LSM: imbed ima calls in the security hooks
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31
32 #include <asm/uaccess.h>
33 #include <asm/cacheflush.h>
34 #include <asm/tlb.h>
35 #include <asm/mmu_context.h>
36
37 #include "internal.h"
38
39 #ifndef arch_mmap_check
40 #define arch_mmap_check(addr, len, flags)       (0)
41 #endif
42
43 #ifndef arch_rebalance_pgtables
44 #define arch_rebalance_pgtables(addr, len)              (addr)
45 #endif
46
47 static void unmap_region(struct mm_struct *mm,
48                 struct vm_area_struct *vma, struct vm_area_struct *prev,
49                 unsigned long start, unsigned long end);
50
51 /*
52  * WARNING: the debugging will use recursive algorithms so never enable this
53  * unless you know what you are doing.
54  */
55 #undef DEBUG_MM_RB
56
57 /* description of effects of mapping type and prot in current implementation.
58  * this is due to the limited x86 page protection hardware.  The expected
59  * behavior is in parens:
60  *
61  * map_type     prot
62  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
63  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
64  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
65  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
66  *              
67  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
68  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
69  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
70  *
71  */
72 pgprot_t protection_map[16] = {
73         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
74         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
75 };
76
77 pgprot_t vm_get_page_prot(unsigned long vm_flags)
78 {
79         return __pgprot(pgprot_val(protection_map[vm_flags &
80                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
81                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
82 }
83 EXPORT_SYMBOL(vm_get_page_prot);
84
85 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
86 int sysctl_overcommit_ratio = 50;       /* default is 50% */
87 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
88 struct percpu_counter vm_committed_as;
89
90 /*
91  * Check that a process has enough memory to allocate a new virtual
92  * mapping. 0 means there is enough memory for the allocation to
93  * succeed and -ENOMEM implies there is not.
94  *
95  * We currently support three overcommit policies, which are set via the
96  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
97  *
98  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
99  * Additional code 2002 Jul 20 by Robert Love.
100  *
101  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
102  *
103  * Note this is a helper function intended to be used by LSMs which
104  * wish to use this logic.
105  */
106 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
107 {
108         unsigned long free, allowed;
109
110         vm_acct_memory(pages);
111
112         /*
113          * Sometimes we want to use more memory than we have
114          */
115         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
116                 return 0;
117
118         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
119                 unsigned long n;
120
121                 free = global_page_state(NR_FILE_PAGES);
122                 free += nr_swap_pages;
123
124                 /*
125                  * Any slabs which are created with the
126                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
127                  * which are reclaimable, under pressure.  The dentry
128                  * cache and most inode caches should fall into this
129                  */
130                 free += global_page_state(NR_SLAB_RECLAIMABLE);
131
132                 /*
133                  * Leave the last 3% for root
134                  */
135                 if (!cap_sys_admin)
136                         free -= free / 32;
137
138                 if (free > pages)
139                         return 0;
140
141                 /*
142                  * nr_free_pages() is very expensive on large systems,
143                  * only call if we're about to fail.
144                  */
145                 n = nr_free_pages();
146
147                 /*
148                  * Leave reserved pages. The pages are not for anonymous pages.
149                  */
150                 if (n <= totalreserve_pages)
151                         goto error;
152                 else
153                         n -= totalreserve_pages;
154
155                 /*
156                  * Leave the last 3% for root
157                  */
158                 if (!cap_sys_admin)
159                         n -= n / 32;
160                 free += n;
161
162                 if (free > pages)
163                         return 0;
164
165                 goto error;
166         }
167
168         allowed = (totalram_pages - hugetlb_total_pages())
169                 * sysctl_overcommit_ratio / 100;
170         /*
171          * Leave the last 3% for root
172          */
173         if (!cap_sys_admin)
174                 allowed -= allowed / 32;
175         allowed += total_swap_pages;
176
177         /* Don't let a single process grow too big:
178            leave 3% of the size of this process for other processes */
179         if (mm)
180                 allowed -= mm->total_vm / 32;
181
182         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
183                 return 0;
184 error:
185         vm_unacct_memory(pages);
186
187         return -ENOMEM;
188 }
189
190 /*
191  * Requires inode->i_mapping->i_mmap_lock
192  */
193 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
194                 struct file *file, struct address_space *mapping)
195 {
196         if (vma->vm_flags & VM_DENYWRITE)
197                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
198         if (vma->vm_flags & VM_SHARED)
199                 mapping->i_mmap_writable--;
200
201         flush_dcache_mmap_lock(mapping);
202         if (unlikely(vma->vm_flags & VM_NONLINEAR))
203                 list_del_init(&vma->shared.vm_set.list);
204         else
205                 vma_prio_tree_remove(vma, &mapping->i_mmap);
206         flush_dcache_mmap_unlock(mapping);
207 }
208
209 /*
210  * Unlink a file-based vm structure from its prio_tree, to hide
211  * vma from rmap and vmtruncate before freeing its page tables.
212  */
213 void unlink_file_vma(struct vm_area_struct *vma)
214 {
215         struct file *file = vma->vm_file;
216
217         if (file) {
218                 struct address_space *mapping = file->f_mapping;
219                 spin_lock(&mapping->i_mmap_lock);
220                 __remove_shared_vm_struct(vma, file, mapping);
221                 spin_unlock(&mapping->i_mmap_lock);
222         }
223 }
224
225 /*
226  * Close a vm structure and free it, returning the next.
227  */
228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
229 {
230         struct vm_area_struct *next = vma->vm_next;
231
232         might_sleep();
233         if (vma->vm_ops && vma->vm_ops->close)
234                 vma->vm_ops->close(vma);
235         if (vma->vm_file) {
236                 fput(vma->vm_file);
237                 if (vma->vm_flags & VM_EXECUTABLE)
238                         removed_exe_file_vma(vma->vm_mm);
239         }
240         mpol_put(vma_policy(vma));
241         kmem_cache_free(vm_area_cachep, vma);
242         return next;
243 }
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         min_brk = mm->end_code;
256 #else
257         min_brk = mm->start_brk;
258 #endif
259         if (brk < min_brk)
260                 goto out;
261
262         /*
263          * Check against rlimit here. If this check is done later after the test
264          * of oldbrk with newbrk then it can escape the test and let the data
265          * segment grow beyond its set limit the in case where the limit is
266          * not page aligned -Ram Gupta
267          */
268         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
269         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
270                         (mm->end_data - mm->start_data) > rlim)
271                 goto out;
272
273         newbrk = PAGE_ALIGN(brk);
274         oldbrk = PAGE_ALIGN(mm->brk);
275         if (oldbrk == newbrk)
276                 goto set_brk;
277
278         /* Always allow shrinking brk. */
279         if (brk <= mm->brk) {
280                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
281                         goto set_brk;
282                 goto out;
283         }
284
285         /* Check against existing mmap mappings. */
286         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
287                 goto out;
288
289         /* Ok, looks good - let it rip. */
290         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
291                 goto out;
292 set_brk:
293         mm->brk = brk;
294 out:
295         retval = mm->brk;
296         up_write(&mm->mmap_sem);
297         return retval;
298 }
299
300 #ifdef DEBUG_MM_RB
301 static int browse_rb(struct rb_root *root)
302 {
303         int i = 0, j;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev)
311                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
312                 if (vma->vm_start < pend)
313                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
314                 if (vma->vm_start > vma->vm_end)
315                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
316                 i++;
317                 pn = nd;
318                 prev = vma->vm_start;
319                 pend = vma->vm_end;
320         }
321         j = 0;
322         for (nd = pn; nd; nd = rb_prev(nd)) {
323                 j++;
324         }
325         if (i != j)
326                 printk("backwards %d, forwards %d\n", j, i), i = 0;
327         return i;
328 }
329
330 void validate_mm(struct mm_struct *mm)
331 {
332         int bug = 0;
333         int i = 0;
334         struct vm_area_struct *tmp = mm->mmap;
335         while (tmp) {
336                 tmp = tmp->vm_next;
337                 i++;
338         }
339         if (i != mm->map_count)
340                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
341         i = browse_rb(&mm->mm_rb);
342         if (i != mm->map_count)
343                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
344         BUG_ON(bug);
345 }
346 #else
347 #define validate_mm(mm) do { } while (0)
348 #endif
349
350 static struct vm_area_struct *
351 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
352                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
353                 struct rb_node ** rb_parent)
354 {
355         struct vm_area_struct * vma;
356         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
357
358         __rb_link = &mm->mm_rb.rb_node;
359         rb_prev = __rb_parent = NULL;
360         vma = NULL;
361
362         while (*__rb_link) {
363                 struct vm_area_struct *vma_tmp;
364
365                 __rb_parent = *__rb_link;
366                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
367
368                 if (vma_tmp->vm_end > addr) {
369                         vma = vma_tmp;
370                         if (vma_tmp->vm_start <= addr)
371                                 break;
372                         __rb_link = &__rb_parent->rb_left;
373                 } else {
374                         rb_prev = __rb_parent;
375                         __rb_link = &__rb_parent->rb_right;
376                 }
377         }
378
379         *pprev = NULL;
380         if (rb_prev)
381                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
382         *rb_link = __rb_link;
383         *rb_parent = __rb_parent;
384         return vma;
385 }
386
387 static inline void
388 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
389                 struct vm_area_struct *prev, struct rb_node *rb_parent)
390 {
391         if (prev) {
392                 vma->vm_next = prev->vm_next;
393                 prev->vm_next = vma;
394         } else {
395                 mm->mmap = vma;
396                 if (rb_parent)
397                         vma->vm_next = rb_entry(rb_parent,
398                                         struct vm_area_struct, vm_rb);
399                 else
400                         vma->vm_next = NULL;
401         }
402 }
403
404 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
405                 struct rb_node **rb_link, struct rb_node *rb_parent)
406 {
407         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
408         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
409 }
410
411 static void __vma_link_file(struct vm_area_struct *vma)
412 {
413         struct file *file;
414
415         file = vma->vm_file;
416         if (file) {
417                 struct address_space *mapping = file->f_mapping;
418
419                 if (vma->vm_flags & VM_DENYWRITE)
420                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
421                 if (vma->vm_flags & VM_SHARED)
422                         mapping->i_mmap_writable++;
423
424                 flush_dcache_mmap_lock(mapping);
425                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
426                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
427                 else
428                         vma_prio_tree_insert(vma, &mapping->i_mmap);
429                 flush_dcache_mmap_unlock(mapping);
430         }
431 }
432
433 static void
434 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435         struct vm_area_struct *prev, struct rb_node **rb_link,
436         struct rb_node *rb_parent)
437 {
438         __vma_link_list(mm, vma, prev, rb_parent);
439         __vma_link_rb(mm, vma, rb_link, rb_parent);
440         __anon_vma_link(vma);
441 }
442
443 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
444                         struct vm_area_struct *prev, struct rb_node **rb_link,
445                         struct rb_node *rb_parent)
446 {
447         struct address_space *mapping = NULL;
448
449         if (vma->vm_file)
450                 mapping = vma->vm_file->f_mapping;
451
452         if (mapping) {
453                 spin_lock(&mapping->i_mmap_lock);
454                 vma->vm_truncate_count = mapping->truncate_count;
455         }
456         anon_vma_lock(vma);
457
458         __vma_link(mm, vma, prev, rb_link, rb_parent);
459         __vma_link_file(vma);
460
461         anon_vma_unlock(vma);
462         if (mapping)
463                 spin_unlock(&mapping->i_mmap_lock);
464
465         mm->map_count++;
466         validate_mm(mm);
467 }
468
469 /*
470  * Helper for vma_adjust in the split_vma insert case:
471  * insert vm structure into list and rbtree and anon_vma,
472  * but it has already been inserted into prio_tree earlier.
