memcg: remove redundant initialization in mem_cgroup_create()
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@redhat.com>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29
30 #include <asm/uaccess.h>
31 #include <asm/cacheflush.h>
32 #include <asm/tlb.h>
33 #include <asm/mmu_context.h>
34
35 #ifndef arch_mmap_check
36 #define arch_mmap_check(addr, len, flags)       (0)
37 #endif
38
39 #ifndef arch_rebalance_pgtables
40 #define arch_rebalance_pgtables(addr, len)              (addr)
41 #endif
42
43 static void unmap_region(struct mm_struct *mm,
44                 struct vm_area_struct *vma, struct vm_area_struct *prev,
45                 unsigned long start, unsigned long end);
46
47 /*
48  * WARNING: the debugging will use recursive algorithms so never enable this
49  * unless you know what you are doing.
50  */
51 #undef DEBUG_MM_RB
52
53 /* description of effects of mapping type and prot in current implementation.
54  * this is due to the limited x86 page protection hardware.  The expected
55  * behavior is in parens:
56  *
57  * map_type     prot
58  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
59  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
60  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
61  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
62  *              
63  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
64  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
65  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
66  *
67  */
68 pgprot_t protection_map[16] = {
69         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
70         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
71 };
72
73 pgprot_t vm_get_page_prot(unsigned long vm_flags)
74 {
75         return protection_map[vm_flags &
76                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)];
77 }
78 EXPORT_SYMBOL(vm_get_page_prot);
79
80 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
81 int sysctl_overcommit_ratio = 50;       /* default is 50% */
82 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
83 atomic_t vm_committed_space = ATOMIC_INIT(0);
84
85 /*
86  * Check that a process has enough memory to allocate a new virtual
87  * mapping. 0 means there is enough memory for the allocation to
88  * succeed and -ENOMEM implies there is not.
89  *
90  * We currently support three overcommit policies, which are set via the
91  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
92  *
93  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
94  * Additional code 2002 Jul 20 by Robert Love.
95  *
96  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
97  *
98  * Note this is a helper function intended to be used by LSMs which
99  * wish to use this logic.
100  */
101 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
102 {
103         unsigned long free, allowed;
104
105         vm_acct_memory(pages);
106
107         /*
108          * Sometimes we want to use more memory than we have
109          */
110         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
111                 return 0;
112
113         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
114                 unsigned long n;
115
116                 free = global_page_state(NR_FILE_PAGES);
117                 free += nr_swap_pages;
118
119                 /*
120                  * Any slabs which are created with the
121                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
122                  * which are reclaimable, under pressure.  The dentry
123                  * cache and most inode caches should fall into this
124                  */
125                 free += global_page_state(NR_SLAB_RECLAIMABLE);
126
127                 /*
128                  * Leave the last 3% for root
129                  */
130                 if (!cap_sys_admin)
131                         free -= free / 32;
132
133                 if (free > pages)
134                         return 0;
135
136                 /*
137                  * nr_free_pages() is very expensive on large systems,
138                  * only call if we're about to fail.
139                  */
140                 n = nr_free_pages();
141
142                 /*
143                  * Leave reserved pages. The pages are not for anonymous pages.
144                  */
145                 if (n <= totalreserve_pages)
146                         goto error;
147                 else
148                         n -= totalreserve_pages;
149
150                 /*
151                  * Leave the last 3% for root
152                  */
153                 if (!cap_sys_admin)
154                         n -= n / 32;
155                 free += n;
156
157                 if (free > pages)
158                         return 0;
159
160                 goto error;
161         }
162
163         allowed = (totalram_pages - hugetlb_total_pages())
164                 * sysctl_overcommit_ratio / 100;
165         /*
166          * Leave the last 3% for root
167          */
168         if (!cap_sys_admin)
169                 allowed -= allowed / 32;
170         allowed += total_swap_pages;
171
172         /* Don't let a single process grow too big:
173            leave 3% of the size of this process for other processes */
174         allowed -= mm->total_vm / 32;
175
176         /*
177          * cast `allowed' as a signed long because vm_committed_space
178          * sometimes has a negative value
179          */
180         if (atomic_read(&vm_committed_space) < (long)allowed)
181                 return 0;
182 error:
183         vm_unacct_memory(pages);
184
185         return -ENOMEM;
186 }
187
188 /*
189  * Requires inode->i_mapping->i_mmap_lock
190  */
191 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
192                 struct file *file, struct address_space *mapping)
193 {
194         if (vma->vm_flags & VM_DENYWRITE)
195                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
196         if (vma->vm_flags & VM_SHARED)
197                 mapping->i_mmap_writable--;
198
199         flush_dcache_mmap_lock(mapping);
200         if (unlikely(vma->vm_flags & VM_NONLINEAR))
201                 list_del_init(&vma->shared.vm_set.list);
202         else
203                 vma_prio_tree_remove(vma, &mapping->i_mmap);
204         flush_dcache_mmap_unlock(mapping);
205 }
206
207 /*
208  * Unlink a file-based vm structure from its prio_tree, to hide
209  * vma from rmap and vmtruncate before freeing its page tables.
210  */
211 void unlink_file_vma(struct vm_area_struct *vma)
212 {
213         struct file *file = vma->vm_file;
214
215         if (file) {
216                 struct address_space *mapping = file->f_mapping;
217                 spin_lock(&mapping->i_mmap_lock);
218                 __remove_shared_vm_struct(vma, file, mapping);
219                 spin_unlock(&mapping->i_mmap_lock);
220         }
221 }
222
223 /*
224  * Close a vm structure and free it, returning the next.
225  */
226 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
227 {
228         struct vm_area_struct *next = vma->vm_next;
229
230         might_sleep();
231         if (vma->vm_ops && vma->vm_ops->close)
232                 vma->vm_ops->close(vma);
233         if (vma->vm_file)
234                 fput(vma->vm_file);
235         mpol_put(vma_policy(vma));
236         kmem_cache_free(vm_area_cachep, vma);
237         return next;
238 }
239
240 asmlinkage unsigned long sys_brk(unsigned long brk)
241 {
242         unsigned long rlim, retval;
243         unsigned long newbrk, oldbrk;
244         struct mm_struct *mm = current->mm;
245
246         down_write(&mm->mmap_sem);
247
248         if (brk < mm->start_brk)
249                 goto out;
250
251         /*
252          * Check against rlimit here. If this check is done later after the test
253          * of oldbrk with newbrk then it can escape the test and let the data
254          * segment grow beyond its set limit the in case where the limit is
255          * not page aligned -Ram Gupta
256          */
257         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
258         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
259                         (mm->end_data - mm->start_data) > rlim)
260                 goto out;
261
262         newbrk = PAGE_ALIGN(brk);
263         oldbrk = PAGE_ALIGN(mm->brk);
264         if (oldbrk == newbrk)
265                 goto set_brk;
266
267         /* Always allow shrinking brk. */
268         if (brk <= mm->brk) {
269                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
270                         goto set_brk;
271                 goto out;
272         }
273
274         /* Check against existing mmap mappings. */
275         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
276                 goto out;
277
278         /* Ok, looks good - let it rip. */
279         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
280                 goto out;
281 set_brk:
282         mm->brk = brk;
283 out:
284         retval = mm->brk;
285         up_write(&mm->mmap_sem);
286         return retval;
287 }
288
289 #ifdef DEBUG_MM_RB
290 static int browse_rb(struct rb_root *root)
291 {
292         int i = 0, j;
293         struct rb_node *nd, *pn = NULL;
294         unsigned long prev = 0, pend = 0;
295
296         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
297                 struct vm_area_struct *vma;
298                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
299                 if (vma->vm_start < prev)
300                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
301                 if (vma->vm_start < pend)
302                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
303                 if (vma->vm_start > vma->vm_end)
304                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
305                 i++;
306                 pn = nd;
307                 prev = vma->vm_start;
308                 pend = vma->vm_end;
309         }
310         j = 0;
311         for (nd = pn; nd; nd = rb_prev(nd)) {
312                 j++;
313         }
314         if (i != j)
315                 printk("backwards %d, forwards %d\n", j, i), i = 0;
316         return i;
317 }
318
319 void validate_mm(struct mm_struct *mm)
320 {
321         int bug = 0;
322         int i = 0;
323         struct vm_area_struct *tmp = mm->mmap;
324         while (tmp) {
325                 tmp = tmp->vm_next;
326                 i++;
327         }
328         if (i != mm->map_count)
329                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
330         i = browse_rb(&mm->mm_rb);
