Merge commit 'v2.6.30-rc6' into perfcounters/core
[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/ima.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/module.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_counter.h>
32
33 #include <asm/uaccess.h>
34 #include <asm/cacheflush.h>
35 #include <asm/tlb.h>
36 #include <asm/mmu_context.h>
37
38 #include "internal.h"
39
40 #ifndef arch_mmap_check
41 #define arch_mmap_check(addr, len, flags)       (0)
42 #endif
43
44 #ifndef arch_rebalance_pgtables
45 #define arch_rebalance_pgtables(addr, len)              (addr)
46 #endif
47
48 static void unmap_region(struct mm_struct *mm,
49                 struct vm_area_struct *vma, struct vm_area_struct *prev,
50                 unsigned long start, unsigned long end);
51
52 /*
53  * WARNING: the debugging will use recursive algorithms so never enable this
54  * unless you know what you are doing.
55  */
56 #undef DEBUG_MM_RB
57
58 /* description of effects of mapping type and prot in current implementation.
59  * this is due to the limited x86 page protection hardware.  The expected
60  * behavior is in parens:
61  *
62  * map_type     prot
63  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
64  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
65  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
66  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
67  *              
68  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
69  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
70  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
71  *
72  */
73 pgprot_t protection_map[16] = {
74         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
75         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
76 };
77
78 pgprot_t vm_get_page_prot(unsigned long vm_flags)
79 {
80         return __pgprot(pgprot_val(protection_map[vm_flags &
81                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
82                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
83 }
84 EXPORT_SYMBOL(vm_get_page_prot);
85
86 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
87 int sysctl_overcommit_ratio = 50;       /* default is 50% */
88 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
89 struct percpu_counter vm_committed_as;
90
91 /*
92  * Check that a process has enough memory to allocate a new virtual
93  * mapping. 0 means there is enough memory for the allocation to
94  * succeed and -ENOMEM implies there is not.
95  *
96  * We currently support three overcommit policies, which are set via the
97  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
98  *
99  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
100  * Additional code 2002 Jul 20 by Robert Love.
101  *
102  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
103  *
104  * Note this is a helper function intended to be used by LSMs which
105  * wish to use this logic.
106  */
107 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
108 {
109         unsigned long free, allowed;
110
111         vm_acct_memory(pages);
112
113         /*
114          * Sometimes we want to use more memory than we have
115          */
116         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
117                 return 0;
118
119         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
120                 unsigned long n;
121
122                 free = global_page_state(NR_FILE_PAGES);
123                 free += nr_swap_pages;
124
125                 /*
126                  * Any slabs which are created with the
127                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
128                  * which are reclaimable, under pressure.  The dentry
129                  * cache and most inode caches should fall into this
130                  */
131                 free += global_page_state(NR_SLAB_RECLAIMABLE);
132
133                 /*
134                  * Leave the last 3% for root
135                  */
136                 if (!cap_sys_admin)
137                         free -= free / 32;
138
139                 if (free > pages)
140                         return 0;
141
142                 /*
143                  * nr_free_pages() is very expensive on large systems,
144                  * only call if we're about to fail.
145                  */
146                 n = nr_free_pages();
147
148                 /*
149                  * Leave reserved pages. The pages are not for anonymous pages.
150                  */
151                 if (n <= totalreserve_pages)
152                         goto error;
153                 else
154                         n -= totalreserve_pages;
155
156                 /*
157                  * Leave the last 3% for root
158                  */
159                 if (!cap_sys_admin)
160                         n -= n / 32;
161                 free += n;
162
163                 if (free > pages)
164                         return 0;
165
166                 goto error;
167         }
168
169         allowed = (totalram_pages - hugetlb_total_pages())
170                 * sysctl_overcommit_ratio / 100;
171         /*
172          * Leave the last 3% for root
173          */
174         if (!cap_sys_admin)
175                 allowed -= allowed / 32;
176         allowed += total_swap_pages;
177
178         /* Don't let a single process grow too big:
179            leave 3% of the size of this process for other processes */
180         if (mm)
181                 allowed -= mm->total_vm / 32;
182
183         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
184                 return 0;
185 error:
186         vm_unacct_memory(pages);
187
188         return -ENOMEM;
189 }
190
191 /*
192  * Requires inode->i_mapping->i_mmap_lock
193  */
194 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
195                 struct file *file, struct address_space *mapping)
196 {
197         if (vma->vm_flags & VM_DENYWRITE)
198                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
199         if (vma->vm_flags & VM_SHARED)
200                 mapping->i_mmap_writable--;
201
202         flush_dcache_mmap_lock(mapping);
203         if (unlikely(vma->vm_flags & VM_NONLINEAR))
204                 list_del_init(&vma->shared.vm_set.list);
205         else
206                 vma_prio_tree_remove(vma, &mapping->i_mmap);
207         flush_dcache_mmap_unlock(mapping);
208 }
209
210 /*
211  * Unlink a file-based vm structure from its prio_tree, to hide
212  * vma from rmap and vmtruncate before freeing its page tables.