473  */
474 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
475 {
476         struct vm_area_struct *__vma, *prev;
477         struct rb_node **rb_link, *rb_parent;
478
479         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
480         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
481         __vma_link(mm, vma, prev, rb_link, rb_parent);
482         mm->map_count++;
483 }
484
485 static inline void
486 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
487                 struct vm_area_struct *prev)
488 {
489         prev->vm_next = vma->vm_next;
490         rb_erase(&vma->vm_rb, &mm->mm_rb);
491         if (mm->mmap_cache == vma)
492                 mm->mmap_cache = prev;
493 }
494
495 /*
496  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
497  * is already present in an i_mmap tree without adjusting the tree.
498  * The following helper function should be used when such adjustments
499  * are necessary.  The "insert" vma (if any) is to be inserted
500  * before we drop the necessary locks.
501  */
502 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
503         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
504 {
505         struct mm_struct *mm = vma->vm_mm;
506         struct vm_area_struct *next = vma->vm_next;
507         struct vm_area_struct *importer = NULL;
508         struct address_space *mapping = NULL;
509         struct prio_tree_root *root = NULL;
510         struct file *file = vma->vm_file;
511         struct anon_vma *anon_vma = NULL;
512         long adjust_next = 0;
513         int remove_next = 0;
514
515         if (next && !insert) {
516                 if (end >= next->vm_end) {
517                         /*
518                          * vma expands, overlapping all the next, and
519                          * perhaps the one after too (mprotect case 6).
520                          */
521 again:                  remove_next = 1 + (end > next->vm_end);
522                         end = next->vm_end;
523                         anon_vma = next->anon_vma;
524                         importer = vma;
525                 } else if (end > next->vm_start) {
526                         /*
527                          * vma expands, overlapping part of the next:
528                          * mprotect case 5 shifting the boundary up.
529                          */
530                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
531                         anon_vma = next->anon_vma;
532                         importer = vma;
533                 } else if (end < vma->vm_end) {
534                         /*
535                          * vma shrinks, and !insert tells it's not
536                          * split_vma inserting another: so it must be
537                          * mprotect case 4 shifting the boundary down.
538                          */
539                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
540                         anon_vma = next->anon_vma;
541                         importer = next;
542                 }
543         }
544
545         if (file) {
546                 mapping = file->f_mapping;
547                 if (!(vma->vm_flags & VM_NONLINEAR))
548                         root = &mapping->i_mmap;
549                 spin_lock(&mapping->i_mmap_lock);
550                 if (importer &&
551                     vma->vm_truncate_count != next->vm_truncate_count) {
552                         /*
553                          * unmap_mapping_range might be in progress:
554                          * ensure that the expanding vma is rescanned.
555                          */
556                         importer->vm_truncate_count = 0;
557                 }
558                 if (insert) {
559                         insert->vm_truncate_count = vma->vm_truncate_count;
560                         /*
561                          * Put into prio_tree now, so instantiated pages
562                          * are visible to arm/parisc __flush_dcache_page
563                          * throughout; but we cannot insert into address
564                          * space until vma start or end is updated.
565                          */
566                         __vma_link_file(insert);
567                 }
568         }
569
570         /*
571          * When changing only vma->vm_end, we don't really need
572          * anon_vma lock.
573          */
574         if (vma->anon_vma && (insert || importer || start != vma->vm_start))
575                 anon_vma = vma->anon_vma;
576         if (anon_vma) {
577                 spin_lock(&anon_vma->lock);
578                 /*
579                  * Easily overlooked: when mprotect shifts the boundary,
580                  * make sure the expanding vma has anon_vma set if the
581                  * shrinking vma had, to cover any anon pages imported.
582                  */
583                 if (importer && !importer->anon_vma) {
584                         importer->anon_vma = anon_vma;
585                         __anon_vma_link(importer);
586                 }
587         }
588
589         if (root) {
590                 flush_dcache_mmap_lock(mapping);
591                 vma_prio_tree_remove(vma, root);
592                 if (adjust_next)
593                         vma_prio_tree_remove(next, root);
594         }
595
596         vma->vm_start = start;
597         vma->vm_end = end;
598         vma->vm_pgoff = pgoff;
599         if (adjust_next) {
600                 next->vm_start += adjust_next << PAGE_SHIFT;
601                 next->vm_pgoff += adjust_next;
602         }
603
604         if (root) {
605                 if (adjust_next)
606                         vma_prio_tree_insert(next, root);
607                 vma_prio_tree_insert(vma, root);
608                 flush_dcache_mmap_unlock(mapping);
609         }
610
611         if (remove_next) {
612                 /*
613                  * vma_merge has merged next into vma, and needs
614                  * us to remove next before dropping the locks.
615                  */
616                 __vma_unlink(mm, next, vma);
617                 if (file)
618                         __remove_shared_vm_struct(next, file, mapping);
619                 if (next->anon_vma)
620                         __anon_vma_merge(vma, next);
621         } else if (insert) {
622                 /*
623                  * split_vma has split insert from vma, and needs
624                  * us to insert it before dropping the locks
625                  * (it may either follow vma or precede it).
626                  */
627                 __insert_vm_struct(mm, insert);
628         }
629
630         if (anon_vma)
631                 spin_unlock(&anon_vma->lock);
632         if (mapping)
633                 spin_unlock(&mapping->i_mmap_lock);
634
635         if (remove_next) {
636                 if (file) {
637                         fput(file);
638                         if (next->vm_flags & VM_EXECUTABLE)
639                                 removed_exe_file_vma(mm);
640                 }
641                 mm->map_count--;
642                 mpol_put(vma_policy(next));
643                 kmem_cache_free(vm_area_cachep, next);
644                 /*
645                  * In mprotect's case 6 (see comments on vma_merge),
646                  * we must remove another next too. It would clutter
647                  * up the code too much to do both in one go.
648                  */
649                 if (remove_next == 2) {
650                         next = vma->vm_next;
651                         goto again;
652                 }
653         }
654
655         validate_mm(mm);
656 }
657
658 /*
659  * If the vma has a ->close operation then the driver probably needs to release
660  * per-vma resources, so we don't attempt to merge those.
661  */
662 static inline int is_mergeable_vma(struct vm_area_struct *vma,
663                         struct file *file, unsigned long vm_flags)
664 {
665         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
666         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
667                 return 0;
668         if (vma->vm_file != file)
669                 return 0;
670         if (vma->vm_ops && vma->vm_ops->close)
671                 return 0;
672         return 1;
673 }
674
675 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
676                                         struct anon_vma *anon_vma2)
677 {
678         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
679 }
680
681 /*
682  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
683  * in front of (at a lower virtual address and file offset than) the vma.
684  *
685  * We cannot merge two vmas if they have differently assigned (non-NULL)
686  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
687  *
688  * We don't check here for the merged mmap wrapping around the end of pagecache
689  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
690  * wrap, nor mmaps which cover the final page at index -1UL.
