hwmon: (lm92) Convert to a new-style i2c driver
[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 __pgprot(pgprot_val(protection_map[vm_flags &
76                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
77                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
78 }
79 EXPORT_SYMBOL(vm_get_page_prot);
80
81 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
82 int sysctl_overcommit_ratio = 50;       /* default is 50% */
83 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
84 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
85
86 /*
87  * Check that a process has enough memory to allocate a new virtual
88  * mapping. 0 means there is enough memory for the allocation to
89  * succeed and -ENOMEM implies there is not.
90  *
91  * We currently support three overcommit policies, which are set via the
92  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
93  *
94  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
95  * Additional code 2002 Jul 20 by Robert Love.
96  *
97  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
98  *
99  * Note this is a helper function intended to be used by LSMs which
100  * wish to use this logic.
101  */
102 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
103 {
104         unsigned long free, allowed;
105
106         vm_acct_memory(pages);
107
108         /*
109          * Sometimes we want to use more memory than we have
110          */
111         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
112                 return 0;
113
114         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
115                 unsigned long n;
116
117                 free = global_page_state(NR_FILE_PAGES);
118                 free += nr_swap_pages;
119
120                 /*
121                  * Any slabs which are created with the
122                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
123                  * which are reclaimable, under pressure.  The dentry
124                  * cache and most inode caches should fall into this
125                  */
126                 free += global_page_state(NR_SLAB_RECLAIMABLE);
127
128                 /*
129                  * Leave the last 3% for root
130                  */
131                 if (!cap_sys_admin)
132                         free -= free / 32;
133
134                 if (free > pages)
135                         return 0;
136
137                 /*
138                  * nr_free_pages() is very expensive on large systems,
139                  * only call if we're about to fail.
140                  */
141                 n = nr_free_pages();
142
143                 /*
144                  * Leave reserved pages. The pages are not for anonymous pages.
145                  */
146                 if (n <= totalreserve_pages)
147                         goto error;
148                 else
149                         n -= totalreserve_pages;
150
151                 /*
152                  * Leave the last 3% for root
153                  */
154                 if (!cap_sys_admin)
155                         n -= n / 32;
156                 free += n;
157
158                 if (free > pages)
159                         return 0;
160
161                 goto error;
162         }
163
164         allowed = (totalram_pages - hugetlb_total_pages())
165                 * sysctl_overcommit_ratio / 100;
166         /*
167          * Leave the last 3% for root
168          */
169         if (!cap_sys_admin)
170                 allowed -= allowed / 32;
171         allowed += total_swap_pages;
172
173         /* Don't let a single process grow too big:
174            leave 3% of the size of this process for other processes */
175         allowed -= mm->total_vm / 32;
176
177         /*
178          * cast `allowed' as a signed long because vm_committed_space
179          * sometimes has a negative value
180          */
181         if (atomic_long_read(&vm_committed_space) < (long)allowed)
182                 return 0;
183 error:
184         vm_unacct_memory(pages);
185
186         return -ENOMEM;
187 }
188
189 /*
190  * Requires inode->i_mapping->i_mmap_lock
191  */
192 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
193                 struct file *file, struct address_space *mapping)
194 {
195         if (vma->vm_flags & VM_DENYWRITE)
196                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
197         if (vma->vm_flags & VM_SHARED)
198                 mapping->i_mmap_writable--;
199
200         flush_dcache_mmap_lock(mapping);
201         if (unlikely(vma->vm_flags & VM_NONLINEAR))
202                 list_del_init(&vma->shared.vm_set.list);
203         else
204                 vma_prio_tree_remove(vma, &mapping->i_mmap);
205         flush_dcache_mmap_unlock(mapping);
206 }
207
208 /*
209  * Unlink a file-based vm structure from its prio_tree, to hide
210  * vma from rmap and vmtruncate before freeing its page tables.
211  */
212 void unlink_file_vma(struct vm_area_struct *vma)
213 {
214         struct file *file = vma->vm_file;
215
216         if (file) {
217                 struct address_space *mapping = file->f_mapping;
218                 spin_lock(&mapping->i_mmap_lock);
219                 __remove_shared_vm_struct(vma, file, mapping);
220                 spin_unlock(&mapping->i_mmap_lock);
221         }
222 }
223
224 /*
225  * Close a vm structure and free it, returning the next.
226  */
227 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
228 {
229         struct vm_area_struct *next = vma->vm_next;
230
231         might_sleep();
232         if (vma->vm_ops && vma->vm_ops->close)
233                 vma->vm_ops->close(vma);
234         if (vma->vm_file) {
235                 fput(vma->vm_file);
236                 if (vma->vm_flags & VM_EXECUTABLE)
237                         removed_exe_file_vma(vma->vm_mm);
238         }
239         mpol_put(vma_policy(vma));
240         kmem_cache_free(vm_area_cachep, vma);
241         return next;
242 }
243
244 asmlinkage unsigned long sys_brk(unsigned long brk)
245 {
246         unsigned long rlim, retval;
247         unsigned long newbrk, oldbrk;
248         struct mm_struct *mm = current->mm;
249         unsigned long min_brk;
250
251         down_write(&mm->mmap_sem);
252
253 #ifdef CONFIG_COMPAT_BRK
254         min_brk = mm->end_code;
255 #else
256         min_brk = mm->start_brk;
257 #endif
258         if (brk < min_brk)
259                 goto out;
260
261         /*
262          * Check against rlimit here. If this check is done later after the test
263          * of oldbrk with newbrk then it can escape the test and let the data
264          * segment grow beyond its set limit the in case where the limit is
265          * not page aligned -Ram Gupta
266          */
267         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
268         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
269                         (mm->end_data - mm->start_data) > rlim)
270                 goto out;
271
272         newbrk = PAGE_ALIGN(brk);
273         oldbrk = PAGE_ALIGN(mm->brk);
274         if (oldbrk == newbrk)
275                 goto set_brk;
276
277         /* Always allow shrinking brk. */
278         if (brk <= mm->brk) {
279                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
280                         goto set_brk;
281                 goto out;
282         }
283
284         /* Check against existing mmap mappings. */
285         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
286                 goto out;
287
288         /* Ok, looks good - let it rip. */
289         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
290                 goto out;
291 set_brk:
292         mm->brk = brk;
293 out:
294         retval = mm->brk;
295         up_write(&mm->mmap_sem);
296         return retval;
297 }
298
299 #ifdef DEBUG_MM_RB
300 static int browse_rb(struct rb_root *root)
301 {
302         int i = 0, j;
303         struct rb_node *nd, *pn = NULL;
304         unsigned long prev = 0, pend = 0;
305
306         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
307                 struct vm_area_struct *vma;
308                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
309                 if (vma->vm_start < prev)
310                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
311                 if (vma->vm_start < pend)
312                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
313                 if (vma->vm_start > vma->vm_end)
314                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
315                 i++;
316                 pn = nd;
317                 prev = vma->vm_start;
318                 pend = vma->vm_end;
319         }
320         j = 0;
321         for (nd = pn; nd; nd = rb_prev(nd)) {
322                 j++;
323         }
324         if (i != j)
325                 printk("backwards %d, forwards %d\n", j, i), i = 0;
326         return i;
327 }
328
329 void validate_mm(struct mm_struct *mm)
330 {
331         int bug = 0;
332         int i = 0;
333         struct vm_area_struct *tmp = mm->mmap;
334         while (tmp) {
335                 tmp = tmp->vm_next;
336                 i++;
337         }
338         if (i != mm->map_count)
339                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
340         i = browse_rb(&mm->mm_rb);
341         if (i != mm->map_count)
342                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
343         BUG_ON(bug);
344 }
345 #else
346 #define validate_mm(mm) do { } while (0)
347 #endif
348
349 static struct vm_area_struct *
350 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
351                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
352                 struct rb_node ** rb_parent)
353 {
354         struct vm_area_struct * vma;
355         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
356
357         __rb_link = &mm->mm_rb.rb_node;
358         rb_prev = __rb_parent = NULL;
359         vma = NULL;
360
361         while (*__rb_link) {
362                 struct vm_area_struct *vma_tmp;
363
364                 __rb_parent = *__rb_link;
365                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
366
367                 if (vma_tmp->vm_end > addr) {
368                         vma = vma_tmp;
369                         if (vma_tmp->vm_start <= addr)
370                                 return vma;
371                         __rb_link = &__rb_parent->rb_left;
372                 } else {
373                         rb_prev = __rb_parent;
374                         __rb_link = &__rb_parent->rb_right;
375                 }
376         }
377
378         *pprev = NULL;
379         if (rb_prev)