331         if (i != mm->map_count)
332                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
333         BUG_ON(bug);
334 }
335 #else
336 #define validate_mm(mm) do { } while (0)
337 #endif
338
339 static struct vm_area_struct *
340 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
341                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
342                 struct rb_node ** rb_parent)
343 {
344         struct vm_area_struct * vma;
345         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
346
347         __rb_link = &mm->mm_rb.rb_node;
348         rb_prev = __rb_parent = NULL;
349         vma = NULL;
350
351         while (*__rb_link) {
352                 struct vm_area_struct *vma_tmp;
353
354                 __rb_parent = *__rb_link;
355                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
356
357                 if (vma_tmp->vm_end > addr) {
358                         vma = vma_tmp;
359                         if (vma_tmp->vm_start <= addr)
360                                 return vma;
361                         __rb_link = &__rb_parent->rb_left;
362                 } else {
363                         rb_prev = __rb_parent;
364                         __rb_link = &__rb_parent->rb_right;
365                 }
366         }
367
368         *pprev = NULL;
369         if (rb_prev)
370                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
371         *rb_link = __rb_link;
372         *rb_parent = __rb_parent;
373         return vma;
374 }
375
376 static inline void
377 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
378                 struct vm_area_struct *prev, struct rb_node *rb_parent)
379 {
380         if (prev) {
381                 vma->vm_next = prev->vm_next;
382                 prev->vm_next = vma;
383         } else {
384                 mm->mmap = vma;
385                 if (rb_parent)
386                         vma->vm_next = rb_entry(rb_parent,
387                                         struct vm_area_struct, vm_rb);
388                 else
389                         vma->vm_next = NULL;
390         }
391 }
392
393 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
394                 struct rb_node **rb_link, struct rb_node *rb_parent)
395 {
396         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
397         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
398 }
399
400 static inline void __vma_link_file(struct vm_area_struct *vma)
401 {
402         struct file * file;
403
404         file = vma->vm_file;
405         if (file) {
406                 struct address_space *mapping = file->f_mapping;
407
408                 if (vma->vm_flags & VM_DENYWRITE)
409                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
410                 if (vma->vm_flags & VM_SHARED)
411                         mapping->i_mmap_writable++;
412
413                 flush_dcache_mmap_lock(mapping);
414                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
415                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
416                 else
417                         vma_prio_tree_insert(vma, &mapping->i_mmap);
418                 flush_dcache_mmap_unlock(mapping);
419         }
420 }
421
422 static void
423 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
424         struct vm_area_struct *prev, struct rb_node **rb_link,
425         struct rb_node *rb_parent)
426 {
427         __vma_link_list(mm, vma, prev, rb_parent);
428         __vma_link_rb(mm, vma, rb_link, rb_parent);
429         __anon_vma_link(vma);
430 }
431
432 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
433                         struct vm_area_struct *prev, struct rb_node **rb_link,
434                         struct rb_node *rb_parent)
435 {
436         struct address_space *mapping = NULL;
437
438         if (vma->vm_file)
439                 mapping = vma->vm_file->f_mapping;
440
441         if (mapping) {
442                 spin_lock(&mapping->i_mmap_lock);
443                 vma->vm_truncate_count = mapping->truncate_count;
444         }
445         anon_vma_lock(vma);
446
447         __vma_link(mm, vma, prev, rb_link, rb_parent);
448         __vma_link_file(vma);
449
450         anon_vma_unlock(vma);
451         if (mapping)
452                 spin_unlock(&mapping->i_mmap_lock);
453
454         mm->map_count++;
455         validate_mm(mm);
456 }
457
458 /*
459  * Helper for vma_adjust in the split_vma insert case:
460  * insert vm structure into list and rbtree and anon_vma,
461  * but it has already been inserted into prio_tree earlier.
462  */
463 static void
464 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
465 {
466         struct vm_area_struct * __vma, * prev;
467         struct rb_node ** rb_link, * rb_parent;
468
469         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
470         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
471         __vma_link(mm, vma, prev, rb_link, rb_parent);
472         mm->map_count++;
473 }
474
475 static inline void
476 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
477                 struct vm_area_struct *prev)
478 {
479         prev->vm_next = vma->vm_next;
480         rb_erase(&vma->vm_rb, &mm->mm_rb);
481         if (mm->mmap_cache == vma)
482                 mm->mmap_cache = prev;
483 }
484
485 /*
486  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
487  * is already present in an i_mmap tree without adjusting the tree.
488  * The following helper function should be used when such adjustments
489  * are necessary.  The "insert" vma (if any) is to be inserted
490  * before we drop the necessary locks.
491  */
492 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
493         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
494 {
495         struct mm_struct *mm = vma->vm_mm;
496         struct vm_area_struct *next = vma->vm_next;
497         struct vm_area_struct *importer = NULL;
498         struct address_space *mapping = NULL;
499         struct prio_tree_root *root = NULL;
500         struct file *file = vma->vm_file;
501         struct anon_vma *anon_vma = NULL;
502         long adjust_next = 0;
503         int remove_next = 0;
504
505         if (next && !insert) {
506                 if (end >= next->vm_end) {
507                         /*
508                          * vma expands, overlapping all the next, and
509                          * perhaps the one after too (mprotect case 6).
510                          */
511 again:                  remove_next = 1 + (end > next->vm_end);
512                         end = next->vm_end;
513                         anon_vma = next->anon_vma;
514                         importer = vma;
515                 } else if (end > next->vm_start) {
516                         /*
517                          * vma expands, overlapping part of the next:
518                          * mprotect case 5 shifting the boundary up.
519                          */
520                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
521                         anon_vma = next->anon_vma;
522                         importer = vma;
523                 } else if (end < vma->vm_end) {
524                         /*
525                          * vma shrinks, and !insert tells it's not
526                          * split_vma inserting another: so it must be
527                          * mprotect case 4 shifting the boundary down.
528                          */
529                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
530                         anon_vma = next->anon_vma;
531                         importer = next;
532                 }
533         }
534
535         if (file) {
536                 mapping = file->f_mapping;
537                 if (!(vma->vm_flags & VM_NONLINEAR))
538                         root = &mapping->i_mmap;
539                 spin_lock(&mapping->i_mmap_lock);
540                 if (importer &&
541                     vma->vm_truncate_count != next->vm_truncate_count) {
542                         /*
543                          * unmap_mapping_range might be in progress:
544                          * ensure that the expanding vma is rescanned.
545                          */
546                         importer->vm_truncate_count = 0;
547                 }
548                 if (insert) {
549                         insert->vm_truncate_count = vma->vm_truncate_count;
550                         /*
551                          * Put into prio_tree now, so instantiated pages
552                          * are visible to arm/parisc __flush_dcache_page
553                          * throughout; but we cannot insert into address
554                          * space until vma start or end is updated.
555                          */
556                         __vma_link_file(insert);
557                 }
558         }
559
560         /*
561          * When changing only vma->vm_end, we don't really need
562          * anon_vma lock: but is that case worth optimizing out?
563          */
564         if (vma->anon_vma)
565                 anon_vma = vma->anon_vma;
566         if (anon_vma) {
567                 spin_lock(&anon_vma->lock);
568                 /*
569                  * Easily overlooked: when mprotect shifts the boundary,
570                  * make sure the expanding vma has anon_vma set if the
571                  * shrinking vma had, to cover any anon pages imported.
572                  */
573                 if (importer && !importer->anon_vma) {
574                         importer->anon_vma = anon_vma;
575                         __anon_vma_link(importer);
576                 }
577         }
578
579         if (root) {
580                 flush_dcache_mmap_lock(mapping);
581                 vma_prio_tree_remove(vma, root);
582                 if (adjust_next)
583                         vma_prio_tree_remove(next, root);
584         }
585
586         vma->vm_start = start;
587         vma->vm_end = end;
588         vma->vm_pgoff = pgoff;
589         if (adjust_next) {
590                 next->vm_start += adjust_next << PAGE_SHIFT;
591                 next->vm_pgoff += adjust_next;
592         }
593
594         if (root) {
595                 if (adjust_next)
596                         vma_prio_tree_insert(next, root);
597                 vma_prio_tree_insert(vma, root);
598                 flush_dcache_mmap_unlock(mapping);
599         }
600
601         if (remove_next) {
602                 /*
603                  * vma_merge has merged next into vma, and needs
604                  * us to remove next before dropping the locks.