691  */
692 static int
693 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
694         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
695 {
696         if (is_mergeable_vma(vma, file, vm_flags) &&
697             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
698                 if (vma->vm_pgoff == vm_pgoff)
699                         return 1;
700         }
701         return 0;
702 }
703
704 /*
705  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
706  * beyond (at a higher virtual address and file offset than) the vma.
707  *
708  * We cannot merge two vmas if they have differently assigned (non-NULL)
709  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
710  */
711 static int
712 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
713         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
714 {
715         if (is_mergeable_vma(vma, file, vm_flags) &&
716             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
717                 pgoff_t vm_pglen;
718                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
719                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
720                         return 1;
721         }
722         return 0;
723 }
724
725 /*
726  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
727  * whether that can be merged with its predecessor or its successor.
728  * Or both (it neatly fills a hole).
729  *
730  * In most cases - when called for mmap, brk or mremap - [addr,end) is
731  * certain not to be mapped by the time vma_merge is called; but when
732  * called for mprotect, it is certain to be already mapped (either at
733  * an offset within prev, or at the start of next), and the flags of
734  * this area are about to be changed to vm_flags - and the no-change
735  * case has already been eliminated.
736  *
737  * The following mprotect cases have to be considered, where AAAA is
738  * the area passed down from mprotect_fixup, never extending beyond one
739  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
740  *
741  *     AAAA             AAAA                AAAA          AAAA
742  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
743  *    cannot merge    might become    might become    might become
744  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
745  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
746  *    mremap move:                                    PPPPNNNNNNNN 8
747  *        AAAA
748  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
749  *    might become    case 1 below    case 2 below    case 3 below
750  *
751  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
752  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
753  */
754 struct vm_area_struct *vma_merge(struct mm_struct *mm,
755                         struct vm_area_struct *prev, unsigned long addr,
756                         unsigned long end, unsigned long vm_flags,
757                         struct anon_vma *anon_vma, struct file *file,
758                         pgoff_t pgoff, struct mempolicy *policy)
759 {
760         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
761         struct vm_area_struct *area, *next;
762
763         /*
764          * We later require that vma->vm_flags == vm_flags,
765          * so this tests vma->vm_flags & VM_SPECIAL, too.
766          */
767         if (vm_flags & VM_SPECIAL)
768                 return NULL;
769
770         if (prev)
771                 next = prev->vm_next;
772         else
773                 next = mm->mmap;
774         area = next;
775         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
776                 next = next->vm_next;
777
778         /*
779          * Can it merge with the predecessor?
780          */
781         if (prev && prev->vm_end == addr &&
782                         mpol_equal(vma_policy(prev), policy) &&
783                         can_vma_merge_after(prev, vm_flags,
784                                                 anon_vma, file, pgoff)) {
785                 /*
786                  * OK, it can.  Can we now merge in the successor as well?
787                  */
788                 if (next && end == next->vm_start &&
789                                 mpol_equal(policy, vma_policy(next)) &&
790                                 can_vma_merge_before(next, vm_flags,
791                                         anon_vma, file, pgoff+pglen) &&
792                                 is_mergeable_anon_vma(prev->anon_vma,
793                                                       next->anon_vma)) {
794                                                         /* cases 1, 6 */
795                         vma_adjust(prev, prev->vm_start,
796                                 next->vm_end, prev->vm_pgoff, NULL);
797                 } else                                  /* cases 2, 5, 7 */
798                         vma_adjust(prev, prev->vm_start,
799                                 end, prev->vm_pgoff, NULL);
800                 return prev;
801         }
802
803         /*
804          * Can this new request be merged in front of next?
805          */
806         if (next && end == next->vm_start &&
807                         mpol_equal(policy, vma_policy(next)) &&
808                         can_vma_merge_before(next, vm_flags,
809                                         anon_vma, file, pgoff+pglen)) {
810                 if (prev && addr < prev->vm_end)        /* case 4 */
811                         vma_adjust(prev, prev->vm_start,
812                                 addr, prev->vm_pgoff, NULL);
813                 else                                    /* cases 3, 8 */
814                         vma_adjust(area, addr, next->vm_end,
815                                 next->vm_pgoff - pglen, NULL);
816                 return area;
817         }
818
819         return NULL;
820 }
821
822 /*
823  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
824  * neighbouring vmas for a suitable anon_vma, before it goes off
825  * to allocate a new anon_vma.  It checks because a repetitive
826  * sequence of mprotects and faults may otherwise lead to distinct
827  * anon_vmas being allocated, preventing vma merge in subsequent
828  * mprotect.
829  */
830 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
831 {
832         struct vm_area_struct *near;
833         unsigned long vm_flags;
834
835         near = vma->vm_next;
836         if (!near)
837                 goto try_prev;
838
839         /*
840          * Since only mprotect tries to remerge vmas, match flags
841          * which might be mprotected into each other later on.
842          * Neither mlock nor madvise tries to remerge at present,
843          * so leave their flags as obstructing a merge.
844          */
845         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
846         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
847
848         if (near->anon_vma && vma->vm_end == near->vm_start &&
849                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
850                         can_vma_merge_before(near, vm_flags,
851                                 NULL, vma->vm_file, vma->vm_pgoff +
852                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
853                 return near->anon_vma;
854 try_prev:
855         /*
856          * It is potentially slow to have to call find_vma_prev here.
857          * But it's only on the first write fault on the vma, not
858          * every time, and we could devise a way to avoid it later
859          * (e.g. stash info in next's anon_vma_node when assigning
860          * an anon_vma, or when trying vma_merge).  Another time.
861          */
862         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
863         if (!near)
864                 goto none;
865
866         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
867         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
868
869         if (near->anon_vma && near->vm_end == vma->vm_start &&
870                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
871                         can_vma_merge_after(near, vm_flags,
872                                 NULL, vma->vm_file, vma->vm_pgoff))
873                 return near->anon_vma;
874 none:
875         /*
876          * There's no absolute need to look only at touching neighbours:
877          * we could search further afield for "compatible" anon_vmas.
878          * But it would probably just be a waste of time searching,
879          * or lead to too many vmas hanging off the same anon_vma.
880          * We're trying to allow mprotect remerging later on,
881          * not trying to minimize memory used for anon_vmas.
882          */
883         return NULL;
884 }
885
886 #ifdef CONFIG_PROC_FS
887 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
888                                                 struct file *file, long pages)
889 {
890         const unsigned long stack_flags
891                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
892
893         if (file) {
894                 mm->shared_vm += pages;
895                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
896                         mm->exec_vm += pages;
897         } else if (flags & stack_flags)
898                 mm->stack_vm += pages;
899         if (flags & (VM_RESERVED|VM_IO))
900                 mm->reserved_vm += pages;
901 }
902 #endif /* CONFIG_PROC_FS */
903
904 /*
905  * The caller must hold down_write(&current->mm->mmap_sem).
906  */
907
908 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
909                         unsigned long len, unsigned long prot,
910                         unsigned long flags, unsigned long pgoff)
911 {
912         struct mm_struct * mm = current->mm;
913         struct inode *inode;
914         unsigned int vm_flags;
915         int error;
916         unsigned long reqprot = prot;
917
918         /*
919          * Does the application expect PROT_READ to imply PROT_EXEC?
920          *
921          * (the exception is when the underlying filesystem is noexec
922          *  mounted, in which case we dont add PROT_EXEC.)
923          */
924         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
925                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
926                         prot |= PROT_EXEC;
927
928         if (!len)
929                 return -EINVAL;
930
931         if (!(flags & MAP_FIXED))
932                 addr = round_hint_to_min(addr);
933
934         error = arch_mmap_check(addr, len, flags);
935         if (error)
936                 return error;
937
938         /* Careful about overflows.. */
939         len = PAGE_ALIGN(len);
940         if (!len || len > TASK_SIZE)
941                 return -ENOMEM;
942
943         /* offset overflow? */
944         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
945                return -EOVERFLOW;
946
947         /* Too many mappings? */
948         if (mm->map_count > sysctl_max_map_count)
949                 return -ENOMEM;
950
951         if (flags & MAP_HUGETLB) {
952                 struct user_struct *user = NULL;
953                 if (file)
954                         return -EINVAL;
955
956                 /*
957                  * VM_NORESERVE is used because the reservations will be
958                  * taken when vm_ops->mmap() is called
959                  * A dummy user value is used because we are not locking
960                  * memory so no accounting is necessary
961                  */
962                 len = ALIGN(len, huge_page_size(&default_hstate));
963                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
964                                                 &user, HUGETLB_ANONHUGE_INODE);
965                 if (IS_ERR(file))
966                         return PTR_ERR(file);
967         }
968
969         /* Obtain the address to map to. we verify (or select) it and ensure
970          * that it represents a valid section of the address space.
971          */
972         addr = get_unmapped_area(file, addr, len, pgoff, flags);
973         if (addr & ~PAGE_MASK)
974                 return addr;
975
976         /* Do simple checking here so the lower-level routines won't have
977          * to. we assume access permissions have been handled by the open
978          * of the memory object, so we don't do any here.
