vm audit: add VM_DONTEXPAND to mmap for drivers that need it
[sfrench/cifs-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@redhat.com>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29
30 #include <asm/uaccess.h>
31 #include <asm/cacheflush.h>
32 #include <asm/tlb.h>
33 #include <asm/mmu_context.h>
34
35 #ifndef arch_mmap_check
36 #define arch_mmap_check(addr, len, flags)       (0)
37 #endif
38
39 static void unmap_region(struct mm_struct *mm,
40                 struct vm_area_struct *vma, struct vm_area_struct *prev,
41                 unsigned long start, unsigned long end);
42
43 /*
44  * WARNING: the debugging will use recursive algorithms so never enable this
45  * unless you know what you are doing.
46  */
47 #undef DEBUG_MM_RB
48
49 /* description of effects of mapping type and prot in current implementation.
50  * this is due to the limited x86 page protection hardware.  The expected
51  * behavior is in parens:
52  *
53  * map_type     prot
54  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
55  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
56  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
57  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
58  *              
59  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
60  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
61  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
62  *
63  */
64 pgprot_t protection_map[16] = {
65         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
66         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
67 };
68
69 pgprot_t vm_get_page_prot(unsigned long vm_flags)
70 {
71         return protection_map[vm_flags &
72                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
73 }
74 EXPORT_SYMBOL(vm_get_page_prot);
75
76 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
77 int sysctl_overcommit_ratio = 50;       /* default is 50% */
78 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
79 atomic_t vm_committed_space = ATOMIC_INIT(0);
80
81 /*
82  * Check that a process has enough memory to allocate a new virtual
83  * mapping. 0 means there is enough memory for the allocation to
84  * succeed and -ENOMEM implies there is not.
85  *
86  * We currently support three overcommit policies, which are set via the
87  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
88  *
89  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
90  * Additional code 2002 Jul 20 by Robert Love.
91  *
92  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
93  *
94  * Note this is a helper function intended to be used by LSMs which
95  * wish to use this logic.
96  */
97 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
98 {
99         unsigned long free, allowed;
100
101         vm_acct_memory(pages);
102
103         /*
104          * Sometimes we want to use more memory than we have
105          */
106         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
107                 return 0;
108
109         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
110                 unsigned long n;
111
112                 free = global_page_state(NR_FILE_PAGES);
113                 free += nr_swap_pages;
114
115                 /*
116                  * Any slabs which are created with the
117                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
118                  * which are reclaimable, under pressure.  The dentry
119                  * cache and most inode caches should fall into this
120                  */
121                 free += global_page_state(NR_SLAB_RECLAIMABLE);
122
123                 /*
124                  * Leave the last 3% for root
125                  */
126                 if (!cap_sys_admin)
127                         free -= free / 32;
128
129                 if (free > pages)
130                         return 0;
131
132                 /*
133                  * nr_free_pages() is very expensive on large systems,
134                  * only call if we're about to fail.
135                  */
136                 n = nr_free_pages();
137
138                 /*
139                  * Leave reserved pages. The pages are not for anonymous pages.
140                  */
141                 if (n <= totalreserve_pages)
142                         goto error;
143                 else
144                         n -= totalreserve_pages;
145
146                 /*
147                  * Leave the last 3% for root
148                  */
149                 if (!cap_sys_admin)
150                         n -= n / 32;
151                 free += n;
152
153                 if (free > pages)
154                         return 0;
155
156                 goto error;
157         }
158
159         allowed = (totalram_pages - hugetlb_total_pages())
160                 * sysctl_overcommit_ratio / 100;
161         /*
162          * Leave the last 3% for root
163          */
164         if (!cap_sys_admin)
165                 allowed -= allowed / 32;
166         allowed += total_swap_pages;
167
168         /* Don't let a single process grow too big:
169            leave 3% of the size of this process for other processes */
170         allowed -= mm->total_vm / 32;
171
172         /*
173          * cast `allowed' as a signed long because vm_committed_space
174          * sometimes has a negative value
175          */
176         if (atomic_read(&vm_committed_space) < (long)allowed)
177                 return 0;
178 error:
179         vm_unacct_memory(pages);
180
181         return -ENOMEM;
182 }
183
184 /*
185  * Requires inode->i_mapping->i_mmap_lock
186  */
187 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
188                 struct file *file, struct address_space *mapping)
189 {
190         if (vma->vm_flags & VM_DENYWRITE)
191                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
192         if (vma->vm_flags & VM_SHARED)
193                 mapping->i_mmap_writable--;
194
195         flush_dcache_mmap_lock(mapping);
196         if (unlikely(vma->vm_flags & VM_NONLINEAR))
197                 list_del_init(&vma->shared.vm_set.list);
198         else
199                 vma_prio_tree_remove(vma, &mapping->i_mmap);
200         flush_dcache_mmap_unlock(mapping);
201 }
202
203 /*
204  * Unlink a file-based vm structure from its prio_tree, to hide
205  * vma from rmap and vmtruncate before freeing its page tables.
206  */
207 void unlink_file_vma(struct vm_area_struct *vma)
208 {
209         struct file *file = vma->vm_file;
210
211         if (file) {
212                 struct address_space *mapping = file->f_mapping;
213                 spin_lock(&mapping->i_mmap_lock);
214                 __remove_shared_vm_struct(vma, file, mapping);
215                 spin_unlock(&mapping->i_mmap_lock);
216         }
217 }
218
219 /*
220  * Close a vm structure and free it, returning the next.
221  */
222 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
223 {
224         struct vm_area_struct *next = vma->vm_next;
225
226         might_sleep();
227         if (vma->vm_ops && vma->vm_ops->close)
228                 vma->vm_ops->close(vma);
229         if (vma->vm_file)
230                 fput(vma->vm_file);
231         mpol_free(vma_policy(vma));
232         kmem_cache_free(vm_area_cachep, vma);
233         return next;
234 }
235
236 asmlinkage unsigned long sys_brk(unsigned long brk)
237 {
238         unsigned long rlim, retval;
239         unsigned long newbrk, oldbrk;
240         struct mm_struct *mm = current->mm;
241
242         down_write(&mm->mmap_sem);
243
244         if (brk < mm->end_code)
245                 goto out;
246
247         /*
248          * Check against rlimit here. If this check is done later after the test
249          * of oldbrk with newbrk then it can escape the test and let the data
250          * segment grow beyond its set limit the in case where the limit is
251          * not page aligned -Ram Gupta
252          */
253         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
254         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
255                         (mm->end_data - mm->start_data) > rlim)
256                 goto out;
257
258         newbrk = PAGE_ALIGN(brk);
259         oldbrk = PAGE_ALIGN(mm->brk);
260         if (oldbrk == newbrk)
261                 goto set_brk;
262
263         /* Always allow shrinking brk. */
264         if (brk <= mm->brk) {
265                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
266                         goto set_brk;
267                 goto out;
268         }
269
270         /* Check against existing mmap mappings. */
271         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
272                 goto out;
273
274         /* Ok, looks good - let it rip. */
275         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
276                 goto out;
277 set_brk:
278         mm->brk = brk;
279 out:
280         retval = mm->brk;
281         up_write(&mm->mmap_sem);
282         return retval;
283 }
284
285 #ifdef DEBUG_MM_RB
286 static int browse_rb(struct rb_root *root)
287 {
288         int i = 0, j;
289         struct rb_node *nd, *pn = NULL;
290         unsigned long prev = 0, pend = 0;
291
292         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
293                 struct vm_area_struct *vma;
294                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
295                 if (vma->vm_start < prev)
296                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
297                 if (vma->vm_start < pend)
298                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
299                 if (vma->vm_start > vma->vm_end)
300                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
301                 i++;
302                 pn = nd;
303                 prev = vma->vm_start;
304                 pend = vma->vm_end;
305         }
306         j = 0;
307         for (nd = pn; nd; nd = rb_prev(nd)) {
308                 j++;
309         }
310         if (i != j)
311                 printk("backwards %d, forwards %d\n", j, i), i = 0;
312         return i;
313 }
314
315 void validate_mm(struct mm_struct *mm)
316 {
317         int bug = 0;
318         int i = 0;
319         struct vm_area_struct *tmp = mm->mmap;
320         while (tmp) {
321                 tmp = tmp->vm_next;
322                 i++;
323         }
324         if (i != mm->map_count)
325                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
326         i = browse_rb(&mm->mm_rb);
327         if (i != mm->map_count)
