307 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1308 unsigned long
1309 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1310                           const unsigned long len, const unsigned long pgoff,
1311                           const unsigned long flags)
1312 {
1313         struct vm_area_struct *vma;
1314         struct mm_struct *mm = current->mm;
1315         unsigned long addr = addr0;
1316
1317         /* requested length too big for entire address space */
1318         if (len > TASK_SIZE)
1319                 return -ENOMEM;
1320
1321         if (flags & MAP_FIXED)
1322                 return addr;
1323
1324         /* requesting a specific address */
1325         if (addr) {
1326                 addr = PAGE_ALIGN(addr);
1327                 vma = find_vma(mm, addr);
1328                 if (TASK_SIZE - len >= addr &&
1329                                 (!vma || addr + len <= vma->vm_start))
1330                         return addr;
1331         }
1332
1333         /* check if free_area_cache is useful for us */
1334         if (len <= mm->cached_hole_size) {
1335                 mm->cached_hole_size = 0;
1336                 mm->free_area_cache = mm->mmap_base;
1337         }
1338
1339         /* either no address requested or can't fit in requested address hole */
1340         addr = mm->free_area_cache;
1341
1342         /* make sure it can fit in the remaining address space */
1343         if (addr > len) {
1344                 vma = find_vma(mm, addr-len);
1345                 if (!vma || addr <= vma->vm_start)
1346                         /* remember the address as a hint for next time */
1347                         return (mm->free_area_cache = addr-len);
1348         }
1349
1350         if (mm->mmap_base < len)
1351                 goto bottomup;
1352
1353         addr = mm->mmap_base-len;
1354
1355         do {
1356                 /*
1357                  * Lookup failure means no vma is above this address,
1358                  * else if new region fits below vma->vm_start,
1359                  * return with success:
1360                  */
1361                 vma = find_vma(mm, addr);
1362                 if (!vma || addr+len <= vma->vm_start)
1363                         /* remember the address as a hint for next time */
1364                         return (mm->free_area_cache = addr);
1365
1366                 /* remember the largest hole we saw so far */
1367                 if (addr + mm->cached_hole_size < vma->vm_start)
1368                         mm->cached_hole_size = vma->vm_start - addr;
1369
1370                 /* try just below the current vma->vm_start */
1371                 addr = vma->vm_start-len;
1372         } while (len < vma->vm_start);
1373
1374 bottomup:
1375         /*
1376          * A failed mmap() very likely causes application failure,
1377          * so fall back to the bottom-up function here. This scenario
1378          * can happen with large stack limits and large mmap()
1379          * allocations.
1380          */
1381         mm->cached_hole_size = ~0UL;
1382         mm->free_area_cache = TASK_UNMAPPED_BASE;
1383         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1384         /*
1385          * Restore the topdown base:
1386          */
1387         mm->free_area_cache = mm->mmap_base;
1388         mm->cached_hole_size = ~0UL;
1389
1390         return addr;
1391 }
1392 #endif
1393
1394 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1395 {
1396         /*
1397          * Is this a new hole at the highest possible address?