605                  */
606                 __vma_unlink(mm, next, vma);
607                 if (file)
608                         __remove_shared_vm_struct(next, file, mapping);
609                 if (next->anon_vma)
610                         __anon_vma_merge(vma, next);
611         } else if (insert) {
612                 /*
613                  * split_vma has split insert from vma, and needs
614                  * us to insert it before dropping the locks
615                  * (it may either follow vma or precede it).
616                  */
617                 __insert_vm_struct(mm, insert);
618         }
619
620         if (anon_vma)
621                 spin_unlock(&anon_vma->lock);
622         if (mapping)
623                 spin_unlock(&mapping->i_mmap_lock);
624
625         if (remove_next) {
626                 if (file)
627                         fput(file);
628                 mm->map_count--;
629                 mpol_put(vma_policy(next));
630                 kmem_cache_free(vm_area_cachep, next);
631                 /*
632                  * In mprotect's case 6 (see comments on vma_merge),
633                  * we must remove another next too. It would clutter
634                  * up the code too much to do both in one go.
635                  */
636                 if (remove_next == 2) {
637                         next = vma->vm_next;
638                         goto again;
639                 }
640         }
641
642         validate_mm(mm);
643 }
644
645 /*
646  * If the vma has a ->close operation then the driver probably needs to release
647  * per-vma resources, so we don't attempt to merge those.
648  */
649 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
650
651 static inline int is_mergeable_vma(struct vm_area_struct *vma,
652                         struct file *file, unsigned long vm_flags)
653 {
654         if (vma->vm_flags != vm_flags)
655                 return 0;
656         if (vma->vm_file != file)
657                 return 0;
658         if (vma->vm_ops && vma->vm_ops->close)
659                 return 0;
660         return 1;
661 }
662
663 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
664                                         struct anon_vma *anon_vma2)
665 {
666         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
667 }
668
669 /*
670  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
671  * in front of (at a lower virtual address and file offset than) the vma.
672  *
673  * We cannot merge two vmas if they have differently assigned (non-NULL)
674  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
675  *
676  * We don't check here for the merged mmap wrapping around the end of pagecache
677  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
678  * wrap, nor mmaps which cover the final page at index -1UL.
679  */
680 static int
681 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
682         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
683 {
684         if (is_mergeable_vma(vma, file, vm_flags) &&
685             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
686                 if (vma->vm_pgoff == vm_pgoff)
687                         return 1;
688         }
689         return 0;
690 }
691
692 /*
693  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
694  * beyond (at a higher virtual address and file offset than) the vma.
695  *
696  * We cannot merge two vmas if they have differently assigned (non-NULL)
697  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
698  */
699 static int
700 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
701         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
702 {
703         if (is_mergeable_vma(vma, file, vm_flags) &&
704             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
705                 pgoff_t vm_pglen;
706                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
707                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
708                         return 1;
709         }
710         return 0;
711 }
712
713 /*
714  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
715  * whether that can be merged with its predecessor or its successor.
716  * Or both (it neatly fills a hole).
717  *
718  * In most cases - when called for mmap, brk or mremap - [addr,end) is
719  * certain not to be mapped by the time vma_merge is called; but when
720  * called for mprotect, it is certain to be already mapped (either at
721  * an offset within prev, or at the start of next), and the flags of
722  * this area are about to be changed to vm_flags - and the no-change
723  * case has already been eliminated.
724  *
725  * The following mprotect cases have to be considered, where AAAA is
726  * the area passed down from mprotect_fixup, never extending beyond one
727  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
728  *
729  *     AAAA             AAAA                AAAA          AAAA
730  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
731  *    cannot merge    might become    might become    might become
732  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
733  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
734  *    mremap move:                                    PPPPNNNNNNNN 8
735  *        AAAA
736  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
737  *    might become    case 1 below    case 2 below    case 3 below
738  *
739  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
740  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
741  */
742 struct vm_area_struct *vma_merge(struct mm_struct *mm,
743                         struct vm_area_struct *prev, unsigned long addr,
744                         unsigned long end, unsigned long vm_flags,
745                         struct anon_vma *anon_vma, struct file *file,
746                         pgoff_t pgoff, struct mempolicy *policy)
747 {
748         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
749         struct vm_area_struct *area, *next;
750
751         /*
752          * We later require that vma->vm_flags == vm_flags,
753          * so this tests vma->vm_flags & VM_SPECIAL, too.
754          */
755         if (vm_flags & VM_SPECIAL)
756                 return NULL;
757
758         if (prev)
759                 next = prev->vm_next;
760         else
761                 next = mm->mmap;
762         area = next;
763         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
764                 next = next->vm_next;
765
766         /*
767          * Can it merge with the predecessor?
768          */
769         if (prev && prev->vm_end == addr &&
770                         mpol_equal(vma_policy(prev), policy) &&
771                         can_vma_merge_after(prev, vm_flags,
772                                                 anon_vma, file, pgoff)) {
773                 /*
774                  * OK, it can.  Can we now merge in the successor as well?
775                  */
776                 if (next && end == next->vm_start &&
777                                 mpol_equal(policy, vma_policy(next)) &&
778                                 can_vma_merge_before(next, vm_flags,
779                                         anon_vma, file, pgoff+pglen) &&
780                                 is_mergeable_anon_vma(prev->anon_vma,
781                                                       next->anon_vma)) {
782                                                         /* cases 1, 6 */
783                         vma_adjust(prev, prev->vm_start,
784                                 next->vm_end, prev->vm_pgoff, NULL);
785                 } else                                  /* cases 2, 5, 7 */
786                         vma_adjust(prev, prev->vm_start,
787                                 end, prev->vm_pgoff, NULL);
788                 return prev;
789         }
790
791         /*
792          * Can this new request be merged in front of next?
793          */
794         if (next && end == next->vm_start &&
795                         mpol_equal(policy, vma_policy(next)) &&
796                         can_vma_merge_before(next, vm_flags,
797                                         anon_vma, file, pgoff+pglen)) {
798                 if (prev && addr < prev->vm_end)        /* case 4 */
799                         vma_adjust(prev, prev->vm_start,
800                                 addr, prev->vm_pgoff, NULL);
801                 else                                    /* cases 3, 8 */
802                         vma_adjust(area, addr, next->vm_end,
803                                 next->vm_pgoff - pglen, NULL);
804                 return area;
805         }
806
807         return NULL;
808 }
809
810 /*
811  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
812  * neighbouring vmas for a suitable anon_vma, before it goes off
813  * to allocate a new anon_vma.  It checks because a repetitive
814  * sequence of mprotects and faults may otherwise lead to distinct
815  * anon_vmas being allocated, preventing vma merge in subsequent
816  * mprotect.
817  */
818 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
819 {
820         struct vm_area_struct *near;
821         unsigned long vm_flags;
822
823         near = vma->vm_next;
824         if (!near)
825                 goto try_prev;
826
827         /*
828          * Since only mprotect tries to remerge vmas, match flags
829          * which might be mprotected into each other later on.
830          * Neither mlock nor madvise tries to remerge at present,
831          * so leave their flags as obstructing a merge.
832          */
833         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
834         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
835
836         if (near->anon_vma && vma->vm_end == near->vm_start &&
837                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
838                         can_vma_merge_before(near, vm_flags,
839                                 NULL, vma->vm_file, vma->vm_pgoff +
840                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
841                 return near->anon_vma;
842 try_prev:
843         /*
844          * It is potentially slow to have to call find_vma_prev here.
845          * But it's only on the first write fault on the vma, not
846          * every time, and we could devise a way to avoid it later
847          * (e.g. stash info in next's anon_vma_node when assigning
848          * an anon_vma, or when trying vma_merge).  Another time.
849          */
850         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
851         if (!near)
852                 goto none;
853
854         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
855         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
856
857         if (near->anon_vma && near->vm_end == vma->vm_start &&
858                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
859                         can_vma_merge_after(near, vm_flags,
860                                 NULL, vma->vm_file, vma->vm_pgoff))
861                 return near->anon_vma;
862 none:
863         /*
864          * There's no absolute need to look only at touching neighbours:
865          * we could search further afield for "compatible" anon_vmas.
866          * But it would probably just be a waste of time searching,
867          * or lead to too many vmas hanging off the same anon_vma.
868          * We're trying to allow mprotect remerging later on,
869          * not trying to minimize memory used for anon_vmas.