328                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
329         BUG_ON(bug);
330 }
331 #else
332 #define validate_mm(mm) do { } while (0)
333 #endif
334
335 static struct vm_area_struct *
336 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
337                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
338                 struct rb_node ** rb_parent)
339 {
340         struct vm_area_struct * vma;
341         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
342
343         __rb_link = &mm->mm_rb.rb_node;
344         rb_prev = __rb_parent = NULL;
345         vma = NULL;
346
347         while (*__rb_link) {
348                 struct vm_area_struct *vma_tmp;
349
350                 __rb_parent = *__rb_link;
351                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
352
353                 if (vma_tmp->vm_end > addr) {
354                         vma = vma_tmp;
355                         if (vma_tmp->vm_start <= addr)
356                                 return vma;
357                         __rb_link = &__rb_parent->rb_left;
358                 } else {
359                         rb_prev = __rb_parent;
360                         __rb_link = &__rb_parent->rb_right;
361                 }
362         }
363
364         *pprev = NULL;
365         if (rb_prev)
366                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
367         *rb_link = __rb_link;
368         *rb_parent = __rb_parent;
369         return vma;
370 }
371
372 static inline void
373 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
374                 struct vm_area_struct *prev, struct rb_node *rb_parent)
375 {
376         if (prev) {
377                 vma->vm_next = prev->vm_next;
378                 prev->vm_next = vma;
379         } else {
380                 mm->mmap = vma;
381                 if (rb_parent)
382                         vma->vm_next = rb_entry(rb_parent,
383                                         struct vm_area_struct, vm_rb);
384                 else
385                         vma->vm_next = NULL;
386         }
387 }
388
389 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
390                 struct rb_node **rb_link, struct rb_node *rb_parent)
391 {
392         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
393         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
394 }
395
396 static inline void __vma_link_file(struct vm_area_struct *vma)
397 {
398         struct file * file;
399
400         file = vma->vm_file;
401         if (file) {
402                 struct address_space *mapping = file->f_mapping;
403
404                 if (vma->vm_flags & VM_DENYWRITE)
405                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
406                 if (vma->vm_flags & VM_SHARED)
407                         mapping->i_mmap_writable++;
408
409                 flush_dcache_mmap_lock(mapping);
410                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
411                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
412                 else
413                         vma_prio_tree_insert(vma, &mapping->i_mmap);
414                 flush_dcache_mmap_unlock(mapping);
415         }
416 }
417
418 static void
419 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
420         struct vm_area_struct *prev, struct rb_node **rb_link,
421         struct rb_node *rb_parent)
422 {
423         __vma_link_list(mm, vma, prev, rb_parent);
424         __vma_link_rb(mm, vma, rb_link, rb_parent);
425         __anon_vma_link(vma);
426 }
427
428 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
429                         struct vm_area_struct *prev, struct rb_node **rb_link,
430                         struct rb_node *rb_parent)
431 {
432         struct address_space *mapping = NULL;
433
434         if (vma->vm_file)
435                 mapping = vma->vm_file->f_mapping;
436
437         if (mapping) {
438                 spin_lock(&mapping->i_mmap_lock);
439                 vma->vm_truncate_count = mapping->truncate_count;
440         }
441         anon_vma_lock(vma);
442
443         __vma_link(mm, vma, prev, rb_link, rb_parent);
444         __vma_link_file(vma);
445
446         anon_vma_unlock(vma);
447         if (mapping)
448                 spin_unlock(&mapping->i_mmap_lock);
449
450         mm->map_count++;
451         validate_mm(mm);
452 }
453
454 /*
455  * Helper for vma_adjust in the split_vma insert case:
456  * insert vm structure into list and rbtree and anon_vma,
457  * but it has already been inserted into prio_tree earlier.
458  */
459 static void
460 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
461 {
462         struct vm_area_struct * __vma, * prev;
463         struct rb_node ** rb_link, * rb_parent;
464
465         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
466         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
467         __vma_link(mm, vma, prev, rb_link, rb_parent);
468         mm->map_count++;
469 }
470
471 static inline void
472 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
473                 struct vm_area_struct *prev)
474 {
475         prev->vm_next = vma->vm_next;
476         rb_erase(&vma->vm_rb, &mm->mm_rb);
477         if (mm->mmap_cache == vma)
478                 mm->mmap_cache = prev;
479 }
480
481 /*
482  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
483  * is already present in an i_mmap tree without adjusting the tree.
484  * The following helper function should be used when such adjustments
485  * are necessary.  The "insert" vma (if any) is to be inserted
486  * before we drop the necessary locks.
487  */
488 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
489         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
490 {
491         struct mm_struct *mm = vma->vm_mm;
492         struct vm_area_struct *next = vma->vm_next;
493         struct vm_area_struct *importer = NULL;
494         struct address_space *mapping = NULL;
495         struct prio_tree_root *root = NULL;
496         struct file *file = vma->vm_file;
497         struct anon_vma *anon_vma = NULL;
498         long adjust_next = 0;
499         int remove_next = 0;
500
501         if (next && !insert) {
502                 if (end >= next->vm_end) {
503                         /*
504                          * vma expands, overlapping all the next, and
505                          * perhaps the one after too (mprotect case 6).
506                          */
507 again:                  remove_next = 1 + (end > next->vm_end);
508                         end = next->vm_end;
509                         anon_vma = next->anon_vma;
510                         importer = vma;
511                 } else if (end > next->vm_start) {
512                         /*
513                          * vma expands, overlapping part of the next:
514                          * mprotect case 5 shifting the boundary up.
515                          */
516                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
517                         anon_vma = next->anon_vma;
518                         importer = vma;
519                 } else if (end < vma->vm_end) {
520                         /*
521                          * vma shrinks, and !insert tells it's not
522                          * split_vma inserting another: so it must be
523                          * mprotect case 4 shifting the boundary down.
524                          */
525                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
526                         anon_vma = next->anon_vma;
527                         importer = next;
528                 }
529         }
530
531         if (file) {
532                 mapping = file->f_mapping;
533                 if (!(vma->vm_flags & VM_NONLINEAR))
534                         root = &mapping->i_mmap;
535                 spin_lock(&mapping->i_mmap_lock);
536                 if (importer &&
537                     vma->vm_truncate_count != next->vm_truncate_count) {
538                         /*
539                          * unmap_mapping_range might be in progress:
540                          * ensure that the expanding vma is rescanned.
541                          */
542                         importer->vm_truncate_count = 0;
543                 }
544                 if (insert) {
545                         insert->vm_truncate_count = vma->vm_truncate_count;
546                         /*
547                          * Put into prio_tree now, so instantiated pages
548                          * are visible to arm/parisc __flush_dcache_page
549                          * throughout; but we cannot insert into address
550                          * space until vma start or end is updated.
551                          */
552                         __vma_link_file(insert);
553                 }
554         }
555
556         /*
557          * When changing only vma->vm_end, we don't really need
558          * anon_vma lock: but is that case worth optimizing out?
559          */
560         if (vma->anon_vma)
561                 anon_vma = vma->anon_vma;
562         if (anon_vma) {
563                 spin_lock(&anon_vma->lock);
564                 /*
565                  * Easily overlooked: when mprotect shifts the boundary,
566                  * make sure the expanding vma has anon_vma set if the
567                  * shrinking vma had, to cover any anon pages imported.
568                  */
569                 if (importer && !importer->anon_vma) {
570                         importer->anon_vma = anon_vma;
571                         __anon_vma_link(importer);
572                 }
573         }
574
575         if (root) {
576                 flush_dcache_mmap_lock(mapping);
577                 vma_prio_tree_remove(vma, root);
578                 if (adjust_next)
579                         vma_prio_tree_remove(next, root);
580         }
581
582         vma->vm_start = start;
583         vma->vm_end = end;
584         vma->vm_pgoff = pgoff;
585         if (adjust_next) {
586                 next->vm_start += adjust_next << PAGE_SHIFT;
587                 next->vm_pgoff += adjust_next;
588         }
589
590         if (root) {
591                 if (adjust_next)
592                         vma_prio_tree_insert(next, root);
593                 vma_prio_tree_insert(vma, root);
594                 flush_dcache_mmap_unlock(mapping);
595         }
596
597         if (remove_next) {
598                 /*
599                  * vma_merge has merged next into vma, and needs
600                  * us to remove next before dropping the locks.