1398          */
1399         if (addr > mm->free_area_cache)
1400                 mm->free_area_cache = addr;
1401
1402         /* dont allow allocations above current base */
1403         if (mm->free_area_cache > mm->mmap_base)
1404                 mm->free_area_cache = mm->mmap_base;
1405 }
1406
1407 unsigned long
1408 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1409                 unsigned long pgoff, unsigned long flags)
1410 {
1411         unsigned long (*get_area)(struct file *, unsigned long,
1412                                   unsigned long, unsigned long, unsigned long);
1413
1414         get_area = current->mm->get_unmapped_area;
1415         if (file && file->f_op && file->f_op->get_unmapped_area)
1416                 get_area = file->f_op->get_unmapped_area;
1417         addr = get_area(file, addr, len, pgoff, flags);
1418         if (IS_ERR_VALUE(addr))
1419                 return addr;
1420
1421         if (addr > TASK_SIZE - len)
1422                 return -ENOMEM;
1423         if (addr & ~PAGE_MASK)
1424                 return -EINVAL;
1425
1426         return addr;
1427 }
1428
1429 EXPORT_SYMBOL(get_unmapped_area);
1430
1431 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1432 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1433 {
1434         struct vm_area_struct *vma = NULL;
1435
1436         if (mm) {
1437                 /* Check the cache first. */
1438                 /* (Cache hit rate is typically around 35%.) */
1439                 vma = mm->mmap_cache;
1440                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1441                         struct rb_node * rb_node;
1442
1443                         rb_node = mm->mm_rb.rb_node;
1444                         vma = NULL;
1445
1446                         while (rb_node) {
1447                                 struct vm_area_struct * vma_tmp;
1448
1449                                 vma_tmp = rb_entry(rb_node,
1450                                                 struct vm_area_struct, vm_rb);
1451
1452                                 if (vma_tmp->vm_end > addr) {
1453                                         vma = vma_tmp;
1454                                         if (vma_tmp->vm_start <= addr)
1455                                                 break;
1456                                         rb_node = rb_node->rb_left;
1457                                 } else
1458                                         rb_node = rb_node->rb_right;
1459                         }
1460                         if (vma)
1461                                 mm->mmap_cache = vma;
1462                 }
1463         }
1464         return vma;
1465 }
1466
1467 EXPORT_SYMBOL(find_vma);
1468
1469 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1470 struct vm_area_struct *
1471 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1472                         struct vm_area_struct **pprev)
1473 {
1474         struct vm_area_struct *vma = NULL, *prev = NULL;
1475         struct rb_node * rb_node;
1476         if (!mm)
1477                 goto out;
1478
1479         /* Guard against addr being lower than the first VMA */
1480         vma = mm->mmap;
1481
1482         /* Go through the RB tree quickly. */
1483         rb_node = mm->mm_rb.rb_node;
1484
1485         while (rb_node) {
1486                 struct vm_area_struct *vma_tmp;
1487                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1488
1489                 if (addr < vma_tmp->vm_end) {
1490                         rb_node = rb_node->rb_left;
1491                 } else {
1492                         prev = vma_tmp;
1493                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1494                                 break;
1495                         rb_node = rb_node->rb_right;
1496                 }
1497         }
1498
1499 out:
1500         *pprev = prev;
1501         return prev ? prev->vm_next : vma;
1502 }
1503
1504 /*
1505  * Verify that the stack growth is acceptable and
1506  * update accounting. This is shared with both the
1507  * grow-up and grow-down cases.
1508  */
1509 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1510 {
1511         struct mm_struct *mm = vma->vm_mm;
1512         struct rlimit *rlim = current->signal->rlim;
1513         unsigned long new_start;
1514
1515         /* address space limit tests */
1516         if (!may_expand_vm(mm, grow))
1517                 return -ENOMEM;
1518
1519         /* Stack limit test */
1520         if (size > rlim[RLIMIT_STACK].rlim_cur)
1521                 return -ENOMEM;
1522
1523         /* mlock limit tests */
1524         if (vma->vm_flags & VM_LOCKED) {
1525                 unsigned long locked;
1526                 unsigned long limit;
1527                 locked = mm->locked_vm + grow;
1528                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1529                 if (locked > limit && !capable(CAP_IPC_LOCK))
1530                         return -ENOMEM;
1531         }
1532
1533         /* Check to ensure the stack will not grow into a hugetlb-only region */
1534         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1535                         vma->vm_end - size;
1536         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1537                 return -EFAULT;
1538
1539         /*
1540          * Overcommit..  This must be the final test, as it will
1541          * update security statistics.
1542          */
1543         if (security_vm_enough_memory(grow))
1544                 return -ENOMEM;
1545
1546         /* Ok, everything looks good - let it rip */
1547         mm->total_vm += grow;
1548         if (vma->vm_flags & VM_LOCKED)
1549                 mm->locked_vm += grow;
1550         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1551         return 0;
1552 }
1553
1554 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1555 /*
1556  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1557  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1558  */
1559 #ifndef CONFIG_IA64
1560 static inline
1561 #endif
1562 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1563 {
1564         int error;
1565
1566         if (!(vma->vm_flags & VM_GROWSUP))
1567                 return -EFAULT;
1568
1569         /*
1570          * We must make sure the anon_vma is allocated
1571          * so that the anon_vma locking is not a noop.
