mm: rearrange exit_mmap() to unlock before arch_exit_mmap
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
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 #include <linux/mmu_notifier.h>
30
31 #include <asm/uaccess.h>
32 #include <asm/cacheflush.h>
33 #include <asm/tlb.h>
34 #include <asm/mmu_context.h>
35
36 #include "internal.h"
37
38 #ifndef arch_mmap_check
39 #define arch_mmap_check(addr, len, flags)       (0)
40 #endif
41
42 #ifndef arch_rebalance_pgtables
43 #define arch_rebalance_pgtables(addr, len)              (addr)
44 #endif
45
46 static void unmap_region(struct mm_struct *mm,
47                 struct vm_area_struct *vma, struct vm_area_struct *prev,
48                 unsigned long start, unsigned long end);
49
50 /*
51  * WARNING: the debugging will use recursive algorithms so never enable this
52  * unless you know what you are doing.
53  */
54 #undef DEBUG_MM_RB
55
56 /* description of effects of mapping type and prot in current implementation.
57  * this is due to the limited x86 page protection hardware.  The expected
58  * behavior is in parens:
59  *
60  * map_type     prot
61  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
62  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
63  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
64  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
65  *              
66  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
67  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
68  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
69  *
70  */
71 pgprot_t protection_map[16] = {
72         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
73         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
74 };
75
76 pgprot_t vm_get_page_prot(unsigned long vm_flags)
77 {
78         return __pgprot(pgprot_val(protection_map[vm_flags &
79                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
80                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
81 }
82 EXPORT_SYMBOL(vm_get_page_prot);
83
84 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
85 int sysctl_overcommit_ratio = 50;       /* default is 50% */
86 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
87 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
88
89 /*
90  * Check that a process has enough memory to allocate a new virtual
91  * mapping. 0 means there is enough memory for the allocation to
92  * succeed and -ENOMEM implies there is not.
93  *
94  * We currently support three overcommit policies, which are set via the
95  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
96  *
97  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
98  * Additional code 2002 Jul 20 by Robert Love.
99  *
100  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
101  *
102  * Note this is a helper function intended to be used by LSMs which
103  * wish to use this logic.
104  */
105 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
106 {
107         unsigned long free, allowed;
108
109         vm_acct_memory(pages);
110
111         /*
112          * Sometimes we want to use more memory than we have
113          */
114         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
115                 return 0;
116
117         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
118                 unsigned long n;
119
120                 free = global_page_state(NR_FILE_PAGES);
121                 free += nr_swap_pages;
122
123                 /*
124                  * Any slabs which are created with the
125                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
126                  * which are reclaimable, under pressure.  The dentry
127                  * cache and most inode caches should fall into this
128                  */
129                 free += global_page_state(NR_SLAB_RECLAIMABLE);
130
131                 /*
132                  * Leave the last 3% for root
133                  */
134                 if (!cap_sys_admin)
135                         free -= free / 32;
136
137                 if (free > pages)
138                         return 0;
139
140                 /*
141                  * nr_free_pages() is very expensive on large systems,
142                  * only call if we're about to fail.
143                  */
144                 n = nr_free_pages();
145
146                 /*
147                  * Leave reserved pages. The pages are not for anonymous pages.
148                  */
149                 if (n <= totalreserve_pages)
150                         goto error;
151                 else
152                         n -= totalreserve_pages;
153
154                 /*
155                  * Leave the last 3% for root
156                  */
157                 if (!cap_sys_admin)
158                         n -= n / 32;
159                 free += n;
160
161                 if (free > pages)
162                         return 0;
163
164                 goto error;
165         }
166
167         allowed = (totalram_pages - hugetlb_total_pages())
168                 * sysctl_overcommit_ratio / 100;
169         /*
170          * Leave the last 3% for root
171          */
172         if (!cap_sys_admin)
173                 allowed -= allowed / 32;
174         allowed += total_swap_pages;
175
176         /* Don't let a single process grow too big:
177            leave 3% of the size of this process for other processes */
178         if (mm)
179                 allowed -= mm->total_vm / 32;
180
181         /*
182          * cast `allowed' as a signed long because vm_committed_space
183          * sometimes has a negative value
184          */
185         if (atomic_long_read(&vm_committed_space) < (long)allowed)
186                 return 0;
187 error:
188         vm_unacct_memory(pages);
189
190         return -ENOMEM;
191 }
192
193 /*
194  * Requires inode->i_mapping->i_mmap_lock
195  */
196 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
197                 struct file *file, struct address_space *mapping)
198 {
199         if (vma->vm_flags & VM_DENYWRITE)
200                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
201         if (vma->vm_flags & VM_SHARED)
202                 mapping->i_mmap_writable--;
203
204         flush_dcache_mmap_lock(mapping);
205         if (unlikely(vma->vm_flags & VM_NONLINEAR))
206                 list_del_init(&vma->shared.vm_set.list);
207         else
208                 vma_prio_tree_remove(vma, &mapping->i_mmap);
209         flush_dcache_mmap_unlock(mapping);
210 }
211
212 /*
213  * Unlink a file-based vm structure from its prio_tree, to hide
214  * vma from rmap and vmtruncate before freeing its page tables.
215  */
216 void unlink_file_vma(struct vm_area_struct *vma)
217 {
218         struct file *file = vma->vm_file;
219
220         if (file) {
221                 struct address_space *mapping = file->f_mapping;
222                 spin_lock(&mapping->i_mmap_lock);
223                 __remove_shared_vm_struct(vma, file, mapping);
224                 spin_unlock(&mapping->i_mmap_lock);
225         }
226 }
227
228 /*
229  * Close a vm structure and free it, returning the next.
230  */
231 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
232 {
233         struct vm_area_struct *next = vma->vm_next;
234
235         might_sleep();
236         if (vma->vm_ops && vma->vm_ops->close)
237                 vma->vm_ops->close(vma);
238         if (vma->vm_file) {
239                 fput(vma->vm_file);
240                 if (vma->vm_flags & VM_EXECUTABLE)
241                         removed_exe_file_vma(vma->vm_mm);
242         }
243         mpol_put(vma_policy(vma));
244         kmem_cache_free(vm_area_cachep, vma);
245         return next;
246 }
247
248 SYSCALL_DEFINE1(brk, unsigned long, brk)
249 {
250         unsigned long rlim, retval;
251         unsigned long newbrk, oldbrk;
252         struct mm_struct *mm = current->mm;
253         unsigned long min_brk;
254
255         down_write(&mm->mmap_sem);
256
257 #ifdef CONFIG_COMPAT_BRK
258         min_brk = mm->end_code;
259 #else
260         min_brk = mm->start_brk;
261 #endif
262         if (brk < min_brk)
263                 goto out;
264
265         /*
266          * Check against rlimit here. If this check is done later after the test
267          * of oldbrk with newbrk then it can escape the test and let the data
268          * segment grow beyond its set limit the in case where the limit is
269          * not page aligned -Ram Gupta
270          */
271         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
272         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
273                         (mm->end_data - mm->start_data) > rlim)
274                 goto out;
275
276         newbrk = PAGE_ALIGN(brk);
277         oldbrk = PAGE_ALIGN(mm->brk);
278         if (oldbrk == newbrk)
279                 goto set_brk;
280
281         /* Always allow shrinking brk. */
282         if (brk <= mm->brk) {
283                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
284                         goto set_brk;
285                 goto out;
286         }
287
288         /* Check against existing mmap mappings. */
289         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
290                 goto out;
291
292         /* Ok, looks good - let it rip. */
293         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
294                 goto out;
295 set_brk:
296         mm->brk = brk;
297 out:
298         retval = mm->brk;
299         up_write(&mm->mmap_sem);
300         return retval;
301 }
302
303 #ifdef DEBUG_MM_RB
304 static int browse_rb(struct rb_root *root)
305 {
306         int i = 0, j;
307         struct rb_node *nd, *pn = NULL;
308         unsigned long prev = 0, pend = 0;
309
310         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
311                 struct vm_area_struct *vma;
312                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
313                 if (vma->vm_start < prev)
314                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
315                 if (vma->vm_start < pend)
316                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
317                 if (vma->vm_start > vma->vm_end)
318                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
319                 i++;
320                 pn = nd;
321                 prev = vma->vm_start;
322                 pend = vma->vm_end;
323         }
324         j = 0;
325         for (nd = pn; nd; nd = rb_prev(nd)) {
326                 j++;
327         }
328         if (i != j)
329                 printk("backwards %d, forwards %d\n", j, i), i = 0;
330         return i;
331 }
332
333 void validate_mm(struct mm_struct *mm)
334 {
335         int bug = 0;
336         int i = 0;
337         struct vm_area_struct *tmp = mm->mmap;
338         while (tmp) {
339                 tmp = tmp->vm_next;
340                 i++;
341         }
342         if (i != mm->map_count)
343                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
344         i = browse_rb(&mm->mm_rb);
345         if (i != mm->map_count)
346                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
347         BUG_ON(bug);
348 }
349 #else
350 #define validate_mm(mm) do { } while (0)
351 #endif
352
353 static struct vm_area_struct *
354 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
355                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
356                 struct rb_node ** rb_parent)
357 {
358         struct vm_area_struct * vma;
359         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
360
361         __rb_link = &mm->mm_rb.rb_node;
362         rb_prev = __rb_parent = NULL;
363         vma = NULL;
364
365         while (*__rb_link) {
366                 struct vm_area_struct *vma_tmp;