601                  */
602                 __vma_unlink(mm, next, vma);
603                 if (file)
604                         __remove_shared_vm_struct(next, file, mapping);
605                 if (next->anon_vma)
606                         __anon_vma_merge(vma, next);
607         } else if (insert) {
608                 /*
609                  * split_vma has split insert from vma, and needs
610                  * us to insert it before dropping the locks
611                  * (it may either follow vma or precede it).
612                  */
613                 __insert_vm_struct(mm, insert);
614         }
615
616         if (anon_vma)
617                 spin_unlock(&anon_vma->lock);
618         if (mapping)
619                 spin_unlock(&mapping->i_mmap_lock);
620
621         if (remove_next) {
622                 if (file)
623                         fput(file);
624                 mm->map_count--;
625                 mpol_free(vma_policy(next));
626                 kmem_cache_free(vm_area_cachep, next);
627                 /*
628                  * In mprotect's case 6 (see comments on vma_merge),
629                  * we must remove another next too. It would clutter
630                  * up the code too much to do both in one go.
631                  */
632                 if (remove_next == 2) {
633                         next = vma->vm_next;
634                         goto again;
635                 }
636         }
637
638         validate_mm(mm);
639 }
640
641 /*
642  * If the vma has a ->close operation then the driver probably needs to release
643  * per-vma resources, so we don't attempt to merge those.
644  */
645 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
646
647 static inline int is_mergeable_vma(struct vm_area_struct *vma,
648                         struct file *file, unsigned long vm_flags)
649 {
650         if (vma->vm_flags != vm_flags)
651                 return 0;
652         if (vma->vm_file != file)
653                 return 0;
654         if (vma->vm_ops && vma->vm_ops->close)
655                 return 0;
656         return 1;
657 }
658
659 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
660                                         struct anon_vma *anon_vma2)
661 {
662         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
663 }
664
665 /*
666  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
667  * in front of (at a lower virtual address and file offset than) the vma.
668  *
669  * We cannot merge two vmas if they have differently assigned (non-NULL)
670  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
671  *
672  * We don't check here for the merged mmap wrapping around the end of pagecache
673  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
674  * wrap, nor mmaps which cover the final page at index -1UL.
675  */
676 static int
677 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
678         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
679 {
680         if (is_mergeable_vma(vma, file, vm_flags) &&
681             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
682                 if (vma->vm_pgoff == vm_pgoff)
683                         return 1;
684         }
685         return 0;
686 }
687
688 /*
689  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
690  * beyond (at a higher virtual address and file offset than) the vma.
691  *
692  * We cannot merge two vmas if they have differently assigned (non-NULL)
693  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
694  */
695 static int
696 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
697         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
698 {
699         if (is_mergeable_vma(vma, file, vm_flags) &&
700             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
701                 pgoff_t vm_pglen;
702                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
703                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
704                         return 1;
705         }
706         return 0;
707 }
708
709 /*
710  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
711  * whether that can be merged with its predecessor or its successor.
712  * Or both (it neatly fills a hole).
713  *
714  * In most cases - when called for mmap, brk or mremap - [addr,end) is
715  * certain not to be mapped by the time vma_merge is called; but when
716  * called for mprotect, it is certain to be already mapped (either at
717  * an offset within prev, or at the start of next), and the flags of
718  * this area are about to be changed to vm_flags - and the no-change
719  * case has already been eliminated.
720  *
721  * The following mprotect cases have to be considered, where AAAA is
722  * the area passed down from mprotect_fixup, never extending beyond one
723  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
724  *
725  *     AAAA             AAAA                AAAA          AAAA
726  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
727  *    cannot merge    might become    might become    might become
728  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
729  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
730  *    mremap move:                                    PPPPNNNNNNNN 8
731  *        AAAA
732  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
733  *    might become    case 1 below    case 2 below    case 3 below
734  *
735  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
736  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
737  */
738 struct vm_area_struct *vma_merge(struct mm_struct *mm,
739                         struct vm_area_struct *prev, unsigned long addr,
740                         unsigned long end, unsigned long vm_flags,
741                         struct anon_vma *anon_vma, struct file *file,
742                         pgoff_t pgoff, struct mempolicy *policy)
743 {
744         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
745         struct vm_area_struct *area, *next;
746
747         /*
748          * We later require that vma->vm_flags == vm_flags,
749          * so this tests vma->vm_flags & VM_SPECIAL, too.
750          */
751         if (vm_flags & VM_SPECIAL)
752                 return NULL;
753
754         if (prev)
755                 next = prev->vm_next;
756         else
757                 next = mm->mmap;
758         area = next;
759         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
760                 next = next->vm_next;
761
762         /*
763          * Can it merge with the predecessor?
764          */
765         if (prev && prev->vm_end == addr &&
766                         mpol_equal(vma_policy(prev), policy) &&
767                         can_vma_merge_after(prev, vm_flags,
768                                                 anon_vma, file, pgoff)) {
769                 /*
770                  * OK, it can.  Can we now merge in the successor as well?
771                  */
772                 if (next && end == next->vm_start &&
773                                 mpol_equal(policy, vma_policy(next)) &&
774                                 can_vma_merge_before(next, vm_flags,
775                                         anon_vma, file, pgoff+pglen) &&
776                                 is_mergeable_anon_vma(prev->anon_vma,
777                                                       next->anon_vma)) {
778                                                         /* cases 1, 6 */
779                         vma_adjust(prev, prev->vm_start,
780                                 next->vm_end, prev->vm_pgoff, NULL);
781                 } else                                  /* cases 2, 5, 7 */
782                         vma_adjust(prev, prev->vm_start,
783                                 end, prev->vm_pgoff, NULL);
784                 return prev;
785         }
786
787         /*
788          * Can this new request be merged in front of next?
789          */
790         if (next && end == next->vm_start &&
791                         mpol_equal(policy, vma_policy(next)) &&
792                         can_vma_merge_before(next, vm_flags,
793                                         anon_vma, file, pgoff+pglen)) {
794                 if (prev && addr < prev->vm_end)        /* case 4 */
795                         vma_adjust(prev, prev->vm_start,
796                                 addr, prev->vm_pgoff, NULL);
797                 else                                    /* cases 3, 8 */
798                         vma_adjust(area, addr, next->vm_end,
799                                 next->vm_pgoff - pglen, NULL);
800                 return area;
801         }
802
803         return NULL;
804 }
805
806 /*
807  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
808  * neighbouring vmas for a suitable anon_vma, before it goes off
809  * to allocate a new anon_vma.  It checks because a repetitive
810  * sequence of mprotects and faults may otherwise lead to distinct
811  * anon_vmas being allocated, preventing vma merge in subsequent
812  * mprotect.
813  */
814 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
815 {
816         struct vm_area_struct *near;
817         unsigned long vm_flags;
818
819         near = vma->vm_next;
820         if (!near)
821                 goto try_prev;
822
823         /*
824          * Since only mprotect tries to remerge vmas, match flags
825          * which might be mprotected into each other later on.
826          * Neither mlock nor madvise tries to remerge at present,
827          * so leave their flags as obstructing a merge.
828          */
829         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
830         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
831
832         if (near->anon_vma && vma->vm_end == near->vm_start &&
833                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
834                         can_vma_merge_before(near, vm_flags,
835                                 NULL, vma->vm_file, vma->vm_pgoff +
836                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
837                 return near->anon_vma;
838 try_prev:
839         /*
840          * It is potentially slow to have to call find_vma_prev here.
841          * But it's only on the first write fault on the vma, not
842          * every time, and we could devise a way to avoid it later
843          * (e.g. stash info in next's anon_vma_node when assigning
844          * an anon_vma, or when trying vma_merge).  Another time.
845          */
846         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
847         if (!near)
848                 goto none;
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 && near->vm_end == vma->vm_start &&
854                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
855                         can_vma_merge_after(near, vm_flags,
856                                 NULL, vma->vm_file, vma->vm_pgoff))
857                 return near->anon_vma;
858 none:
859         /*
860          * There's no absolute need to look only at touching neighbours:
861          * we could search further afield for "compatible" anon_vmas.
862          * But it would probably just be a waste of time searching,
863          * or lead to too many vmas hanging off the same anon_vma.
864          * We're trying to allow mprotect remerging later on,
865          * not trying to minimize memory used for anon_vmas.