979          */
980         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
981                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
982
983         if (flags & MAP_LOCKED)
984                 if (!can_do_mlock())
985                         return -EPERM;
986
987         /* mlock MCL_FUTURE? */
988         if (vm_flags & VM_LOCKED) {
989                 unsigned long locked, lock_limit;
990                 locked = len >> PAGE_SHIFT;
991                 locked += mm->locked_vm;
992                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
993                 lock_limit >>= PAGE_SHIFT;
994                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
995                         return -EAGAIN;
996         }
997
998         inode = file ? file->f_path.dentry->d_inode : NULL;
999
1000         if (file) {
1001                 switch (flags & MAP_TYPE) {
1002                 case MAP_SHARED:
1003                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1004                                 return -EACCES;
1005
1006                         /*
1007                          * Make sure we don't allow writing to an append-only
1008                          * file..
1009                          */
1010                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1011                                 return -EACCES;
1012
1013                         /*
1014                          * Make sure there are no mandatory locks on the file.
1015                          */
1016                         if (locks_verify_locked(inode))
1017                                 return -EAGAIN;
1018
1019                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1020                         if (!(file->f_mode & FMODE_WRITE))
1021                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1022
1023                         /* fall through */
1024                 case MAP_PRIVATE:
1025                         if (!(file->f_mode & FMODE_READ))
1026                                 return -EACCES;
1027                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1028                                 if (vm_flags & VM_EXEC)
1029                                         return -EPERM;
1030                                 vm_flags &= ~VM_MAYEXEC;
1031                         }
1032
1033                         if (!file->f_op || !file->f_op->mmap)
1034                                 return -ENODEV;
1035                         break;
1036
1037                 default:
1038                         return -EINVAL;
1039                 }
1040         } else {
1041                 switch (flags & MAP_TYPE) {
1042                 case MAP_SHARED:
1043                         /*
1044                          * Ignore pgoff.
1045                          */
1046                         pgoff = 0;
1047                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1048                         break;
1049                 case MAP_PRIVATE:
1050                         /*
1051                          * Set pgoff according to addr for anon_vma.
1052                          */
1053                         pgoff = addr >> PAGE_SHIFT;
1054                         break;
1055                 default:
1056                         return -EINVAL;
1057                 }
1058         }
1059
1060         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1061         if (error)
1062                 return error;
1063
1064         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1065 }
1066 EXPORT_SYMBOL(do_mmap_pgoff);
1067
1068 /*
1069  * Some shared mappigns will want the pages marked read-only
1070  * to track write events. If so, we'll downgrade vm_page_prot
1071  * to the private version (using protection_map[] without the
1072  * VM_SHARED bit).
1073  */
1074 int vma_wants_writenotify(struct vm_area_struct *vma)
1075 {
1076         unsigned int vm_flags = vma->vm_flags;
1077
1078         /* If it was private or non-writable, the write bit is already clear */
1079         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1080                 return 0;
1081
1082         /* The backer wishes to know when pages are first written to? */
1083         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1084                 return 1;
1085
1086         /* The open routine did something to the protections already? */
1087         if (pgprot_val(vma->vm_page_prot) !=
1088             pgprot_val(vm_get_page_prot(vm_flags)))
1089                 return 0;
1090
1091         /* Specialty mapping? */
1092         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1093                 return 0;
1094
1095         /* Can the mapping track the dirty pages? */
1096         return vma->vm_file && vma->vm_file->f_mapping &&
1097                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1098 }
1099
1100 /*
1101  * We account for memory if it's a private writeable mapping,
1102  * not hugepages and VM_NORESERVE wasn't set.
1103  */
1104 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1105 {
1106         /*
1107          * hugetlb has its own accounting separate from the core VM
1108          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1109          */
1110         if (file && is_file_hugepages(file))
1111                 return 0;
1112
1113         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1114 }
1115
1116 unsigned long mmap_region(struct file *file, unsigned long addr,
1117                           unsigned long len, unsigned long flags,
1118                           unsigned int vm_flags, unsigned long pgoff)
1119 {
1120         struct mm_struct *mm = current->mm;
1121         struct vm_area_struct *vma, *prev;
1122         int correct_wcount = 0;
1123         int error;
1124         struct rb_node **rb_link, *rb_parent;
1125         unsigned long charged = 0;
1126         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1127
1128         /* Clear old maps */
1129         error = -ENOMEM;
1130 munmap_back:
1131         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1132         if (vma && vma->vm_start < addr + len) {
1133                 if (do_munmap(mm, addr, len))
1134                         return -ENOMEM;
1135                 goto munmap_back;
1136         }
1137
1138         /* Check against address space limit. */
1139         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1140                 return -ENOMEM;
1141
1142         /*
1143          * Set 'VM_NORESERVE' if we should not account for the
1144          * memory use of this mapping.
1145          */
1146         if ((flags & MAP_NORESERVE)) {
1147                 /* We honor MAP_NORESERVE if allowed to overcommit */
1148                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1149                         vm_flags |= VM_NORESERVE;
1150
1151                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1152                 if (file && is_file_hugepages(file))
1153                         vm_flags |= VM_NORESERVE;
1154         }
1155
1156         /*
1157          * Private writable mapping: check memory availability
1158          */
1159         if (accountable_mapping(file, vm_flags)) {
1160                 charged = len >> PAGE_SHIFT;
1161                 if (security_vm_enough_memory(charged))
1162                         return -ENOMEM;
1163                 vm_flags |= VM_ACCOUNT;
1164         }
1165
1166         /*
1167          * Can we just expand an old mapping?
1168          */
1169         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1170         if (vma)
1171                 goto out;
1172
1173         /*
1174          * Determine the object being mapped and call the appropriate
1175          * specific mapper. the address has already been validated, but
1176          * not unmapped, but the maps are removed from the list.
1177          */
1178         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1179         if (!vma) {
1180                 error = -ENOMEM;
1181                 goto unacct_error;
1182         }
1183
1184         vma->vm_mm = mm;
1185         vma->vm_start = addr;
1186         vma->vm_end = addr + len;
1187         vma->vm_flags = vm_flags;
1188         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1189         vma->vm_pgoff = pgoff;
1190
1191         if (file) {
1192                 error = -EINVAL;
1193                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1194                         goto free_vma;
1195                 if (vm_flags & VM_DENYWRITE) {
1196                         error = deny_write_access(file);
1197                         if (error)
1198                                 goto free_vma;
1199                         correct_wcount = 1;
1200                 }
1201                 vma->vm_file = file;
1202                 get_file(file);
1203                 error = file->f_op->mmap(file, vma);
1204                 if (error)
1205                         goto unmap_and_free_vma;
1206                 if (vm_flags & VM_EXECUTABLE)
1207                         added_exe_file_vma(mm);
1208
1209                 /* Can addr have changed??
1210                  *
1211                  * Answer: Yes, several device drivers can do it in their
1212                  *         f_op->mmap method. -DaveM
1213                  */
1214                 addr = vma->vm_start;
1215                 pgoff = vma->vm_pgoff;
1216                 vm_flags = vma->vm_flags;
1217         } else if (vm_flags & VM_SHARED) {
1218                 error = shmem_zero_setup(vma);
1219                 if (error)
1220                         goto free_vma;
1221         }
1222
1223         if (vma_wants_writenotify(vma))
1224                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1225
1226         vma_link(mm, vma, prev, rb_link, rb_parent);
1227         file = vma->vm_file;
1228
1229         /* Once vma denies write, undo our temporary denial count */
1230         if (correct_wcount)
1231                 atomic_inc(&inode->i_writecount);
1232 out:
1233         perf_event_mmap(vma);
1234
1235         mm->total_vm += len >> PAGE_SHIFT;
1236         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1237         if (vm_flags & VM_LOCKED) {
1238                 /*
1239                  * makes pages present; downgrades, drops, reacquires mmap_sem
1240                  */
1241                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1242                 if (nr_pages < 0)
1243                         return nr_pages;        /* vma gone! */
1244                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1245         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1246                 make_pages_present(addr, addr + len);
1247         return addr;
1248
1249 unmap_and_free_vma:
1250         if (correct_wcount)
1251                 atomic_inc(&inode->i_writecount);
1252         vma->vm_file = NULL;
1253         fput(file);
1254
1255         /* Undo any partial mapping done by a device driver. */
1256         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1257         charged = 0;
1258 free_vma:
1259         kmem_cache_free(vm_area_cachep, vma);
1260 unacct_error:
1261         if (charged)
1262                 vm_unacct_memory(charged);
1263         return error;
1264 }
1265
1266 /* Get an address range which is currently unmapped.
1267  * For shmat() with addr=0.
1268  *
1269  * Ugly calling convention alert:
1270  * Return value with the low bits set means error value,
1271  * ie
1272  *      if (ret & ~PAGE_MASK)
1273  *              error = ret;
1274  *
1275  * This function "knows" that -ENOMEM has the bits set.