1572          */
1573         if (unlikely(anon_vma_prepare(vma)))
1574                 return -ENOMEM;
1575         anon_vma_lock(vma);
1576
1577         /*
1578          * vma->vm_start/vm_end cannot change under us because the caller
1579          * is required to hold the mmap_sem in read mode.  We need the
1580          * anon_vma lock to serialize against concurrent expand_stacks.
1581          * Also guard against wrapping around to address 0.
1582          */
1583         if (address < PAGE_ALIGN(address+4))
1584                 address = PAGE_ALIGN(address+4);
1585         else {
1586                 anon_vma_unlock(vma);
1587                 return -ENOMEM;
1588         }
1589         error = 0;
1590
1591         /* Somebody else might have raced and expanded it already */
1592         if (address > vma->vm_end) {
1593                 unsigned long size, grow;
1594
1595                 size = address - vma->vm_start;
1596                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1597
1598                 error = acct_stack_growth(vma, size, grow);
1599                 if (!error)
1600                         vma->vm_end = address;
1601         }
1602         anon_vma_unlock(vma);
1603         return error;
1604 }
1605 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1606
1607 /*
1608  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1609  */
1610 static inline int expand_downwards(struct vm_area_struct *vma,
1611                                    unsigned long address)
1612 {
1613         int error;
1614
1615         /*
1616          * We must make sure the anon_vma is allocated
1617          * so that the anon_vma locking is not a noop.
1618          */
1619         if (unlikely(anon_vma_prepare(vma)))
1620                 return -ENOMEM;
1621
1622         address &= PAGE_MASK;
1623         error = security_file_mmap(0, 0, 0, 0, address, 1);
1624         if (error)
1625                 return error;
1626
1627         anon_vma_lock(vma);
1628
1629         /*
1630          * vma->vm_start/vm_end cannot change under us because the caller
1631          * is required to hold the mmap_sem in read mode.  We need the
1632          * anon_vma lock to serialize against concurrent expand_stacks.
1633          */
1634
1635         /* Somebody else might have raced and expanded it already */
1636         if (address < vma->vm_start) {
1637                 unsigned long size, grow;
1638
1639                 size = vma->vm_end - address;
1640                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1641
1642                 error = acct_stack_growth(vma, size, grow);
1643                 if (!error) {
1644                         vma->vm_start = address;
1645                         vma->vm_pgoff -= grow;
1646                 }
1647         }
1648         anon_vma_unlock(vma);
1649         return error;
1650 }
1651
1652 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1653 {
1654         return expand_downwards(vma, address);
1655 }
1656
1657 #ifdef CONFIG_STACK_GROWSUP
1658 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1659 {
1660         return expand_upwards(vma, address);
1661 }
1662
1663 struct vm_area_struct *
1664 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1665 {
1666         struct vm_area_struct *vma, *prev;
1667
1668         addr &= PAGE_MASK;
1669         vma = find_vma_prev(mm, addr, &prev);
1670         if (vma && (vma->vm_start <= addr))
1671                 return vma;
1672         if (!prev || expand_stack(prev, addr))
1673                 return NULL;
1674         if (prev->vm_flags & VM_LOCKED)
1675                 make_pages_present(addr, prev->vm_end);
1676         return prev;
1677 }
1678 #else
1679 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1680 {
1681         return expand_downwards(vma, address);
1682 }
1683
1684 struct vm_area_struct *
1685 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1686 {
1687         struct vm_area_struct * vma;
1688         unsigned long start;
1689
1690         addr &= PAGE_MASK;
1691         vma = find_vma(mm,addr);
1692         if (!vma)
1693                 return NULL;
1694         if (vma->vm_start <= addr)
1695                 return vma;
1696         if (!(vma->vm_flags & VM_GROWSDOWN))
1697                 return NULL;
1698         start = vma->vm_start;
1699         if (expand_stack(vma, addr))
1700                 return NULL;
1701         if (vma->vm_flags & VM_LOCKED)
1702                 make_pages_present(addr, start);
1703         return vma;
1704 }
1705 #endif
1706
1707 /*
1708  * Ok - we have the memory areas we should free on the vma list,
1709  * so release them, and do the vma updates.