866          */
867         return NULL;
868 }
869
870 #ifdef CONFIG_PROC_FS
871 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
872                                                 struct file *file, long pages)
873 {
874         const unsigned long stack_flags
875                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
876
877         if (file) {
878                 mm->shared_vm += pages;
879                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
880                         mm->exec_vm += pages;
881         } else if (flags & stack_flags)
882                 mm->stack_vm += pages;
883         if (flags & (VM_RESERVED|VM_IO))
884                 mm->reserved_vm += pages;
885 }
886 #endif /* CONFIG_PROC_FS */
887
888 /*
889  * The caller must hold down_write(current->mm->mmap_sem).
890  */
891
892 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
893                         unsigned long len, unsigned long prot,
894                         unsigned long flags, unsigned long pgoff)
895 {
896         struct mm_struct * mm = current->mm;
897         struct inode *inode;
898         unsigned int vm_flags;
899         int error;
900         int accountable = 1;
901         unsigned long reqprot = prot;
902
903         /*
904          * Does the application expect PROT_READ to imply PROT_EXEC?
905          *
906          * (the exception is when the underlying filesystem is noexec
907          *  mounted, in which case we dont add PROT_EXEC.)
908          */
909         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
910                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
911                         prot |= PROT_EXEC;
912
913         if (!len)
914                 return -EINVAL;
915
916         if (!(flags & MAP_FIXED))
917                 addr = round_hint_to_min(addr);
918
919         error = arch_mmap_check(addr, len, flags);
920         if (error)
921                 return error;
922
923         /* Careful about overflows.. */
924         len = PAGE_ALIGN(len);
925         if (!len || len > TASK_SIZE)
926                 return -ENOMEM;
927
928         /* offset overflow? */
929         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
930                return -EOVERFLOW;
931
932         /* Too many mappings? */
933         if (mm->map_count > sysctl_max_map_count)
934                 return -ENOMEM;
935
936         /* Obtain the address to map to. we verify (or select) it and ensure
937          * that it represents a valid section of the address space.
938          */
939         addr = get_unmapped_area(file, addr, len, pgoff, flags);
940         if (addr & ~PAGE_MASK)
941                 return addr;
942
943         /* Do simple checking here so the lower-level routines won't have
944          * to. we assume access permissions have been handled by the open
945          * of the memory object, so we don't do any here.
946          */
947         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
948                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
949
950         if (flags & MAP_LOCKED) {
951                 if (!can_do_mlock())
952                         return -EPERM;
953                 vm_flags |= VM_LOCKED;
954         }
955         /* mlock MCL_FUTURE? */
956         if (vm_flags & VM_LOCKED) {
957                 unsigned long locked, lock_limit;
958                 locked = len >> PAGE_SHIFT;
959                 locked += mm->locked_vm;
960                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
961                 lock_limit >>= PAGE_SHIFT;
962                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
963                         return -EAGAIN;
964         }
965
966         inode = file ? file->f_path.dentry->d_inode : NULL;
967
968         if (file) {
969                 switch (flags & MAP_TYPE) {
970                 case MAP_SHARED:
971                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
972                                 return -EACCES;
973
974                         /*
975                          * Make sure we don't allow writing to an append-only
976                          * file..
977                          */
978                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
979                                 return -EACCES;
980
981                         /*
982                          * Make sure there are no mandatory locks on the file.
983                          */
984                         if (locks_verify_locked(inode))
985                                 return -EAGAIN;
986
987                         vm_flags |= VM_SHARED | VM_MAYSHARE;
988                         if (!(file->f_mode & FMODE_WRITE))
989                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
990
991                         /* fall through */
992                 case MAP_PRIVATE:
993                         if (!(file->f_mode & FMODE_READ))
994                                 return -EACCES;
995                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
996                                 if (vm_flags & VM_EXEC)
997                                         return -EPERM;
998                                 vm_flags &= ~VM_MAYEXEC;
999                         }
1000                         if (is_file_hugepages(file))
1001                                 accountable = 0;
1002
1003                         if (!file->f_op || !file->f_op->mmap)
1004                                 return -ENODEV;
1005                         break;
1006
1007                 default:
1008                         return -EINVAL;
1009                 }
1010         } else {
1011                 switch (flags & MAP_TYPE) {
1012                 case MAP_SHARED:
1013                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1014                         break;
1015                 case MAP_PRIVATE:
1016                         /*
1017                          * Set pgoff according to addr for anon_vma.
1018                          */
1019                         pgoff = addr >> PAGE_SHIFT;
1020                         break;
1021                 default:
1022                         return -EINVAL;
1023                 }
1024         }
1025
1026         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1027         if (error)
1028                 return error;
1029
1030         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1031                            accountable);
1032 }
1033 EXPORT_SYMBOL(do_mmap_pgoff);
1034
1035 /*
1036  * Some shared mappigns will want the pages marked read-only
1037  * to track write events. If so, we'll downgrade vm_page_prot
1038  * to the private version (using protection_map[] without the
1039  * VM_SHARED bit).
1040  */
1041 int vma_wants_writenotify(struct vm_area_struct *vma)
1042 {
1043         unsigned int vm_flags = vma->vm_flags;
1044
1045         /* If it was private or non-writable, the write bit is already clear */
1046         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1047                 return 0;
1048
1049         /* The backer wishes to know when pages are first written to? */
1050         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1051                 return 1;
1052
1053         /* The open routine did something to the protections already? */
1054         if (pgprot_val(vma->vm_page_prot) !=
1055             pgprot_val(vm_get_page_prot(vm_flags)))
1056                 return 0;
1057
1058         /* Specialty mapping? */
1059         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1060                 return 0;
1061
1062         /* Can the mapping track the dirty pages? */
1063         return vma->vm_file && vma->vm_file->f_mapping &&
1064                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1065 }
1066
1067
1068 unsigned long mmap_region(struct file *file, unsigned long addr,
1069                           unsigned long len, unsigned long flags,
1070                           unsigned int vm_flags, unsigned long pgoff,
1071                           int accountable)
1072 {
1073         struct mm_struct *mm = current->mm;
1074         struct vm_area_struct *vma, *prev;
1075         int correct_wcount = 0;
1076         int error;
1077         struct rb_node **rb_link, *rb_parent;
1078         unsigned long charged = 0;
1079         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1080
1081         /* Clear old maps */
1082         error = -ENOMEM;
1083 munmap_back:
1084         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1085         if (vma && vma->vm_start < addr + len) {
1086                 if (do_munmap(mm, addr, len))
1087                         return -ENOMEM;
1088                 goto munmap_back;
1089         }
1090
1091         /* Check against address space limit. */
1092         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1093                 return -ENOMEM;
1094
1095         if (accountable && (!(flags & MAP_NORESERVE) ||
1096                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1097                 if (vm_flags & VM_SHARED) {
1098                         /* Check memory availability in shmem_file_setup? */
1099                         vm_flags |= VM_ACCOUNT;
1100                 } else if (vm_flags & VM_WRITE) {
1101                         /*
1102                          * Private writable mapping: check memory availability
1103                          */
1104                         charged = len >> PAGE_SHIFT;
1105                         if (security_vm_enough_memory(charged))
1106                                 return -ENOMEM;
1107                         vm_flags |= VM_ACCOUNT;
1108                 }
1109         }
1110
1111         /*
1112          * Can we just expand an old private anonymous mapping?
1113          * The VM_SHARED test is necessary because shmem_zero_setup
1114          * will create the file object for a shared anonymous map below.
1115          */
1116         if (!file && !(vm_flags & VM_SHARED) &&
1117             vma_merge(mm, prev, addr, addr + len, vm_flags,
1118                                         NULL, NULL, pgoff, NULL))
1119                 goto out;
1120
1121         /*
1122          * Determine the object being mapped and call the appropriate
1123          * specific mapper. the address has already been validated, but
1124          * not unmapped, but the maps are removed from the list.
1125          */
1126         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1127         if (!vma) {
1128                 error = -ENOMEM;
1129                 goto unacct_error;
1130         }
1131
1132         vma->vm_mm = mm;
1133         vma->vm_start = addr;
1134         vma->vm_end = addr + len;
1135         vma->vm_flags = vm_flags;
1136         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1137         vma->vm_pgoff = pgoff;
1138
1139         if (file) {
1140                 error = -EINVAL;
1141                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1142                         goto free_vma;
1143                 if (vm_flags & VM_DENYWRITE) {
1144                         error = deny_write_access(file);
1145                         if (error)
1146                                 goto free_vma;
1147                         correct_wcount = 1;
1148                 }
1149                 vma->vm_file = file;
1150                 get_file(file);
1151                 error = file->f_op->mmap(file, vma);
1152                 if (error)
1153                         goto unmap_and_free_vma;
1154         } else if (vm_flags & VM_SHARED) {
1155                 error = shmem_zero_setup(vma);
1156                 if (error)
1157                         goto free_vma;
1158         }
1159
1160         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1161          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1162          * that memory reservation must be checked; but that reservation
1163          * belongs to shared memory object, not to vma: so now clear it.
