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