380                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
381         *rb_link = __rb_link;
382         *rb_parent = __rb_parent;
383         return vma;
384 }
385
386 static inline void
387 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
388                 struct vm_area_struct *prev, struct rb_node *rb_parent)
389 {
390         if (prev) {
391                 vma->vm_next = prev->vm_next;
392                 prev->vm_next = vma;
393         } else {
394                 mm->mmap = vma;
395                 if (rb_parent)
396                         vma->vm_next = rb_entry(rb_parent,
397                                         struct vm_area_struct, vm_rb);
398                 else
399                         vma->vm_next = NULL;
400         }
401 }
402
403 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
404                 struct rb_node **rb_link, struct rb_node *rb_parent)
405 {
406         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
407         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
408 }
409
410 static inline void __vma_link_file(struct vm_area_struct *vma)
411 {
412         struct file * file;
413
414         file = vma->vm_file;
415         if (file) {
416                 struct address_space *mapping = file->f_mapping;
417
418                 if (vma->vm_flags & VM_DENYWRITE)
419                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
420                 if (vma->vm_flags & VM_SHARED)
421                         mapping->i_mmap_writable++;
422
423                 flush_dcache_mmap_lock(mapping);
424                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
425                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
426                 else
427                         vma_prio_tree_insert(vma, &mapping->i_mmap);
428                 flush_dcache_mmap_unlock(mapping);
429         }
430 }
431
432 static void
433 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
434         struct vm_area_struct *prev, struct rb_node **rb_link,
435         struct rb_node *rb_parent)
436 {
437         __vma_link_list(mm, vma, prev, rb_parent);
438         __vma_link_rb(mm, vma, rb_link, rb_parent);
439         __anon_vma_link(vma);
440 }
441
442 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
443                         struct vm_area_struct *prev, struct rb_node **rb_link,
444                         struct rb_node *rb_parent)
445 {
446         struct address_space *mapping = NULL;
447
448         if (vma->vm_file)
449                 mapping = vma->vm_file->f_mapping;
450
451         if (mapping) {
452                 spin_lock(&mapping->i_mmap_lock);
453                 vma->vm_truncate_count = mapping->truncate_count;
454         }
455         anon_vma_lock(vma);
456
457         __vma_link(mm, vma, prev, rb_link, rb_parent);
458         __vma_link_file(vma);
459
460         anon_vma_unlock(vma);
461         if (mapping)
462                 spin_unlock(&mapping->i_mmap_lock);
463
464         mm->map_count++;
465         validate_mm(mm);
466 }
467
468 /*
469  * Helper for vma_adjust in the split_vma insert case:
470  * insert vm structure into list and rbtree and anon_vma,
471  * but it has already been inserted into prio_tree earlier.
472  */
473 static void
474 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
475 {
476         struct vm_area_struct * __vma, * prev;
477         struct rb_node ** rb_link, * rb_parent;
478
479         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
480         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
481         __vma_link(mm, vma, prev, rb_link, rb_parent);
482         mm->map_count++;
483 }
484
485 static inline void
486 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
487                 struct vm_area_struct *prev)
488 {
489         prev->vm_next = vma->vm_next;
490         rb_erase(&vma->vm_rb, &mm->mm_rb);
491         if (mm->mmap_cache == vma)
492                 mm->mmap_cache = prev;
493 }
494
495 /*
496  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
497  * is already present in an i_mmap tree without adjusting the tree.
498  * The following helper function should be used when such adjustments
499  * are necessary.  The "insert" vma (if any) is to be inserted
500  * before we drop the necessary locks.
501  */
502 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
503         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
504 {
505         struct mm_struct *mm = vma->vm_mm;
506         struct vm_area_struct *next = vma->vm_next;
507         struct vm_area_struct *importer = NULL;
508         struct address_space *mapping = NULL;
509         struct prio_tree_root *root = NULL;
510         struct file *file = vma->vm_file;
511         struct anon_vma *anon_vma = NULL;
512         long adjust_next = 0;
513         int remove_next = 0;
514
515         if (next && !insert) {
516                 if (end >= next->vm_end) {
517                         /*
518                          * vma expands, overlapping all the next, and
519                          * perhaps the one after too (mprotect case 6).
520                          */
521 again:                  remove_next = 1 + (end > next->vm_end);
522                         end = next->vm_end;
523                         anon_vma = next->anon_vma;
524                         importer = vma;
525                 } else if (end > next->vm_start) {
526                         /*
527                          * vma expands, overlapping part of the next:
528                          * mprotect case 5 shifting the boundary up.
529                          */
530                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
531                         anon_vma = next->anon_vma;
532                         importer = vma;
533                 } else if (end < vma->vm_end) {
534                         /*
535                          * vma shrinks, and !insert tells it's not
536                          * split_vma inserting another: so it must be
537                          * mprotect case 4 shifting the boundary down.
538                          */
539                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
540                         anon_vma = next->anon_vma;
541                         importer = next;
542                 }
543         }
544
545         if (file) {
546                 mapping = file->f_mapping;
547                 if (!(vma->vm_flags & VM_NONLINEAR))
548                         root = &mapping->i_mmap;
549                 spin_lock(&mapping->i_mmap_lock);
550                 if (importer &&
551                     vma->vm_truncate_count != next->vm_truncate_count) {
552                         /*
553                          * unmap_mapping_range might be in progress:
554                          * ensure that the expanding vma is rescanned.
555                          */
556                         importer->vm_truncate_count = 0;
557                 }
558                 if (insert) {
559                         insert->vm_truncate_count = vma->vm_truncate_count;
560                         /*
561                          * Put into prio_tree now, so instantiated pages
562                          * are visible to arm/parisc __flush_dcache_page
563                          * throughout; but we cannot insert into address
564                          * space until vma start or end is updated.
565                          */
566                         __vma_link_file(insert);
567                 }
568         }
569
570         /*
571          * When changing only vma->vm_end, we don't really need
572          * anon_vma lock: but is that case worth optimizing out?
573          */
574         if (vma->anon_vma)
575                 anon_vma = vma->anon_vma;
576         if (anon_vma) {
577                 spin_lock(&anon_vma->lock);
578                 /*
579                  * Easily overlooked: when mprotect shifts the boundary,
580                  * make sure the expanding vma has anon_vma set if the
581                  * shrinking vma had, to cover any anon pages imported.
582                  */
583                 if (importer && !importer->anon_vma) {
584                         importer->anon_vma = anon_vma;
585                         __anon_vma_link(importer);
586                 }
587         }
588
589         if (root) {
590                 flush_dcache_mmap_lock(mapping);
591                 vma_prio_tree_remove(vma, root);
592                 if (adjust_next)
593                         vma_prio_tree_remove(next, root);
594         }
595
596         vma->vm_start = start;
597         vma->vm_end = end;
598         vma->vm_pgoff = pgoff;
599         if (adjust_next) {
600                 next->vm_start += adjust_next << PAGE_SHIFT;
601                 next->vm_pgoff += adjust_next;
602         }
603
604         if (root) {
605                 if (adjust_next)
606                         vma_prio_tree_insert(next, root);
607                 vma_prio_tree_insert(vma, root);
608                 flush_dcache_mmap_unlock(mapping);
609         }
610
611         if (remove_next) {
612                 /*
613                  * vma_merge has merged next into vma, and needs
614                  * us to remove next before dropping the locks.
615                  */
616                 __vma_unlink(mm, next, vma);
617                 if (file)
618                         __remove_shared_vm_struct(next, file, mapping);
619                 if (next->anon_vma)
620                         __anon_vma_merge(vma, next);
621         } else if (insert) {
622                 /*
623                  * split_vma has split insert from vma, and needs
624                  * us to insert it before dropping the locks
625                  * (it may either follow vma or precede it).
626                  */
627                 __insert_vm_struct(mm, insert);
628         }
629
630         if (anon_vma)
631                 spin_unlock(&anon_vma->lock);
632         if (mapping)
633                 spin_unlock(&mapping->i_mmap_lock);
634
635         if (remove_next) {
636                 if (file) {
637                         fput(file);
638                         if (next->vm_flags & VM_EXECUTABLE)
639                                 removed_exe_file_vma(mm);
640                 }
641                 mm->map_count--;
642                 mpol_put(vma_policy(next));
643                 kmem_cache_free(vm_area_cachep, next);
644                 /*
645                  * In mprotect's case 6 (see comments on vma_merge),
646                  * we must remove another next too. It would clutter
647                  * up the code too much to do both in one go.