213  */
214 void unlink_file_vma(struct vm_area_struct *vma)
215 {
216         struct file *file = vma->vm_file;
217
218         if (file) {
219                 struct address_space *mapping = file->f_mapping;
220                 spin_lock(&mapping->i_mmap_lock);
221                 __remove_shared_vm_struct(vma, file, mapping);
222                 spin_unlock(&mapping->i_mmap_lock);
223         }
224 }
225
226 /*
227  * Close a vm structure and free it, returning the next.
228  */
229 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
230 {
231         struct vm_area_struct *next = vma->vm_next;
232
233         might_sleep();
234         if (vma->vm_ops && vma->vm_ops->close)
235                 vma->vm_ops->close(vma);
236         if (vma->vm_file) {
237                 fput(vma->vm_file);
238                 if (vma->vm_flags & VM_EXECUTABLE)
239                         removed_exe_file_vma(vma->vm_mm);
240         }
241         mpol_put(vma_policy(vma));
242         kmem_cache_free(vm_area_cachep, vma);
243         return next;
244 }
245
246 SYSCALL_DEFINE1(brk, unsigned long, brk)
247 {
248         unsigned long rlim, retval;
249         unsigned long newbrk, oldbrk;
250         struct mm_struct *mm = current->mm;
251         unsigned long min_brk;
252
253         down_write(&mm->mmap_sem);
254
255 #ifdef CONFIG_COMPAT_BRK
256         min_brk = mm->end_code;
257 #else
258         min_brk = mm->start_brk;
259 #endif
260         if (brk < min_brk)
261                 goto out;
262
263         /*
264          * Check against rlimit here. If this check is done later after the test
265          * of oldbrk with newbrk then it can escape the test and let the data
266          * segment grow beyond its set limit the in case where the limit is
267          * not page aligned -Ram Gupta
268          */
269         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
270         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
271                         (mm->end_data - mm->start_data) > rlim)
272                 goto out;
273
274         newbrk = PAGE_ALIGN(brk);
275         oldbrk = PAGE_ALIGN(mm->brk);
276         if (oldbrk == newbrk)
277                 goto set_brk;
278
279         /* Always allow shrinking brk. */
280         if (brk <= mm->brk) {
281                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
282                         goto set_brk;
283                 goto out;
284         }
285
286         /* Check against existing mmap mappings. */
287         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
288                 goto out;
289
290         /* Ok, looks good - let it rip. */
291         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
292                 goto out;
293 set_brk:
294         mm->brk = brk;
295 out:
296         retval = mm->brk;
297         up_write(&mm->mmap_sem);
298         return retval;
299 }
300
301 #ifdef DEBUG_MM_RB
302 static int browse_rb(struct rb_root *root)
303 {
304         int i = 0, j;
305         struct rb_node *nd, *pn = NULL;
306         unsigned long prev = 0, pend = 0;
307
308         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
309                 struct vm_area_struct *vma;
310                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
311                 if (vma->vm_start < prev)
312                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
313                 if (vma->vm_start < pend)
314                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
315                 if (vma->vm_start > vma->vm_end)
316                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
317                 i++;
318                 pn = nd;
319                 prev = vma->vm_start;
320                 pend = vma->vm_end;
321         }
322         j = 0;
323         for (nd = pn; nd; nd = rb_prev(nd)) {
324                 j++;
325         }
326         if (i != j)
327                 printk("backwards %d, forwards %d\n", j, i), i = 0;
328         return i;
329 }
330
331 void validate_mm(struct mm_struct *mm)
332 {
333         int bug = 0;
334         int i = 0;
335         struct vm_area_struct *tmp = mm->mmap;
336         while (tmp) {
337                 tmp = tmp->vm_next;
338                 i++;
339         }
340         if (i != mm->map_count)
341                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
342         i = browse_rb(&mm->mm_rb);
343         if (i != mm->map_count)
344                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
345         BUG_ON(bug);
346 }
347 #else
348 #define validate_mm(mm) do { } while (0)
349 #endif
350
351 static struct vm_area_struct *
352 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
353                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
354                 struct rb_node ** rb_parent)
355 {
356         struct vm_area_struct * vma;
357         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