367
368                 __rb_parent = *__rb_link;
369                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
370
371                 if (vma_tmp->vm_end > addr) {
372                         vma = vma_tmp;
373                         if (vma_tmp->vm_start <= addr)
374                                 break;
375                         __rb_link = &__rb_parent->rb_left;
376                 } else {
377                         rb_prev = __rb_parent;
378                         __rb_link = &__rb_parent->rb_right;
379                 }
380         }
381
382         *pprev = NULL;
383         if (rb_prev)
384                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
385         *rb_link = __rb_link;
386         *rb_parent = __rb_parent;
387         return vma;
388 }
389
390 static inline void
391 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
392                 struct vm_area_struct *prev, struct rb_node *rb_parent)
393 {
394         if (prev) {
395                 vma->vm_next = prev->vm_next;
396                 prev->vm_next = vma;
397         } else {
398                 mm->mmap = vma;
399                 if (rb_parent)
400                         vma->vm_next = rb_entry(rb_parent,
401                                         struct vm_area_struct, vm_rb);
402                 else
403                         vma->vm_next = NULL;
404         }
405 }
406
407 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
408                 struct rb_node **rb_link, struct rb_node *rb_parent)
409 {
410         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
411         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
412 }
413
414 static void __vma_link_file(struct vm_area_struct *vma)
415 {
416         struct file *file;
417
418         file = vma->vm_file;
419         if (file) {
420                 struct address_space *mapping = file->f_mapping;
421
422                 if (vma->vm_flags & VM_DENYWRITE)
423                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
424                 if (vma->vm_flags & VM_SHARED)
425                         mapping->i_mmap_writable++;
426
427                 flush_dcache_mmap_lock(mapping);
428                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
429                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
430                 else
431                         vma_prio_tree_insert(vma, &mapping->i_mmap);
432                 flush_dcache_mmap_unlock(mapping);
433         }
434 }
435
436 static void
437 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
438         struct vm_area_struct *prev, struct rb_node **rb_link,
439         struct rb_node *rb_parent)
440 {
441         __vma_link_list(mm, vma, prev, rb_parent);
442         __vma_link_rb(mm, vma, rb_link, rb_parent);
443         __anon_vma_link(vma);
444 }
445
446 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
447                         struct vm_area_struct *prev, struct rb_node **rb_link,
448                         struct rb_node *rb_parent)
449 {
450         struct address_space *mapping = NULL;
451
452         if (vma->vm_file)
453                 mapping = vma->vm_file->f_mapping;
454
455         if (mapping) {
456                 spin_lock(&mapping->i_mmap_lock);
457                 vma->vm_truncate_count = mapping->truncate_count;
458         }
459         anon_vma_lock(vma);
460
461         __vma_link(mm, vma, prev, rb_link, rb_parent);
462         __vma_link_file(vma);
463
464         anon_vma_unlock(vma);
465         if (mapping)
466                 spin_unlock(&mapping->i_mmap_lock);
467
468         mm->map_count++;
469         validate_mm(mm);
470 }
471
472 /*
473  * Helper for vma_adjust in the split_vma insert case:
474  * insert vm structure into list and rbtree and anon_vma,
475  * but it has already been inserted into prio_tree earlier.
476  */
477 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
478 {
479         struct vm_area_struct *__vma, *prev;
480         struct rb_node **rb_link, *rb_parent;
481
482         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
483         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
484         __vma_link(mm, vma, prev, rb_link, rb_parent);
485         mm->map_count++;
486 }
487
488 static inline void
489 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
490                 struct vm_area_struct *prev)
491 {
492         prev->vm_next = vma->vm_next;
493         rb_erase(&vma->vm_rb, &mm->mm_rb);
494         if (mm->mmap_cache == vma)
495                 mm->mmap_cache = prev;
496 }
497
498 /*
499  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
500  * is already present in an i_mmap tree without adjusting the tree.
501  * The following helper function should be used when such adjustments
502  * are necessary.  The "insert" vma (if any) is to be inserted
503  * before we drop the necessary locks.
504  */
505 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
506         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
507 {
508         struct mm_struct *mm = vma->vm_mm;
509         struct vm_area_struct *next = vma->vm_next;
510         struct vm_area_struct *importer = NULL;
511         struct address_space *mapping = NULL;
512         struct prio_tree_root *root = NULL;
513         struct file *file = vma->vm_file;
514         struct anon_vma *anon_vma = NULL;
515         long adjust_next = 0;
516         int remove_next = 0;
517
518         if (next && !insert) {
519                 if (end >= next->vm_end) {
520                         /*
521                          * vma expands, overlapping all the next, and
522                          * perhaps the one after too (mprotect case 6).
523                          */
524 again:                  remove_next = 1 + (end > next->vm_end);
525                         end = next->vm_end;
526                         anon_vma = next->anon_vma;
527                         importer = vma;
528                 } else if (end > next->vm_start) {
529                         /*
530                          * vma expands, overlapping part of the next:
531                          * mprotect case 5 shifting the boundary up.
532                          */
533                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
534                         anon_vma = next->anon_vma;
535                         importer = vma;
536                 } else if (end < vma->vm_end) {
537                         /*
538                          * vma shrinks, and !insert tells it's not
539                          * split_vma inserting another: so it must be
540                          * mprotect case 4 shifting the boundary down.
541                          */
542                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
543                         anon_vma = next->anon_vma;
544                         importer = next;
545                 }
546         }
547
548         if (file) {
549                 mapping = file->f_mapping;
550                 if (!(vma->vm_flags & VM_NONLINEAR))
551                         root = &mapping->i_mmap;
552                 spin_lock(&mapping->i_mmap_lock);
553                 if (importer &&
554                     vma->vm_truncate_count != next->vm_truncate_count) {
555                         /*
556                          * unmap_mapping_range might be in progress:
557                          * ensure that the expanding vma is rescanned.
558                          */
559                         importer->vm_truncate_count = 0;
560                 }
561                 if (insert) {
562                         insert->vm_truncate_count = vma->vm_truncate_count;
563                         /*
564                          * Put into prio_tree now, so instantiated pages
565                          * are visible to arm/parisc __flush_dcache_page
566                          * throughout; but we cannot insert into address
567                          * space until vma start or end is updated.
568                          */
569                         __vma_link_file(insert);
570                 }
571         }
572
573         /*
574          * When changing only vma->vm_end, we don't really need
575          * anon_vma lock: but is that case worth optimizing out?
576          */
577         if (vma->anon_vma)
578                 anon_vma = vma->anon_vma;
579         if (anon_vma) {
580                 spin_lock(&anon_vma->lock);
581                 /*
582                  * Easily overlooked: when mprotect shifts the boundary,
583                  * make sure the expanding vma has anon_vma set if the
584                  * shrinking vma had, to cover any anon pages imported.
585                  */
586                 if (importer && !importer->anon_vma) {
587                         importer->anon_vma = anon_vma;
588                         __anon_vma_link(importer);
589                 }
590         }
591
592         if (root) {
593                 flush_dcache_mmap_lock(mapping);
594                 vma_prio_tree_remove(vma, root);
595                 if (adjust_next)
596                         vma_prio_tree_remove(next, root);
597         }
598
599         vma->vm_start = start;
600         vma->vm_end = end;
601         vma->vm_pgoff = pgoff;
602         if (adjust_next) {
603                 next->vm_start += adjust_next << PAGE_SHIFT;
604                 next->vm_pgoff += adjust_next;
605         }
606
607         if (root) {
608                 if (adjust_next)
609                         vma_prio_tree_insert(next, root);
610                 vma_prio_tree_insert(vma, root);
611                 flush_dcache_mmap_unlock(mapping);
612         }
613
614         if (remove_next) {
615                 /*
616                  * vma_merge has merged next into vma, and needs
617                  * us to remove next before dropping the locks.
618                  */
619                 __vma_unlink(mm, next, vma);
620                 if (file)
621                         __remove_shared_vm_struct(next, file, mapping);
622                 if (next->anon_vma)
623                         __anon_vma_merge(vma, next);
624         } else if (insert) {
625                 /*
626                  * split_vma has split insert from vma, and needs
627                  * us to insert it before dropping the locks
628                  * (it may either follow vma or precede it).
629                  */
630                 __insert_vm_struct(mm, insert);
631         }
632
633         if (anon_vma)
634                 spin_unlock(&anon_vma->lock);
635         if (mapping)
636                 spin_unlock(&mapping->i_mmap_lock);
637
638         if (remove_next) {
639                 if (file) {
640                         fput(file);
641                         if (next->vm_flags & VM_EXECUTABLE)
642                                 removed_exe_file_vma(mm);
643                 }
644                 mm->map_count--;
645                 mpol_put(vma_policy(next));
646                 kmem_cache_free(vm_area_cachep, next);
647                 /*
648                  * In mprotect's case 6 (see comments on vma_merge),
649                  * we must remove another next too. It would clutter
650                  * up the code too much to do both in one go.
651                  */
652                 if (remove_next == 2) {
653                         next = vma->vm_next;
654                         goto again;
655                 }
656         }
657
658         validate_mm(mm);
659 }
660
661 /* Flags that can be inherited from an existing mapping when merging */
662 #define VM_MERGEABLE_FLAGS (VM_CAN_NONLINEAR)
663
664 /*
665  * If the vma has a ->close operation then the driver probably needs to release
666  * per-vma resources, so we don't attempt to merge those.