648                  */
649                 if (remove_next == 2) {
650                         next = vma->vm_next;
651                         goto again;
652                 }
653         }
654
655         validate_mm(mm);
656 }
657
658 /*
659  * If the vma has a ->close operation then the driver probably needs to release
660  * per-vma resources, so we don't attempt to merge those.
661  */
662 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
663
664 static inline int is_mergeable_vma(struct vm_area_struct *vma,
665                         struct file *file, unsigned long vm_flags)
666 {
667         if (vma->vm_flags != vm_flags)
668                 return 0;
669         if (vma->vm_file != file)
670                 return 0;
671         if (vma->vm_ops && vma->vm_ops->close)
672                 return 0;
673         return 1;
674 }
675
676 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
677                                         struct anon_vma *anon_vma2)
678 {
679         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
680 }
681
682 /*
683  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
684  * in front of (at a lower virtual address and file offset than) the vma.
685  *
686  * We cannot merge two vmas if they have differently assigned (non-NULL)
687  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
688  *
689  * We don't check here for the merged mmap wrapping around the end of pagecache
690  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
691  * wrap, nor mmaps which cover the final page at index -1UL.
692  */
693 static int
694 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
695         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
696 {
697         if (is_mergeable_vma(vma, file, vm_flags) &&
698             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
699                 if (vma->vm_pgoff == vm_pgoff)
700                         return 1;
701         }
702         return 0;
703 }
704
705 /*
706  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
707  * beyond (at a higher virtual address and file offset than) the vma.
708  *
709  * We cannot merge two vmas if they have differently assigned (non-NULL)
710  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
711  */
712 static int
713 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
714         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
715 {
716         if (is_mergeable_vma(vma, file, vm_flags) &&
717             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
718                 pgoff_t vm_pglen;
719                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
720                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
721                         return 1;
722         }
723         return 0;
724 }
725
726 /*
727  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
728  * whether that can be merged with its predecessor or its successor.
729  * Or both (it neatly fills a hole).
730  *
731  * In most cases - when called for mmap, brk or mremap - [addr,end) is
732  * certain not to be mapped by the time vma_merge is called; but when
733  * called for mprotect, it is certain to be already mapped (either at
734  * an offset within prev, or at the start of next), and the flags of
735  * this area are about to be changed to vm_flags - and the no-change
736  * case has already been eliminated.
737  *
738  * The following mprotect cases have to be considered, where AAAA is
739  * the area passed down from mprotect_fixup, never extending beyond one
740  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
741  *
742  *     AAAA             AAAA                AAAA          AAAA
743  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
744  *    cannot merge    might become    might become    might become
745  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
746  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
747  *    mremap move:                                    PPPPNNNNNNNN 8
748  *        AAAA
749  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
750  *    might become    case 1 below    case 2 below    case 3 below
751  *
752  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
753  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
754  */
755 struct vm_area_struct *vma_merge(struct mm_struct *mm,
756                         struct vm_area_struct *prev, unsigned long addr,
757                         unsigned long end, unsigned long vm_flags,
758                         struct anon_vma *anon_vma, struct file *file,
759                         pgoff_t pgoff, struct mempolicy *policy)
760 {
761         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
762         struct vm_area_struct *area, *next;
763
764         /*
765          * We later require that vma->vm_flags == vm_flags,
766          * so this tests vma->vm_flags & VM_SPECIAL, too.
767          */
768         if (vm_flags & VM_SPECIAL)
769                 return NULL;
770
771         if (prev)
772                 next = prev->vm_next;
773         else
774                 next = mm->mmap;
775         area = next;
776         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
777                 next = next->vm_next;
778
779         /*
780          * Can it merge with the predecessor?
781          */
782         if (prev && prev->vm_end == addr &&
783                         mpol_equal(vma_policy(prev), policy) &&
784                         can_vma_merge_after(prev, vm_flags,
785                                                 anon_vma, file, pgoff)) {
786                 /*
787                  * OK, it can.  Can we now merge in the successor as well?
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                                 is_mergeable_anon_vma(prev->anon_vma,
794                                                       next->anon_vma)) {
795                                                         /* cases 1, 6 */
796                         vma_adjust(prev, prev->vm_start,
797                                 next->vm_end, prev->vm_pgoff, NULL);
798                 } else                                  /* cases 2, 5, 7 */
799                         vma_adjust(prev, prev->vm_start,
800                                 end, prev->vm_pgoff, NULL);
801                 return prev;
802         }
803
804         /*
805          * Can this new request be merged in front of next?
806          */
807         if (next && end == next->vm_start &&
808                         mpol_equal(policy, vma_policy(next)) &&
809                         can_vma_merge_before(next, vm_flags,
810                                         anon_vma, file, pgoff+pglen)) {
811                 if (prev && addr < prev->vm_end)        /* case 4 */
812                         vma_adjust(prev, prev->vm_start,
813                                 addr, prev->vm_pgoff, NULL);
814                 else                                    /* cases 3, 8 */
815                         vma_adjust(area, addr, next->vm_end,
816                                 next->vm_pgoff - pglen, NULL);
817                 return area;
818         }
819
820         return NULL;
821 }
822
823 /*
824  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
825  * neighbouring vmas for a suitable anon_vma, before it goes off
826  * to allocate a new anon_vma.  It checks because a repetitive
827  * sequence of mprotects and faults may otherwise lead to distinct
828  * anon_vmas being allocated, preventing vma merge in subsequent
829  * mprotect.
830  */
831 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
832 {
833         struct vm_area_struct *near;
834         unsigned long vm_flags;
835
836         near = vma->vm_next;
837         if (!near)
838                 goto try_prev;
839
840         /*
841          * Since only mprotect tries to remerge vmas, match flags
842          * which might be mprotected into each other later on.
843          * Neither mlock nor madvise tries to remerge at present,
844          * so leave their flags as obstructing a merge.
845          */
846         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
847         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
848
849         if (near->anon_vma && vma->vm_end == near->vm_start &&
850                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
851                         can_vma_merge_before(near, vm_flags,
852                                 NULL, vma->vm_file, vma->vm_pgoff +
853                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
854                 return near->anon_vma;
855 try_prev:
856         /*
857          * It is potentially slow to have to call find_vma_prev here.
858          * But it's only on the first write fault on the vma, not
859          * every time, and we could devise a way to avoid it later
860          * (e.g. stash info in next's anon_vma_node when assigning
861          * an anon_vma, or when trying vma_merge).  Another time.
862          */
863         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
864         if (!near)
865                 goto none;
866
867         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
868         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
869
870         if (near->anon_vma && near->vm_end == vma->vm_start &&
871                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
872                         can_vma_merge_after(near, vm_flags,
873                                 NULL, vma->vm_file, vma->vm_pgoff))
874                 return near->anon_vma;
875 none:
876         /*
877          * There's no absolute need to look only at touching neighbours:
878          * we could search further afield for "compatible" anon_vmas.
879          * But it would probably just be a waste of time searching,
880          * or lead to too many vmas hanging off the same anon_vma.
881          * We're trying to allow mprotect remerging later on,
882          * not trying to minimize memory used for anon_vmas.
883          */
884         return NULL;
885 }
886
887 #ifdef CONFIG_PROC_FS
888 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
889                                                 struct file *file, long pages)
890 {
891         const unsigned long stack_flags
892                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
893
894         if (file) {
895                 mm->shared_vm += pages;
896                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
897                         mm->exec_vm += pages;
898         } else if (flags & stack_flags)
899                 mm->stack_vm += pages;
900         if (flags & (VM_RESERVED|VM_IO))
901                 mm->reserved_vm += pages;
902 }
903 #endif /* CONFIG_PROC_FS */
904
905 /*
906  * The caller must hold down_write(current->mm->mmap_sem).
907  */
908
909 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
910                         unsigned long len, unsigned long prot,
911                         unsigned long flags, unsigned long pgoff)
912 {
913         struct mm_struct * mm = current->mm;
914         struct inode *inode;
915         unsigned int vm_flags;
916         int error;
917         int accountable = 1;
918         unsigned long reqprot = prot;
919
920         /*
921          * Does the application expect PROT_READ to imply PROT_EXEC?