870          */
871         return NULL;
872 }
873
874 #ifdef CONFIG_PROC_FS
875 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
876                                                 struct file *file, long pages)
877 {
878         const unsigned long stack_flags
879                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
880
881         if (file) {
882                 mm->shared_vm += pages;
883                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
884                         mm->exec_vm += pages;
885         } else if (flags & stack_flags)
886                 mm->stack_vm += pages;
887         if (flags & (VM_RESERVED|VM_IO))
888                 mm->reserved_vm += pages;
889 }
890 #endif /* CONFIG_PROC_FS */
891
892 /*
893  * The caller must hold down_write(current->mm->mmap_sem).
894  */
895
896 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
897                         unsigned long len, unsigned long prot,
898                         unsigned long flags, unsigned long pgoff)
899 {
900         struct mm_struct * mm = current->mm;
901         struct inode *inode;
902         unsigned int vm_flags;
903         int error;
904         int accountable = 1;
905         unsigned long reqprot = prot;
906
907         /*
908          * Does the application expect PROT_READ to imply PROT_EXEC?
909          *
910          * (the exception is when the underlying filesystem is noexec
911          *  mounted, in which case we dont add PROT_EXEC.)
912          */
913         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
914                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
915                         prot |= PROT_EXEC;
916
917         if (!len)
918                 return -EINVAL;
919
920         if (!(flags & MAP_FIXED))
921                 addr = round_hint_to_min(addr);
922
923         error = arch_mmap_check(addr, len, flags);
924         if (error)
925                 return error;
926
927         /* Careful about overflows.. */
928         len = PAGE_ALIGN(len);
929         if (!len || len > TASK_SIZE)
930                 return -ENOMEM;
931
932         /* offset overflow? */
933         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
934                return -EOVERFLOW;
935
936         /* Too many mappings? */
937         if (mm->map_count > sysctl_max_map_count)
938                 return -ENOMEM;
939
940         /* Obtain the address to map to. we verify (or select) it and ensure
941          * that it represents a valid section of the address space.
942          */
943         addr = get_unmapped_area(file, addr, len, pgoff, flags);
944         if (addr & ~PAGE_MASK)
945                 return addr;
946
947         /* Do simple checking here so the lower-level routines won't have
948          * to. we assume access permissions have been handled by the open
949          * of the memory object, so we don't do any here.
950          */
951         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
952                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
953
954         if (flags & MAP_LOCKED) {
955                 if (!can_do_mlock())
956                         return -EPERM;
957                 vm_flags |= VM_LOCKED;
958         }
959         /* mlock MCL_FUTURE? */
960         if (vm_flags & VM_LOCKED) {
961                 unsigned long locked, lock_limit;
962                 locked = len >> PAGE_SHIFT;
963                 locked += mm->locked_vm;
964                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
965                 lock_limit >>= PAGE_SHIFT;
966                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
967                         return -EAGAIN;
968         }
969
970         inode = file ? file->f_path.dentry->d_inode : NULL;
971
972         if (file) {
973                 switch (flags & MAP_TYPE) {
974                 case MAP_SHARED:
975                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
976                                 return -EACCES;
977
978                         /*
979                          * Make sure we don't allow writing to an append-only
980                          * file..
981                          */
982                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
983                                 return -EACCES;
984
985                         /*
986                          * Make sure there are no mandatory locks on the file.
987                          */
988                         if (locks_verify_locked(inode))
989                                 return -EAGAIN;
990
991                         vm_flags |= VM_SHARED | VM_MAYSHARE;
992                         if (!(file->f_mode & FMODE_WRITE))
993                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
994
995                         /* fall through */
996                 case MAP_PRIVATE:
997                         if (!(file->f_mode & FMODE_READ))
998                                 return -EACCES;
999                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1000                                 if (vm_flags & VM_EXEC)
1001                                         return -EPERM;
1002                                 vm_flags &= ~VM_MAYEXEC;
1003                         }
1004                         if (is_file_hugepages(file))
1005                                 accountable = 0;
1006
1007                         if (!file->f_op || !file->f_op->mmap)
1008                                 return -ENODEV;
1009                         break;
1010
1011                 default:
1012                         return -EINVAL;
1013                 }
1014         } else {
1015                 switch (flags & MAP_TYPE) {
1016                 case MAP_SHARED:
1017                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1018                         break;
1019                 case MAP_PRIVATE:
1020                         /*
1021                          * Set pgoff according to addr for anon_vma.
1022                          */
1023                         pgoff = addr >> PAGE_SHIFT;
1024                         break;
1025                 default:
1026                         return -EINVAL;
1027                 }
1028         }
1029
1030         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1031         if (error)
1032                 return error;
1033
1034         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1035                            accountable);
1036 }
1037 EXPORT_SYMBOL(do_mmap_pgoff);
1038
1039 /*
1040  * Some shared mappigns will want the pages marked read-only
1041  * to track write events. If so, we'll downgrade vm_page_prot
1042  * to the private version (using protection_map[] without the
1043  * VM_SHARED bit).
1044  */
1045 int vma_wants_writenotify(struct vm_area_struct *vma)
1046 {
1047         unsigned int vm_flags = vma->vm_flags;
1048
1049         /* If it was private or non-writable, the write bit is already clear */
1050         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1051                 return 0;
1052
1053         /* The backer wishes to know when pages are first written to? */
1054         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1055                 return 1;
1056
1057         /* The open routine did something to the protections already? */
1058         if (pgprot_val(vma->vm_page_prot) !=
1059             pgprot_val(vm_get_page_prot(vm_flags)))
1060                 return 0;
1061
1062         /* Specialty mapping? */
1063         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1064                 return 0;
1065
1066         /* Can the mapping track the dirty pages? */
1067         return vma->vm_file && vma->vm_file->f_mapping &&
1068                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1069 }
1070
1071 unsigned long mmap_region(struct file *file, unsigned long addr,
1072                           unsigned long len, unsigned long flags,
1073                           unsigned int vm_flags, unsigned long pgoff,
1074                           int accountable)
1075 {
1076         struct mm_struct *mm = current->mm;
1077         struct vm_area_struct *vma, *prev;
1078         int correct_wcount = 0;
1079         int error;
1080         struct rb_node **rb_link, *rb_parent;
1081         unsigned long charged = 0;
1082         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1083
1084         /* Clear old maps */
1085         error = -ENOMEM;
1086 munmap_back:
1087         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1088         if (vma && vma->vm_start < addr + len) {
1089                 if (do_munmap(mm, addr, len))
1090                         return -ENOMEM;
1091                 goto munmap_back;
1092         }
1093
1094         /* Check against address space limit. */
1095         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1096                 return -ENOMEM;
1097
1098         if (accountable && (!(flags & MAP_NORESERVE) ||
1099                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1100                 if (vm_flags & VM_SHARED) {
1101                         /* Check memory availability in shmem_file_setup? */
1102                         vm_flags |= VM_ACCOUNT;
1103                 } else if (vm_flags & VM_WRITE) {
1104                         /*
1105                          * Private writable mapping: check memory availability
1106                          */
1107                         charged = len >> PAGE_SHIFT;
1108                         if (security_vm_enough_memory(charged))
1109                                 return -ENOMEM;
1110                         vm_flags |= VM_ACCOUNT;
1111                 }
1112         }
1113
1114         /*
1115          * Can we just expand an old private anonymous mapping?
1116          * The VM_SHARED test is necessary because shmem_zero_setup
1117          * will create the file object for a shared anonymous map below.
1118          */
1119         if (!file && !(vm_flags & VM_SHARED) &&
1120             vma_merge(mm, prev, addr, addr + len, vm_flags,
1121                                         NULL, NULL, pgoff, NULL))
1122                 goto out;
1123
1124         /*
1125          * Determine the object being mapped and call the appropriate
1126          * specific mapper. the address has already been validated, but
1127          * not unmapped, but the maps are removed from the list.
1128          */
1129         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1130         if (!vma) {
1131                 error = -ENOMEM;
1132                 goto unacct_error;
1133         }
1134
1135         vma->vm_mm = mm;
1136         vma->vm_start = addr;
1137         vma->vm_end = addr + len;
1138         vma->vm_flags = vm_flags;
1139         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1140         vma->vm_pgoff = pgoff;
1141
1142         if (file) {
1143                 error = -EINVAL;
1144                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1145                         goto free_vma;
1146                 if (vm_flags & VM_DENYWRITE) {
1147                         error = deny_write_access(file);
1148                         if (error)
1149                                 goto free_vma;
1150                         correct_wcount = 1;
1151                 }
1152                 vma->vm_file = file;
1153                 get_file(file);
1154                 error = file->f_op->mmap(file, vma);
1155                 if (error)
1156                         goto unmap_and_free_vma;
1157         } else if (vm_flags & VM_SHARED) {
1158                 error = shmem_zero_setup(vma);
1159                 if (error)
1160                         goto free_vma;
1161         }
1162
1163         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1164          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1165          * that memory reservation must be checked; but that reservation
1166          * belongs to shared memory object, not to vma: so now clear it.