667  */
668 static inline int is_mergeable_vma(struct vm_area_struct *vma,
669                         struct file *file, unsigned long vm_flags)
670 {
671         if ((vma->vm_flags ^ vm_flags) & ~VM_MERGEABLE_FLAGS)
672                 return 0;
673         if (vma->vm_file != file)
674                 return 0;
675         if (vma->vm_ops && vma->vm_ops->close)
676                 return 0;
677         return 1;
678 }
679
680 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
681                                         struct anon_vma *anon_vma2)
682 {
683         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
684 }
685
686 /*
687  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
688  * in front of (at a lower virtual address and file offset than) the vma.
689  *
690  * We cannot merge two vmas if they have differently assigned (non-NULL)
691  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
692  *
693  * We don't check here for the merged mmap wrapping around the end of pagecache
694  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
695  * wrap, nor mmaps which cover the final page at index -1UL.
696  */
697 static int
698 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
699         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
700 {
701         if (is_mergeable_vma(vma, file, vm_flags) &&
702             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
703                 if (vma->vm_pgoff == vm_pgoff)
704                         return 1;
705         }
706         return 0;
707 }
708
709 /*
710  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
711  * beyond (at a higher virtual address and file offset than) the vma.
712  *
713  * We cannot merge two vmas if they have differently assigned (non-NULL)
714  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
715  */
716 static int
717 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
718         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
719 {
720         if (is_mergeable_vma(vma, file, vm_flags) &&
721             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
722                 pgoff_t vm_pglen;
723                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
724                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
725                         return 1;
726         }
727         return 0;
728 }
729
730 /*
731  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
732  * whether that can be merged with its predecessor or its successor.
733  * Or both (it neatly fills a hole).
734  *
735  * In most cases - when called for mmap, brk or mremap - [addr,end) is
736  * certain not to be mapped by the time vma_merge is called; but when
737  * called for mprotect, it is certain to be already mapped (either at
738  * an offset within prev, or at the start of next), and the flags of
739  * this area are about to be changed to vm_flags - and the no-change
740  * case has already been eliminated.
741  *
742  * The following mprotect cases have to be considered, where AAAA is
743  * the area passed down from mprotect_fixup, never extending beyond one
744  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
745  *
746  *     AAAA             AAAA                AAAA          AAAA
747  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
748  *    cannot merge    might become    might become    might become
749  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
750  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
751  *    mremap move:                                    PPPPNNNNNNNN 8
752  *        AAAA
753  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
754  *    might become    case 1 below    case 2 below    case 3 below
755  *
756  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
757  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
758  */
759 struct vm_area_struct *vma_merge(struct mm_struct *mm,
760                         struct vm_area_struct *prev, unsigned long addr,
761                         unsigned long end, unsigned long vm_flags,
762                         struct anon_vma *anon_vma, struct file *file,
763                         pgoff_t pgoff, struct mempolicy *policy)
764 {
765         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
766         struct vm_area_struct *area, *next;
767
768         /*
769          * We later require that vma->vm_flags == vm_flags,
770          * so this tests vma->vm_flags & VM_SPECIAL, too.
771          */
772         if (vm_flags & VM_SPECIAL)
773                 return NULL;
774
775         if (prev)
776                 next = prev->vm_next;
777         else
778                 next = mm->mmap;
779         area = next;
780         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
781                 next = next->vm_next;
782
783         /*
784          * Can it merge with the predecessor?
785          */
786         if (prev && prev->vm_end == addr &&
787                         mpol_equal(vma_policy(prev), policy) &&
788                         can_vma_merge_after(prev, vm_flags,
789                                                 anon_vma, file, pgoff)) {
790                 /*
791                  * OK, it can.  Can we now merge in the successor as well?
792                  */
793                 if (next && end == next->vm_start &&
794                                 mpol_equal(policy, vma_policy(next)) &&
795                                 can_vma_merge_before(next, vm_flags,
796                                         anon_vma, file, pgoff+pglen) &&
797                                 is_mergeable_anon_vma(prev->anon_vma,
798                                                       next->anon_vma)) {
799                                                         /* cases 1, 6 */
800                         vma_adjust(prev, prev->vm_start,
801                                 next->vm_end, prev->vm_pgoff, NULL);
802                 } else                                  /* cases 2, 5, 7 */
803                         vma_adjust(prev, prev->vm_start,
804                                 end, prev->vm_pgoff, NULL);
805                 return prev;
806         }
807
808         /*
809          * Can this new request be merged in front of next?
810          */
811         if (next && end == next->vm_start &&
812                         mpol_equal(policy, vma_policy(next)) &&
813                         can_vma_merge_before(next, vm_flags,
814                                         anon_vma, file, pgoff+pglen)) {
815                 if (prev && addr < prev->vm_end)        /* case 4 */
816                         vma_adjust(prev, prev->vm_start,
817                                 addr, prev->vm_pgoff, NULL);
818                 else                                    /* cases 3, 8 */
819                         vma_adjust(area, addr, next->vm_end,
820                                 next->vm_pgoff - pglen, NULL);
821                 return area;
822         }
823
824         return NULL;
825 }
826
827 /*
828  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
829  * neighbouring vmas for a suitable anon_vma, before it goes off
830  * to allocate a new anon_vma.  It checks because a repetitive
831  * sequence of mprotects and faults may otherwise lead to distinct
832  * anon_vmas being allocated, preventing vma merge in subsequent
833  * mprotect.
834  */
835 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
836 {
837         struct vm_area_struct *near;
838         unsigned long vm_flags;
839
840         near = vma->vm_next;
841         if (!near)
842                 goto try_prev;
843
844         /*
845          * Since only mprotect tries to remerge vmas, match flags
846          * which might be mprotected into each other later on.
847          * Neither mlock nor madvise tries to remerge at present,
848          * so leave their flags as obstructing a merge.
849          */
850         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
851         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
852
853         if (near->anon_vma && vma->vm_end == near->vm_start &&
854                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
855                         can_vma_merge_before(near, vm_flags,
856                                 NULL, vma->vm_file, vma->vm_pgoff +
857                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
858                 return near->anon_vma;
859 try_prev:
860         /*
861          * It is potentially slow to have to call find_vma_prev here.
862          * But it's only on the first write fault on the vma, not
863          * every time, and we could devise a way to avoid it later
864          * (e.g. stash info in next's anon_vma_node when assigning
865          * an anon_vma, or when trying vma_merge).  Another time.
866          */
867         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
868         if (!near)
869                 goto none;
870
871         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
872         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
873
874         if (near->anon_vma && near->vm_end == vma->vm_start &&
875                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
876                         can_vma_merge_after(near, vm_flags,
877                                 NULL, vma->vm_file, vma->vm_pgoff))
878                 return near->anon_vma;
879 none:
880         /*
881          * There's no absolute need to look only at touching neighbours:
882          * we could search further afield for "compatible" anon_vmas.
883          * But it would probably just be a waste of time searching,
884          * or lead to too many vmas hanging off the same anon_vma.
885          * We're trying to allow mprotect remerging later on,
886          * not trying to minimize memory used for anon_vmas.
887          */
888         return NULL;
889 }
890
891 #ifdef CONFIG_PROC_FS
892 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
893                                                 struct file *file, long pages)
894 {
895         const unsigned long stack_flags
896                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
897
898         if (file) {
899                 mm->shared_vm += pages;
900                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
901                         mm->exec_vm += pages;
902         } else if (flags & stack_flags)
903                 mm->stack_vm += pages;
904         if (flags & (VM_RESERVED|VM_IO))
905                 mm->reserved_vm += pages;
906 }
907 #endif /* CONFIG_PROC_FS */
908
909 /*
910  * The caller must hold down_write(current->mm->mmap_sem).
911  */
912
913 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
914                         unsigned long len, unsigned long prot,
915                         unsigned long flags, unsigned long pgoff)
916 {
917         struct mm_struct * mm = current->mm;
918         struct inode *inode;
919         unsigned int vm_flags;
920         int error;
921         unsigned long reqprot = prot;
922
923         /*
924          * Does the application expect PROT_READ to imply PROT_EXEC?
925          *
926          * (the exception is when the underlying filesystem is noexec
927          *  mounted, in which case we dont add PROT_EXEC.)
928          */
929         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
930                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
931                         prot |= PROT_EXEC;
932
933         if (!len)
934                 return -EINVAL;
935
936         if (!(flags & MAP_FIXED))
937                 addr = round_hint_to_min(addr);
938
939         error = arch_mmap_check(addr, len, flags);
940         if (error)
941                 return error;
942
943         /* Careful about overflows.. */
944         len = PAGE_ALIGN(len);
945         if (!len || len > TASK_SIZE)
946                 return -ENOMEM;
947
948         /* offset overflow? */
949         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
950                return -EOVERFLOW;
951
952         /* Too many mappings? */
953         if (mm->map_count > sysctl_max_map_count)
954                 return -ENOMEM;
955
956         /* Obtain the address to map to. we verify (or select) it and ensure
957          * that it represents a valid section of the address space.
958          */
959         addr = get_unmapped_area(file, addr, len, pgoff, flags);
960         if (addr & ~PAGE_MASK)
961                 return addr;
962
963         /* Do simple checking here so the lower-level routines won't have
964          * to. we assume access permissions have been handled by the open
965          * of the memory object, so we don't do any here.