358
359         __rb_link = &mm->mm_rb.rb_node;
360         rb_prev = __rb_parent = NULL;
361         vma = NULL;
362
363         while (*__rb_link) {
364                 struct vm_area_struct *vma_tmp;
365
366                 __rb_parent = *__rb_link;
367                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
368
369                 if (vma_tmp->vm_end > addr) {
370                         vma = vma_tmp;
371                         if (vma_tmp->vm_start <= addr)
372                                 break;
373                         __rb_link = &__rb_parent->rb_left;
374                 } else {
375                         rb_prev = __rb_parent;
376                         __rb_link = &__rb_parent->rb_right;
377                 }
378         }
379
380         *pprev = NULL;
381         if (rb_prev)
382                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
383         *rb_link = __rb_link;
384         *rb_parent = __rb_parent;
385         return vma;
386 }
387
388 static inline void
389 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
390                 struct vm_area_struct *prev, struct rb_node *rb_parent)
391 {
392         if (prev) {
393                 vma->vm_next = prev->vm_next;
394                 prev->vm_next = vma;
395         } else {
396                 mm->mmap = vma;
397                 if (rb_parent)
398                         vma->vm_next = rb_entry(rb_parent,
399                                         struct vm_area_struct, vm_rb);
400                 else
401                         vma->vm_next = NULL;
402         }
403 }
404
405 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
406                 struct rb_node **rb_link, struct rb_node *rb_parent)
407 {
408         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
409         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
410 }
411
412 static void __vma_link_file(struct vm_area_struct *vma)
413 {
414         struct file *file;
415
416         file = vma->vm_file;
417         if (file) {
418                 struct address_space *mapping = file->f_mapping;
419
420                 if (vma->vm_flags & VM_DENYWRITE)
421                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
422                 if (vma->vm_flags & VM_SHARED)
423                         mapping->i_mmap_writable++;
424
425                 flush_dcache_mmap_lock(mapping);
426                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
427                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
428                 else
429                         vma_prio_tree_insert(vma, &mapping->i_mmap);
430                 flush_dcache_mmap_unlock(mapping);
431         }
432 }
433
434 static void
435 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
436         struct vm_area_struct *prev, struct rb_node **rb_link,
437         struct rb_node *rb_parent)
438 {
439         __vma_link_list(mm, vma, prev, rb_parent);
440         __vma_link_rb(mm, vma, rb_link, rb_parent);
441         __anon_vma_link(vma);
442 }
443
444 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
445                         struct vm_area_struct *prev, struct rb_node **rb_link,
446                         struct rb_node *rb_parent)
447 {
448         struct address_space *mapping = NULL;
449
450         if (vma->vm_file)
451                 mapping = vma->vm_file->f_mapping;
452
453         if (mapping) {
454                 spin_lock(&mapping->i_mmap_lock);
455                 vma->vm_truncate_count = mapping->truncate_count;
456         }
457         anon_vma_lock(vma);
458
459         __vma_link(mm, vma, prev, rb_link, rb_parent);
460         __vma_link_file(vma);
461
462         anon_vma_unlock(vma);
463         if (mapping)
464                 spin_unlock(&mapping->i_mmap_lock);
465
466         mm->map_count++;
467         validate_mm(mm);
468 }
469
470 /*
471  * Helper for vma_adjust in the split_vma insert case:
472  * insert vm structure into list and rbtree and anon_vma,
473  * but it has already been inserted into prio_tree earlier.
474  */
475 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
476 {
477         struct vm_area_struct *__vma, *prev;
478         struct rb_node **rb_link, *rb_parent;
479
480         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
481         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
482         __vma_link(mm, vma, prev, rb_link, rb_parent);
483         mm->map_count++;
484 }
485
486 static inline void
487 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
488                 struct vm_area_struct *prev)
489 {
490         prev->vm_next = vma->vm_next;
491         rb_erase(&vma->vm_rb, &mm->mm_rb);
492         if (mm->mmap_cache == vma)
493                 mm->mmap_cache = prev;
494 }
495
496 /*
497  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
498  * is already present in an i_mmap tree without adjusting the tree.