922          *
923          * (the exception is when the underlying filesystem is noexec
924          *  mounted, in which case we dont add PROT_EXEC.)
925          */
926         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
927                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
928                         prot |= PROT_EXEC;
929
930         if (!len)
931                 return -EINVAL;
932
933         if (!(flags & MAP_FIXED))
934                 addr = round_hint_to_min(addr);
935
936         error = arch_mmap_check(addr, len, flags);
937         if (error)
938                 return error;
939
940         /* Careful about overflows.. */
941         len = PAGE_ALIGN(len);
942         if (!len || len > TASK_SIZE)
943                 return -ENOMEM;
944
945         /* offset overflow? */
946         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
947                return -EOVERFLOW;
948
949         /* Too many mappings? */
950         if (mm->map_count > sysctl_max_map_count)
951                 return -ENOMEM;
952
953         /* Obtain the address to map to. we verify (or select) it and ensure
954          * that it represents a valid section of the address space.
955          */
956         addr = get_unmapped_area(file, addr, len, pgoff, flags);
957         if (addr & ~PAGE_MASK)
958                 return addr;
959
960         /* Do simple checking here so the lower-level routines won't have
961          * to. we assume access permissions have been handled by the open
962          * of the memory object, so we don't do any here.
963          */
964         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
965                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
966
967         if (flags & MAP_LOCKED) {
968                 if (!can_do_mlock())
969                         return -EPERM;
970                 vm_flags |= VM_LOCKED;
971         }
972         /* mlock MCL_FUTURE? */
973         if (vm_flags & VM_LOCKED) {
974                 unsigned long locked, lock_limit;
975                 locked = len >> PAGE_SHIFT;
976                 locked += mm->locked_vm;
977                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
978                 lock_limit >>= PAGE_SHIFT;
979                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
980                         return -EAGAIN;
981         }
982
983         inode = file ? file->f_path.dentry->d_inode : NULL;
984
985         if (file) {
986                 switch (flags & MAP_TYPE) {
987                 case MAP_SHARED:
988                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
989                                 return -EACCES;
990
991                         /*
992                          * Make sure we don't allow writing to an append-only
993                          * file..
994                          */
995                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
996                                 return -EACCES;
997
998                         /*
999                          * Make sure there are no mandatory locks on the file.
1000                          */
1001                         if (locks_verify_locked(inode))
1002                                 return -EAGAIN;
1003
1004                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1005                         if (!(file->f_mode & FMODE_WRITE))
1006                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1007
1008                         /* fall through */
1009                 case MAP_PRIVATE:
1010                         if (!(file->f_mode & FMODE_READ))
1011                                 return -EACCES;
1012                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1013                                 if (vm_flags & VM_EXEC)
1014                                         return -EPERM;
1015                                 vm_flags &= ~VM_MAYEXEC;
1016                         }
1017                         if (is_file_hugepages(file))
1018                                 accountable = 0;
1019
1020                         if (!file->f_op || !file->f_op->mmap)
1021                                 return -ENODEV;
1022                         break;
1023
1024                 default:
1025                         return -EINVAL;
1026                 }
1027         } else {
1028                 switch (flags & MAP_TYPE) {
1029                 case MAP_SHARED:
1030                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1031                         break;
1032                 case MAP_PRIVATE:
1033                         /*
1034                          * Set pgoff according to addr for anon_vma.
1035                          */
1036                         pgoff = addr >> PAGE_SHIFT;
1037                         break;
1038                 default:
1039                         return -EINVAL;
1040                 }
1041         }
1042
1043         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1044         if (error)
1045                 return error;
1046
1047         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1048                            accountable);
1049 }
1050 EXPORT_SYMBOL(do_mmap_pgoff);
1051
1052 /*
1053  * Some shared mappigns will want the pages marked read-only
1054  * to track write events. If so, we'll downgrade vm_page_prot
1055  * to the private version (using protection_map[] without the
1056  * VM_SHARED bit).
1057  */
1058 int vma_wants_writenotify(struct vm_area_struct *vma)
1059 {
1060         unsigned int vm_flags = vma->vm_flags;
1061
1062         /* If it was private or non-writable, the write bit is already clear */
1063         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1064                 return 0;
1065
1066         /* The backer wishes to know when pages are first written to? */
1067         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1068                 return 1;
1069
1070         /* The open routine did something to the protections already? */
1071         if (pgprot_val(vma->vm_page_prot) !=
1072             pgprot_val(vm_get_page_prot(vm_flags)))
1073                 return 0;
1074
1075         /* Specialty mapping? */
1076         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1077                 return 0;
1078
1079         /* Can the mapping track the dirty pages? */
1080         return vma->vm_file && vma->vm_file->f_mapping &&
1081                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1082 }
1083
1084 unsigned long mmap_region(struct file *file, unsigned long addr,
1085                           unsigned long len, unsigned long flags,
1086                           unsigned int vm_flags, unsigned long pgoff,
1087                           int accountable)
1088 {
1089         struct mm_struct *mm = current->mm;
1090         struct vm_area_struct *vma, *prev;
1091         int correct_wcount = 0;
1092         int error;
1093         struct rb_node **rb_link, *rb_parent;
1094         unsigned long charged = 0;
1095         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1096
1097         /* Clear old maps */
1098         error = -ENOMEM;
1099 munmap_back:
1100         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1101         if (vma && vma->vm_start < addr + len) {
1102                 if (do_munmap(mm, addr, len))
1103                         return -ENOMEM;
1104                 goto munmap_back;
1105         }
1106
1107         /* Check against address space limit. */
1108         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1109                 return -ENOMEM;
1110
1111         if (accountable && (!(flags & MAP_NORESERVE) ||
1112                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1113                 if (vm_flags & VM_SHARED) {
1114                         /* Check memory availability in shmem_file_setup? */
1115                         vm_flags |= VM_ACCOUNT;
1116                 } else if (vm_flags & VM_WRITE) {
1117                         /*
1118                          * Private writable mapping: check memory availability
1119                          */
1120                         charged = len >> PAGE_SHIFT;
1121                         if (security_vm_enough_memory(charged))
1122                                 return -ENOMEM;
1123                         vm_flags |= VM_ACCOUNT;
1124                 }
1125         }
1126
1127         /*
1128          * Can we just expand an old private anonymous mapping?
1129          * The VM_SHARED test is necessary because shmem_zero_setup
1130          * will create the file object for a shared anonymous map below.
1131          */
1132         if (!file && !(vm_flags & VM_SHARED) &&
1133             vma_merge(mm, prev, addr, addr + len, vm_flags,
1134                                         NULL, NULL, pgoff, NULL))
1135                 goto out;
1136
1137         /*
1138          * Determine the object being mapped and call the appropriate
1139          * specific mapper. the address has already been validated, but
1140          * not unmapped, but the maps are removed from the list.
1141          */
1142         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1143         if (!vma) {
1144                 error = -ENOMEM;
1145                 goto unacct_error;
1146         }
1147
1148         vma->vm_mm = mm;
1149         vma->vm_start = addr;
1150         vma->vm_end = addr + len;
1151         vma->vm_flags = vm_flags;
1152         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1153         vma->vm_pgoff = pgoff;
1154
1155         if (file) {
1156                 error = -EINVAL;
1157                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1158                         goto free_vma;
1159                 if (vm_flags & VM_DENYWRITE) {
1160                         error = deny_write_access(file);
1161                         if (error)
1162                                 goto free_vma;
1163                         correct_wcount = 1;
1164                 }
1165                 vma->vm_file = file;
1166                 get_file(file);
1167                 error = file->f_op->mmap(file, vma);
1168                 if (error)
1169                         goto unmap_and_free_vma;
1170                 if (vm_flags & VM_EXECUTABLE)
1171                         added_exe_file_vma(mm);
1172         } else if (vm_flags & VM_SHARED) {
1173                 error = shmem_zero_setup(vma);
1174                 if (error)
1175                         goto free_vma;
1176         }
1177
1178         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1179          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1180          * that memory reservation must be checked; but that reservation
1181          * belongs to shared memory object, not to vma: so now clear it.
1182          */
1183         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1184                 vma->vm_flags &= ~VM_ACCOUNT;
1185
1186         /* Can addr have changed??