966          */
967         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
968                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
969
970         if (flags & MAP_LOCKED) {
971                 if (!can_do_mlock())
972                         return -EPERM;
973                 vm_flags |= VM_LOCKED;
974         }
975
976         /* mlock MCL_FUTURE? */
977         if (vm_flags & VM_LOCKED) {
978                 unsigned long locked, lock_limit;
979                 locked = len >> PAGE_SHIFT;
980                 locked += mm->locked_vm;
981                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
982                 lock_limit >>= PAGE_SHIFT;
983                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
984                         return -EAGAIN;
985         }
986
987         inode = file ? file->f_path.dentry->d_inode : NULL;
988
989         if (file) {
990                 switch (flags & MAP_TYPE) {
991                 case MAP_SHARED:
992                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
993                                 return -EACCES;
994
995                         /*
996                          * Make sure we don't allow writing to an append-only
997                          * file..
998                          */
999                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1000                                 return -EACCES;
1001
1002                         /*
1003                          * Make sure there are no mandatory locks on the file.
1004                          */
1005                         if (locks_verify_locked(inode))
1006                                 return -EAGAIN;
1007
1008                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1009                         if (!(file->f_mode & FMODE_WRITE))
1010                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1011
1012                         /* fall through */
1013                 case MAP_PRIVATE:
1014                         if (!(file->f_mode & FMODE_READ))
1015                                 return -EACCES;
1016                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1017                                 if (vm_flags & VM_EXEC)
1018                                         return -EPERM;
1019                                 vm_flags &= ~VM_MAYEXEC;
1020                         }
1021
1022                         if (!file->f_op || !file->f_op->mmap)
1023                                 return -ENODEV;
1024                         break;
1025
1026                 default:
1027                         return -EINVAL;
1028                 }
1029         } else {
1030                 switch (flags & MAP_TYPE) {
1031                 case MAP_SHARED:
1032                         /*
1033                          * Ignore pgoff.
1034                          */
1035                         pgoff = 0;
1036                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1037                         break;
1038                 case MAP_PRIVATE:
1039                         /*
1040                          * Set pgoff according to addr for anon_vma.
1041                          */
1042                         pgoff = addr >> PAGE_SHIFT;
1043                         break;
1044                 default:
1045                         return -EINVAL;
1046                 }
1047         }
1048
1049         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1050         if (error)
1051                 return error;
1052
1053         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1054 }
1055 EXPORT_SYMBOL(do_mmap_pgoff);
1056
1057 /*
1058  * Some shared mappigns will want the pages marked read-only
1059  * to track write events. If so, we'll downgrade vm_page_prot
1060  * to the private version (using protection_map[] without the
1061  * VM_SHARED bit).
1062  */
1063 int vma_wants_writenotify(struct vm_area_struct *vma)
1064 {
1065         unsigned int vm_flags = vma->vm_flags;
1066
1067         /* If it was private or non-writable, the write bit is already clear */
1068         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1069                 return 0;
1070
1071         /* The backer wishes to know when pages are first written to? */
1072         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1073                 return 1;
1074
1075         /* The open routine did something to the protections already? */
1076         if (pgprot_val(vma->vm_page_prot) !=
1077             pgprot_val(vm_get_page_prot(vm_flags)))
1078                 return 0;
1079
1080         /* Specialty mapping? */
1081         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1082                 return 0;
1083
1084         /* Can the mapping track the dirty pages? */
1085         return vma->vm_file && vma->vm_file->f_mapping &&
1086                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1087 }
1088
1089 /*
1090  * We account for memory if it's a private writeable mapping,
1091  * not hugepages and VM_NORESERVE wasn't set.
1092  */
1093 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1094 {
1095         /*
1096          * hugetlb has its own accounting separate from the core VM
1097          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1098          */
1099         if (file && is_file_hugepages(file))
1100                 return 0;
1101
1102         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1103 }
1104
1105 unsigned long mmap_region(struct file *file, unsigned long addr,
1106                           unsigned long len, unsigned long flags,
1107                           unsigned int vm_flags, unsigned long pgoff)
1108 {
1109         struct mm_struct *mm = current->mm;
1110         struct vm_area_struct *vma, *prev;
1111         int correct_wcount = 0;
1112         int error;
1113         struct rb_node **rb_link, *rb_parent;
1114         unsigned long charged = 0;
1115         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1116
1117         /* Clear old maps */
1118         error = -ENOMEM;
1119 munmap_back:
1120         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1121         if (vma && vma->vm_start < addr + len) {
1122                 if (do_munmap(mm, addr, len))
1123                         return -ENOMEM;
1124                 goto munmap_back;
1125         }
1126
1127         /* Check against address space limit. */
1128         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1129                 return -ENOMEM;
1130
1131         /*
1132          * Set 'VM_NORESERVE' if we should not account for the
1133          * memory use of this mapping.
1134          */
1135         if ((flags & MAP_NORESERVE)) {
1136                 /* We honor MAP_NORESERVE if allowed to overcommit */
1137                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1138                         vm_flags |= VM_NORESERVE;
1139
1140                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1141                 if (file && is_file_hugepages(file))
1142                         vm_flags |= VM_NORESERVE;
1143         }
1144
1145         /*
1146          * Private writable mapping: check memory availability
1147          */
1148         if (accountable_mapping(file, vm_flags)) {
1149                 charged = len >> PAGE_SHIFT;
1150                 if (security_vm_enough_memory(charged))
1151                         return -ENOMEM;
1152                 vm_flags |= VM_ACCOUNT;
1153         }
1154
1155         /*
1156          * Can we just expand an old mapping?
1157          */
1158         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1159         if (vma)
1160                 goto out;
1161
1162         /*
1163          * Determine the object being mapped and call the appropriate
1164          * specific mapper. the address has already been validated, but
1165          * not unmapped, but the maps are removed from the list.
1166          */
1167         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1168         if (!vma) {
1169                 error = -ENOMEM;
1170                 goto unacct_error;
1171         }
1172
1173         vma->vm_mm = mm;
1174         vma->vm_start = addr;
1175         vma->vm_end = addr + len;
1176         vma->vm_flags = vm_flags;
1177         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1178         vma->vm_pgoff = pgoff;
1179
1180         if (file) {
1181                 error = -EINVAL;
1182                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1183                         goto free_vma;
1184                 if (vm_flags & VM_DENYWRITE) {
1185                         error = deny_write_access(file);
1186                         if (error)
1187                                 goto free_vma;
1188                         correct_wcount = 1;
1189                 }
1190                 vma->vm_file = file;
1191                 get_file(file);
1192                 error = file->f_op->mmap(file, vma);
1193                 if (error)
1194                         goto unmap_and_free_vma;
1195                 if (vm_flags & VM_EXECUTABLE)
1196                         added_exe_file_vma(mm);
1197         } else if (vm_flags & VM_SHARED) {
1198                 error = shmem_zero_setup(vma);
1199                 if (error)
1200                         goto free_vma;
1201         }
1202
1203         /* Can addr have changed??
1204          *
1205          * Answer: Yes, several device drivers can do it in their
1206          *         f_op->mmap method. -DaveM
1207          */
1208         addr = vma->vm_start;
1209         pgoff = vma->vm_pgoff;
1210         vm_flags = vma->vm_flags;
1211
1212         if (vma_wants_writenotify(vma))
1213                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1214
1215         vma_link(mm, vma, prev, rb_link, rb_parent);
1216         file = vma->vm_file;
1217
1218         /* Once vma denies write, undo our temporary denial count */
1219         if (correct_wcount)
1220                 atomic_inc(&inode->i_writecount);
1221 out:
1222         mm->total_vm += len >> PAGE_SHIFT;
1223         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1224         if (vm_flags & VM_LOCKED) {
1225                 /*
1226                  * makes pages present; downgrades, drops, reacquires mmap_sem
1227                  */
1228                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1229                 if (nr_pages < 0)
1230                         return nr_pages;        /* vma gone! */
1231                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1232         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1233                 make_pages_present(addr, addr + len);
1234         return addr;
1235
1236 unmap_and_free_vma:
1237         if (correct_wcount)
1238                 atomic_inc(&inode->i_writecount);
1239         vma->vm_file = NULL;
1240         fput(file);
1241
1242         /* Undo any partial mapping done by a device driver. */
1243         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1244         charged = 0;
1245 free_vma:
1246         kmem_cache_free(vm_area_cachep, vma);
1247 unacct_error:
1248         if (charged)
1249                 vm_unacct_memory(charged);
1250         return error;
1251 }
1252
1253 /* Get an address range which is currently unmapped.
1254  * For shmat() with addr=0.
1255  *
1256  * Ugly calling convention alert:
1257  * Return value with the low bits set means error value,
1258  * ie
1259  *      if (ret & ~PAGE_MASK)
1260  *              error = ret;
1261  *
1262  * This function "knows" that -ENOMEM has the bits set.