1276  */
1277 #ifndef HAVE_ARCH_UNMAPPED_AREA
1278 unsigned long
1279 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1280                 unsigned long len, unsigned long pgoff, unsigned long flags)
1281 {
1282         struct mm_struct *mm = current->mm;
1283         struct vm_area_struct *vma;
1284         unsigned long start_addr;
1285
1286         if (len > TASK_SIZE)
1287                 return -ENOMEM;
1288
1289         if (flags & MAP_FIXED)
1290                 return addr;
1291
1292         if (addr) {
1293                 addr = PAGE_ALIGN(addr);
1294                 vma = find_vma(mm, addr);
1295                 if (TASK_SIZE - len >= addr &&
1296                     (!vma || addr + len <= vma->vm_start))
1297                         return addr;
1298         }
1299         if (len > mm->cached_hole_size) {
1300                 start_addr = addr = mm->free_area_cache;
1301         } else {
1302                 start_addr = addr = TASK_UNMAPPED_BASE;
1303                 mm->cached_hole_size = 0;
1304         }
1305
1306 full_search:
1307         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1308                 /* At this point:  (!vma || addr < vma->vm_end). */
1309                 if (TASK_SIZE - len < addr) {
1310                         /*
1311                          * Start a new search - just in case we missed
1312                          * some holes.
1313                          */
1314                         if (start_addr != TASK_UNMAPPED_BASE) {
1315                                 addr = TASK_UNMAPPED_BASE;
1316                                 start_addr = addr;
1317                                 mm->cached_hole_size = 0;
1318                                 goto full_search;
1319                         }
1320                         return -ENOMEM;
1321                 }
1322                 if (!vma || addr + len <= vma->vm_start) {
1323                         /*
1324                          * Remember the place where we stopped the search:
1325                          */
1326                         mm->free_area_cache = addr + len;
1327                         return addr;
1328                 }
1329                 if (addr + mm->cached_hole_size < vma->vm_start)
1330                         mm->cached_hole_size = vma->vm_start - addr;
1331                 addr = vma->vm_end;
1332         }
1333 }
1334 #endif  
1335
1336 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1337 {
1338         /*
1339          * Is this a new hole at the lowest possible address?
1340          */
1341         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1342                 mm->free_area_cache = addr;
1343                 mm->cached_hole_size = ~0UL;
1344         }
1345 }
1346
1347 /*
1348  * This mmap-allocator allocates new areas top-down from below the
1349  * stack's low limit (the base):
1350  */
1351 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1352 unsigned long
1353 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1354                           const unsigned long len, const unsigned long pgoff,
1355                           const unsigned long flags)
1356 {
1357         struct vm_area_struct *vma;
1358         struct mm_struct *mm = current->mm;
1359         unsigned long addr = addr0;
1360
1361         /* requested length too big for entire address space */
1362         if (len > TASK_SIZE)
1363                 return -ENOMEM;
1364
1365         if (flags & MAP_FIXED)
1366                 return addr;
1367
1368         /* requesting a specific address */
1369         if (addr) {
1370                 addr = PAGE_ALIGN(addr);
1371                 vma = find_vma(mm, addr);
1372                 if (TASK_SIZE - len >= addr &&
1373                                 (!vma || addr + len <= vma->vm_start))
1374                         return addr;
1375         }
1376
1377         /* check if free_area_cache is useful for us */
1378         if (len <= mm->cached_hole_size) {
1379                 mm->cached_hole_size = 0;
1380                 mm->free_area_cache = mm->mmap_base;
1381         }
1382
1383         /* either no address requested or can't fit in requested address hole */
1384         addr = mm->free_area_cache;
1385
1386         /* make sure it can fit in the remaining address space */
1387         if (addr > len) {
1388                 vma = find_vma(mm, addr-len);
1389                 if (!vma || addr <= vma->vm_start)
1390                         /* remember the address as a hint for next time */
1391                         return (mm->free_area_cache = addr-len);
1392         }
1393
1394         if (mm->mmap_base < len)
1395                 goto bottomup;
1396
1397         addr = mm->mmap_base-len;
1398
1399         do {
1400                 /*
1401                  * Lookup failure means no vma is above this address,
1402                  * else if new region fits below vma->vm_start,
1403                  * return with success:
1404                  */
1405                 vma = find_vma(mm, addr);
1406                 if (!vma || addr+len <= vma->vm_start)
1407                         /* remember the address as a hint for next time */
1408                         return (mm->free_area_cache = addr);
1409
1410                 /* remember the largest hole we saw so far */
1411                 if (addr + mm->cached_hole_size < vma->vm_start)
1412                         mm->cached_hole_size = vma->vm_start - addr;
1413
1414                 /* try just below the current vma->vm_start */
1415                 addr = vma->vm_start-len;
1416         } while (len < vma->vm_start);
1417
1418 bottomup:
1419         /*
1420          * A failed mmap() very likely causes application failure,
1421          * so fall back to the bottom-up function here. This scenario
1422          * can happen with large stack limits and large mmap()
1423          * allocations.
1424          */
1425         mm->cached_hole_size = ~0UL;
1426         mm->free_area_cache = TASK_UNMAPPED_BASE;
1427         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1428         /*
1429          * Restore the topdown base:
1430          */
1431         mm->free_area_cache = mm->mmap_base;
1432         mm->cached_hole_size = ~0UL;
1433
1434         return addr;
1435 }
1436 #endif
1437
1438 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1439 {
1440         /*
1441          * Is this a new hole at the highest possible address?
1442          */
1443         if (addr > mm->free_area_cache)
1444                 mm->free_area_cache = addr;
1445
1446         /* dont allow allocations above current base */
1447         if (mm->free_area_cache > mm->mmap_base)
1448                 mm->free_area_cache = mm->mmap_base;
1449 }
1450
1451 unsigned long
1452 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1453                 unsigned long pgoff, unsigned long flags)
1454 {
1455         unsigned long (*get_area)(struct file *, unsigned long,
1456                                   unsigned long, unsigned long, unsigned long);
1457
1458         get_area = current->mm->get_unmapped_area;
1459         if (file && file->f_op && file->f_op->get_unmapped_area)
1460                 get_area = file->f_op->get_unmapped_area;
1461         addr = get_area(file, addr, len, pgoff, flags);
1462         if (IS_ERR_VALUE(addr))
1463                 return addr;
1464
1465         if (addr > TASK_SIZE - len)
1466                 return -ENOMEM;
1467         if (addr & ~PAGE_MASK)
1468                 return -EINVAL;
1469
1470         return arch_rebalance_pgtables(addr, len);
1471 }
1472
1473 EXPORT_SYMBOL(get_unmapped_area);
1474
1475 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1476 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1477 {
1478         struct vm_area_struct *vma = NULL;
1479
1480         if (mm) {
1481                 /* Check the cache first. */
1482                 /* (Cache hit rate is typically around 35%.) */
1483                 vma = mm->mmap_cache;
1484                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1485                         struct rb_node * rb_node;
1486
1487                         rb_node = mm->mm_rb.rb_node;
1488                         vma = NULL;
1489
1490                         while (rb_node) {
1491                                 struct vm_area_struct * vma_tmp;
1492
1493                                 vma_tmp = rb_entry(rb_node,
1494                                                 struct vm_area_struct, vm_rb);
1495
1496                                 if (vma_tmp->vm_end > addr) {
1497                                         vma = vma_tmp;
1498                                         if (vma_tmp->vm_start <= addr)
1499                                                 break;
1500                                         rb_node = rb_node->rb_left;
1501                                 } else
1502                                         rb_node = rb_node->rb_right;
1503                         }
1504                         if (vma)
1505                                 mm->mmap_cache = vma;
1506                 }
1507         }
1508         return vma;
1509 }
1510
1511 EXPORT_SYMBOL(find_vma);
1512
1513 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1514 struct vm_area_struct *
1515 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1516                         struct vm_area_struct **pprev)
1517 {
1518         struct vm_area_struct *vma = NULL, *prev = NULL;
1519         struct rb_node *rb_node;
1520         if (!mm)
1521                 goto out;
1522
1523         /* Guard against addr being lower than the first VMA */
1524         vma = mm->mmap;
1525
1526         /* Go through the RB tree quickly. */
1527         rb_node = mm->mm_rb.rb_node;
1528
1529         while (rb_node) {
1530                 struct vm_area_struct *vma_tmp;
1531                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1532
1533                 if (addr < vma_tmp->vm_end) {
1534                         rb_node = rb_node->rb_left;
1535                 } else {
1536                         prev = vma_tmp;
1537                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1538                                 break;
1539                         rb_node = rb_node->rb_right;
1540                 }
1541         }
1542
1543 out:
1544         *pprev = prev;
1545         return prev ? prev->vm_next : vma;
1546 }
1547
1548 /*
1549  * Verify that the stack growth is acceptable and
1550  * update accounting. This is shared with both the
1551  * grow-up and grow-down cases.
1552  */
1553 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1554 {
1555         struct mm_struct *mm = vma->vm_mm;
1556         struct rlimit *rlim = current->signal->rlim;
1557         unsigned long new_start;
1558
1559         /* address space limit tests */
1560         if (!may_expand_vm(mm, grow))
1561                 return -ENOMEM;
1562
1563         /* Stack limit test */
1564         if (size > rlim[RLIMIT_STACK].rlim_cur)
1565                 return -ENOMEM;
1566
1567         /* mlock limit tests */
1568         if (vma->vm_flags & VM_LOCKED) {
1569                 unsigned long locked;
1570                 unsigned long limit;
1571                 locked = mm->locked_vm + grow;
1572                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1573                 if (locked > limit && !capable(CAP_IPC_LOCK))
1574                         return -ENOMEM;
1575         }
1576
1577         /* Check to ensure the stack will not grow into a hugetlb-only region */
1578         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1579                         vma->vm_end - size;
1580         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1581                 return -EFAULT;
1582
1583         /*
1584          * Overcommit..  This must be the final test, as it will
1585          * update security statistics.