1710  *
1711  * Called with the mm semaphore held.
1712  */
1713 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1714 {
1715         /* Update high watermark before we lower total_vm */
1716         update_hiwater_vm(mm);
1717         do {
1718                 long nrpages = vma_pages(vma);
1719
1720                 mm->total_vm -= nrpages;
1721                 if (vma->vm_flags & VM_LOCKED)
1722                         mm->locked_vm -= nrpages;
1723                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1724                 vma = remove_vma(vma);
1725         } while (vma);
1726         validate_mm(mm);
1727 }
1728
1729 /*
1730  * Get rid of page table information in the indicated region.
1731  *
1732  * Called with the mm semaphore held.
1733  */
1734 static void unmap_region(struct mm_struct *mm,
1735                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1736                 unsigned long start, unsigned long end)
1737 {
1738         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1739         struct mmu_gather *tlb;
1740         unsigned long nr_accounted = 0;
1741
1742         lru_add_drain();
1743         tlb = tlb_gather_mmu(mm, 0);
1744         update_hiwater_rss(mm);
1745         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1746         vm_unacct_memory(nr_accounted);
1747         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1748                                  next? next->vm_start: 0);
1749         tlb_finish_mmu(tlb, start, end);
1750 }
1751
1752 /*
1753  * Create a list of vma's touched by the unmap, removing them from the mm's
1754  * vma list as we go..
1755  */
1756 static void
1757 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1758         struct vm_area_struct *prev, unsigned long end)
1759 {
1760         struct vm_area_struct **insertion_point;
1761         struct vm_area_struct *tail_vma = NULL;
1762         unsigned long addr;
1763
1764         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1765         do {
1766                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1767                 mm->map_count--;
1768                 tail_vma = vma;
1769                 vma = vma->vm_next;
1770         } while (vma && vma->vm_start < end);
1771         *insertion_point = vma;
1772         tail_vma->vm_next = NULL;
1773         if (mm->unmap_area == arch_unmap_area)
1774                 addr = prev ? prev->vm_end : mm->mmap_base;
1775         else
1776                 addr = vma ?  vma->vm_start : mm->mmap_base;
1777         mm->unmap_area(mm, addr);
1778         mm->mmap_cache = NULL;          /* Kill the cache. */
1779 }
1780
1781 /*
1782  * Split a vma into two pieces at address 'addr', a new vma is allocated
1783  * either for the first part or the tail.
1784  */
1785 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1786               unsigned long addr, int new_below)
1787 {
1788         struct mempolicy *pol;
1789         struct vm_area_struct *new;
1790
1791         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1792                 return -EINVAL;
1793
1794         if (mm->map_count >= sysctl_max_map_count)
1795                 return -ENOMEM;
1796
1797         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1798         if (!new)
1799                 return -ENOMEM;
1800
1801         /* most fields are the same, copy all, and then fixup */
1802         *new = *vma;
1803
1804         if (new_below)
1805                 new->vm_end = addr;
1806         else {
1807                 new->vm_start = addr;
1808                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1809         }
1810
1811         pol = mpol_copy(vma_policy(vma));
1812         if (IS_ERR(pol)) {
1813                 kmem_cache_free(vm_area_cachep, new);
1814                 return PTR_ERR(pol);
1815         }
1816         vma_set_policy(new, pol);
1817
1818         if (new->vm_file)
1819                 get_file(new->vm_file);
1820
1821         if (new->vm_ops && new->vm_ops->open)
1822                 new->vm_ops->open(new);
1823
1824         if (new_below)
1825                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1826                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1827         else
1828                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1829
1830         return 0;
1831 }
1832
1833 /* Munmap is split into 2 main parts -- this part which finds
1834  * what needs doing, and the areas themselves, which do the
1835  * work.  This now handles partial unmappings.