1263  */
1264 #ifndef HAVE_ARCH_UNMAPPED_AREA
1265 unsigned long
1266 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1267                 unsigned long len, unsigned long pgoff, unsigned long flags)
1268 {
1269         struct mm_struct *mm = current->mm;
1270         struct vm_area_struct *vma;
1271         unsigned long start_addr;
1272
1273         if (len > TASK_SIZE)
1274                 return -ENOMEM;
1275
1276         if (flags & MAP_FIXED)
1277                 return addr;
1278
1279         if (addr) {
1280                 addr = PAGE_ALIGN(addr);
1281                 vma = find_vma(mm, addr);
1282                 if (TASK_SIZE - len >= addr &&
1283                     (!vma || addr + len <= vma->vm_start))
1284                         return addr;
1285         }
1286         if (len > mm->cached_hole_size) {
1287                 start_addr = addr = mm->free_area_cache;
1288         } else {
1289                 start_addr = addr = TASK_UNMAPPED_BASE;
1290                 mm->cached_hole_size = 0;
1291         }
1292
1293 full_search:
1294         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1295                 /* At this point:  (!vma || addr < vma->vm_end). */
1296                 if (TASK_SIZE - len < addr) {
1297                         /*
1298                          * Start a new search - just in case we missed
1299                          * some holes.
1300                          */
1301                         if (start_addr != TASK_UNMAPPED_BASE) {
1302                                 addr = TASK_UNMAPPED_BASE;
1303                                 start_addr = addr;
1304                                 mm->cached_hole_size = 0;
1305                                 goto full_search;
1306                         }
1307                         return -ENOMEM;
1308                 }
1309                 if (!vma || addr + len <= vma->vm_start) {
1310                         /*
1311                          * Remember the place where we stopped the search:
1312                          */
1313                         mm->free_area_cache = addr + len;
1314                         return addr;
1315                 }
1316                 if (addr + mm->cached_hole_size < vma->vm_start)
1317                         mm->cached_hole_size = vma->vm_start - addr;
1318                 addr = vma->vm_end;
1319         }
1320 }
1321 #endif  
1322
1323 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1324 {
1325         /*
1326          * Is this a new hole at the lowest possible address?
1327          */
1328         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1329                 mm->free_area_cache = addr;
1330                 mm->cached_hole_size = ~0UL;
1331         }
1332 }
1333
1334 /*
1335  * This mmap-allocator allocates new areas top-down from below the
1336  * stack's low limit (the base):
1337  */
1338 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1339 unsigned long
1340 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1341                           const unsigned long len, const unsigned long pgoff,
1342                           const unsigned long flags)
1343 {
1344         struct vm_area_struct *vma;
1345         struct mm_struct *mm = current->mm;
1346         unsigned long addr = addr0;
1347
1348         /* requested length too big for entire address space */
1349         if (len > TASK_SIZE)
1350                 return -ENOMEM;
1351
1352         if (flags & MAP_FIXED)
1353                 return addr;
1354
1355         /* requesting a specific address */
1356         if (addr) {
1357                 addr = PAGE_ALIGN(addr);
1358                 vma = find_vma(mm, addr);
1359                 if (TASK_SIZE - len >= addr &&
1360                                 (!vma || addr + len <= vma->vm_start))
1361                         return addr;
1362         }
1363
1364         /* check if free_area_cache is useful for us */
1365         if (len <= mm->cached_hole_size) {
1366                 mm->cached_hole_size = 0;
1367                 mm->free_area_cache = mm->mmap_base;
1368         }
1369
1370         /* either no address requested or can't fit in requested address hole */
1371         addr = mm->free_area_cache;
1372
1373         /* make sure it can fit in the remaining address space */
1374         if (addr > len) {
1375                 vma = find_vma(mm, addr-len);
1376                 if (!vma || addr <= vma->vm_start)
1377                         /* remember the address as a hint for next time */
1378                         return (mm->free_area_cache = addr-len);
1379         }
1380
1381         if (mm->mmap_base < len)
1382                 goto bottomup;
1383
1384         addr = mm->mmap_base-len;
1385
1386         do {
1387                 /*
1388                  * Lookup failure means no vma is above this address,
1389                  * else if new region fits below vma->vm_start,
1390                  * return with success:
1391                  */
1392                 vma = find_vma(mm, addr);
1393                 if (!vma || addr+len <= vma->vm_start)
1394                         /* remember the address as a hint for next time */
1395                         return (mm->free_area_cache = addr);
1396
1397                 /* remember the largest hole we saw so far */
1398                 if (addr + mm->cached_hole_size < vma->vm_start)
1399                         mm->cached_hole_size = vma->vm_start - addr;
1400
1401                 /* try just below the current vma->vm_start */
1402                 addr = vma->vm_start-len;
1403         } while (len < vma->vm_start);
1404
1405 bottomup:
1406         /*
1407          * A failed mmap() very likely causes application failure,
1408          * so fall back to the bottom-up function here. This scenario
1409          * can happen with large stack limits and large mmap()
1410          * allocations.
1411          */
1412         mm->cached_hole_size = ~0UL;
1413         mm->free_area_cache = TASK_UNMAPPED_BASE;
1414         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1415         /*
1416          * Restore the topdown base:
1417          */
1418         mm->free_area_cache = mm->mmap_base;
1419         mm->cached_hole_size = ~0UL;
1420
1421         return addr;
1422 }
1423 #endif
1424
1425 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1426 {
1427         /*
1428          * Is this a new hole at the highest possible address?
1429          */
1430         if (addr > mm->free_area_cache)
1431                 mm->free_area_cache = addr;
1432
1433         /* dont allow allocations above current base */
1434         if (mm->free_area_cache > mm->mmap_base)
1435                 mm->free_area_cache = mm->mmap_base;
1436 }
1437
1438 unsigned long
1439 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1440                 unsigned long pgoff, unsigned long flags)
1441 {
1442         unsigned long (*get_area)(struct file *, unsigned long,
1443                                   unsigned long, unsigned long, unsigned long);
1444
1445         get_area = current->mm->get_unmapped_area;
1446         if (file && file->f_op && file->f_op->get_unmapped_area)
1447                 get_area = file->f_op->get_unmapped_area;
1448         addr = get_area(file, addr, len, pgoff, flags);
1449         if (IS_ERR_VALUE(addr))
1450                 return addr;
1451
1452         if (addr > TASK_SIZE - len)
1453                 return -ENOMEM;
1454         if (addr & ~PAGE_MASK)
1455                 return -EINVAL;
1456
1457         return arch_rebalance_pgtables(addr, len);
1458 }
1459
1460 EXPORT_SYMBOL(get_unmapped_area);
1461
1462 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1463 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1464 {
1465         struct vm_area_struct *vma = NULL;
1466
1467         if (mm) {
1468                 /* Check the cache first. */
1469                 /* (Cache hit rate is typically around 35%.) */
1470                 vma = mm->mmap_cache;
1471                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1472                         struct rb_node * rb_node;
1473
1474                         rb_node = mm->mm_rb.rb_node;
1475                         vma = NULL;
1476
1477                         while (rb_node) {
1478                                 struct vm_area_struct * vma_tmp;
1479
1480                                 vma_tmp = rb_entry(rb_node,
1481                                                 struct vm_area_struct, vm_rb);
1482
1483                                 if (vma_tmp->vm_end > addr) {
1484                                         vma = vma_tmp;
1485                                         if (vma_tmp->vm_start <= addr)
1486                                                 break;
1487                                         rb_node = rb_node->rb_left;
1488                                 } else
1489                                         rb_node = rb_node->rb_right;
1490                         }
1491                         if (vma)
1492                                 mm->mmap_cache = vma;
1493                 }
1494         }
1495         return vma;
1496 }
1497
1498 EXPORT_SYMBOL(find_vma);
1499
1500 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1501 struct vm_area_struct *
1502 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1503                         struct vm_area_struct **pprev)
1504 {
1505         struct vm_area_struct *vma = NULL, *prev = NULL;
1506         struct rb_node *rb_node;
1507         if (!mm)
1508                 goto out;
1509
1510         /* Guard against addr being lower than the first VMA */
1511         vma = mm->mmap;
1512
1513         /* Go through the RB tree quickly. */
1514         rb_node = mm->mm_rb.rb_node;
1515
1516         while (rb_node) {
1517                 struct vm_area_struct *vma_tmp;
1518                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1519
1520                 if (addr < vma_tmp->vm_end) {
1521                         rb_node = rb_node->rb_left;
1522                 } else {
1523                         prev = vma_tmp;
1524                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1525                                 break;
1526                         rb_node = rb_node->rb_right;
1527                 }
1528         }
1529
1530 out:
1531         *pprev = prev;
1532         return prev ? prev->vm_next : vma;
1533 }
1534
1535 /*
1536  * Verify that the stack growth is acceptable and
1537  * update accounting. This is shared with both the
1538  * grow-up and grow-down cases.
1539  */
1540 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1541 {
1542         struct mm_struct *mm = vma->vm_mm;
1543         struct rlimit *rlim = current->signal->rlim;
1544         unsigned long new_start;
1545
1546         /* address space limit tests */
1547         if (!may_expand_vm(mm, grow))
1548                 return -ENOMEM;
1549
1550         /* Stack limit test */
1551         if (size > rlim[RLIMIT_STACK].rlim_cur)
1552                 return -ENOMEM;
1553
1554         /* mlock limit tests */
1555         if (vma->vm_flags & VM_LOCKED) {
1556                 unsigned long locked;
1557                 unsigned long limit;
1558                 locked = mm->locked_vm + grow;
1559                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1560                 if (locked > limit && !capable(CAP_IPC_LOCK))
1561                         return -ENOMEM;
1562         }
1563
1564         /* Check to ensure the stack will not grow into a hugetlb-only region */
1565         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1566                         vma->vm_end - size;
1567         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1568                 return -EFAULT;
1569
1570         /*
1571          * Overcommit..  This must be the final test, as it will
1572          * update security statistics.