1164          */
1165         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1166                 vma->vm_flags &= ~VM_ACCOUNT;
1167
1168         /* Can addr have changed??
1169          *
1170          * Answer: Yes, several device drivers can do it in their
1171          *         f_op->mmap method. -DaveM
1172          */
1173         addr = vma->vm_start;
1174         pgoff = vma->vm_pgoff;
1175         vm_flags = vma->vm_flags;
1176
1177         if (vma_wants_writenotify(vma))
1178                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1179
1180         if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1181                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1182                 file = vma->vm_file;
1183                 vma_link(mm, vma, prev, rb_link, rb_parent);
1184                 if (correct_wcount)
1185                         atomic_inc(&inode->i_writecount);
1186         } else {
1187                 if (file) {
1188                         if (correct_wcount)
1189                                 atomic_inc(&inode->i_writecount);
1190                         fput(file);
1191                 }
1192                 mpol_free(vma_policy(vma));
1193                 kmem_cache_free(vm_area_cachep, vma);
1194         }
1195 out:    
1196         mm->total_vm += len >> PAGE_SHIFT;
1197         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1198         if (vm_flags & VM_LOCKED) {
1199                 mm->locked_vm += len >> PAGE_SHIFT;
1200                 make_pages_present(addr, addr + len);
1201         }
1202         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1203                 make_pages_present(addr, addr + len);
1204         return addr;
1205
1206 unmap_and_free_vma:
1207         if (correct_wcount)
1208                 atomic_inc(&inode->i_writecount);
1209         vma->vm_file = NULL;
1210         fput(file);
1211
1212         /* Undo any partial mapping done by a device driver. */
1213         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1214         charged = 0;
1215 free_vma:
1216         kmem_cache_free(vm_area_cachep, vma);
1217 unacct_error:
1218         if (charged)
1219                 vm_unacct_memory(charged);
1220         return error;
1221 }
1222
1223 /* Get an address range which is currently unmapped.
1224  * For shmat() with addr=0.
1225  *
1226  * Ugly calling convention alert:
1227  * Return value with the low bits set means error value,
1228  * ie
1229  *      if (ret & ~PAGE_MASK)
1230  *              error = ret;
1231  *
1232  * This function "knows" that -ENOMEM has the bits set.
1233  */
1234 #ifndef HAVE_ARCH_UNMAPPED_AREA
1235 unsigned long
1236 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1237                 unsigned long len, unsigned long pgoff, unsigned long flags)
1238 {
1239         struct mm_struct *mm = current->mm;
1240         struct vm_area_struct *vma;
1241         unsigned long start_addr;
1242
1243         if (len > TASK_SIZE)
1244                 return -ENOMEM;
1245
1246         if (flags & MAP_FIXED)
1247                 return addr;
1248
1249         if (addr) {
1250                 addr = PAGE_ALIGN(addr);
1251                 vma = find_vma(mm, addr);
1252                 if (TASK_SIZE - len >= addr &&
1253                     (!vma || addr + len <= vma->vm_start))
1254                         return addr;
1255         }
1256         if (len > mm->cached_hole_size) {
1257                 start_addr = addr = mm->free_area_cache;
1258         } else {
1259                 start_addr = addr = TASK_UNMAPPED_BASE;
1260                 mm->cached_hole_size = 0;
1261         }
1262
1263 full_search:
1264         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1265                 /* At this point:  (!vma || addr < vma->vm_end). */
1266                 if (TASK_SIZE - len < addr) {
1267                         /*
1268                          * Start a new search - just in case we missed
1269                          * some holes.
1270                          */
1271                         if (start_addr != TASK_UNMAPPED_BASE) {
1272                                 addr = TASK_UNMAPPED_BASE;
1273                                 start_addr = addr;
1274                                 mm->cached_hole_size = 0;
1275                                 goto full_search;
1276                         }
1277                         return -ENOMEM;
1278                 }
1279                 if (!vma || addr + len <= vma->vm_start) {
1280                         /*
1281                          * Remember the place where we stopped the search:
1282                          */
1283                         mm->free_area_cache = addr + len;
1284                         return addr;
1285                 }
1286                 if (addr + mm->cached_hole_size < vma->vm_start)
1287                         mm->cached_hole_size = vma->vm_start - addr;
1288                 addr = vma->vm_end;
1289         }
1290 }
1291 #endif  
1292
1293 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1294 {
1295         /*
1296          * Is this a new hole at the lowest possible address?
1297          */
1298         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1299                 mm->free_area_cache = addr;
1300                 mm->cached_hole_size = ~0UL;
1301         }
1302 }
1303
1304 /*
1305  * This mmap-allocator allocates new areas top-down from below the
1306  * stack's low limit (the base):
1307  */
1308 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1309 unsigned long
1310 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1311                           const unsigned long len, const unsigned long pgoff,
1312                           const unsigned long flags)
1313 {
1314         struct vm_area_struct *vma;
1315         struct mm_struct *mm = current->mm;
1316         unsigned long addr = addr0;
1317
1318         /* requested length too big for entire address space */
1319         if (len > TASK_SIZE)
1320                 return -ENOMEM;
1321
1322         if (flags & MAP_FIXED)
1323                 return addr;
1324
1325         /* requesting a specific address */
1326         if (addr) {
1327                 addr = PAGE_ALIGN(addr);
1328                 vma = find_vma(mm, addr);
1329                 if (TASK_SIZE - len >= addr &&
1330                                 (!vma || addr + len <= vma->vm_start))
1331                         return addr;
1332         }
1333
1334         /* check if free_area_cache is useful for us */
1335         if (len <= mm->cached_hole_size) {
1336                 mm->cached_hole_size = 0;
1337                 mm->free_area_cache = mm->mmap_base;
1338         }
1339
1340         /* either no address requested or can't fit in requested address hole */
1341         addr = mm->free_area_cache;
1342
1343         /* make sure it can fit in the remaining address space */
1344         if (addr > len) {
1345                 vma = find_vma(mm, addr-len);
1346                 if (!vma || addr <= vma->vm_start)
1347                         /* remember the address as a hint for next time */
1348                         return (mm->free_area_cache = addr-len);
1349         }
1350
1351         if (mm->mmap_base < len)
1352                 goto bottomup;
1353
1354         addr = mm->mmap_base-len;
1355
1356         do {
1357                 /*
1358                  * Lookup failure means no vma is above this address,
1359                  * else if new region fits below vma->vm_start,
1360                  * return with success:
1361                  */
1362                 vma = find_vma(mm, addr);
1363                 if (!vma || addr+len <= vma->vm_start)
1364                         /* remember the address as a hint for next time */
1365                         return (mm->free_area_cache = addr);
1366
1367                 /* remember the largest hole we saw so far */
1368                 if (addr + mm->cached_hole_size < vma->vm_start)
1369                         mm->cached_hole_size = vma->vm_start - addr;
1370
1371                 /* try just below the current vma->vm_start */
1372                 addr = vma->vm_start-len;
1373         } while (len < vma->vm_start);
1374
1375 bottomup:
1376         /*
1377          * A failed mmap() very likely causes application failure,
1378          * so fall back to the bottom-up function here. This scenario
1379          * can happen with large stack limits and large mmap()
1380          * allocations.
1381          */
1382         mm->cached_hole_size = ~0UL;
1383         mm->free_area_cache = TASK_UNMAPPED_BASE;
1384         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1385         /*
1386          * Restore the topdown base:
1387          */
1388         mm->free_area_cache = mm->mmap_base;
1389         mm->cached_hole_size = ~0UL;
1390
1391         return addr;
1392 }
1393 #endif
1394
1395 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1396 {
1397         /*
1398          * Is this a new hole at the highest possible address?