1167          */
1168         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1169                 vma->vm_flags &= ~VM_ACCOUNT;
1170
1171         /* Can addr have changed??
1172          *
1173          * Answer: Yes, several device drivers can do it in their
1174          *         f_op->mmap method. -DaveM
1175          */
1176         addr = vma->vm_start;
1177         pgoff = vma->vm_pgoff;
1178         vm_flags = vma->vm_flags;
1179
1180         if (vma_wants_writenotify(vma))
1181                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1182
1183         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1184                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1185                 mpol_put(vma_policy(vma));
1186                 kmem_cache_free(vm_area_cachep, vma);
1187                 fput(file);
1188         } else {
1189                 vma_link(mm, vma, prev, rb_link, rb_parent);
1190                 file = vma->vm_file;
1191         }
1192
1193         /* Once vma denies write, undo our temporary denial count */
1194         if (correct_wcount)
1195                 atomic_inc(&inode->i_writecount);
1196 out:
1197         mm->total_vm += len >> PAGE_SHIFT;
1198         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1199         if (vm_flags & VM_LOCKED) {
1200                 mm->locked_vm += len >> PAGE_SHIFT;
1201                 make_pages_present(addr, addr + len);
1202         }
1203         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1204                 make_pages_present(addr, addr + len);
1205         return addr;
1206
1207 unmap_and_free_vma:
1208         if (correct_wcount)
1209                 atomic_inc(&inode->i_writecount);
1210         vma->vm_file = NULL;
1211         fput(file);
1212
1213         /* Undo any partial mapping done by a device driver. */
1214         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1215         charged = 0;
1216 free_vma:
1217         kmem_cache_free(vm_area_cachep, vma);
1218 unacct_error:
1219         if (charged)
1220                 vm_unacct_memory(charged);
1221         return error;
1222 }
1223
1224 /* Get an address range which is currently unmapped.
1225  * For shmat() with addr=0.
1226  *
1227  * Ugly calling convention alert:
1228  * Return value with the low bits set means error value,
1229  * ie
1230  *      if (ret & ~PAGE_MASK)
1231  *              error = ret;
1232  *
1233  * This function "knows" that -ENOMEM has the bits set.
1234  */
1235 #ifndef HAVE_ARCH_UNMAPPED_AREA
1236 unsigned long
1237 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1238                 unsigned long len, unsigned long pgoff, unsigned long flags)
1239 {
1240         struct mm_struct *mm = current->mm;
1241         struct vm_area_struct *vma;
1242         unsigned long start_addr;
1243
1244         if (len > TASK_SIZE)
1245                 return -ENOMEM;
1246
1247         if (flags & MAP_FIXED)
1248                 return addr;
1249
1250         if (addr) {
1251                 addr = PAGE_ALIGN(addr);
1252                 vma = find_vma(mm, addr);
1253                 if (TASK_SIZE - len >= addr &&
1254                     (!vma || addr + len <= vma->vm_start))
1255                         return addr;
1256         }
1257         if (len > mm->cached_hole_size) {
1258                 start_addr = addr = mm->free_area_cache;
1259         } else {
1260                 start_addr = addr = TASK_UNMAPPED_BASE;
1261                 mm->cached_hole_size = 0;
1262         }
1263
1264 full_search:
1265         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1266                 /* At this point:  (!vma || addr < vma->vm_end). */
1267                 if (TASK_SIZE - len < addr) {
1268                         /*
1269                          * Start a new search - just in case we missed
1270                          * some holes.
1271                          */
1272                         if (start_addr != TASK_UNMAPPED_BASE) {
1273                                 addr = TASK_UNMAPPED_BASE;
1274                                 start_addr = addr;
1275                                 mm->cached_hole_size = 0;
1276                                 goto full_search;
1277                         }
1278                         return -ENOMEM;
1279                 }
1280                 if (!vma || addr + len <= vma->vm_start) {
1281                         /*
1282                          * Remember the place where we stopped the search:
1283                          */
1284                         mm->free_area_cache = addr + len;
1285                         return addr;
1286                 }
1287                 if (addr + mm->cached_hole_size < vma->vm_start)
1288                         mm->cached_hole_size = vma->vm_start - addr;
1289                 addr = vma->vm_end;
1290         }
1291 }
1292 #endif  
1293
1294 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1295 {
1296         /*
1297          * Is this a new hole at the lowest possible address?
1298          */
1299         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1300                 mm->free_area_cache = addr;
1301                 mm->cached_hole_size = ~0UL;
1302         }
1303 }
1304
1305 /*
1306  * This mmap-allocator allocates new areas top-down from below the
1307  * stack's low limit (the base):
1308  */
1309 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1310 unsigned long
1311 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1312                           const unsigned long len, const unsigned long pgoff,
1313                           const unsigned long flags)
1314 {
1315         struct vm_area_struct *vma;
1316         struct mm_struct *mm = current->mm;
1317         unsigned long addr = addr0;
1318
1319         /* requested length too big for entire address space */
1320         if (len > TASK_SIZE)
1321                 return -ENOMEM;
1322
1323         if (flags & MAP_FIXED)
1324                 return addr;
1325
1326         /* requesting a specific address */
1327         if (addr) {
1328                 addr = PAGE_ALIGN(addr);
1329                 vma = find_vma(mm, addr);
1330                 if (TASK_SIZE - len >= addr &&
1331                                 (!vma || addr + len <= vma->vm_start))
1332                         return addr;
1333         }
1334
1335         /* check if free_area_cache is useful for us */
1336         if (len <= mm->cached_hole_size) {
1337                 mm->cached_hole_size = 0;
1338                 mm->free_area_cache = mm->mmap_base;
1339         }
1340
1341         /* either no address requested or can't fit in requested address hole */
1342         addr = mm->free_area_cache;
1343
1344         /* make sure it can fit in the remaining address space */
1345         if (addr > len) {
1346                 vma = find_vma(mm, addr-len);
1347                 if (!vma || addr <= vma->vm_start)
1348                         /* remember the address as a hint for next time */
1349                         return (mm->free_area_cache = addr-len);
1350         }
1351
1352         if (mm->mmap_base < len)
1353                 goto bottomup;
1354
1355         addr = mm->mmap_base-len;
1356
1357         do {
1358                 /*
1359                  * Lookup failure means no vma is above this address,
1360                  * else if new region fits below vma->vm_start,
1361                  * return with success:
1362                  */
1363                 vma = find_vma(mm, addr);
1364                 if (!vma || addr+len <= vma->vm_start)
1365                         /* remember the address as a hint for next time */
1366                         return (mm->free_area_cache = addr);
1367
1368                 /* remember the largest hole we saw so far */
1369                 if (addr + mm->cached_hole_size < vma->vm_start)
1370                         mm->cached_hole_size = vma->vm_start - addr;
1371
1372                 /* try just below the current vma->vm_start */
1373                 addr = vma->vm_start-len;
1374         } while (len < vma->vm_start);
1375
1376 bottomup:
1377         /*
1378          * A failed mmap() very likely causes application failure,
1379          * so fall back to the bottom-up function here. This scenario
1380          * can happen with large stack limits and large mmap()
1381          * allocations.
1382          */
1383         mm->cached_hole_size = ~0UL;
1384         mm->free_area_cache = TASK_UNMAPPED_BASE;
1385         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1386         /*
1387          * Restore the topdown base:
1388          */
1389         mm->free_area_cache = mm->mmap_base;
1390         mm->cached_hole_size = ~0UL;
1391
1392         return addr;
1393 }
1394 #endif
1395
1396 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1397 {
1398         /*
1399          * Is this a new hole at the highest possible address?