1187          *
1188          * Answer: Yes, several device drivers can do it in their
1189          *         f_op->mmap method. -DaveM
1190          */
1191         addr = vma->vm_start;
1192         pgoff = vma->vm_pgoff;
1193         vm_flags = vma->vm_flags;
1194
1195         if (vma_wants_writenotify(vma))
1196                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1197
1198         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1199                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1200                 mpol_put(vma_policy(vma));
1201                 kmem_cache_free(vm_area_cachep, vma);
1202                 fput(file);
1203                 if (vm_flags & VM_EXECUTABLE)
1204                         removed_exe_file_vma(mm);
1205         } else {
1206                 vma_link(mm, vma, prev, rb_link, rb_parent);
1207                 file = vma->vm_file;
1208         }
1209
1210         /* Once vma denies write, undo our temporary denial count */
1211         if (correct_wcount)
1212                 atomic_inc(&inode->i_writecount);
1213 out:
1214         mm->total_vm += len >> PAGE_SHIFT;
1215         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1216         if (vm_flags & VM_LOCKED) {
1217                 mm->locked_vm += len >> PAGE_SHIFT;
1218                 make_pages_present(addr, addr + len);
1219         }
1220         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1221                 make_pages_present(addr, addr + len);
1222         return addr;
1223
1224 unmap_and_free_vma:
1225         if (correct_wcount)
1226                 atomic_inc(&inode->i_writecount);
1227         vma->vm_file = NULL;
1228         fput(file);
1229
1230         /* Undo any partial mapping done by a device driver. */
1231         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1232         charged = 0;
1233 free_vma:
1234         kmem_cache_free(vm_area_cachep, vma);
1235 unacct_error:
1236         if (charged)
1237                 vm_unacct_memory(charged);
1238         return error;
1239 }
1240
1241 /* Get an address range which is currently unmapped.
1242  * For shmat() with addr=0.
1243  *
1244  * Ugly calling convention alert:
1245  * Return value with the low bits set means error value,
1246  * ie
1247  *      if (ret & ~PAGE_MASK)
1248  *              error = ret;
1249  *
1250  * This function "knows" that -ENOMEM has the bits set.
1251  */
1252 #ifndef HAVE_ARCH_UNMAPPED_AREA
1253 unsigned long
1254 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1255                 unsigned long len, unsigned long pgoff, unsigned long flags)
1256 {
1257         struct mm_struct *mm = current->mm;
1258         struct vm_area_struct *vma;
1259         unsigned long start_addr;
1260
1261         if (len > TASK_SIZE)
1262                 return -ENOMEM;
1263
1264         if (flags & MAP_FIXED)
1265                 return addr;
1266
1267         if (addr) {
1268                 addr = PAGE_ALIGN(addr);
1269                 vma = find_vma(mm, addr);
1270                 if (TASK_SIZE - len >= addr &&
1271                     (!vma || addr + len <= vma->vm_start))
1272                         return addr;
1273         }
1274         if (len > mm->cached_hole_size) {
1275                 start_addr = addr = mm->free_area_cache;
1276         } else {
1277                 start_addr = addr = TASK_UNMAPPED_BASE;
1278                 mm->cached_hole_size = 0;
1279         }
1280
1281 full_search:
1282         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1283                 /* At this point:  (!vma || addr < vma->vm_end). */
1284                 if (TASK_SIZE - len < addr) {
1285                         /*
1286                          * Start a new search - just in case we missed
1287                          * some holes.
1288                          */
1289                         if (start_addr != TASK_UNMAPPED_BASE) {
1290                                 addr = TASK_UNMAPPED_BASE;
1291                                 start_addr = addr;
1292                                 mm->cached_hole_size = 0;
1293                                 goto full_search;
1294                         }
1295                         return -ENOMEM;
1296                 }
1297                 if (!vma || addr + len <= vma->vm_start) {
1298                         /*
1299                          * Remember the place where we stopped the search:
1300                          */
1301                         mm->free_area_cache = addr + len;
1302                         return addr;
1303                 }
1304                 if (addr + mm->cached_hole_size < vma->vm_start)
1305                         mm->cached_hole_size = vma->vm_start - addr;
1306                 addr = vma->vm_end;
1307         }
1308 }
1309 #endif  
1310
1311 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1312 {
1313         /*
1314          * Is this a new hole at the lowest possible address?
1315          */
1316         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1317                 mm->free_area_cache = addr;
1318                 mm->cached_hole_size = ~0UL;
1319         }
1320 }
1321
1322 /*
1323  * This mmap-allocator allocates new areas top-down from below the
1324  * stack's low limit (the base):
1325  */
1326 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1327 unsigned long
1328 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1329                           const unsigned long len, const unsigned long pgoff,
1330                           const unsigned long flags)
1331 {
1332         struct vm_area_struct *vma;
1333         struct mm_struct *mm = current->mm;
1334         unsigned long addr = addr0;
1335
1336         /* requested length too big for entire address space */
1337         if (len > TASK_SIZE)
1338                 return -ENOMEM;
1339
1340         if (flags & MAP_FIXED)
1341                 return addr;
1342
1343         /* requesting a specific address */
1344         if (addr) {
1345                 addr = PAGE_ALIGN(addr);
1346                 vma = find_vma(mm, addr);
1347                 if (TASK_SIZE - len >= addr &&
1348                                 (!vma || addr + len <= vma->vm_start))
1349                         return addr;
1350         }
1351
1352         /* check if free_area_cache is useful for us */
1353         if (len <= mm->cached_hole_size) {
1354                 mm->cached_hole_size = 0;
1355                 mm->free_area_cache = mm->mmap_base;
1356         }
1357
1358         /* either no address requested or can't fit in requested address hole */
1359         addr = mm->free_area_cache;
1360
1361         /* make sure it can fit in the remaining address space */
1362         if (addr > len) {
1363                 vma = find_vma(mm, addr-len);
1364                 if (!vma || addr <= vma->vm_start)
1365                         /* remember the address as a hint for next time */
1366                         return (mm->free_area_cache = addr-len);
1367         }
1368
1369         if (mm->mmap_base < len)
1370                 goto bottomup;
1371
1372         addr = mm->mmap_base-len;
1373
1374         do {
1375                 /*
1376                  * Lookup failure means no vma is above this address,
1377                  * else if new region fits below vma->vm_start,
1378                  * return with success:
1379                  */
1380                 vma = find_vma(mm, addr);
1381                 if (!vma || addr+len <= vma->vm_start)
1382                         /* remember the address as a hint for next time */
1383                         return (mm->free_area_cache = addr);
1384
1385                 /* remember the largest hole we saw so far */
1386                 if (addr + mm->cached_hole_size < vma->vm_start)
1387                         mm->cached_hole_size = vma->vm_start - addr;
1388
1389                 /* try just below the current vma->vm_start */
1390                 addr = vma->vm_start-len;
1391         } while (len < vma->vm_start);
1392
1393 bottomup:
1394         /*
1395          * A failed mmap() very likely causes application failure,
1396          * so fall back to the bottom-up function here. This scenario
1397          * can happen with large stack limits and large mmap()
1398          * allocations.
1399          */
1400         mm->cached_hole_size = ~0UL;
1401         mm->free_area_cache = TASK_UNMAPPED_BASE;
1402         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1403         /*
1404          * Restore the topdown base:
1405          */
1406         mm->free_area_cache = mm->mmap_base;
1407         mm->cached_hole_size = ~0UL;
1408
1409         return addr;
1410 }
1411 #endif
1412
1413 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1414 {
1415         /*
1416          * Is this a new hole at the highest possible address?