1399          */
1400         if (addr > mm->free_area_cache)
1401                 mm->free_area_cache = addr;
1402
1403         /* dont allow allocations above current base */
1404         if (mm->free_area_cache > mm->mmap_base)
1405                 mm->free_area_cache = mm->mmap_base;
1406 }
1407
1408 unsigned long
1409 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1410                 unsigned long pgoff, unsigned long flags)
1411 {
1412         unsigned long (*get_area)(struct file *, unsigned long,
1413                                   unsigned long, unsigned long, unsigned long);
1414
1415         get_area = current->mm->get_unmapped_area;
1416         if (file && file->f_op && file->f_op->get_unmapped_area)
1417                 get_area = file->f_op->get_unmapped_area;
1418         addr = get_area(file, addr, len, pgoff, flags);
1419         if (IS_ERR_VALUE(addr))
1420                 return addr;
1421
1422         if (addr > TASK_SIZE - len)
1423                 return -ENOMEM;
1424         if (addr & ~PAGE_MASK)
1425                 return -EINVAL;
1426
1427         return addr;
1428 }
1429
1430 EXPORT_SYMBOL(get_unmapped_area);
1431
1432 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1433 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1434 {
1435         struct vm_area_struct *vma = NULL;
1436
1437         if (mm) {
1438                 /* Check the cache first. */
1439                 /* (Cache hit rate is typically around 35%.) */
1440                 vma = mm->mmap_cache;
1441                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1442                         struct rb_node * rb_node;
1443
1444                         rb_node = mm->mm_rb.rb_node;
1445                         vma = NULL;
1446
1447                         while (rb_node) {
1448                                 struct vm_area_struct * vma_tmp;
1449
1450                                 vma_tmp = rb_entry(rb_node,
1451                                                 struct vm_area_struct, vm_rb);
1452
1453                                 if (vma_tmp->vm_end > addr) {
1454                                         vma = vma_tmp;
1455                                         if (vma_tmp->vm_start <= addr)
1456                                                 break;
1457                                         rb_node = rb_node->rb_left;
1458                                 } else
1459                                         rb_node = rb_node->rb_right;
1460                         }
1461                         if (vma)
1462                                 mm->mmap_cache = vma;
1463                 }
1464         }
1465         return vma;
1466 }
1467
1468 EXPORT_SYMBOL(find_vma);
1469
1470 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1471 struct vm_area_struct *
1472 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1473                         struct vm_area_struct **pprev)
1474 {
1475         struct vm_area_struct *vma = NULL, *prev = NULL;
1476         struct rb_node * rb_node;
1477         if (!mm)
1478                 goto out;
1479
1480         /* Guard against addr being lower than the first VMA */
1481         vma = mm->mmap;
1482
1483         /* Go through the RB tree quickly. */
1484         rb_node = mm->mm_rb.rb_node;
1485
1486         while (rb_node) {
1487                 struct vm_area_struct *vma_tmp;
1488                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1489
1490                 if (addr < vma_tmp->vm_end) {
1491                         rb_node = rb_node->rb_left;
1492                 } else {
1493                         prev = vma_tmp;
1494                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1495                                 break;
1496                         rb_node = rb_node->rb_right;
1497                 }
1498         }
1499
1500 out:
1501         *pprev = prev;
1502         return prev ? prev->vm_next : vma;
1503 }
1504
1505 /*
1506  * Verify that the stack growth is acceptable and
1507  * update accounting. This is shared with both the
1508  * grow-up and grow-down cases.
1509  */
1510 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1511 {
1512         struct mm_struct *mm = vma->vm_mm;
1513         struct rlimit *rlim = current->signal->rlim;
1514         unsigned long new_start;
1515
1516         /* address space limit tests */
1517         if (!may_expand_vm(mm, grow))
1518                 return -ENOMEM;
1519
1520         /* Stack limit test */
1521         if (size > rlim[RLIMIT_STACK].rlim_cur)
1522                 return -ENOMEM;
1523
1524         /* mlock limit tests */
1525         if (vma->vm_flags & VM_LOCKED) {
1526                 unsigned long locked;
1527                 unsigned long limit;
1528                 locked = mm->locked_vm + grow;
1529                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1530                 if (locked > limit && !capable(CAP_IPC_LOCK))
1531                         return -ENOMEM;
1532         }
1533
1534         /* Check to ensure the stack will not grow into a hugetlb-only region */
1535         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1536                         vma->vm_end - size;
1537         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1538                 return -EFAULT;
1539
1540         /*
1541          * Overcommit..  This must be the final test, as it will
1542          * update security statistics.
1543          */
1544         if (security_vm_enough_memory(grow))
1545                 return -ENOMEM;
1546
1547         /* Ok, everything looks good - let it rip */
1548         mm->total_vm += grow;
1549         if (vma->vm_flags & VM_LOCKED)
1550                 mm->locked_vm += grow;
1551         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1552         return 0;
1553 }
1554
1555 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1556 /*
1557  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1558  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1559  */
1560 #ifndef CONFIG_IA64
1561 static inline
1562 #endif
1563 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1564 {
1565         int error;
1566
1567         if (!(vma->vm_flags & VM_GROWSUP))
1568                 return -EFAULT;
1569
1570         /*
1571          * We must make sure the anon_vma is allocated
1572          * so that the anon_vma locking is not a noop.
1573          */
1574         if (unlikely(anon_vma_prepare(vma)))
1575                 return -ENOMEM;
1576         anon_vma_lock(vma);
1577
1578         /*
1579          * vma->vm_start/vm_end cannot change under us because the caller
1580          * is required to hold the mmap_sem in read mode.  We need the
1581          * anon_vma lock to serialize against concurrent expand_stacks.
1582          * Also guard against wrapping around to address 0.
1583          */
1584         if (address < PAGE_ALIGN(address+4))
1585                 address = PAGE_ALIGN(address+4);
1586         else {
1587                 anon_vma_unlock(vma);
1588                 return -ENOMEM;
1589         }
1590         error = 0;
1591
1592         /* Somebody else might have raced and expanded it already */
1593         if (address > vma->vm_end) {
1594                 unsigned long size, grow;
1595
1596                 size = address - vma->vm_start;
1597                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1598
1599                 error = acct_stack_growth(vma, size, grow);
1600                 if (!error)
1601                         vma->vm_end = address;
1602         }
1603         anon_vma_unlock(vma);
1604         return error;
1605 }
1606 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1607
1608 /*
1609  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1610  */
1611 static inline int expand_downwards(struct vm_area_struct *vma,
1612                                    unsigned long address)
1613 {
1614         int error;
1615
1616         /*
1617          * We must make sure the anon_vma is allocated
1618          * so that the anon_vma locking is not a noop.
1619          */
1620         if (unlikely(anon_vma_prepare(vma)))
1621                 return -ENOMEM;
1622
1623         address &= PAGE_MASK;
1624         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1625         if (error)
1626                 return error;
1627
1628         anon_vma_lock(vma);
1629
1630         /*
1631          * vma->vm_start/vm_end cannot change under us because the caller
1632          * is required to hold the mmap_sem in read mode.  We need the
1633          * anon_vma lock to serialize against concurrent expand_stacks.
1634          */
1635
1636         /* Somebody else might have raced and expanded it already */
1637         if (address < vma->vm_start) {
1638                 unsigned long size, grow;
1639
1640                 size = vma->vm_end - address;
1641                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1642
1643                 error = acct_stack_growth(vma, size, grow);
1644                 if (!error) {
1645                         vma->vm_start = address;
1646                         vma->vm_pgoff -= grow;
1647                 }
1648         }
1649         anon_vma_unlock(vma);
1650         return error;
1651 }
1652
1653 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1654 {
1655         return expand_downwards(vma, address);
1656 }
1657
1658 #ifdef CONFIG_STACK_GROWSUP
1659 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1660 {
1661         return expand_upwards(vma, address);
1662 }
1663
1664 struct vm_area_struct *
1665 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1666 {
1667         struct vm_area_struct *vma, *prev;
1668
1669         addr &= PAGE_MASK;
1670         vma = find_vma_prev(mm, addr, &prev);
1671         if (vma && (vma->vm_start <= addr))
1672                 return vma;
1673         if (!prev || expand_stack(prev, addr))
1674                 return NULL;
1675         if (prev->vm_flags & VM_LOCKED)
1676                 make_pages_present(addr, prev->vm_end);
1677         return prev;
1678 }
1679 #else
1680 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1681 {
1682         return expand_downwards(vma, address);
1683 }
1684
1685 struct vm_area_struct *
1686 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1687 {
1688         struct vm_area_struct * vma;
1689         unsigned long start;
1690
1691         addr &= PAGE_MASK;
1692         vma = find_vma(mm,addr);
1693         if (!vma)
1694                 return NULL;
1695         if (vma->vm_start <= addr)
1696                 return vma;
1697         if (!(vma->vm_flags & VM_GROWSDOWN))
1698                 return NULL;
1699         start = vma->vm_start;
1700         if (expand_stack(vma, addr))
1701                 return NULL;
1702         if (vma->vm_flags & VM_LOCKED)
1703                 make_pages_present(addr, start);
1704         return vma;
1705 }
1706 #endif
1707
1708 /*
1709  * Ok - we have the memory areas we should free on the vma list,
1710  * so release them, and do the vma updates.