1400          */
1401         if (addr > mm->free_area_cache)
1402                 mm->free_area_cache = addr;
1403
1404         /* dont allow allocations above current base */
1405         if (mm->free_area_cache > mm->mmap_base)
1406                 mm->free_area_cache = mm->mmap_base;
1407 }
1408
1409 unsigned long
1410 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1411                 unsigned long pgoff, unsigned long flags)
1412 {
1413         unsigned long (*get_area)(struct file *, unsigned long,
1414                                   unsigned long, unsigned long, unsigned long);
1415
1416         get_area = current->mm->get_unmapped_area;
1417         if (file && file->f_op && file->f_op->get_unmapped_area)
1418                 get_area = file->f_op->get_unmapped_area;
1419         addr = get_area(file, addr, len, pgoff, flags);
1420         if (IS_ERR_VALUE(addr))
1421                 return addr;
1422
1423         if (addr > TASK_SIZE - len)
1424                 return -ENOMEM;
1425         if (addr & ~PAGE_MASK)
1426                 return -EINVAL;
1427
1428         return arch_rebalance_pgtables(addr, len);
1429 }
1430
1431 EXPORT_SYMBOL(get_unmapped_area);
1432
1433 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1434 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1435 {
1436         struct vm_area_struct *vma = NULL;
1437
1438         if (mm) {
1439                 /* Check the cache first. */
1440                 /* (Cache hit rate is typically around 35%.) */
1441                 vma = mm->mmap_cache;
1442                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1443                         struct rb_node * rb_node;
1444
1445                         rb_node = mm->mm_rb.rb_node;
1446                         vma = NULL;
1447
1448                         while (rb_node) {
1449                                 struct vm_area_struct * vma_tmp;
1450
1451                                 vma_tmp = rb_entry(rb_node,
1452                                                 struct vm_area_struct, vm_rb);
1453
1454                                 if (vma_tmp->vm_end > addr) {
1455                                         vma = vma_tmp;
1456                                         if (vma_tmp->vm_start <= addr)
1457                                                 break;
1458                                         rb_node = rb_node->rb_left;
1459                                 } else
1460                                         rb_node = rb_node->rb_right;
1461                         }
1462                         if (vma)
1463                                 mm->mmap_cache = vma;
1464                 }
1465         }
1466         return vma;
1467 }
1468
1469 EXPORT_SYMBOL(find_vma);
1470
1471 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1472 struct vm_area_struct *
1473 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1474                         struct vm_area_struct **pprev)
1475 {
1476         struct vm_area_struct *vma = NULL, *prev = NULL;
1477         struct rb_node * rb_node;
1478         if (!mm)
1479                 goto out;
1480
1481         /* Guard against addr being lower than the first VMA */
1482         vma = mm->mmap;
1483
1484         /* Go through the RB tree quickly. */
1485         rb_node = mm->mm_rb.rb_node;
1486
1487         while (rb_node) {
1488                 struct vm_area_struct *vma_tmp;
1489                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1490
1491                 if (addr < vma_tmp->vm_end) {
1492                         rb_node = rb_node->rb_left;
1493                 } else {
1494                         prev = vma_tmp;
1495                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1496                                 break;
1497                         rb_node = rb_node->rb_right;
1498                 }
1499         }
1500
1501 out:
1502         *pprev = prev;
1503         return prev ? prev->vm_next : vma;
1504 }
1505
1506 /*
1507  * Verify that the stack growth is acceptable and
1508  * update accounting. This is shared with both the
1509  * grow-up and grow-down cases.
1510  */
1511 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1512 {
1513         struct mm_struct *mm = vma->vm_mm;
1514         struct rlimit *rlim = current->signal->rlim;
1515         unsigned long new_start;
1516
1517         /* address space limit tests */
1518         if (!may_expand_vm(mm, grow))
1519                 return -ENOMEM;
1520
1521         /* Stack limit test */
1522         if (size > rlim[RLIMIT_STACK].rlim_cur)
1523                 return -ENOMEM;
1524
1525         /* mlock limit tests */
1526         if (vma->vm_flags & VM_LOCKED) {
1527                 unsigned long locked;
1528                 unsigned long limit;
1529                 locked = mm->locked_vm + grow;
1530                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1531                 if (locked > limit && !capable(CAP_IPC_LOCK))
1532                         return -ENOMEM;
1533         }
1534
1535         /* Check to ensure the stack will not grow into a hugetlb-only region */
1536         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1537                         vma->vm_end - size;
1538         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1539                 return -EFAULT;
1540
1541         /*
1542          * Overcommit..  This must be the final test, as it will
1543          * update security statistics.
1544          */
1545         if (security_vm_enough_memory(grow))
1546                 return -ENOMEM;
1547
1548         /* Ok, everything looks good - let it rip */
1549         mm->total_vm += grow;
1550         if (vma->vm_flags & VM_LOCKED)
1551                 mm->locked_vm += grow;
1552         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1553         return 0;
1554 }
1555
1556 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1557 /*
1558  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1559  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1560  */
1561 #ifndef CONFIG_IA64
1562 static inline
1563 #endif
1564 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1565 {
1566         int error;
1567
1568         if (!(vma->vm_flags & VM_GROWSUP))
1569                 return -EFAULT;
1570
1571         /*
1572          * We must make sure the anon_vma is allocated
1573          * so that the anon_vma locking is not a noop.
1574          */
1575         if (unlikely(anon_vma_prepare(vma)))
1576                 return -ENOMEM;
1577         anon_vma_lock(vma);
1578
1579         /*
1580          * vma->vm_start/vm_end cannot change under us because the caller
1581          * is required to hold the mmap_sem in read mode.  We need the
1582          * anon_vma lock to serialize against concurrent expand_stacks.
1583          * Also guard against wrapping around to address 0.
1584          */
1585         if (address < PAGE_ALIGN(address+4))
1586                 address = PAGE_ALIGN(address+4);
1587         else {
1588                 anon_vma_unlock(vma);
1589                 return -ENOMEM;
1590         }
1591         error = 0;
1592
1593         /* Somebody else might have raced and expanded it already */
1594         if (address > vma->vm_end) {
1595                 unsigned long size, grow;
1596
1597                 size = address - vma->vm_start;
1598                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1599
1600                 error = acct_stack_growth(vma, size, grow);
1601                 if (!error)
1602                         vma->vm_end = address;
1603         }
1604         anon_vma_unlock(vma);
1605         return error;
1606 }
1607 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1608
1609 /*
1610  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1611  */
1612 static inline int expand_downwards(struct vm_area_struct *vma,
1613                                    unsigned long address)
1614 {
1615         int error;
1616
1617         /*
1618          * We must make sure the anon_vma is allocated
1619          * so that the anon_vma locking is not a noop.
1620          */
1621         if (unlikely(anon_vma_prepare(vma)))
1622                 return -ENOMEM;
1623
1624         address &= PAGE_MASK;
1625         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1626         if (error)
1627                 return error;
1628
1629         anon_vma_lock(vma);
1630
1631         /*
1632          * vma->vm_start/vm_end cannot change under us because the caller
1633          * is required to hold the mmap_sem in read mode.  We need the
1634          * anon_vma lock to serialize against concurrent expand_stacks.
1635          */
1636
1637         /* Somebody else might have raced and expanded it already */
1638         if (address < vma->vm_start) {
1639                 unsigned long size, grow;
1640
1641                 size = vma->vm_end - address;
1642                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1643
1644                 error = acct_stack_growth(vma, size, grow);
1645                 if (!error) {
1646                         vma->vm_start = address;
1647                         vma->vm_pgoff -= grow;
1648                 }
1649         }
1650         anon_vma_unlock(vma);
1651         return error;
1652 }
1653
1654 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1655 {
1656         return expand_downwards(vma, address);
1657 }
1658
1659 #ifdef CONFIG_STACK_GROWSUP
1660 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1661 {
1662         return expand_upwards(vma, address);
1663 }
1664
1665 struct vm_area_struct *
1666 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1667 {
1668         struct vm_area_struct *vma, *prev;
1669
1670         addr &= PAGE_MASK;
1671         vma = find_vma_prev(mm, addr, &prev);
1672         if (vma && (vma->vm_start <= addr))
1673                 return vma;
1674         if (!prev || expand_stack(prev, addr))
1675                 return NULL;
1676         if (prev->vm_flags & VM_LOCKED)
1677                 make_pages_present(addr, prev->vm_end);
1678         return prev;
1679 }
1680 #else
1681 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1682 {
1683         return expand_downwards(vma, address);
1684 }
1685
1686 struct vm_area_struct *
1687 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1688 {
1689         struct vm_area_struct * vma;
1690         unsigned long start;
1691
1692         addr &= PAGE_MASK;
1693         vma = find_vma(mm,addr);
1694         if (!vma)
1695                 return NULL;
1696         if (vma->vm_start <= addr)
1697                 return vma;
1698         if (!(vma->vm_flags & VM_GROWSDOWN))
1699                 return NULL;
1700         start = vma->vm_start;
1701         if (expand_stack(vma, addr))
1702                 return NULL;
1703         if (vma->vm_flags & VM_LOCKED)
1704                 make_pages_present(addr, start);
1705         return vma;
1706 }
1707 #endif
1708
1709 /*
1710  * Ok - we have the memory areas we should free on the vma list,
1711  * so release them, and do the vma updates.