1836  * Jeremy Fitzhardinge <jeremy@goop.org>
1837  */
1838 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1839 {
1840         unsigned long end;
1841         struct vm_area_struct *vma, *prev, *last;
1842
1843         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1844                 return -EINVAL;
1845
1846         if ((len = PAGE_ALIGN(len)) == 0)
1847                 return -EINVAL;
1848
1849         /* Find the first overlapping VMA */
1850         vma = find_vma_prev(mm, start, &prev);
1851         if (!vma)
1852                 return 0;
1853         /* we have  start < vma->vm_end  */
1854
1855         /* if it doesn't overlap, we have nothing.. */
1856         end = start + len;
1857         if (vma->vm_start >= end)
1858                 return 0;
1859
1860         /*
1861          * If we need to split any vma, do it now to save pain later.
1862          *
1863          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1864          * unmapped vm_area_struct will remain in use: so lower split_vma
1865          * places tmp vma above, and higher split_vma places tmp vma below.
1866          */
1867         if (start > vma->vm_start) {
1868                 int error = split_vma(mm, vma, start, 0);
1869                 if (error)
1870                         return error;
1871                 prev = vma;
1872         }
1873
1874         /* Does it split the last one? */
1875         last = find_vma(mm, end);
1876         if (last && end > last->vm_start) {
1877                 int error = split_vma(mm, last, end, 1);
1878                 if (error)
1879                         return error;
1880         }
1881         vma = prev? prev->vm_next: mm->mmap;
1882
1883         /*
1884          * Remove the vma's, and unmap the actual pages
1885          */
1886         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1887         unmap_region(mm, vma, prev, start, end);
1888
1889         /* Fix up all other VM information */
1890         remove_vma_list(mm, vma);
1891
1892         return 0;
1893 }
1894
1895 EXPORT_SYMBOL(do_munmap);
1896
1897 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1898 {
1899         int ret;
1900         struct mm_struct *mm = current->mm;
1901
1902         profile_munmap(addr);
1903
1904         down_write(&mm->mmap_sem);
1905         ret = do_munmap(mm, addr, len);
1906         up_write(&mm->mmap_sem);
1907         return ret;
1908 }
1909
1910 static inline void verify_mm_writelocked(struct mm_struct *mm)
1911 {
1912 #ifdef CONFIG_DEBUG_VM
1913         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1914                 WARN_ON(1);
1915                 up_read(&mm->mmap_sem);
1916         }
1917 #endif
1918 }
1919
1920 /*
1921  *  this is really a simplified "do_mmap".  it only handles
1922  *  anonymous maps.  eventually we may be able to do some
1923  *  brk-specific accounting here.
1924  */
1925 unsigned long do_brk(unsigned long addr, unsigned long len)
1926 {
1927         struct mm_struct * mm = current->mm;
1928         struct vm_area_struct * vma, * prev;
1929         unsigned long flags;
1930         struct rb_node ** rb_link, * rb_parent;
1931         pgoff_t pgoff = addr >> PAGE_SHIFT;
1932         int error;
1933
1934         len = PAGE_ALIGN(len);
1935         if (!len)
1936                 return addr;
1937
1938         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1939                 return -EINVAL;
1940
1941         if (is_hugepage_only_range(mm, addr, len))
1942                 return -EINVAL;
1943
1944         error = security_file_mmap(0, 0, 0, 0, addr, 1);
1945         if (error)
1946                 return error;
1947
1948         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1949
1950         error = arch_mmap_check(addr, len, flags);
1951         if (error)
1952                 return error;
1953
1954         /*
1955          * mlock MCL_FUTURE?
1956          */
1957         if (mm->def_flags & VM_LOCKED) {
1958                 unsigned long locked, lock_limit;
1959                 locked = len >> PAGE_SHIFT;
1960                 locked += mm->locked_vm;
1961                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1962                 lock_limit >>= PAGE_SHIFT;
1963                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1964                         return -EAGAIN;
1965         }
1966
1967         /*
1968          * mm->mmap_sem is required to protect against another thread
1969          * changing the mappings in case we sleep.