1711  *
1712  * Called with the mm semaphore held.
1713  */
1714 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1715 {
1716         /* Update high watermark before we lower total_vm */
1717         update_hiwater_vm(mm);
1718         do {
1719                 long nrpages = vma_pages(vma);
1720
1721                 mm->total_vm -= nrpages;
1722                 if (vma->vm_flags & VM_LOCKED)
1723                         mm->locked_vm -= nrpages;
1724                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1725                 vma = remove_vma(vma);
1726         } while (vma);
1727         validate_mm(mm);
1728 }
1729
1730 /*
1731  * Get rid of page table information in the indicated region.
1732  *
1733  * Called with the mm semaphore held.
1734  */
1735 static void unmap_region(struct mm_struct *mm,
1736                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1737                 unsigned long start, unsigned long end)
1738 {
1739         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1740         struct mmu_gather *tlb;
1741         unsigned long nr_accounted = 0;
1742
1743         lru_add_drain();
1744         tlb = tlb_gather_mmu(mm, 0);
1745         update_hiwater_rss(mm);
1746         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1747         vm_unacct_memory(nr_accounted);
1748         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1749                                  next? next->vm_start: 0);
1750         tlb_finish_mmu(tlb, start, end);
1751 }
1752
1753 /*
1754  * Create a list of vma's touched by the unmap, removing them from the mm's
1755  * vma list as we go..
1756  */
1757 static void
1758 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1759         struct vm_area_struct *prev, unsigned long end)
1760 {
1761         struct vm_area_struct **insertion_point;
1762         struct vm_area_struct *tail_vma = NULL;
1763         unsigned long addr;
1764
1765         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1766         do {
1767                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1768                 mm->map_count--;
1769                 tail_vma = vma;
1770                 vma = vma->vm_next;
1771         } while (vma && vma->vm_start < end);
1772         *insertion_point = vma;
1773         tail_vma->vm_next = NULL;
1774         if (mm->unmap_area == arch_unmap_area)
1775                 addr = prev ? prev->vm_end : mm->mmap_base;
1776         else
1777                 addr = vma ?  vma->vm_start : mm->mmap_base;
1778         mm->unmap_area(mm, addr);
1779         mm->mmap_cache = NULL;          /* Kill the cache. */
1780 }
1781
1782 /*
1783  * Split a vma into two pieces at address 'addr', a new vma is allocated
1784  * either for the first part or the tail.
1785  */
1786 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1787               unsigned long addr, int new_below)
1788 {
1789         struct mempolicy *pol;
1790         struct vm_area_struct *new;
1791
1792         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1793                 return -EINVAL;
1794
1795         if (mm->map_count >= sysctl_max_map_count)
1796                 return -ENOMEM;
1797
1798         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1799         if (!new)
1800                 return -ENOMEM;
1801
1802         /* most fields are the same, copy all, and then fixup */
1803         *new = *vma;
1804
1805         if (new_below)
1806                 new->vm_end = addr;
1807         else {
1808                 new->vm_start = addr;
1809                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1810         }
1811
1812         pol = mpol_copy(vma_policy(vma));
1813         if (IS_ERR(pol)) {
1814                 kmem_cache_free(vm_area_cachep, new);
1815                 return PTR_ERR(pol);
1816         }
1817         vma_set_policy(new, pol);
1818
1819         if (new->vm_file)
1820                 get_file(new->vm_file);
1821
1822         if (new->vm_ops && new->vm_ops->open)
1823                 new->vm_ops->open(new);
1824
1825         if (new_below)
1826                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1827                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1828         else
1829                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1830
1831         return 0;
1832 }
1833
1834 /* Munmap is split into 2 main parts -- this part which finds
1835  * what needs doing, and the areas themselves, which do the
1836  * work.  This now handles partial unmappings.
1837  * Jeremy Fitzhardinge <jeremy@goop.org>
1838  */
1839 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1840 {
1841         unsigned long end;
1842         struct vm_area_struct *vma, *prev, *last;
1843
1844         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1845                 return -EINVAL;
1846
1847         if ((len = PAGE_ALIGN(len)) == 0)
1848                 return -EINVAL;
1849
1850         /* Find the first overlapping VMA */
1851         vma = find_vma_prev(mm, start, &prev);
1852         if (!vma)
1853                 return 0;
1854         /* we have  start < vma->vm_end  */
1855
1856         /* if it doesn't overlap, we have nothing.. */
1857         end = start + len;
1858         if (vma->vm_start >= end)
1859                 return 0;
1860
1861         /*
1862          * If we need to split any vma, do it now to save pain later.
1863          *
1864          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1865          * unmapped vm_area_struct will remain in use: so lower split_vma
1866          * places tmp vma above, and higher split_vma places tmp vma below.
1867          */
1868         if (start > vma->vm_start) {
1869                 int error = split_vma(mm, vma, start, 0);
1870                 if (error)
1871                         return error;
1872                 prev = vma;
1873         }
1874
1875         /* Does it split the last one? */
1876         last = find_vma(mm, end);
1877         if (last && end > last->vm_start) {
1878                 int error = split_vma(mm, last, end, 1);
1879                 if (error)
1880                         return error;
1881         }
1882         vma = prev? prev->vm_next: mm->mmap;
1883
1884         /*
1885          * Remove the vma's, and unmap the actual pages
1886          */
1887         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1888         unmap_region(mm, vma, prev, start, end);
1889
1890         /* Fix up all other VM information */
1891         remove_vma_list(mm, vma);
1892
1893         return 0;
1894 }
1895
1896 EXPORT_SYMBOL(do_munmap);
1897
1898 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1899 {
1900         int ret;
1901         struct mm_struct *mm = current->mm;
1902
1903         profile_munmap(addr);
1904
1905         down_write(&mm->mmap_sem);
1906         ret = do_munmap(mm, addr, len);
1907         up_write(&mm->mmap_sem);
1908         return ret;
1909 }
1910
1911 static inline void verify_mm_writelocked(struct mm_struct *mm)
1912 {
1913 #ifdef CONFIG_DEBUG_VM
1914         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1915                 WARN_ON(1);
1916                 up_read(&mm->mmap_sem);
1917         }
1918 #endif
1919 }
1920
1921 /*
1922  *  this is really a simplified "do_mmap".  it only handles
1923  *  anonymous maps.  eventually we may be able to do some
1924  *  brk-specific accounting here.
1925  */
1926 unsigned long do_brk(unsigned long addr, unsigned long len)
1927 {
1928         struct mm_struct * mm = current->mm;
1929         struct vm_area_struct * vma, * prev;
1930         unsigned long flags;
1931         struct rb_node ** rb_link, * rb_parent;
1932         pgoff_t pgoff = addr >> PAGE_SHIFT;
1933         int error;
1934
1935         len = PAGE_ALIGN(len);
1936         if (!len)
1937                 return addr;
1938
1939         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1940                 return -EINVAL;
1941
1942         if (is_hugepage_only_range(mm, addr, len))
1943                 return -EINVAL;
1944
1945         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1946         if (error)
1947                 return error;
1948
1949         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1950
1951         error = arch_mmap_check(addr, len, flags);
1952         if (error)
1953                 return error;
1954
1955         /*
1956          * mlock MCL_FUTURE?
1957          */
1958         if (mm->def_flags & VM_LOCKED) {
1959                 unsigned long locked, lock_limit;
1960                 locked = len >> PAGE_SHIFT;
1961                 locked += mm->locked_vm;
1962                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1963                 lock_limit >>= PAGE_SHIFT;
1964                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1965                         return -EAGAIN;
1966         }
1967
1968         /*
1969          * mm->mmap_sem is required to protect against another thread
1970          * changing the mappings in case we sleep.