1417          */
1418         if (addr > mm->free_area_cache)
1419                 mm->free_area_cache = addr;
1420
1421         /* dont allow allocations above current base */
1422         if (mm->free_area_cache > mm->mmap_base)
1423                 mm->free_area_cache = mm->mmap_base;
1424 }
1425
1426 unsigned long
1427 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1428                 unsigned long pgoff, unsigned long flags)
1429 {
1430         unsigned long (*get_area)(struct file *, unsigned long,
1431                                   unsigned long, unsigned long, unsigned long);
1432
1433         get_area = current->mm->get_unmapped_area;
1434         if (file && file->f_op && file->f_op->get_unmapped_area)
1435                 get_area = file->f_op->get_unmapped_area;
1436         addr = get_area(file, addr, len, pgoff, flags);
1437         if (IS_ERR_VALUE(addr))
1438                 return addr;
1439
1440         if (addr > TASK_SIZE - len)
1441                 return -ENOMEM;
1442         if (addr & ~PAGE_MASK)
1443                 return -EINVAL;
1444
1445         return arch_rebalance_pgtables(addr, len);
1446 }
1447
1448 EXPORT_SYMBOL(get_unmapped_area);
1449
1450 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1451 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1452 {
1453         struct vm_area_struct *vma = NULL;
1454
1455         if (mm) {
1456                 /* Check the cache first. */
1457                 /* (Cache hit rate is typically around 35%.) */
1458                 vma = mm->mmap_cache;
1459                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1460                         struct rb_node * rb_node;
1461
1462                         rb_node = mm->mm_rb.rb_node;
1463                         vma = NULL;
1464
1465                         while (rb_node) {
1466                                 struct vm_area_struct * vma_tmp;
1467
1468                                 vma_tmp = rb_entry(rb_node,
1469                                                 struct vm_area_struct, vm_rb);
1470
1471                                 if (vma_tmp->vm_end > addr) {
1472                                         vma = vma_tmp;
1473                                         if (vma_tmp->vm_start <= addr)
1474                                                 break;
1475                                         rb_node = rb_node->rb_left;
1476                                 } else
1477                                         rb_node = rb_node->rb_right;
1478                         }
1479                         if (vma)
1480                                 mm->mmap_cache = vma;
1481                 }
1482         }
1483         return vma;
1484 }
1485
1486 EXPORT_SYMBOL(find_vma);
1487
1488 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1489 struct vm_area_struct *
1490 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1491                         struct vm_area_struct **pprev)
1492 {
1493         struct vm_area_struct *vma = NULL, *prev = NULL;
1494         struct rb_node * rb_node;
1495         if (!mm)
1496                 goto out;
1497
1498         /* Guard against addr being lower than the first VMA */
1499         vma = mm->mmap;
1500
1501         /* Go through the RB tree quickly. */
1502         rb_node = mm->mm_rb.rb_node;
1503
1504         while (rb_node) {
1505                 struct vm_area_struct *vma_tmp;
1506                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1507
1508                 if (addr < vma_tmp->vm_end) {
1509                         rb_node = rb_node->rb_left;
1510                 } else {
1511                         prev = vma_tmp;
1512                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1513                                 break;
1514                         rb_node = rb_node->rb_right;
1515                 }
1516         }
1517
1518 out:
1519         *pprev = prev;
1520         return prev ? prev->vm_next : vma;
1521 }
1522
1523 /*
1524  * Verify that the stack growth is acceptable and
1525  * update accounting. This is shared with both the
1526  * grow-up and grow-down cases.
1527  */
1528 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1529 {
1530         struct mm_struct *mm = vma->vm_mm;
1531         struct rlimit *rlim = current->signal->rlim;
1532         unsigned long new_start;
1533
1534         /* address space limit tests */
1535         if (!may_expand_vm(mm, grow))
1536                 return -ENOMEM;
1537
1538         /* Stack limit test */
1539         if (size > rlim[RLIMIT_STACK].rlim_cur)
1540                 return -ENOMEM;
1541
1542         /* mlock limit tests */
1543         if (vma->vm_flags & VM_LOCKED) {
1544                 unsigned long locked;
1545                 unsigned long limit;
1546                 locked = mm->locked_vm + grow;
1547                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1548                 if (locked > limit && !capable(CAP_IPC_LOCK))
1549                         return -ENOMEM;
1550         }
1551
1552         /* Check to ensure the stack will not grow into a hugetlb-only region */
1553         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1554                         vma->vm_end - size;
1555         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1556                 return -EFAULT;
1557
1558         /*
1559          * Overcommit..  This must be the final test, as it will
1560          * update security statistics.
1561          */
1562         if (security_vm_enough_memory(grow))
1563                 return -ENOMEM;
1564
1565         /* Ok, everything looks good - let it rip */
1566         mm->total_vm += grow;
1567         if (vma->vm_flags & VM_LOCKED)
1568                 mm->locked_vm += grow;
1569         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1570         return 0;
1571 }
1572
1573 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1574 /*
1575  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1576  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1577  */
1578 #ifndef CONFIG_IA64
1579 static inline
1580 #endif
1581 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1582 {
1583         int error;
1584
1585         if (!(vma->vm_flags & VM_GROWSUP))
1586                 return -EFAULT;
1587
1588         /*
1589          * We must make sure the anon_vma is allocated
1590          * so that the anon_vma locking is not a noop.
1591          */
1592         if (unlikely(anon_vma_prepare(vma)))
1593                 return -ENOMEM;
1594         anon_vma_lock(vma);
1595
1596         /*
1597          * vma->vm_start/vm_end cannot change under us because the caller
1598          * is required to hold the mmap_sem in read mode.  We need the
1599          * anon_vma lock to serialize against concurrent expand_stacks.
1600          * Also guard against wrapping around to address 0.
1601          */
1602         if (address < PAGE_ALIGN(address+4))
1603                 address = PAGE_ALIGN(address+4);
1604         else {
1605                 anon_vma_unlock(vma);
1606                 return -ENOMEM;
1607         }
1608         error = 0;
1609
1610         /* Somebody else might have raced and expanded it already */
1611         if (address > vma->vm_end) {
1612                 unsigned long size, grow;
1613
1614                 size = address - vma->vm_start;
1615                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1616
1617                 error = acct_stack_growth(vma, size, grow);
1618                 if (!error)
1619                         vma->vm_end = address;
1620         }
1621         anon_vma_unlock(vma);
1622         return error;
1623 }
1624 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1625
1626 /*
1627  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1628  */
1629 static inline int expand_downwards(struct vm_area_struct *vma,
1630                                    unsigned long address)
1631 {
1632         int error;
1633
1634         /*
1635          * We must make sure the anon_vma is allocated
1636          * so that the anon_vma locking is not a noop.
1637          */
1638         if (unlikely(anon_vma_prepare(vma)))
1639                 return -ENOMEM;
1640
1641         address &= PAGE_MASK;
1642         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1643         if (error)
1644                 return error;
1645
1646         anon_vma_lock(vma);
1647
1648         /*
1649          * vma->vm_start/vm_end cannot change under us because the caller
1650          * is required to hold the mmap_sem in read mode.  We need the
1651          * anon_vma lock to serialize against concurrent expand_stacks.
1652          */
1653
1654         /* Somebody else might have raced and expanded it already */
1655         if (address < vma->vm_start) {
1656                 unsigned long size, grow;
1657
1658                 size = vma->vm_end - address;
1659                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1660
1661                 error = acct_stack_growth(vma, size, grow);
1662                 if (!error) {
1663                         vma->vm_start = address;
1664                         vma->vm_pgoff -= grow;
1665                 }
1666         }
1667         anon_vma_unlock(vma);
1668         return error;
1669 }
1670
1671 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1672 {
1673         return expand_downwards(vma, address);
1674 }
1675
1676 #ifdef CONFIG_STACK_GROWSUP
1677 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1678 {
1679         return expand_upwards(vma, address);
1680 }
1681
1682 struct vm_area_struct *
1683 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1684 {
1685         struct vm_area_struct *vma, *prev;
1686
1687         addr &= PAGE_MASK;
1688         vma = find_vma_prev(mm, addr, &prev);
1689         if (vma && (vma->vm_start <= addr))
1690                 return vma;
1691         if (!prev || expand_stack(prev, addr))
1692                 return NULL;
1693         if (prev->vm_flags & VM_LOCKED)
1694                 make_pages_present(addr, prev->vm_end);
1695         return prev;
1696 }
1697 #else
1698 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1699 {
1700         return expand_downwards(vma, address);
1701 }
1702
1703 struct vm_area_struct *
1704 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1705 {
1706         struct vm_area_struct * vma;
1707         unsigned long start;
1708
1709         addr &= PAGE_MASK;
1710         vma = find_vma(mm,addr);
1711         if (!vma)
1712                 return NULL;
1713         if (vma->vm_start <= addr)
1714                 return vma;
1715         if (!(vma->vm_flags & VM_GROWSDOWN))
1716                 return NULL;
1717         start = vma->vm_start;
1718         if (expand_stack(vma, addr))
1719                 return NULL;
1720         if (vma->vm_flags & VM_LOCKED)
1721                 make_pages_present(addr, start);
1722         return vma;
1723 }
1724 #endif
1725
1726 /*
1727  * Ok - we have the memory areas we should free on the vma list,
1728  * so release them, and do the vma updates.