1712  *
1713  * Called with the mm semaphore held.
1714  */
1715 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1716 {
1717         /* Update high watermark before we lower total_vm */
1718         update_hiwater_vm(mm);
1719         do {
1720                 long nrpages = vma_pages(vma);
1721
1722                 mm->total_vm -= nrpages;
1723                 if (vma->vm_flags & VM_LOCKED)
1724                         mm->locked_vm -= nrpages;
1725                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1726                 vma = remove_vma(vma);
1727         } while (vma);
1728         validate_mm(mm);
1729 }
1730
1731 /*
1732  * Get rid of page table information in the indicated region.
1733  *
1734  * Called with the mm semaphore held.
1735  */
1736 static void unmap_region(struct mm_struct *mm,
1737                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1738                 unsigned long start, unsigned long end)
1739 {
1740         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1741         struct mmu_gather *tlb;
1742         unsigned long nr_accounted = 0;
1743
1744         lru_add_drain();
1745         tlb = tlb_gather_mmu(mm, 0);
1746         update_hiwater_rss(mm);
1747         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1748         vm_unacct_memory(nr_accounted);
1749         free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1750                                  next? next->vm_start: 0);
1751         tlb_finish_mmu(tlb, start, end);
1752 }
1753
1754 /*
1755  * Create a list of vma's touched by the unmap, removing them from the mm's
1756  * vma list as we go..
1757  */
1758 static void
1759 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1760         struct vm_area_struct *prev, unsigned long end)
1761 {
1762         struct vm_area_struct **insertion_point;
1763         struct vm_area_struct *tail_vma = NULL;
1764         unsigned long addr;
1765
1766         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1767         do {
1768                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1769                 mm->map_count--;
1770                 tail_vma = vma;
1771                 vma = vma->vm_next;
1772         } while (vma && vma->vm_start < end);
1773         *insertion_point = vma;
1774         tail_vma->vm_next = NULL;
1775         if (mm->unmap_area == arch_unmap_area)
1776                 addr = prev ? prev->vm_end : mm->mmap_base;
1777         else
1778                 addr = vma ?  vma->vm_start : mm->mmap_base;
1779         mm->unmap_area(mm, addr);
1780         mm->mmap_cache = NULL;          /* Kill the cache. */
1781 }
1782
1783 /*
1784  * Split a vma into two pieces at address 'addr', a new vma is allocated
1785  * either for the first part or the tail.
1786  */
1787 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1788               unsigned long addr, int new_below)
1789 {
1790         struct mempolicy *pol;
1791         struct vm_area_struct *new;
1792
1793         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1794                 return -EINVAL;
1795
1796         if (mm->map_count >= sysctl_max_map_count)
1797                 return -ENOMEM;
1798
1799         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1800         if (!new)
1801                 return -ENOMEM;
1802
1803         /* most fields are the same, copy all, and then fixup */
1804         *new = *vma;
1805
1806         if (new_below)
1807                 new->vm_end = addr;
1808         else {
1809                 new->vm_start = addr;
1810                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1811         }
1812
1813         pol = mpol_dup(vma_policy(vma));
1814         if (IS_ERR(pol)) {
1815                 kmem_cache_free(vm_area_cachep, new);
1816                 return PTR_ERR(pol);
1817         }
1818         vma_set_policy(new, pol);
1819
1820         if (new->vm_file)
1821                 get_file(new->vm_file);
1822
1823         if (new->vm_ops && new->vm_ops->open)
1824                 new->vm_ops->open(new);
1825
1826         if (new_below)
1827                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1828                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1829         else
1830                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1831
1832         return 0;
1833 }
1834
1835 /* Munmap is split into 2 main parts -- this part which finds
1836  * what needs doing, and the areas themselves, which do the
1837  * work.  This now handles partial unmappings.
1838  * Jeremy Fitzhardinge <jeremy@goop.org>
1839  */
1840 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1841 {
1842         unsigned long end;
1843         struct vm_area_struct *vma, *prev, *last;
1844
1845         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1846                 return -EINVAL;
1847
1848         if ((len = PAGE_ALIGN(len)) == 0)
1849                 return -EINVAL;
1850
1851         /* Find the first overlapping VMA */
1852         vma = find_vma_prev(mm, start, &prev);
1853         if (!vma)
1854                 return 0;
1855         /* we have  start < vma->vm_end  */
1856
1857         /* if it doesn't overlap, we have nothing.. */
1858         end = start + len;
1859         if (vma->vm_start >= end)
1860                 return 0;
1861
1862         /*
1863          * If we need to split any vma, do it now to save pain later.
1864          *
1865          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1866          * unmapped vm_area_struct will remain in use: so lower split_vma
1867          * places tmp vma above, and higher split_vma places tmp vma below.
1868          */
1869         if (start > vma->vm_start) {
1870                 int error = split_vma(mm, vma, start, 0);
1871                 if (error)
1872                         return error;
1873                 prev = vma;
1874         }
1875
1876         /* Does it split the last one? */
1877         last = find_vma(mm, end);
1878         if (last && end > last->vm_start) {
1879                 int error = split_vma(mm, last, end, 1);
1880                 if (error)
1881                         return error;
1882         }
1883         vma = prev? prev->vm_next: mm->mmap;
1884
1885         /*
1886          * Remove the vma's, and unmap the actual pages
1887          */
1888         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1889         unmap_region(mm, vma, prev, start, end);
1890
1891         /* Fix up all other VM information */
1892         remove_vma_list(mm, vma);
1893
1894         return 0;
1895 }
1896
1897 EXPORT_SYMBOL(do_munmap);
1898
1899 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1900 {
1901         int ret;
1902         struct mm_struct *mm = current->mm;
1903
1904         profile_munmap(addr);
1905
1906         down_write(&mm->mmap_sem);
1907         ret = do_munmap(mm, addr, len);
1908         up_write(&mm->mmap_sem);
1909         return ret;
1910 }
1911
1912 static inline void verify_mm_writelocked(struct mm_struct *mm)
1913 {
1914 #ifdef CONFIG_DEBUG_VM
1915         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1916                 WARN_ON(1);
1917                 up_read(&mm->mmap_sem);
1918         }
1919 #endif
1920 }
1921
1922 /*
1923  *  this is really a simplified "do_mmap".  it only handles
1924  *  anonymous maps.  eventually we may be able to do some
1925  *  brk-specific accounting here.
1926  */
1927 unsigned long do_brk(unsigned long addr, unsigned long len)
1928 {
1929         struct mm_struct * mm = current->mm;
1930         struct vm_area_struct * vma, * prev;
1931         unsigned long flags;
1932         struct rb_node ** rb_link, * rb_parent;
1933         pgoff_t pgoff = addr >> PAGE_SHIFT;
1934         int error;
1935
1936         len = PAGE_ALIGN(len);
1937         if (!len)
1938                 return addr;
1939
1940         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1941                 return -EINVAL;
1942
1943         if (is_hugepage_only_range(mm, addr, len))
1944                 return -EINVAL;
1945
1946         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1947         if (error)
1948                 return error;
1949
1950         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1951
1952         error = arch_mmap_check(addr, len, flags);
1953         if (error)
1954                 return error;
1955
1956         /*
1957          * mlock MCL_FUTURE?
1958          */
1959         if (mm->def_flags & VM_LOCKED) {
1960                 unsigned long locked, lock_limit;
1961                 locked = len >> PAGE_SHIFT;
1962                 locked += mm->locked_vm;
1963                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1964                 lock_limit >>= PAGE_SHIFT;
1965                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1966                         return -EAGAIN;
1967         }
1968
1969         /*
1970          * mm->mmap_sem is required to protect against another thread
1971          * changing the mappings in case we sleep.