1971          */
1972         verify_mm_writelocked(mm);
1973
1974         /*
1975          * Clear old maps.  this also does some error checking for us
1976          */
1977  munmap_back:
1978         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1979         if (vma && vma->vm_start < addr + len) {
1980                 if (do_munmap(mm, addr, len))
1981                         return -ENOMEM;
1982                 goto munmap_back;
1983         }
1984
1985         /* Check against address space limits *after* clearing old maps... */
1986         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1987                 return -ENOMEM;
1988
1989         if (mm->map_count > sysctl_max_map_count)
1990                 return -ENOMEM;
1991
1992         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1993                 return -ENOMEM;
1994
1995         /* Can we just expand an old private anonymous mapping? */
1996         if (vma_merge(mm, prev, addr, addr + len, flags,
1997                                         NULL, NULL, pgoff, NULL))
1998                 goto out;
1999
2000         /*
2001          * create a vma struct for an anonymous mapping
2002          */
2003         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2004         if (!vma) {
2005                 vm_unacct_memory(len >> PAGE_SHIFT);
2006                 return -ENOMEM;
2007         }
2008
2009         vma->vm_mm = mm;
2010         vma->vm_start = addr;
2011         vma->vm_end = addr + len;
2012         vma->vm_pgoff = pgoff;
2013         vma->vm_flags = flags;
2014         vma->vm_page_prot = vm_get_page_prot(flags);
2015         vma_link(mm, vma, prev, rb_link, rb_parent);
2016 out:
2017         mm->total_vm += len >> PAGE_SHIFT;
2018         if (flags & VM_LOCKED) {
2019                 mm->locked_vm += len >> PAGE_SHIFT;
2020                 make_pages_present(addr, addr + len);
2021         }
2022         return addr;
2023 }
2024
2025 EXPORT_SYMBOL(do_brk);
2026
2027 /* Release all mmaps. */
2028 void exit_mmap(struct mm_struct *mm)
2029 {
2030         struct mmu_gather *tlb;
2031         struct vm_area_struct *vma = mm->mmap;
2032         unsigned long nr_accounted = 0;
2033         unsigned long end;
2034
2035         /* mm's last user has gone, and its about to be pulled down */
2036         arch_exit_mmap(mm);
2037
2038         lru_add_drain();
2039         flush_cache_mm(mm);
2040         tlb = tlb_gather_mmu(mm, 1);
2041         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2042         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2043         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2044         vm_unacct_memory(nr_accounted);
2045         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2046         tlb_finish_mmu(tlb, 0, end);
2047
2048         /*
2049          * Walk the list again, actually closing and freeing it,
2050          * with preemption enabled, without holding any MM locks.
2051          */
2052         while (vma)
2053                 vma = remove_vma(vma);
2054
2055         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2056 }
2057
2058 /* Insert vm structure into process list sorted by address
2059  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2060  * then i_mmap_lock is taken here.
2061  */
2062 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2063 {
2064         struct vm_area_struct * __vma, * prev;
2065         struct rb_node ** rb_link, * rb_parent;
2066
2067         /*
2068          * The vm_pgoff of a purely anonymous vma should be irrelevant
2069          * until its first write fault, when page's anon_vma and index
2070          * are set.  But now set the vm_pgoff it will almost certainly
2071          * end up with (unless mremap moves it elsewhere before that
2072          * first wfault), so /proc/pid/maps tells a consistent story.
2073          *
2074          * By setting it to reflect the virtual start address of the
2075          * vma, merges and splits can happen in a seamless way, just
2076          * using the existing file pgoff checks and manipulations.
2077          * Similarly in do_mmap_pgoff and in do_brk.
2078          */
2079         if (!vma->vm_file) {
2080                 BUG_ON(vma->anon_vma);
2081                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2082         }
2083         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2084         if (__vma && __vma->vm_start < vma->vm_end)
2085                 return -ENOMEM;
2086         if ((vma->vm_flags & VM_ACCOUNT) &&
2087              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2088                 return -ENOMEM;
2089         vma_link(mm, vma, prev, rb_link, rb_parent);
2090         return 0;
2091 }
2092
2093 /*
2094  * Copy the vma structure to a new location in the same mm,
2095  * prior to moving page table entries, to effect an mremap move.
2096  */
2097 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2098         unsigned long addr, unsigned long len, pgoff_t pgoff)
2099 {
2100         struct vm_area_struct *vma = *vmap;
2101         unsigned long vma_start = vma->vm_start;
2102         struct mm_struct *mm = vma->vm_mm;
2103         struct vm_area_struct *new_vma, *prev;
2104         struct rb_node **rb_link, *rb_parent;
2105         struct mempolicy *pol;
2106
2107         /*
2108          * If anonymous vma has not yet been faulted, update new pgoff
2109          * to match new location, to increase its chance of merging.
2110          */
2111         if (!vma->vm_file && !vma->anon_vma)
2112                 pgoff = addr >> PAGE_SHIFT;
2113
2114         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2115         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2116                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2117         if (new_vma) {
2118                 /*
2119                  * Source vma may have been merged into new_vma
2120                  */
2121                 if (vma_start >= new_vma->vm_start &&
2122                     vma_start < new_vma->vm_end)
2123                         *vmap = new_vma;
2124         } else {
2125                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2126                 if (new_vma) {
2127                         *new_vma = *vma;
2128                         pol = mpol_copy(vma_policy(vma));
2129                         if (IS_ERR(pol)) {
2130                                 kmem_cache_free(vm_area_cachep, new_vma);
2131                                 return NULL;
2132                         }
2133                         vma_set_policy(new_vma, pol);
2134                         new_vma->vm_start = addr;
2135                         new_vma->vm_end = addr + len;
2136                         new_vma->vm_pgoff = pgoff;
2137                         if (new_vma->vm_file)
2138                                 get_file(new_vma->vm_file);
2139                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2140                                 new_vma->vm_ops->open(new_vma);
2141                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2142                 }
2143         }
2144         return new_vma;
2145 }
2146
2147 /*
2148  * Return true if the calling process may expand its vm space by the passed
2149  * number of pages
2150  */
2151 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2152 {
2153         unsigned long cur = mm->total_vm;       /* pages */
2154         unsigned long lim;
2155
2156         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2157
2158         if (cur + npages > lim)
2159                 return 0;
2160         return 1;
2161 }
2162
2163
2164 static struct page *special_mapping_nopage(struct vm_area_struct *vma,
2165                                            unsigned long address, int *type)
2166 {
2167         struct page **pages;
2168
2169         BUG_ON(address < vma->vm_start || address >= vma->vm_end);
2170
2171         address -= vma->vm_start;
2172         for (pages = vma->vm_private_data; address > 0 && *pages; ++pages)
2173                 address -= PAGE_SIZE;
2174
2175         if (*pages) {
2176                 struct page *page = *pages;
2177                 get_page(page);
2178                 return page;
2179         }
2180
2181         return NOPAGE_SIGBUS;
2182 }
2183
2184 /*
2185  * Having a close hook prevents vma merging regardless of flags.
2186  */
2187 static void special_mapping_close(struct vm_area_struct *vma)
2188 {
2189 }
2190
2191 static struct vm_operations_struct special_mapping_vmops = {
2192         .close = special_mapping_close,
2193         .nopage = special_mapping_nopage,
2194 };
2195
2196 /*
2197  * Called with mm->mmap_sem held for writing.
2198  * Insert a new vma covering the given region, with the given flags.
2199  * Its pages are supplied by the given array of struct page *.
2200  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2201  * The region past the last page supplied will always produce SIGBUS.
2202  * The array pointer and the pages it points to are assumed to stay alive
2203  * for as long as this mapping might exist.
2204  */
2205 int install_special_mapping(struct mm_struct *mm,
2206                             unsigned long addr, unsigned long len,
2207                             unsigned long vm_flags, struct page **pages)
2208 {
2209         struct vm_area_struct *vma;
2210
2211         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2212         if (unlikely(vma == NULL))
2213                 return -ENOMEM;
2214
2215         vma->vm_mm = mm;
2216         vma->vm_start = addr;
2217         vma->vm_end = addr + len;
2218
2219         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2220         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2221
2222         vma->vm_ops = &special_mapping_vmops;
2223         vma->vm_private_data = pages;
2224
2225         if (unlikely(insert_vm_struct(mm, vma))) {
2226                 kmem_cache_free(vm_area_cachep, vma);
2227                 return -ENOMEM;
2228         }
2229
2230         mm->total_vm += len >> PAGE_SHIFT;
2231
2232         return 0;
2233 }