1729  *
1730  * Called with the mm semaphore held.
1731  */
1732 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1733 {
1734         /* Update high watermark before we lower total_vm */
1735         update_hiwater_vm(mm);
1736         do {
1737                 long nrpages = vma_pages(vma);
1738
1739                 mm->total_vm -= nrpages;
1740                 if (vma->vm_flags & VM_LOCKED)
1741                         mm->locked_vm -= nrpages;
1742                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1743                 vma = remove_vma(vma);
1744         } while (vma);
1745         validate_mm(mm);
1746 }
1747
1748 /*
1749  * Get rid of page table information in the indicated region.
1750  *
1751  * Called with the mm semaphore held.
1752  */
1753 static void unmap_region(struct mm_struct *mm,
1754                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1755                 unsigned long start, unsigned long end)
1756 {
1757         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1758         struct mmu_gather *tlb;
1759         unsigned long nr_accounted = 0;
1760
1761         lru_add_drain();
1762         tlb = tlb_gather_mmu(mm, 0);
1763         update_hiwater_rss(mm);
1764         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1765         vm_unacct_memory(nr_accounted);
1766         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1767                                  next? next->vm_start: 0);
1768         tlb_finish_mmu(tlb, start, end);
1769 }
1770
1771 /*
1772  * Create a list of vma's touched by the unmap, removing them from the mm's
1773  * vma list as we go..
1774  */
1775 static void
1776 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1777         struct vm_area_struct *prev, unsigned long end)
1778 {
1779         struct vm_area_struct **insertion_point;
1780         struct vm_area_struct *tail_vma = NULL;
1781         unsigned long addr;
1782
1783         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1784         do {
1785                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1786                 mm->map_count--;
1787                 tail_vma = vma;
1788                 vma = vma->vm_next;
1789         } while (vma && vma->vm_start < end);
1790         *insertion_point = vma;
1791         tail_vma->vm_next = NULL;
1792         if (mm->unmap_area == arch_unmap_area)
1793                 addr = prev ? prev->vm_end : mm->mmap_base;
1794         else
1795                 addr = vma ?  vma->vm_start : mm->mmap_base;
1796         mm->unmap_area(mm, addr);
1797         mm->mmap_cache = NULL;          /* Kill the cache. */
1798 }
1799
1800 /*
1801  * Split a vma into two pieces at address 'addr', a new vma is allocated
1802  * either for the first part or the tail.
1803  */
1804 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1805               unsigned long addr, int new_below)
1806 {
1807         struct mempolicy *pol;
1808         struct vm_area_struct *new;
1809
1810         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1811                 return -EINVAL;
1812
1813         if (mm->map_count >= sysctl_max_map_count)
1814                 return -ENOMEM;
1815
1816         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1817         if (!new)
1818                 return -ENOMEM;
1819
1820         /* most fields are the same, copy all, and then fixup */
1821         *new = *vma;
1822
1823         if (new_below)
1824                 new->vm_end = addr;
1825         else {
1826                 new->vm_start = addr;
1827                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1828         }
1829
1830         pol = mpol_dup(vma_policy(vma));
1831         if (IS_ERR(pol)) {
1832                 kmem_cache_free(vm_area_cachep, new);
1833                 return PTR_ERR(pol);
1834         }
1835         vma_set_policy(new, pol);
1836
1837         if (new->vm_file) {
1838                 get_file(new->vm_file);
1839                 if (vma->vm_flags & VM_EXECUTABLE)
1840                         added_exe_file_vma(mm);
1841         }
1842
1843         if (new->vm_ops && new->vm_ops->open)
1844                 new->vm_ops->open(new);
1845
1846         if (new_below)
1847                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1848                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1849         else
1850                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1851
1852         return 0;
1853 }
1854
1855 /* Munmap is split into 2 main parts -- this part which finds
1856  * what needs doing, and the areas themselves, which do the
1857  * work.  This now handles partial unmappings.
1858  * Jeremy Fitzhardinge <jeremy@goop.org>
1859  */
1860 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1861 {
1862         unsigned long end;
1863         struct vm_area_struct *vma, *prev, *last;
1864
1865         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1866                 return -EINVAL;
1867
1868         if ((len = PAGE_ALIGN(len)) == 0)
1869                 return -EINVAL;
1870
1871         /* Find the first overlapping VMA */
1872         vma = find_vma_prev(mm, start, &prev);
1873         if (!vma)
1874                 return 0;
1875         /* we have  start < vma->vm_end  */
1876
1877         /* if it doesn't overlap, we have nothing.. */
1878         end = start + len;
1879         if (vma->vm_start >= end)
1880                 return 0;
1881
1882         /*
1883          * If we need to split any vma, do it now to save pain later.
1884          *
1885          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1886          * unmapped vm_area_struct will remain in use: so lower split_vma
1887          * places tmp vma above, and higher split_vma places tmp vma below.
1888          */
1889         if (start > vma->vm_start) {
1890                 int error = split_vma(mm, vma, start, 0);
1891                 if (error)
1892                         return error;
1893                 prev = vma;
1894         }
1895
1896         /* Does it split the last one? */
1897         last = find_vma(mm, end);
1898         if (last && end > last->vm_start) {
1899                 int error = split_vma(mm, last, end, 1);
1900                 if (error)
1901                         return error;
1902         }
1903         vma = prev? prev->vm_next: mm->mmap;
1904
1905         /*
1906          * Remove the vma's, and unmap the actual pages
1907          */
1908         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1909         unmap_region(mm, vma, prev, start, end);
1910
1911         /* Fix up all other VM information */
1912         remove_vma_list(mm, vma);
1913
1914         return 0;
1915 }
1916
1917 EXPORT_SYMBOL(do_munmap);
1918
1919 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1920 {
1921         int ret;
1922         struct mm_struct *mm = current->mm;
1923
1924         profile_munmap(addr);
1925
1926         down_write(&mm->mmap_sem);
1927         ret = do_munmap(mm, addr, len);
1928         up_write(&mm->mmap_sem);
1929         return ret;
1930 }
1931
1932 static inline void verify_mm_writelocked(struct mm_struct *mm)
1933 {
1934 #ifdef CONFIG_DEBUG_VM
1935         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1936                 WARN_ON(1);
1937                 up_read(&mm->mmap_sem);
1938         }
1939 #endif
1940 }
1941
1942 /*
1943  *  this is really a simplified "do_mmap".  it only handles
1944  *  anonymous maps.  eventually we may be able to do some
1945  *  brk-specific accounting here.
1946  */
1947 unsigned long do_brk(unsigned long addr, unsigned long len)
1948 {
1949         struct mm_struct * mm = current->mm;
1950         struct vm_area_struct * vma, * prev;
1951         unsigned long flags;
1952         struct rb_node ** rb_link, * rb_parent;
1953         pgoff_t pgoff = addr >> PAGE_SHIFT;
1954         int error;
1955
1956         len = PAGE_ALIGN(len);
1957         if (!len)
1958                 return addr;
1959
1960         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1961                 return -EINVAL;
1962
1963         if (is_hugepage_only_range(mm, addr, len))
1964                 return -EINVAL;
1965
1966         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1967         if (error)
1968                 return error;
1969
1970         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1971
1972         error = arch_mmap_check(addr, len, flags);
1973         if (error)
1974                 return error;
1975
1976         /*
1977          * mlock MCL_FUTURE?
1978          */
1979         if (mm->def_flags & VM_LOCKED) {
1980                 unsigned long locked, lock_limit;
1981                 locked = len >> PAGE_SHIFT;
1982                 locked += mm->locked_vm;
1983                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1984                 lock_limit >>= PAGE_SHIFT;
1985                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1986                         return -EAGAIN;
1987         }
1988
1989         /*
1990          * mm->mmap_sem is required to protect against another thread
1991          * changing the mappings in case we sleep.