1586          */
1587         if (security_vm_enough_memory_mm(mm, grow))
1588                 return -ENOMEM;
1589
1590         /* Ok, everything looks good - let it rip */
1591         mm->total_vm += grow;
1592         if (vma->vm_flags & VM_LOCKED)
1593                 mm->locked_vm += grow;
1594         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1595         return 0;
1596 }
1597
1598 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1599 /*
1600  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1601  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1602  */
1603 #ifndef CONFIG_IA64
1604 static
1605 #endif
1606 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1607 {
1608         int error;
1609
1610         if (!(vma->vm_flags & VM_GROWSUP))
1611                 return -EFAULT;
1612
1613         /*
1614          * We must make sure the anon_vma is allocated
1615          * so that the anon_vma locking is not a noop.
1616          */
1617         if (unlikely(anon_vma_prepare(vma)))
1618                 return -ENOMEM;
1619         anon_vma_lock(vma);
1620
1621         /*
1622          * vma->vm_start/vm_end cannot change under us because the caller
1623          * is required to hold the mmap_sem in read mode.  We need the
1624          * anon_vma lock to serialize against concurrent expand_stacks.
1625          * Also guard against wrapping around to address 0.
1626          */
1627         if (address < PAGE_ALIGN(address+4))
1628                 address = PAGE_ALIGN(address+4);
1629         else {
1630                 anon_vma_unlock(vma);
1631                 return -ENOMEM;
1632         }
1633         error = 0;
1634
1635         /* Somebody else might have raced and expanded it already */
1636         if (address > vma->vm_end) {
1637                 unsigned long size, grow;
1638
1639                 size = address - vma->vm_start;
1640                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1641
1642                 error = acct_stack_growth(vma, size, grow);
1643                 if (!error)
1644                         vma->vm_end = address;
1645         }
1646         anon_vma_unlock(vma);
1647         return error;
1648 }
1649 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1650
1651 /*
1652  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1653  */
1654 static int expand_downwards(struct vm_area_struct *vma,
1655                                    unsigned long address)
1656 {
1657         int error;
1658
1659         /*
1660          * We must make sure the anon_vma is allocated
1661          * so that the anon_vma locking is not a noop.
1662          */
1663         if (unlikely(anon_vma_prepare(vma)))
1664                 return -ENOMEM;
1665
1666         address &= PAGE_MASK;
1667         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1668         if (error)
1669                 return error;
1670
1671         anon_vma_lock(vma);
1672
1673         /*
1674          * vma->vm_start/vm_end cannot change under us because the caller
1675          * is required to hold the mmap_sem in read mode.  We need the
1676          * anon_vma lock to serialize against concurrent expand_stacks.
1677          */
1678
1679         /* Somebody else might have raced and expanded it already */
1680         if (address < vma->vm_start) {
1681                 unsigned long size, grow;
1682
1683                 size = vma->vm_end - address;
1684                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1685
1686                 error = acct_stack_growth(vma, size, grow);
1687                 if (!error) {
1688                         vma->vm_start = address;
1689                         vma->vm_pgoff -= grow;
1690                 }
1691         }
1692         anon_vma_unlock(vma);
1693         return error;
1694 }
1695
1696 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1697 {
1698         return expand_downwards(vma, address);
1699 }
1700
1701 #ifdef CONFIG_STACK_GROWSUP
1702 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1703 {
1704         return expand_upwards(vma, address);
1705 }
1706
1707 struct vm_area_struct *
1708 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1709 {
1710         struct vm_area_struct *vma, *prev;
1711
1712         addr &= PAGE_MASK;
1713         vma = find_vma_prev(mm, addr, &prev);
1714         if (vma && (vma->vm_start <= addr))
1715                 return vma;
1716         if (!prev || expand_stack(prev, addr))
1717                 return NULL;
1718         if (prev->vm_flags & VM_LOCKED) {
1719                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1720                         return NULL;    /* vma gone! */
1721         }
1722         return prev;
1723 }
1724 #else
1725 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1726 {
1727         return expand_downwards(vma, address);
1728 }
1729
1730 struct vm_area_struct *
1731 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1732 {
1733         struct vm_area_struct * vma;
1734         unsigned long start;
1735
1736         addr &= PAGE_MASK;
1737         vma = find_vma(mm,addr);
1738         if (!vma)
1739                 return NULL;
1740         if (vma->vm_start <= addr)
1741                 return vma;
1742         if (!(vma->vm_flags & VM_GROWSDOWN))
1743                 return NULL;
1744         start = vma->vm_start;
1745         if (expand_stack(vma, addr))
1746                 return NULL;
1747         if (vma->vm_flags & VM_LOCKED) {
1748                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1749                         return NULL;    /* vma gone! */
1750         }
1751         return vma;
1752 }
1753 #endif
1754
1755 /*
1756  * Ok - we have the memory areas we should free on the vma list,
1757  * so release them, and do the vma updates.
1758  *
1759  * Called with the mm semaphore held.
1760  */
1761 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1762 {
1763         /* Update high watermark before we lower total_vm */
1764         update_hiwater_vm(mm);
1765         do {
1766                 long nrpages = vma_pages(vma);
1767
1768                 mm->total_vm -= nrpages;
1769                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1770                 vma = remove_vma(vma);
1771         } while (vma);
1772         validate_mm(mm);
1773 }
1774
1775 /*
1776  * Get rid of page table information in the indicated region.
1777  *
1778  * Called with the mm semaphore held.
1779  */
1780 static void unmap_region(struct mm_struct *mm,
1781                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1782                 unsigned long start, unsigned long end)
1783 {
1784         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1785         struct mmu_gather *tlb;
1786         unsigned long nr_accounted = 0;
1787
1788         lru_add_drain();
1789         tlb = tlb_gather_mmu(mm, 0);
1790         update_hiwater_rss(mm);
1791         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1792         vm_unacct_memory(nr_accounted);
1793         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1794                                  next? next->vm_start: 0);
1795         tlb_finish_mmu(tlb, start, end);
1796 }
1797
1798 /*
1799  * Create a list of vma's touched by the unmap, removing them from the mm's
1800  * vma list as we go..
1801  */
1802 static void
1803 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1804         struct vm_area_struct *prev, unsigned long end)
1805 {
1806         struct vm_area_struct **insertion_point;
1807         struct vm_area_struct *tail_vma = NULL;
1808         unsigned long addr;
1809
1810         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1811         do {
1812                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1813                 mm->map_count--;
1814                 tail_vma = vma;
1815                 vma = vma->vm_next;
1816         } while (vma && vma->vm_start < end);
1817         *insertion_point = vma;
1818         tail_vma->vm_next = NULL;
1819         if (mm->unmap_area == arch_unmap_area)
1820                 addr = prev ? prev->vm_end : mm->mmap_base;
1821         else
1822                 addr = vma ?  vma->vm_start : mm->mmap_base;
1823         mm->unmap_area(mm, addr);
1824         mm->mmap_cache = NULL;          /* Kill the cache. */
1825 }
1826
1827 /*
1828  * Split a vma into two pieces at address 'addr', a new vma is allocated
1829  * either for the first part or the tail.
1830  */
1831 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1832               unsigned long addr, int new_below)
1833 {
1834         struct mempolicy *pol;
1835         struct vm_area_struct *new;
1836
1837         if (is_vm_hugetlb_page(vma) && (addr &
1838                                         ~(huge_page_mask(hstate_vma(vma)))))
1839                 return -EINVAL;
1840
1841         if (mm->map_count >= sysctl_max_map_count)
1842                 return -ENOMEM;
1843
1844         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1845         if (!new)
1846                 return -ENOMEM;
1847
1848         /* most fields are the same, copy all, and then fixup */
1849         *new = *vma;
1850
1851         if (new_below)
1852                 new->vm_end = addr;
1853         else {
1854                 new->vm_start = addr;
1855                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1856         }
1857
1858         pol = mpol_dup(vma_policy(vma));
1859         if (IS_ERR(pol)) {
1860                 kmem_cache_free(vm_area_cachep, new);
1861                 return PTR_ERR(pol);
1862         }
1863         vma_set_policy(new, pol);
1864
1865         if (new->vm_file) {
1866                 get_file(new->vm_file);
1867                 if (vma->vm_flags & VM_EXECUTABLE)
1868                         added_exe_file_vma(mm);
1869         }
1870
1871         if (new->vm_ops && new->vm_ops->open)
1872                 new->vm_ops->open(new);
1873
1874         if (new_below)
1875                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1876                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1877         else
1878                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1879
1880         return 0;
1881 }
1882
1883 /* Munmap is split into 2 main parts -- this part which finds
1884  * what needs doing, and the areas themselves, which do the
1885  * work.  This now handles partial unmappings.
1886  * Jeremy Fitzhardinge <jeremy@goop.org>
1887  */
1888 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1889 {
1890         unsigned long end;
1891         struct vm_area_struct *vma, *prev, *last;
1892
1893         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1894                 return -EINVAL;
1895
1896         if ((len = PAGE_ALIGN(len)) == 0)
1897                 return -EINVAL;
1898
1899         /* Find the first overlapping VMA */
1900         vma = find_vma_prev(mm, start, &prev);
1901         if (!vma)
1902                 return 0;
1903         /* we have  start < vma->vm_end  */
1904
1905         /* if it doesn't overlap, we have nothing.. */
1906         end = start + len;
1907         if (vma->vm_start >= end)
1908                 return 0;
1909
1910         /*
1911          * If we need to split any vma, do it now to save pain later.