1970          */
1971         verify_mm_writelocked(mm);
1972
1973         /*
1974          * Clear old maps.  this also does some error checking for us
1975          */
1976  munmap_back:
1977         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1978         if (vma && vma->vm_start < addr + len) {
1979                 if (do_munmap(mm, addr, len))
1980                         return -ENOMEM;
1981                 goto munmap_back;
1982         }
1983
1984         /* Check against address space limits *after* clearing old maps... */
1985         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1986                 return -ENOMEM;
1987
1988         if (mm->map_count > sysctl_max_map_count)
1989                 return -ENOMEM;
1990
1991         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1992                 return -ENOMEM;
1993
1994         /* Can we just expand an old private anonymous mapping? */
1995         if (vma_merge(mm, prev, addr, addr + len, flags,
1996                                         NULL, NULL, pgoff, NULL))
1997                 goto out;
1998
1999         /*
2000          * create a vma struct for an anonymous mapping
2001          */
2002         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2003         if (!vma) {
2004                 vm_unacct_memory(len >> PAGE_SHIFT);
2005                 return -ENOMEM;
2006         }
2007
2008         vma->vm_mm = mm;
2009         vma->vm_start = addr;
2010         vma->vm_end = addr + len;
2011         vma->vm_pgoff = pgoff;
2012         vma->vm_flags = flags;
2013         vma->vm_page_prot = vm_get_page_prot(flags);
2014         vma_link(mm, vma, prev, rb_link, rb_parent);
2015 out:
2016         mm->total_vm += len >> PAGE_SHIFT;
2017         if (flags & VM_LOCKED) {
2018                 mm->locked_vm += len >> PAGE_SHIFT;
2019                 make_pages_present(addr, addr + len);
2020         }
2021         return addr;
2022 }
2023
2024 EXPORT_SYMBOL(do_brk);
2025
2026 /* Release all mmaps. */
2027 void exit_mmap(struct mm_struct *mm)
2028 {
2029         struct mmu_gather *tlb;
2030         struct vm_area_struct *vma = mm->mmap;
2031         unsigned long nr_accounted = 0;
2032         unsigned long end;
2033
2034         /* mm's last user has gone, and its about to be pulled down */
2035         arch_exit_mmap(mm);
2036
2037         lru_add_drain();
2038         flush_cache_mm(mm);
2039         tlb = tlb_gather_mmu(mm, 1);
2040         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2041         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2042         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2043         vm_unacct_memory(nr_accounted);
2044         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2045         tlb_finish_mmu(tlb, 0, end);
2046
2047         /*
2048          * Walk the list again, actually closing and freeing it,
2049          * with preemption enabled, without holding any MM locks.
2050          */
2051         while (vma)
2052                 vma = remove_vma(vma);
2053
2054         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2055 }
2056
2057 /* Insert vm structure into process list sorted by address
2058  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2059  * then i_mmap_lock is taken here.
2060  */
2061 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2062 {
2063         struct vm_area_struct * __vma, * prev;
2064         struct rb_node ** rb_link, * rb_parent;
2065
2066         /*
2067          * The vm_pgoff of a purely anonymous vma should be irrelevant
2068          * until its first write fault, when page's anon_vma and index
2069          * are set.  But now set the vm_pgoff it will almost certainly
2070          * end up with (unless mremap moves it elsewhere before that
2071          * first wfault), so /proc/pid/maps tells a consistent story.
2072          *
2073          * By setting it to reflect the virtual start address of the
2074          * vma, merges and splits can happen in a seamless way, just
2075          * using the existing file pgoff checks and manipulations.
2076          * Similarly in do_mmap_pgoff and in do_brk.
2077          */
2078         if (!vma->vm_file) {
2079                 BUG_ON(vma->anon_vma);
2080                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2081         }
2082         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2083         if (__vma && __vma->vm_start < vma->vm_end)
2084                 return -ENOMEM;
2085         if ((vma->vm_flags & VM_ACCOUNT) &&
2086              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2087                 return -ENOMEM;
2088         vma_link(mm, vma, prev, rb_link, rb_parent);
2089         return 0;
2090 }
2091
2092 /*
2093  * Copy the vma structure to a new location in the same mm,
2094  * prior to moving page table entries, to effect an mremap move.