1992          */
1993         verify_mm_writelocked(mm);
1994
1995         /*
1996          * Clear old maps.  this also does some error checking for us
1997          */
1998  munmap_back:
1999         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2000         if (vma && vma->vm_start < addr + len) {
2001                 if (do_munmap(mm, addr, len))
2002                         return -ENOMEM;
2003                 goto munmap_back;
2004         }
2005
2006         /* Check against address space limits *after* clearing old maps... */
2007         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2008                 return -ENOMEM;
2009
2010         if (mm->map_count > sysctl_max_map_count)
2011                 return -ENOMEM;
2012
2013         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2014                 return -ENOMEM;
2015
2016         /* Can we just expand an old private anonymous mapping? */
2017         if (vma_merge(mm, prev, addr, addr + len, flags,
2018                                         NULL, NULL, pgoff, NULL))
2019                 goto out;
2020
2021         /*
2022          * create a vma struct for an anonymous mapping
2023          */
2024         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2025         if (!vma) {
2026                 vm_unacct_memory(len >> PAGE_SHIFT);
2027                 return -ENOMEM;
2028         }
2029
2030         vma->vm_mm = mm;
2031         vma->vm_start = addr;
2032         vma->vm_end = addr + len;
2033         vma->vm_pgoff = pgoff;
2034         vma->vm_flags = flags;
2035         vma->vm_page_prot = vm_get_page_prot(flags);
2036         vma_link(mm, vma, prev, rb_link, rb_parent);
2037 out:
2038         mm->total_vm += len >> PAGE_SHIFT;
2039         if (flags & VM_LOCKED) {
2040                 mm->locked_vm += len >> PAGE_SHIFT;
2041                 make_pages_present(addr, addr + len);
2042         }
2043         return addr;
2044 }
2045
2046 EXPORT_SYMBOL(do_brk);
2047
2048 /* Release all mmaps. */
2049 void exit_mmap(struct mm_struct *mm)
2050 {
2051         struct mmu_gather *tlb;
2052         struct vm_area_struct *vma = mm->mmap;
2053         unsigned long nr_accounted = 0;
2054         unsigned long end;
2055
2056         /* mm's last user has gone, and its about to be pulled down */
2057         arch_exit_mmap(mm);
2058
2059         lru_add_drain();
2060         flush_cache_mm(mm);
2061         tlb = tlb_gather_mmu(mm, 1);
2062         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2063         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2064         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2065         vm_unacct_memory(nr_accounted);
2066         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2067         tlb_finish_mmu(tlb, 0, end);
2068
2069         /*
2070          * Walk the list again, actually closing and freeing it,
2071          * with preemption enabled, without holding any MM locks.
2072          */
2073         while (vma)
2074                 vma = remove_vma(vma);
2075
2076         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2077 }
2078
2079 /* Insert vm structure into process list sorted by address
2080  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2081  * then i_mmap_lock is taken here.
2082  */
2083 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2084 {
2085         struct vm_area_struct * __vma, * prev;
2086         struct rb_node ** rb_link, * rb_parent;
2087
2088         /*
2089          * The vm_pgoff of a purely anonymous vma should be irrelevant
2090          * until its first write fault, when page's anon_vma and index
2091          * are set.  But now set the vm_pgoff it will almost certainly
2092          * end up with (unless mremap moves it elsewhere before that
2093          * first wfault), so /proc/pid/maps tells a consistent story.
2094          *
2095          * By setting it to reflect the virtual start address of the
2096          * vma, merges and splits can happen in a seamless way, just
2097          * using the existing file pgoff checks and manipulations.
2098          * Similarly in do_mmap_pgoff and in do_brk.
2099          */
2100         if (!vma->vm_file) {
2101                 BUG_ON(vma->anon_vma);
2102                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2103         }
2104         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2105         if (__vma && __vma->vm_start < vma->vm_end)
2106                 return -ENOMEM;
2107         if ((vma->vm_flags & VM_ACCOUNT) &&
2108              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2109                 return -ENOMEM;
2110         vma_link(mm, vma, prev, rb_link, rb_parent);
2111         return 0;
2112 }
2113
2114 /*
2115  * Copy the vma structure to a new location in the same mm,
2116  * prior to moving page table entries, to effect an mremap move.
2117  */
2118 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2119         unsigned long addr, unsigned long len, pgoff_t pgoff)
2120 {
2121         struct vm_area_struct *vma = *vmap;
2122         unsigned long vma_start = vma->vm_start;
2123         struct mm_struct *mm = vma->vm_mm;
2124         struct vm_area_struct *new_vma, *prev;
2125         struct rb_node **rb_link, *rb_parent;
2126         struct mempolicy *pol;
2127
2128         /*
2129          * If anonymous vma has not yet been faulted, update new pgoff
2130          * to match new location, to increase its chance of merging.
2131          */
2132         if (!vma->vm_file && !vma->anon_vma)
2133                 pgoff = addr >> PAGE_SHIFT;
2134
2135         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2136         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2137                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2138         if (new_vma) {
2139                 /*
2140                  * Source vma may have been merged into new_vma
2141                  */
2142                 if (vma_start >= new_vma->vm_start &&
2143                     vma_start < new_vma->vm_end)
2144                         *vmap = new_vma;
2145         } else {
2146                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2147                 if (new_vma) {
2148                         *new_vma = *vma;
2149                         pol = mpol_dup(vma_policy(vma));
2150                         if (IS_ERR(pol)) {
2151                                 kmem_cache_free(vm_area_cachep, new_vma);
2152                                 return NULL;
2153                         }
2154                         vma_set_policy(new_vma, pol);
2155                         new_vma->vm_start = addr;
2156                         new_vma->vm_end = addr + len;
2157                         new_vma->vm_pgoff = pgoff;
2158                         if (new_vma->vm_file) {
2159                                 get_file(new_vma->vm_file);
2160                                 if (vma->vm_flags & VM_EXECUTABLE)
2161                                         added_exe_file_vma(mm);
2162                         }
2163                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2164                                 new_vma->vm_ops->open(new_vma);
2165                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2166                 }
2167         }
2168         return new_vma;
2169 }
2170
2171 /*
2172  * Return true if the calling process may expand its vm space by the passed
2173  * number of pages
2174  */
2175 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2176 {
2177         unsigned long cur = mm->total_vm;       /* pages */
2178         unsigned long lim;
2179
2180         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2181
2182         if (cur + npages > lim)
2183                 return 0;
2184         return 1;
2185 }
2186
2187
2188 static int special_mapping_fault(struct vm_area_struct *vma,
2189                                 struct vm_fault *vmf)
2190 {
2191         pgoff_t pgoff;
2192         struct page **pages;
2193
2194         /*
2195          * special mappings have no vm_file, and in that case, the mm
2196          * uses vm_pgoff internally. So we have to subtract it from here.
2197          * We are allowed to do this because we are the mm; do not copy
2198          * this code into drivers!
2199          */
2200         pgoff = vmf->pgoff - vma->vm_pgoff;
2201
2202         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2203                 pgoff--;
2204
2205         if (*pages) {
2206                 struct page *page = *pages;
2207                 get_page(page);
2208                 vmf->page = page;
2209                 return 0;
2210         }
2211
2212         return VM_FAULT_SIGBUS;
2213 }
2214
2215 /*
2216  * Having a close hook prevents vma merging regardless of flags.
2217  */
2218 static void special_mapping_close(struct vm_area_struct *vma)
2219 {
2220 }
2221
2222 static struct vm_operations_struct special_mapping_vmops = {
2223         .close = special_mapping_close,
2224         .fault = special_mapping_fault,
2225 };
2226
2227 /*
2228  * Called with mm->mmap_sem held for writing.
2229  * Insert a new vma covering the given region, with the given flags.
2230  * Its pages are supplied by the given array of struct page *.
2231  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2232  * The region past the last page supplied will always produce SIGBUS.
2233  * The array pointer and the pages it points to are assumed to stay alive
2234  * for as long as this mapping might exist.
2235  */
2236 int install_special_mapping(struct mm_struct *mm,
2237                             unsigned long addr, unsigned long len,
2238                             unsigned long vm_flags, struct page **pages)
2239 {
2240         struct vm_area_struct *vma;
2241
2242         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2243         if (unlikely(vma == NULL))
2244                 return -ENOMEM;
2245
2246         vma->vm_mm = mm;
2247         vma->vm_start = addr;
2248         vma->vm_end = addr + len;
2249
2250         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2251         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2252
2253         vma->vm_ops = &special_mapping_vmops;
2254         vma->vm_private_data = pages;
2255
2256         if (unlikely(insert_vm_struct(mm, vma))) {
2257                 kmem_cache_free(vm_area_cachep, vma);
2258                 return -ENOMEM;
2259         }
2260
2261         mm->total_vm += len >> PAGE_SHIFT;
2262
2263         return 0;
2264 }