1912          *
1913          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1914          * unmapped vm_area_struct will remain in use: so lower split_vma
1915          * places tmp vma above, and higher split_vma places tmp vma below.
1916          */
1917         if (start > vma->vm_start) {
1918                 int error = split_vma(mm, vma, start, 0);
1919                 if (error)
1920                         return error;
1921                 prev = vma;
1922         }
1923
1924         /* Does it split the last one? */
1925         last = find_vma(mm, end);
1926         if (last && end > last->vm_start) {
1927                 int error = split_vma(mm, last, end, 1);
1928                 if (error)
1929                         return error;
1930         }
1931         vma = prev? prev->vm_next: mm->mmap;
1932
1933         /*
1934          * unlock any mlock()ed ranges before detaching vmas
1935          */
1936         if (mm->locked_vm) {
1937                 struct vm_area_struct *tmp = vma;
1938                 while (tmp && tmp->vm_start < end) {
1939                         if (tmp->vm_flags & VM_LOCKED) {
1940                                 mm->locked_vm -= vma_pages(tmp);
1941                                 munlock_vma_pages_all(tmp);
1942                         }
1943                         tmp = tmp->vm_next;
1944                 }
1945         }
1946
1947         /*
1948          * Remove the vma's, and unmap the actual pages
1949          */
1950         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1951         unmap_region(mm, vma, prev, start, end);
1952
1953         /* Fix up all other VM information */
1954         remove_vma_list(mm, vma);
1955
1956         return 0;
1957 }
1958
1959 EXPORT_SYMBOL(do_munmap);
1960
1961 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1962 {
1963         int ret;
1964         struct mm_struct *mm = current->mm;
1965
1966         profile_munmap(addr);
1967
1968         down_write(&mm->mmap_sem);
1969         ret = do_munmap(mm, addr, len);
1970         up_write(&mm->mmap_sem);
1971         return ret;
1972 }
1973
1974 static inline void verify_mm_writelocked(struct mm_struct *mm)
1975 {
1976 #ifdef CONFIG_DEBUG_VM
1977         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1978                 WARN_ON(1);
1979                 up_read(&mm->mmap_sem);
1980         }
1981 #endif
1982 }
1983
1984 /*
1985  *  this is really a simplified "do_mmap".  it only handles
1986  *  anonymous maps.  eventually we may be able to do some
1987  *  brk-specific accounting here.
1988  */
1989 unsigned long do_brk(unsigned long addr, unsigned long len)
1990 {
1991         struct mm_struct * mm = current->mm;
1992         struct vm_area_struct * vma, * prev;
1993         unsigned long flags;
1994         struct rb_node ** rb_link, * rb_parent;
1995         pgoff_t pgoff = addr >> PAGE_SHIFT;
1996         int error;
1997
1998         len = PAGE_ALIGN(len);
1999         if (!len)
2000                 return addr;
2001
2002         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
2003                 return -EINVAL;
2004
2005         if (is_hugepage_only_range(mm, addr, len))
2006                 return -EINVAL;
2007
2008         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2009         if (error)
2010                 return error;
2011
2012         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2013
2014         error = arch_mmap_check(addr, len, flags);
2015         if (error)
2016                 return error;
2017
2018         /*
2019          * mlock MCL_FUTURE?
2020          */
2021         if (mm->def_flags & VM_LOCKED) {
2022                 unsigned long locked, lock_limit;
2023                 locked = len >> PAGE_SHIFT;
2024                 locked += mm->locked_vm;
2025                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2026                 lock_limit >>= PAGE_SHIFT;
2027                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2028                         return -EAGAIN;
2029         }
2030
2031         /*
2032          * mm->mmap_sem is required to protect against another thread
2033          * changing the mappings in case we sleep.
2034          */
2035         verify_mm_writelocked(mm);
2036
2037         /*
2038          * Clear old maps.  this also does some error checking for us
2039          */
2040  munmap_back:
2041         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2042         if (vma && vma->vm_start < addr + len) {
2043                 if (do_munmap(mm, addr, len))
2044                         return -ENOMEM;
2045                 goto munmap_back;
2046         }
2047
2048         /* Check against address space limits *after* clearing old maps... */
2049         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2050                 return -ENOMEM;
2051
2052         if (mm->map_count > sysctl_max_map_count)
2053                 return -ENOMEM;
2054
2055         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2056                 return -ENOMEM;
2057
2058         /* Can we just expand an old private anonymous mapping? */
2059         vma = vma_merge(mm, prev, addr, addr + len, flags,
2060                                         NULL, NULL, pgoff, NULL);
2061         if (vma)
2062                 goto out;
2063
2064         /*
2065          * create a vma struct for an anonymous mapping
2066          */
2067         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2068         if (!vma) {
2069                 vm_unacct_memory(len >> PAGE_SHIFT);
2070                 return -ENOMEM;
2071         }
2072
2073         vma->vm_mm = mm;
2074         vma->vm_start = addr;
2075         vma->vm_end = addr + len;
2076         vma->vm_pgoff = pgoff;
2077         vma->vm_flags = flags;
2078         vma->vm_page_prot = vm_get_page_prot(flags);
2079         vma_link(mm, vma, prev, rb_link, rb_parent);
2080 out:
2081         mm->total_vm += len >> PAGE_SHIFT;
2082         if (flags & VM_LOCKED) {
2083                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2084                         mm->locked_vm += (len >> PAGE_SHIFT);
2085         }
2086         return addr;
2087 }
2088
2089 EXPORT_SYMBOL(do_brk);
2090
2091 /* Release all mmaps. */
2092 void exit_mmap(struct mm_struct *mm)
2093 {
2094         struct mmu_gather *tlb;
2095         struct vm_area_struct *vma;
2096         unsigned long nr_accounted = 0;
2097         unsigned long end;
2098
2099         /* mm's last user has gone, and its about to be pulled down */
2100         mmu_notifier_release(mm);
2101
2102         if (mm->locked_vm) {
2103                 vma = mm->mmap;
2104                 while (vma) {
2105                         if (vma->vm_flags & VM_LOCKED)
2106                                 munlock_vma_pages_all(vma);
2107                         vma = vma->vm_next;
2108                 }
2109         }
2110
2111         arch_exit_mmap(mm);
2112
2113         vma = mm->mmap;
2114         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2115                 return;
2116
2117         lru_add_drain();
2118         flush_cache_mm(mm);
2119         tlb = tlb_gather_mmu(mm, 1);
2120         /* update_hiwater_rss(mm) here? but nobody should be looking */
2121         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2122         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2123         vm_unacct_memory(nr_accounted);
2124
2125         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2126         tlb_finish_mmu(tlb, 0, end);
2127
2128         /*
2129          * Walk the list again, actually closing and freeing it,
2130          * with preemption enabled, without holding any MM locks.
2131          */
2132         while (vma)
2133                 vma = remove_vma(vma);
2134
2135         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2136 }
2137
2138 /* Insert vm structure into process list sorted by address
2139  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2140  * then i_mmap_lock is taken here.
2141  */
2142 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2143 {
2144         struct vm_area_struct * __vma, * prev;
2145         struct rb_node ** rb_link, * rb_parent;
2146
2147         /*
2148          * The vm_pgoff of a purely anonymous vma should be irrelevant
2149          * until its first write fault, when page's anon_vma and index
2150          * are set.  But now set the vm_pgoff it will almost certainly
2151          * end up with (unless mremap moves it elsewhere before that
2152          * first wfault), so /proc/pid/maps tells a consistent story.
2153          *
2154          * By setting it to reflect the virtual start address of the
2155          * vma, merges and splits can happen in a seamless way, just
2156          * using the existing file pgoff checks and manipulations.
2157          * Similarly in do_mmap_pgoff and in do_brk.
2158          */
2159         if (!vma->vm_file) {
2160                 BUG_ON(vma->anon_vma);
2161                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2162         }
2163         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2164         if (__vma && __vma->vm_start < vma->vm_end)
2165                 return -ENOMEM;
2166         if ((vma->vm_flags & VM_ACCOUNT) &&
2167              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2168                 return -ENOMEM;
2169         vma_link(mm, vma, prev, rb_link, rb_parent);
2170         return 0;
2171 }
2172
2173 /*
2174  * Copy the vma structure to a new location in the same mm,
2175  * prior to moving page table entries, to effect an mremap move.
2176  */
2177 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2178         unsigned long addr, unsigned long len, pgoff_t pgoff)
2179 {
2180         struct vm_area_struct *vma = *vmap;
2181         unsigned long vma_start = vma->vm_start;
2182         struct mm_struct *mm = vma->vm_mm;
2183         struct vm_area_struct *new_vma, *prev;
2184         struct rb_node **rb_link, *rb_parent;
2185         struct mempolicy *pol;
2186
2187         /*
2188          * If anonymous vma has not yet been faulted, update new pgoff
2189          * to match new location, to increase its chance of merging.