2095  */
2096 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2097         unsigned long addr, unsigned long len, pgoff_t pgoff)
2098 {
2099         struct vm_area_struct *vma = *vmap;
2100         unsigned long vma_start = vma->vm_start;
2101         struct mm_struct *mm = vma->vm_mm;
2102         struct vm_area_struct *new_vma, *prev;
2103         struct rb_node **rb_link, *rb_parent;
2104         struct mempolicy *pol;
2105
2106         /*
2107          * If anonymous vma has not yet been faulted, update new pgoff
2108          * to match new location, to increase its chance of merging.
2109          */
2110         if (!vma->vm_file && !vma->anon_vma)
2111                 pgoff = addr >> PAGE_SHIFT;
2112
2113         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2114         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2115                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2116         if (new_vma) {
2117                 /*
2118                  * Source vma may have been merged into new_vma
2119                  */
2120                 if (vma_start >= new_vma->vm_start &&
2121                     vma_start < new_vma->vm_end)
2122                         *vmap = new_vma;
2123         } else {
2124                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2125                 if (new_vma) {
2126                         *new_vma = *vma;
2127                         pol = mpol_copy(vma_policy(vma));
2128                         if (IS_ERR(pol)) {
2129                                 kmem_cache_free(vm_area_cachep, new_vma);
2130                                 return NULL;
2131                         }
2132                         vma_set_policy(new_vma, pol);
2133                         new_vma->vm_start = addr;
2134                         new_vma->vm_end = addr + len;
2135                         new_vma->vm_pgoff = pgoff;
2136                         if (new_vma->vm_file)
2137                                 get_file(new_vma->vm_file);
2138                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2139                                 new_vma->vm_ops->open(new_vma);
2140                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2141                 }
2142         }
2143         return new_vma;
2144 }
2145
2146 /*
2147  * Return true if the calling process may expand its vm space by the passed
2148  * number of pages
2149  */
2150 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2151 {
2152         unsigned long cur = mm->total_vm;       /* pages */
2153         unsigned long lim;
2154
2155         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2156
2157         if (cur + npages > lim)
2158                 return 0;
2159         return 1;
2160 }
2161
2162
2163 static struct page *special_mapping_nopage(struct vm_area_struct *vma,
2164                                            unsigned long address, int *type)
2165 {
2166         struct page **pages;
2167
2168         BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2169
2170         address -= vma->vm_start;
2171         for (pages = vma->vm_private_data; address > 0 && *pages; ++pages)
2172                 address -= PAGE_SIZE;
2173
2174         if (*pages) {
2175                 struct page *page = *pages;
2176                 get_page(page);
2177                 return page;
2178         }
2179
2180         return NOPAGE_SIGBUS;
2181 }
2182
2183 /*
2184  * Having a close hook prevents vma merging regardless of flags.
2185  */
2186 static void special_mapping_close(struct vm_area_struct *vma)
2187 {
2188 }
2189
2190 static struct vm_operations_struct special_mapping_vmops = {
2191         .close = special_mapping_close,
2192         .nopage = special_mapping_nopage,
2193 };
2194
2195 /*
2196  * Called with mm->mmap_sem held for writing.
2197  * Insert a new vma covering the given region, with the given flags.
2198  * Its pages are supplied by the given array of struct page *.
2199  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2200  * The region past the last page supplied will always produce SIGBUS.
2201  * The array pointer and the pages it points to are assumed to stay alive
2202  * for as long as this mapping might exist.
2203  */
2204 int install_special_mapping(struct mm_struct *mm,
2205                             unsigned long addr, unsigned long len,
2206                             unsigned long vm_flags, struct page **pages)
2207 {
2208         struct vm_area_struct *vma;
2209
2210         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2211         if (unlikely(vma == NULL))
2212                 return -ENOMEM;
2213
2214         vma->vm_mm = mm;
2215         vma->vm_start = addr;
2216         vma->vm_end = addr + len;
2217
2218         vma->vm_flags = vm_flags | mm->def_flags;
2219         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2220
2221         vma->vm_ops = &special_mapping_vmops;
2222         vma->vm_private_data = pages;
2223
2224         if (unlikely(insert_vm_struct(mm, vma))) {
2225                 kmem_cache_free(vm_area_cachep, vma);
2226                 return -ENOMEM;
2227         }
2228
2229         mm->total_vm += len >> PAGE_SHIFT;
2230
2231         return 0;
2232 }