1972          */
1973         verify_mm_writelocked(mm);
1974
1975         /*
1976          * Clear old maps.  this also does some error checking for us
1977          */
1978  munmap_back:
1979         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1980         if (vma && vma->vm_start < addr + len) {
1981                 if (do_munmap(mm, addr, len))
1982                         return -ENOMEM;
1983                 goto munmap_back;
1984         }
1985
1986         /* Check against address space limits *after* clearing old maps... */
1987         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1988                 return -ENOMEM;
1989
1990         if (mm->map_count > sysctl_max_map_count)
1991                 return -ENOMEM;
1992
1993         if (security_vm_enough_memory(len >> PAGE_SHIFT))
1994                 return -ENOMEM;
1995
1996         /* Can we just expand an old private anonymous mapping? */
1997         if (vma_merge(mm, prev, addr, addr + len, flags,
1998                                         NULL, NULL, pgoff, NULL))
1999                 goto out;
2000
2001         /*
2002          * create a vma struct for an anonymous mapping
2003          */
2004         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2005         if (!vma) {
2006                 vm_unacct_memory(len >> PAGE_SHIFT);
2007                 return -ENOMEM;
2008         }
2009
2010         vma->vm_mm = mm;
2011         vma->vm_start = addr;
2012         vma->vm_end = addr + len;
2013         vma->vm_pgoff = pgoff;
2014         vma->vm_flags = flags;
2015         vma->vm_page_prot = vm_get_page_prot(flags);
2016         vma_link(mm, vma, prev, rb_link, rb_parent);
2017 out:
2018         mm->total_vm += len >> PAGE_SHIFT;
2019         if (flags & VM_LOCKED) {
2020                 mm->locked_vm += len >> PAGE_SHIFT;
2021                 make_pages_present(addr, addr + len);
2022         }
2023         return addr;
2024 }
2025
2026 EXPORT_SYMBOL(do_brk);
2027
2028 /* Release all mmaps. */
2029 void exit_mmap(struct mm_struct *mm)
2030 {
2031         struct mmu_gather *tlb;
2032         struct vm_area_struct *vma = mm->mmap;
2033         unsigned long nr_accounted = 0;
2034         unsigned long end;
2035
2036         /* mm's last user has gone, and its about to be pulled down */
2037         arch_exit_mmap(mm);
2038
2039         lru_add_drain();
2040         flush_cache_mm(mm);
2041         tlb = tlb_gather_mmu(mm, 1);
2042         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2043         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2044         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2045         vm_unacct_memory(nr_accounted);
2046         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2047         tlb_finish_mmu(tlb, 0, end);
2048
2049         /*
2050          * Walk the list again, actually closing and freeing it,
2051          * with preemption enabled, without holding any MM locks.
2052          */
2053         while (vma)
2054                 vma = remove_vma(vma);
2055
2056         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2057 }
2058
2059 /* Insert vm structure into process list sorted by address
2060  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2061  * then i_mmap_lock is taken here.
2062  */
2063 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2064 {
2065         struct vm_area_struct * __vma, * prev;
2066         struct rb_node ** rb_link, * rb_parent;
2067
2068         /*
2069          * The vm_pgoff of a purely anonymous vma should be irrelevant
2070          * until its first write fault, when page's anon_vma and index
2071          * are set.  But now set the vm_pgoff it will almost certainly
2072          * end up with (unless mremap moves it elsewhere before that
2073          * first wfault), so /proc/pid/maps tells a consistent story.
2074          *
2075          * By setting it to reflect the virtual start address of the
2076          * vma, merges and splits can happen in a seamless way, just
2077          * using the existing file pgoff checks and manipulations.
2078          * Similarly in do_mmap_pgoff and in do_brk.
2079          */
2080         if (!vma->vm_file) {
2081                 BUG_ON(vma->anon_vma);
2082                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2083         }
2084         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2085         if (__vma && __vma->vm_start < vma->vm_end)
2086                 return -ENOMEM;
2087         if ((vma->vm_flags & VM_ACCOUNT) &&
2088              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2089                 return -ENOMEM;
2090         vma_link(mm, vma, prev, rb_link, rb_parent);
2091         return 0;
2092 }
2093
2094 /*
2095  * Copy the vma structure to a new location in the same mm,
2096  * prior to moving page table entries, to effect an mremap move.
2097  */
2098 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2099         unsigned long addr, unsigned long len, pgoff_t pgoff)
2100 {
2101         struct vm_area_struct *vma = *vmap;
2102         unsigned long vma_start = vma->vm_start;
2103         struct mm_struct *mm = vma->vm_mm;
2104         struct vm_area_struct *new_vma, *prev;
2105         struct rb_node **rb_link, *rb_parent;
2106         struct mempolicy *pol;
2107
2108         /*
2109          * If anonymous vma has not yet been faulted, update new pgoff
2110          * to match new location, to increase its chance of merging.
2111          */
2112         if (!vma->vm_file && !vma->anon_vma)
2113                 pgoff = addr >> PAGE_SHIFT;
2114
2115         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2116         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2117                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2118         if (new_vma) {
2119                 /*
2120                  * Source vma may have been merged into new_vma
2121                  */
2122                 if (vma_start >= new_vma->vm_start &&
2123                     vma_start < new_vma->vm_end)
2124                         *vmap = new_vma;
2125         } else {
2126                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2127                 if (new_vma) {
2128                         *new_vma = *vma;
2129                         pol = mpol_dup(vma_policy(vma));
2130                         if (IS_ERR(pol)) {
2131                                 kmem_cache_free(vm_area_cachep, new_vma);
2132                                 return NULL;
2133                         }
2134                         vma_set_policy(new_vma, pol);
2135                         new_vma->vm_start = addr;
2136                         new_vma->vm_end = addr + len;
2137                         new_vma->vm_pgoff = pgoff;
2138                         if (new_vma->vm_file)
2139                                 get_file(new_vma->vm_file);
2140                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2141                                 new_vma->vm_ops->open(new_vma);
2142                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2143                 }
2144         }
2145         return new_vma;
2146 }
2147
2148 /*
2149  * Return true if the calling process may expand its vm space by the passed
2150  * number of pages
2151  */
2152 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2153 {
2154         unsigned long cur = mm->total_vm;       /* pages */
2155         unsigned long lim;
2156
2157         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2158
2159         if (cur + npages > lim)
2160                 return 0;
2161         return 1;
2162 }
2163
2164
2165 static int special_mapping_fault(struct vm_area_struct *vma,
2166                                 struct vm_fault *vmf)
2167 {
2168         pgoff_t pgoff;
2169         struct page **pages;
2170
2171         /*
2172          * special mappings have no vm_file, and in that case, the mm
2173          * uses vm_pgoff internally. So we have to subtract it from here.
2174          * We are allowed to do this because we are the mm; do not copy
2175          * this code into drivers!
2176          */
2177         pgoff = vmf->pgoff - vma->vm_pgoff;
2178
2179         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2180                 pgoff--;
2181
2182         if (*pages) {
2183                 struct page *page = *pages;
2184                 get_page(page);
2185                 vmf->page = page;
2186                 return 0;
2187         }
2188
2189         return VM_FAULT_SIGBUS;
2190 }
2191
2192 /*
2193  * Having a close hook prevents vma merging regardless of flags.
2194  */
2195 static void special_mapping_close(struct vm_area_struct *vma)
2196 {
2197 }
2198
2199 static struct vm_operations_struct special_mapping_vmops = {
2200         .close = special_mapping_close,
2201         .fault = special_mapping_fault,
2202 };
2203
2204 /*
2205  * Called with mm->mmap_sem held for writing.
2206  * Insert a new vma covering the given region, with the given flags.
2207  * Its pages are supplied by the given array of struct page *.
2208  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2209  * The region past the last page supplied will always produce SIGBUS.
2210  * The array pointer and the pages it points to are assumed to stay alive
2211  * for as long as this mapping might exist.
2212  */
2213 int install_special_mapping(struct mm_struct *mm,
2214                             unsigned long addr, unsigned long len,
2215                             unsigned long vm_flags, struct page **pages)
2216 {
2217         struct vm_area_struct *vma;
2218
2219         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2220         if (unlikely(vma == NULL))
2221                 return -ENOMEM;
2222
2223         vma->vm_mm = mm;
2224         vma->vm_start = addr;
2225         vma->vm_end = addr + len;
2226
2227         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2228         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2229
2230         vma->vm_ops = &special_mapping_vmops;
2231         vma->vm_private_data = pages;
2232
2233         if (unlikely(insert_vm_struct(mm, vma))) {
2234                 kmem_cache_free(vm_area_cachep, vma);
2235                 return -ENOMEM;
2236         }
2237
2238         mm->total_vm += len >> PAGE_SHIFT;
2239
2240         return 0;
2241 }