2190          */
2191         if (!vma->vm_file && !vma->anon_vma)
2192                 pgoff = addr >> PAGE_SHIFT;
2193
2194         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2195         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2196                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2197         if (new_vma) {
2198                 /*
2199                  * Source vma may have been merged into new_vma
2200                  */
2201                 if (vma_start >= new_vma->vm_start &&
2202                     vma_start < new_vma->vm_end)
2203                         *vmap = new_vma;
2204         } else {
2205                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2206                 if (new_vma) {
2207                         *new_vma = *vma;
2208                         pol = mpol_dup(vma_policy(vma));
2209                         if (IS_ERR(pol)) {
2210                                 kmem_cache_free(vm_area_cachep, new_vma);
2211                                 return NULL;
2212                         }
2213                         vma_set_policy(new_vma, pol);
2214                         new_vma->vm_start = addr;
2215                         new_vma->vm_end = addr + len;
2216                         new_vma->vm_pgoff = pgoff;
2217                         if (new_vma->vm_file) {
2218                                 get_file(new_vma->vm_file);
2219                                 if (vma->vm_flags & VM_EXECUTABLE)
2220                                         added_exe_file_vma(mm);
2221                         }
2222                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2223                                 new_vma->vm_ops->open(new_vma);
2224                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2225                 }
2226         }
2227         return new_vma;
2228 }
2229
2230 /*
2231  * Return true if the calling process may expand its vm space by the passed
2232  * number of pages
2233  */
2234 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2235 {
2236         unsigned long cur = mm->total_vm;       /* pages */
2237         unsigned long lim;
2238
2239         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2240
2241         if (cur + npages > lim)
2242                 return 0;
2243         return 1;
2244 }
2245
2246
2247 static int special_mapping_fault(struct vm_area_struct *vma,
2248                                 struct vm_fault *vmf)
2249 {
2250         pgoff_t pgoff;
2251         struct page **pages;
2252
2253         /*
2254          * special mappings have no vm_file, and in that case, the mm
2255          * uses vm_pgoff internally. So we have to subtract it from here.
2256          * We are allowed to do this because we are the mm; do not copy
2257          * this code into drivers!
2258          */
2259         pgoff = vmf->pgoff - vma->vm_pgoff;
2260
2261         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2262                 pgoff--;
2263
2264         if (*pages) {
2265                 struct page *page = *pages;
2266                 get_page(page);
2267                 vmf->page = page;
2268                 return 0;
2269         }
2270
2271         return VM_FAULT_SIGBUS;
2272 }
2273
2274 /*
2275  * Having a close hook prevents vma merging regardless of flags.
2276  */
2277 static void special_mapping_close(struct vm_area_struct *vma)
2278 {
2279 }
2280
2281 static const struct vm_operations_struct special_mapping_vmops = {
2282         .close = special_mapping_close,
2283         .fault = special_mapping_fault,
2284 };
2285
2286 /*
2287  * Called with mm->mmap_sem held for writing.
2288  * Insert a new vma covering the given region, with the given flags.
2289  * Its pages are supplied by the given array of struct page *.
2290  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2291  * The region past the last page supplied will always produce SIGBUS.
2292  * The array pointer and the pages it points to are assumed to stay alive
2293  * for as long as this mapping might exist.
2294  */
2295 int install_special_mapping(struct mm_struct *mm,
2296                             unsigned long addr, unsigned long len,
2297                             unsigned long vm_flags, struct page **pages)
2298 {
2299         struct vm_area_struct *vma;
2300
2301         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2302         if (unlikely(vma == NULL))
2303                 return -ENOMEM;
2304
2305         vma->vm_mm = mm;
2306         vma->vm_start = addr;
2307         vma->vm_end = addr + len;
2308
2309         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2310         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2311
2312         vma->vm_ops = &special_mapping_vmops;
2313         vma->vm_private_data = pages;
2314
2315         if (unlikely(insert_vm_struct(mm, vma))) {
2316                 kmem_cache_free(vm_area_cachep, vma);
2317                 return -ENOMEM;
2318         }
2319
2320         mm->total_vm += len >> PAGE_SHIFT;
2321
2322         perf_event_mmap(vma);
2323
2324         return 0;
2325 }
2326
2327 static DEFINE_MUTEX(mm_all_locks_mutex);
2328
2329 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2330 {
2331         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2332                 /*
2333                  * The LSB of head.next can't change from under us
2334                  * because we hold the mm_all_locks_mutex.
2335                  */
2336                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2337                 /*
2338                  * We can safely modify head.next after taking the
2339                  * anon_vma->lock. If some other vma in this mm shares
2340                  * the same anon_vma we won't take it again.
2341                  *
2342                  * No need of atomic instructions here, head.next
2343                  * can't change from under us thanks to the
2344                  * anon_vma->lock.
2345                  */
2346                 if (__test_and_set_bit(0, (unsigned long *)
2347                                        &anon_vma->head.next))
2348                         BUG();
2349         }
2350 }
2351
2352 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2353 {
2354         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2355                 /*
2356                  * AS_MM_ALL_LOCKS can't change from under us because
2357                  * we hold the mm_all_locks_mutex.
2358                  *
2359                  * Operations on ->flags have to be atomic because
2360                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2361                  * mm_all_locks_mutex, there may be other cpus
2362                  * changing other bitflags in parallel to us.
2363                  */
2364                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2365                         BUG();
2366                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2367         }
2368 }
2369
2370 /*
2371  * This operation locks against the VM for all pte/vma/mm related
2372  * operations that could ever happen on a certain mm. This includes
2373  * vmtruncate, try_to_unmap, and all page faults.
2374  *
2375  * The caller must take the mmap_sem in write mode before calling
2376  * mm_take_all_locks(). The caller isn't allowed to release the
2377  * mmap_sem until mm_drop_all_locks() returns.
2378  *
2379  * mmap_sem in write mode is required in order to block all operations
2380  * that could modify pagetables and free pages without need of
2381  * altering the vma layout (for example populate_range() with
2382  * nonlinear vmas). It's also needed in write mode to avoid new
2383  * anon_vmas to be associated with existing vmas.
2384  *
2385  * A single task can't take more than one mm_take_all_locks() in a row
2386  * or it would deadlock.
2387  *
2388  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2389  * mapping->flags avoid to take the same lock twice, if more than one
2390  * vma in this mm is backed by the same anon_vma or address_space.
2391  *
2392  * We can take all the locks in random order because the VM code
2393  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2394  * takes more than one of them in a row. Secondly we're protected
2395  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2396  *
2397  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2398  * that may have to take thousand of locks.
2399  *
2400  * mm_take_all_locks() can fail if it's interrupted by signals.
2401  */
2402 int mm_take_all_locks(struct mm_struct *mm)
2403 {
2404         struct vm_area_struct *vma;
2405         int ret = -EINTR;
2406
2407         BUG_ON(down_read_trylock(&mm->mmap_sem));
2408
2409         mutex_lock(&mm_all_locks_mutex);
2410
2411         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2412                 if (signal_pending(current))
2413                         goto out_unlock;
2414                 if (vma->vm_file && vma->vm_file->f_mapping)
2415                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2416         }
2417
2418         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2419                 if (signal_pending(current))
2420                         goto out_unlock;
2421                 if (vma->anon_vma)
2422                         vm_lock_anon_vma(mm, vma->anon_vma);
2423         }
2424
2425         ret = 0;
2426
2427 out_unlock:
2428         if (ret)
2429                 mm_drop_all_locks(mm);
2430
2431         return ret;
2432 }
2433
2434 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2435 {
2436         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2437                 /*
2438                  * The LSB of head.next can't change to 0 from under
2439                  * us because we hold the mm_all_locks_mutex.
2440                  *
2441                  * We must however clear the bitflag before unlocking
2442                  * the vma so the users using the anon_vma->head will
2443                  * never see our bitflag.
2444                  *
2445                  * No need of atomic instructions here, head.next
2446                  * can't change from under us until we release the
2447                  * anon_vma->lock.
2448                  */
2449                 if (!__test_and_clear_bit(0, (unsigned long *)
2450                                           &anon_vma->head.next))
2451                         BUG();
2452                 spin_unlock(&anon_vma->lock);
2453         }
2454 }
2455
2456 static void vm_unlock_mapping(struct address_space *mapping)
2457 {
2458         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2459                 /*
2460                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2461                  * because we hold the mm_all_locks_mutex.
2462                  */
2463                 spin_unlock(&mapping->i_mmap_lock);
2464                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2465                                         &mapping->flags))
2466                         BUG();
2467         }
2468 }
2469
2470 /*
2471  * The mmap_sem cannot be released by the caller until
2472  * mm_drop_all_locks() returns.
2473  */
2474 void mm_drop_all_locks(struct mm_struct *mm)
2475 {
2476         struct vm_area_struct *vma;
2477
2478         BUG_ON(down_read_trylock(&mm->mmap_sem));
2479         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2480
2481         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2482                 if (vma->anon_vma)
2483                         vm_unlock_anon_vma(vma->anon_vma);
2484                 if (vma->vm_file && vma->vm_file->f_mapping)
2485                         vm_unlock_mapping(vma->vm_file->f_mapping);
2486         }
2487
2488         mutex_unlock(&mm_all_locks_mutex);
2489 }
2490
2491 /*
2492  * initialise the VMA slab
2493  */
2494 void __init mmap_init(void)
2495 {
2496         int ret;
2497
2498         ret = percpu_counter_init(&vm_committed_as, 0);
2499         VM_BUG_ON(ret);
2500 }