1573          */
1574         if (security_vm_enough_memory(grow))
1575                 return -ENOMEM;
1576
1577         /* Ok, everything looks good - let it rip */
1578         mm->total_vm += grow;
1579         if (vma->vm_flags & VM_LOCKED)
1580                 mm->locked_vm += grow;
1581         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1582         return 0;
1583 }
1584
1585 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1586 /*
1587  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1588  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1589  */
1590 #ifndef CONFIG_IA64
1591 static
1592 #endif
1593 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1594 {
1595         int error;
1596
1597         if (!(vma->vm_flags & VM_GROWSUP))
1598                 return -EFAULT;
1599
1600         /*
1601          * We must make sure the anon_vma is allocated
1602          * so that the anon_vma locking is not a noop.
1603          */
1604         if (unlikely(anon_vma_prepare(vma)))
1605                 return -ENOMEM;
1606         anon_vma_lock(vma);
1607
1608         /*
1609          * vma->vm_start/vm_end cannot change under us because the caller
1610          * is required to hold the mmap_sem in read mode.  We need the
1611          * anon_vma lock to serialize against concurrent expand_stacks.
1612          * Also guard against wrapping around to address 0.
1613          */
1614         if (address < PAGE_ALIGN(address+4))
1615                 address = PAGE_ALIGN(address+4);
1616         else {
1617                 anon_vma_unlock(vma);
1618                 return -ENOMEM;
1619         }
1620         error = 0;
1621
1622         /* Somebody else might have raced and expanded it already */
1623         if (address > vma->vm_end) {
1624                 unsigned long size, grow;
1625
1626                 size = address - vma->vm_start;
1627                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1628
1629                 error = acct_stack_growth(vma, size, grow);
1630                 if (!error)
1631                         vma->vm_end = address;
1632         }
1633         anon_vma_unlock(vma);
1634         return error;
1635 }
1636 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1637
1638 /*
1639  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1640  */
1641 static int expand_downwards(struct vm_area_struct *vma,
1642                                    unsigned long address)
1643 {
1644         int error;
1645
1646         /*
1647          * We must make sure the anon_vma is allocated
1648          * so that the anon_vma locking is not a noop.
1649          */
1650         if (unlikely(anon_vma_prepare(vma)))
1651                 return -ENOMEM;
1652
1653         address &= PAGE_MASK;
1654         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1655         if (error)
1656                 return error;
1657
1658         anon_vma_lock(vma);
1659
1660         /*
1661          * vma->vm_start/vm_end cannot change under us because the caller
1662          * is required to hold the mmap_sem in read mode.  We need the
1663          * anon_vma lock to serialize against concurrent expand_stacks.
1664          */
1665
1666         /* Somebody else might have raced and expanded it already */
1667         if (address < vma->vm_start) {
1668                 unsigned long size, grow;
1669
1670                 size = vma->vm_end - address;
1671                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1672
1673                 error = acct_stack_growth(vma, size, grow);
1674                 if (!error) {
1675                         vma->vm_start = address;
1676                         vma->vm_pgoff -= grow;
1677                 }
1678         }
1679         anon_vma_unlock(vma);
1680         return error;
1681 }
1682
1683 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1684 {
1685         return expand_downwards(vma, address);
1686 }
1687
1688 #ifdef CONFIG_STACK_GROWSUP
1689 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1690 {
1691         return expand_upwards(vma, address);
1692 }
1693
1694 struct vm_area_struct *
1695 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1696 {
1697         struct vm_area_struct *vma, *prev;
1698
1699         addr &= PAGE_MASK;
1700         vma = find_vma_prev(mm, addr, &prev);
1701         if (vma && (vma->vm_start <= addr))
1702                 return vma;
1703         if (!prev || expand_stack(prev, addr))
1704                 return NULL;
1705         if (prev->vm_flags & VM_LOCKED) {
1706                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1707                         return NULL;    /* vma gone! */
1708         }
1709         return prev;
1710 }
1711 #else
1712 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1713 {
1714         return expand_downwards(vma, address);
1715 }
1716
1717 struct vm_area_struct *
1718 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1719 {
1720         struct vm_area_struct * vma;
1721         unsigned long start;
1722
1723         addr &= PAGE_MASK;
1724         vma = find_vma(mm,addr);
1725         if (!vma)
1726                 return NULL;
1727         if (vma->vm_start <= addr)
1728                 return vma;
1729         if (!(vma->vm_flags & VM_GROWSDOWN))
1730                 return NULL;
1731         start = vma->vm_start;
1732         if (expand_stack(vma, addr))
1733                 return NULL;
1734         if (vma->vm_flags & VM_LOCKED) {
1735                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1736                         return NULL;    /* vma gone! */
1737         }
1738         return vma;
1739 }
1740 #endif
1741
1742 /*
1743  * Ok - we have the memory areas we should free on the vma list,
1744  * so release them, and do the vma updates.
1745  *
1746  * Called with the mm semaphore held.
1747  */
1748 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1749 {
1750         /* Update high watermark before we lower total_vm */
1751         update_hiwater_vm(mm);
1752         do {
1753                 long nrpages = vma_pages(vma);
1754
1755                 mm->total_vm -= nrpages;
1756                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1757                 vma = remove_vma(vma);
1758         } while (vma);
1759         validate_mm(mm);
1760 }
1761
1762 /*
1763  * Get rid of page table information in the indicated region.
1764  *
1765  * Called with the mm semaphore held.
1766  */
1767 static void unmap_region(struct mm_struct *mm,
1768                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1769                 unsigned long start, unsigned long end)
1770 {
1771         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1772         struct mmu_gather *tlb;
1773         unsigned long nr_accounted = 0;
1774
1775         lru_add_drain();
1776         tlb = tlb_gather_mmu(mm, 0);
1777         update_hiwater_rss(mm);
1778         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1779         vm_unacct_memory(nr_accounted);
1780         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1781                                  next? next->vm_start: 0);
1782         tlb_finish_mmu(tlb, start, end);
1783 }
1784
1785 /*
1786  * Create a list of vma's touched by the unmap, removing them from the mm's
1787  * vma list as we go..
1788  */
1789 static void
1790 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1791         struct vm_area_struct *prev, unsigned long end)
1792 {
1793         struct vm_area_struct **insertion_point;
1794         struct vm_area_struct *tail_vma = NULL;
1795         unsigned long addr;
1796
1797         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1798         do {
1799                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1800                 mm->map_count--;
1801                 tail_vma = vma;
1802                 vma = vma->vm_next;
1803         } while (vma && vma->vm_start < end);
1804         *insertion_point = vma;
1805         tail_vma->vm_next = NULL;
1806         if (mm->unmap_area == arch_unmap_area)
1807                 addr = prev ? prev->vm_end : mm->mmap_base;
1808         else
1809                 addr = vma ?  vma->vm_start : mm->mmap_base;
1810         mm->unmap_area(mm, addr);
1811         mm->mmap_cache = NULL;          /* Kill the cache. */
1812 }
1813
1814 /*
1815  * Split a vma into two pieces at address 'addr', a new vma is allocated
1816  * either for the first part or the tail.
1817  */
1818 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1819               unsigned long addr, int new_below)
1820 {
1821         struct mempolicy *pol;
1822         struct vm_area_struct *new;
1823
1824         if (is_vm_hugetlb_page(vma) && (addr &
1825                                         ~(huge_page_mask(hstate_vma(vma)))))
1826                 return -EINVAL;
1827
1828         if (mm->map_count >= sysctl_max_map_count)
1829                 return -ENOMEM;
1830
1831         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1832         if (!new)
1833                 return -ENOMEM;
1834
1835         /* most fields are the same, copy all, and then fixup */
1836         *new = *vma;
1837
1838         if (new_below)
1839                 new->vm_end = addr;
1840         else {
1841                 new->vm_start = addr;
1842                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1843         }
1844
1845         pol = mpol_dup(vma_policy(vma));
1846         if (IS_ERR(pol)) {
1847                 kmem_cache_free(vm_area_cachep, new);
1848                 return PTR_ERR(pol);
1849         }
1850         vma_set_policy(new, pol);
1851
1852         if (new->vm_file) {
1853                 get_file(new->vm_file);
1854                 if (vma->vm_flags & VM_EXECUTABLE)
1855                         added_exe_file_vma(mm);
1856         }
1857
1858         if (new->vm_ops && new->vm_ops->open)
1859                 new->vm_ops->open(new);
1860
1861         if (new_below)
1862                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1863                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1864         else
1865                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1866
1867         return 0;
1868 }
1869
1870 /* Munmap is split into 2 main parts -- this part which finds
1871  * what needs doing, and the areas themselves, which do the
1872  * work.  This now handles partial unmappings.
1873  * Jeremy Fitzhardinge <jeremy@goop.org>
1874  */
1875 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1876 {
1877         unsigned long end;
1878         struct vm_area_struct *vma, *prev, *last;
1879
1880         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1881                 return -EINVAL;
1882
1883         if ((len = PAGE_ALIGN(len)) == 0)
1884                 return -EINVAL;
1885
1886         /* Find the first overlapping VMA */
1887         vma = find_vma_prev(mm, start, &prev);
1888         if (!vma)
1889                 return 0;
1890         /* we have  start < vma->vm_end  */
1891
1892         /* if it doesn't overlap, we have nothing.. */
1893         end = start + len;
1894         if (vma->vm_start >= end)
1895                 return 0;
1896
1897         /*
1898          * If we need to split any vma, do it now to save pain later.
1899          *
1900          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1901          * unmapped vm_area_struct will remain in use: so lower split_vma
1902          * places tmp vma above, and higher split_vma places tmp vma below.
1903          */
1904         if (start > vma->vm_start) {
1905                 int error = split_vma(mm, vma, start, 0);
1906                 if (error)
1907                         return error;
1908                 prev = vma;
1909         }
1910
1911         /* Does it split the last one? */
1912         last = find_vma(mm, end);
1913         if (last && end > last->vm_start) {
1914                 int error = split_vma(mm, last, end, 1);
1915                 if (error)
1916                         return error;
1917         }
1918         vma = prev? prev->vm_next: mm->mmap;
1919
1920         /*
1921          * unlock any mlock()ed ranges before detaching vmas
1922          */
1923         if (mm->locked_vm) {
1924                 struct vm_area_struct *tmp = vma;
1925                 while (tmp && tmp->vm_start < end) {
1926                         if (tmp->vm_flags & VM_LOCKED) {
1927                                 mm->locked_vm -= vma_pages(tmp);
1928                                 munlock_vma_pages_all(tmp);
1929                         }
1930                         tmp = tmp->vm_next;
1931                 }
1932         }
1933
1934         /*
1935          * Remove the vma's, and unmap the actual pages
1936          */
1937         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1938         unmap_region(mm, vma, prev, start, end);
1939
1940         /* Fix up all other VM information */
1941         remove_vma_list(mm, vma);
1942
1943         return 0;
1944 }
1945
1946 EXPORT_SYMBOL(do_munmap);
1947
1948 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1949 {
1950         int ret;
1951         struct mm_struct *mm = current->mm;
1952
1953         profile_munmap(addr);
1954
1955         down_write(&mm->mmap_sem);
1956         ret = do_munmap(mm, addr, len);
1957         up_write(&mm->mmap_sem);
1958         return ret;
1959 }
1960
1961 static inline void verify_mm_writelocked(struct mm_struct *mm)
1962 {
1963 #ifdef CONFIG_DEBUG_VM
1964         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1965                 WARN_ON(1);
1966                 up_read(&mm->mmap_sem);
1967         }
1968 #endif
1969 }
1970
1971 /*
1972  *  this is really a simplified "do_mmap".  it only handles
1973  *  anonymous maps.  eventually we may be able to do some
1974  *  brk-specific accounting here.
1975  */
1976 unsigned long do_brk(unsigned long addr, unsigned long len)
1977 {
1978         struct mm_struct * mm = current->mm;
1979         struct vm_area_struct * vma, * prev;
1980         unsigned long flags;
1981         struct rb_node ** rb_link, * rb_parent;
1982         pgoff_t pgoff = addr >> PAGE_SHIFT;
1983         int error;
1984
1985         len = PAGE_ALIGN(len);
1986         if (!len)
1987                 return addr;
1988
1989         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1990                 return -EINVAL;
1991
1992         if (is_hugepage_only_range(mm, addr, len))
1993                 return -EINVAL;
1994
1995         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1996         if (error)
1997                 return error;
1998
1999         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2000
2001         error = arch_mmap_check(addr, len, flags);
2002         if (error)
2003                 return error;
2004
2005         /*
2006          * mlock MCL_FUTURE?
2007          */
2008         if (mm->def_flags & VM_LOCKED) {
2009                 unsigned long locked, lock_limit;
2010                 locked = len >> PAGE_SHIFT;
2011                 locked += mm->locked_vm;
2012                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2013                 lock_limit >>= PAGE_SHIFT;
2014                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2015                         return -EAGAIN;
2016         }
2017
2018         /*
2019          * mm->mmap_sem is required to protect against another thread
2020          * changing the mappings in case we sleep.
2021          */
2022         verify_mm_writelocked(mm);
2023
2024         /*
2025          * Clear old maps.  this also does some error checking for us
2026          */
2027  munmap_back:
2028         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2029         if (vma && vma->vm_start < addr + len) {
2030                 if (do_munmap(mm, addr, len))
2031                         return -ENOMEM;
2032                 goto munmap_back;
2033         }
2034
2035         /* Check against address space limits *after* clearing old maps... */
2036         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2037                 return -ENOMEM;
2038
2039         if (mm->map_count > sysctl_max_map_count)
2040                 return -ENOMEM;
2041
2042         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2043                 return -ENOMEM;
2044
2045         /* Can we just expand an old private anonymous mapping? */
2046         vma = vma_merge(mm, prev, addr, addr + len, flags,
2047                                         NULL, NULL, pgoff, NULL);
2048         if (vma)
2049                 goto out;
2050
2051         /*
2052          * create a vma struct for an anonymous mapping
2053          */
2054         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2055         if (!vma) {
2056                 vm_unacct_memory(len >> PAGE_SHIFT);
2057                 return -ENOMEM;
2058         }
2059
2060         vma->vm_mm = mm;
2061         vma->vm_start = addr;
2062         vma->vm_end = addr + len;
2063         vma->vm_pgoff = pgoff;
2064         vma->vm_flags = flags;
2065         vma->vm_page_prot = vm_get_page_prot(flags);
2066         vma_link(mm, vma, prev, rb_link, rb_parent);
2067 out:
2068         mm->total_vm += len >> PAGE_SHIFT;
2069         if (flags & VM_LOCKED) {
2070                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2071                         mm->locked_vm += (len >> PAGE_SHIFT);
2072         }
2073         return addr;
2074 }
2075
2076 EXPORT_SYMBOL(do_brk);
2077
2078 /* Release all mmaps. */
2079 void exit_mmap(struct mm_struct *mm)
2080 {
2081         struct mmu_gather *tlb;
2082         struct vm_area_struct *vma;
2083         unsigned long nr_accounted = 0;
2084         unsigned long end;
2085
2086         /* mm's last user has gone, and its about to be pulled down */
2087         mmu_notifier_release(mm);
2088
2089         if (mm->locked_vm) {
2090                 vma = mm->mmap;
2091                 while (vma) {
2092                         if (vma->vm_flags & VM_LOCKED)
2093                                 munlock_vma_pages_all(vma);
2094                         vma = vma->vm_next;
2095                 }
2096         }
2097
2098         arch_exit_mmap(mm);
2099
2100         vma = mm->mmap;
2101         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2102                 return;
2103
2104         lru_add_drain();
2105         flush_cache_mm(mm);
2106         tlb = tlb_gather_mmu(mm, 1);
2107         /* update_hiwater_rss(mm) here? but nobody should be looking */
2108         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2109         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2110         vm_unacct_memory(nr_accounted);
2111         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2112         tlb_finish_mmu(tlb, 0, end);
2113
2114         /*
2115          * Walk the list again, actually closing and freeing it,
2116          * with preemption enabled, without holding any MM locks.
2117          */
2118         while (vma)
2119                 vma = remove_vma(vma);
2120
2121         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2122 }
2123
2124 /* Insert vm structure into process list sorted by address
2125  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2126  * then i_mmap_lock is taken here.
2127  */
2128 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2129 {
2130         struct vm_area_struct * __vma, * prev;
2131         struct rb_node ** rb_link, * rb_parent;
2132
2133         /*
2134          * The vm_pgoff of a purely anonymous vma should be irrelevant
2135          * until its first write fault, when page's anon_vma and index
2136          * are set.  But now set the vm_pgoff it will almost certainly
2137          * end up with (unless mremap moves it elsewhere before that
2138          * first wfault), so /proc/pid/maps tells a consistent story.
2139          *
2140          * By setting it to reflect the virtual start address of the
2141          * vma, merges and splits can happen in a seamless way, just
2142          * using the existing file pgoff checks and manipulations.
2143          * Similarly in do_mmap_pgoff and in do_brk.
2144          */
2145         if (!vma->vm_file) {
2146                 BUG_ON(vma->anon_vma);
2147                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2148         }
2149         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2150         if (__vma && __vma->vm_start < vma->vm_end)
2151                 return -ENOMEM;
2152         if ((vma->vm_flags & VM_ACCOUNT) &&
2153              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2154                 return -ENOMEM;
2155         vma_link(mm, vma, prev, rb_link, rb_parent);
2156         return 0;
2157 }
2158
2159 /*
2160  * Copy the vma structure to a new location in the same mm,
2161  * prior to moving page table entries, to effect an mremap move.
2162  */
2163 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2164         unsigned long addr, unsigned long len, pgoff_t pgoff)
2165 {
2166         struct vm_area_struct *vma = *vmap;
2167         unsigned long vma_start = vma->vm_start;
2168         struct mm_struct *mm = vma->vm_mm;
2169         struct vm_area_struct *new_vma, *prev;
2170         struct rb_node **rb_link, *rb_parent;
2171         struct mempolicy *pol;
2172
2173         /*
2174          * If anonymous vma has not yet been faulted, update new pgoff
2175          * to match new location, to increase its chance of merging.
2176          */
2177         if (!vma->vm_file && !vma->anon_vma)
2178                 pgoff = addr >> PAGE_SHIFT;
2179
2180         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2181         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2182                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2183         if (new_vma) {
2184                 /*
2185                  * Source vma may have been merged into new_vma
2186                  */
2187                 if (vma_start >= new_vma->vm_start &&
2188                     vma_start < new_vma->vm_end)
2189                         *vmap = new_vma;
2190         } else {
2191                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2192                 if (new_vma) {
2193                         *new_vma = *vma;
2194                         pol = mpol_dup(vma_policy(vma));
2195                         if (IS_ERR(pol)) {
2196                                 kmem_cache_free(vm_area_cachep, new_vma);
2197                                 return NULL;
2198                         }
2199                         vma_set_policy(new_vma, pol);
2200                         new_vma->vm_start = addr;
2201                         new_vma->vm_end = addr + len;
2202                         new_vma->vm_pgoff = pgoff;
2203                         if (new_vma->vm_file) {
2204                                 get_file(new_vma->vm_file);
2205                                 if (vma->vm_flags & VM_EXECUTABLE)
2206                                         added_exe_file_vma(mm);
2207                         }
2208                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2209                                 new_vma->vm_ops->open(new_vma);
2210                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2211                 }
2212         }
2213         return new_vma;
2214 }
2215
2216 /*
2217  * Return true if the calling process may expand its vm space by the passed
2218  * number of pages
2219  */
2220 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2221 {
2222         unsigned long cur = mm->total_vm;       /* pages */
2223         unsigned long lim;
2224
2225         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2226
2227         if (cur + npages > lim)
2228                 return 0;
2229         return 1;
2230 }
2231
2232
2233 static int special_mapping_fault(struct vm_area_struct *vma,
2234                                 struct vm_fault *vmf)
2235 {
2236         pgoff_t pgoff;
2237         struct page **pages;
2238
2239         /*
2240          * special mappings have no vm_file, and in that case, the mm
2241          * uses vm_pgoff internally. So we have to subtract it from here.
2242          * We are allowed to do this because we are the mm; do not copy
2243          * this code into drivers!
2244          */
2245         pgoff = vmf->pgoff - vma->vm_pgoff;
2246
2247         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2248                 pgoff--;
2249
2250         if (*pages) {
2251                 struct page *page = *pages;
2252                 get_page(page);
2253                 vmf->page = page;
2254                 return 0;
2255         }
2256
2257         return VM_FAULT_SIGBUS;
2258 }
2259
2260 /*
2261  * Having a close hook prevents vma merging regardless of flags.
2262  */
2263 static void special_mapping_close(struct vm_area_struct *vma)
2264 {
2265 }
2266
2267 static struct vm_operations_struct special_mapping_vmops = {
2268         .close = special_mapping_close,
2269         .fault = special_mapping_fault,
2270 };
2271
2272 /*
2273  * Called with mm->mmap_sem held for writing.
2274  * Insert a new vma covering the given region, with the given flags.
2275  * Its pages are supplied by the given array of struct page *.
2276  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2277  * The region past the last page supplied will always produce SIGBUS.
2278  * The array pointer and the pages it points to are assumed to stay alive
2279  * for as long as this mapping might exist.
2280  */
2281 int install_special_mapping(struct mm_struct *mm,
2282                             unsigned long addr, unsigned long len,
2283                             unsigned long vm_flags, struct page **pages)
2284 {
2285         struct vm_area_struct *vma;
2286
2287         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2288         if (unlikely(vma == NULL))
2289                 return -ENOMEM;
2290
2291         vma->vm_mm = mm;
2292         vma->vm_start = addr;
2293         vma->vm_end = addr + len;
2294
2295         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2296         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2297
2298         vma->vm_ops = &special_mapping_vmops;
2299         vma->vm_private_data = pages;
2300
2301         if (unlikely(insert_vm_struct(mm, vma))) {
2302                 kmem_cache_free(vm_area_cachep, vma);
2303                 return -ENOMEM;
2304         }
2305
2306         mm->total_vm += len >> PAGE_SHIFT;
2307
2308         return 0;
2309 }
2310
2311 static DEFINE_MUTEX(mm_all_locks_mutex);
2312
2313 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2314 {
2315         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2316                 /*
2317                  * The LSB of head.next can't change from under us
2318                  * because we hold the mm_all_locks_mutex.
2319                  */
2320                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2321                 /*
2322                  * We can safely modify head.next after taking the
2323                  * anon_vma->lock. If some other vma in this mm shares
2324                  * the same anon_vma we won't take it again.
2325                  *
2326                  * No need of atomic instructions here, head.next
2327                  * can't change from under us thanks to the
2328                  * anon_vma->lock.
2329                  */
2330                 if (__test_and_set_bit(0, (unsigned long *)
2331                                        &anon_vma->head.next))
2332                         BUG();
2333         }
2334 }
2335
2336 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2337 {
2338         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2339                 /*
2340                  * AS_MM_ALL_LOCKS can't change from under us because
2341                  * we hold the mm_all_locks_mutex.
2342                  *
2343                  * Operations on ->flags have to be atomic because
2344                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2345                  * mm_all_locks_mutex, there may be other cpus
2346                  * changing other bitflags in parallel to us.
2347                  */
2348                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2349                         BUG();
2350                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2351         }
2352 }
2353
2354 /*
2355  * This operation locks against the VM for all pte/vma/mm related
2356  * operations that could ever happen on a certain mm. This includes
2357  * vmtruncate, try_to_unmap, and all page faults.
2358  *
2359  * The caller must take the mmap_sem in write mode before calling
2360  * mm_take_all_locks(). The caller isn't allowed to release the
2361  * mmap_sem until mm_drop_all_locks() returns.
2362  *
2363  * mmap_sem in write mode is required in order to block all operations
2364  * that could modify pagetables and free pages without need of
2365  * altering the vma layout (for example populate_range() with
2366  * nonlinear vmas). It's also needed in write mode to avoid new
2367  * anon_vmas to be associated with existing vmas.
2368  *
2369  * A single task can't take more than one mm_take_all_locks() in a row
2370  * or it would deadlock.
2371  *
2372  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2373  * mapping->flags avoid to take the same lock twice, if more than one
2374  * vma in this mm is backed by the same anon_vma or address_space.
2375  *
2376  * We can take all the locks in random order because the VM code
2377  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2378  * takes more than one of them in a row. Secondly we're protected
2379  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2380  *
2381  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2382  * that may have to take thousand of locks.
2383  *
2384  * mm_take_all_locks() can fail if it's interrupted by signals.
2385  */
2386 int mm_take_all_locks(struct mm_struct *mm)
2387 {
2388         struct vm_area_struct *vma;
2389         int ret = -EINTR;
2390
2391         BUG_ON(down_read_trylock(&mm->mmap_sem));
2392
2393         mutex_lock(&mm_all_locks_mutex);
2394
2395         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2396                 if (signal_pending(current))
2397                         goto out_unlock;
2398                 if (vma->vm_file && vma->vm_file->f_mapping)
2399                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2400         }
2401
2402         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2403                 if (signal_pending(current))
2404                         goto out_unlock;
2405                 if (vma->anon_vma)
2406                         vm_lock_anon_vma(mm, vma->anon_vma);
2407         }
2408
2409         ret = 0;
2410
2411 out_unlock:
2412         if (ret)
2413                 mm_drop_all_locks(mm);
2414
2415         return ret;
2416 }
2417
2418 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2419 {
2420         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2421                 /*
2422                  * The LSB of head.next can't change to 0 from under
2423                  * us because we hold the mm_all_locks_mutex.
2424                  *
2425                  * We must however clear the bitflag before unlocking
2426                  * the vma so the users using the anon_vma->head will
2427                  * never see our bitflag.
2428                  *
2429                  * No need of atomic instructions here, head.next
2430                  * can't change from under us until we release the
2431                  * anon_vma->lock.
2432                  */
2433                 if (!__test_and_clear_bit(0, (unsigned long *)
2434                                           &anon_vma->head.next))
2435                         BUG();
2436                 spin_unlock(&anon_vma->lock);
2437         }
2438 }
2439
2440 static void vm_unlock_mapping(struct address_space *mapping)
2441 {
2442         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2443                 /*
2444                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2445                  * because we hold the mm_all_locks_mutex.
2446                  */
2447                 spin_unlock(&mapping->i_mmap_lock);
2448                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2449                                         &mapping->flags))
2450                         BUG();
2451         }
2452 }
2453
2454 /*
2455  * The mmap_sem cannot be released by the caller until
2456  * mm_drop_all_locks() returns.
2457  */
2458 void mm_drop_all_locks(struct mm_struct *mm)
2459 {
2460         struct vm_area_struct *vma;
2461
2462         BUG_ON(down_read_trylock(&mm->mmap_sem));
2463         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2464
2465         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2466                 if (vma->anon_vma)
2467                         vm_unlock_anon_vma(vma->anon_vma);
2468                 if (vma->vm_file && vma->vm_file->f_mapping)
2469                         vm_unlock_mapping(vma->vm_file->f_mapping);
2470         }
2471
2472         mutex_unlock(&mm_all_locks_mutex);
2473 }
2474
2475 /*
2476  * initialise the VMA slab
2477  */
2478 void __init mmap_init(void)
2479 {
2480         vm_area_cachep = kmem_cache_create("vm_area_struct",
2481                         sizeof(struct vm_area_struct), 0,
2482                         SLAB_PANIC, NULL);
2483 }