499  * The following helper function should be used when such adjustments
500  * are necessary.  The "insert" vma (if any) is to be inserted
501  * before we drop the necessary locks.
502  */
503 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
504         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
505 {
506         struct mm_struct *mm = vma->vm_mm;
507         struct vm_area_struct *next = vma->vm_next;
508         struct vm_area_struct *importer = NULL;
509         struct address_space *mapping = NULL;
510         struct prio_tree_root *root = NULL;
511         struct file *file = vma->vm_file;
512         struct anon_vma *anon_vma = NULL;
513         long adjust_next = 0;
514         int remove_next = 0;
515
516         if (next && !insert) {
517                 if (end >= next->vm_end) {
518                         /*
519                          * vma expands, overlapping all the next, and
520                          * perhaps the one after too (mprotect case 6).
521                          */
522 again:                  remove_next = 1 + (end > next->vm_end);
523                         end = next->vm_end;
524                         anon_vma = next->anon_vma;
525                         importer = vma;
526                 } else if (end > next->vm_start) {
527                         /*
528                          * vma expands, overlapping part of the next:
529                          * mprotect case 5 shifting the boundary up.
530                          */
531                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
532                         anon_vma = next->anon_vma;
533                         importer = vma;
534                 } else if (end < vma->vm_end) {
535                         /*
536                          * vma shrinks, and !insert tells it's not
537                          * split_vma inserting another: so it must be
538                          * mprotect case 4 shifting the boundary down.
539                          */
540                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
541                         anon_vma = next->anon_vma;
542                         importer = next;
543                 }
544         }
545
546         if (file) {
547                 mapping = file->f_mapping;
548                 if (!(vma->vm_flags & VM_NONLINEAR))
549                         root = &mapping->i_mmap;
550                 spin_lock(&mapping->i_mmap_lock);
551                 if (importer &&
552                     vma->vm_truncate_count != next->vm_truncate_count) {
553                         /*
554                          * unmap_mapping_range might be in progress:
555                          * ensure that the expanding vma is rescanned.
556                          */
557                         importer->vm_truncate_count = 0;
558                 }
559                 if (insert) {
560                         insert->vm_truncate_count = vma->vm_truncate_count;
561                         /*
562                          * Put into prio_tree now, so instantiated pages
563                          * are visible to arm/parisc __flush_dcache_page
564                          * throughout; but we cannot insert into address
565                          * space until vma start or end is updated.
566                          */
567                         __vma_link_file(insert);
568                 }
569         }
570
571         /*
572          * When changing only vma->vm_end, we don't really need
573          * anon_vma lock: but is that case worth optimizing out?
574          */
575         if (vma->anon_vma)
576                 anon_vma = vma->anon_vma;
577         if (anon_vma) {
578                 spin_lock(&anon_vma->lock);
579                 /*
580                  * Easily overlooked: when mprotect shifts the boundary,
581                  * make sure the expanding vma has anon_vma set if the
582                  * shrinking vma had, to cover any anon pages imported.
583                  */
584                 if (importer && !importer->anon_vma) {
585                         importer->anon_vma = anon_vma;
586                         __anon_vma_link(importer);
587                 }
588         }
589
590         if (root) {
591                 flush_dcache_mmap_lock(mapping);
592                 vma_prio_tree_remove(vma, root);
593                 if (adjust_next)
594                         vma_prio_tree_remove(next, root);
595         }
596
597         vma->vm_start = start;
598         vma->vm_end = end;
599         vma->vm_pgoff = pgoff;
600         if (adjust_next) {
601                 next->vm_start += adjust_next << PAGE_SHIFT;
602                 next->vm_pgoff += adjust_next;
603         }
604
605         if (root) {
606                 if (adjust_next)
607                         vma_prio_tree_insert(next, root);
608                 vma_prio_tree_insert(vma, root);
609                 flush_dcache_mmap_unlock(mapping);
610         }
611
612         if (remove_next) {
613                 /*
614                  * vma_merge has merged next into vma, and needs
615                  * us to remove next before dropping the locks.
616                  */
617                 __vma_unlink(mm, next, vma);
618                 if (file)
619                         __remove_shared_vm_struct(next, file, mapping);
620                 if (next->anon_vma)
621                         __anon_vma_merge(vma, next);
622         } else if (insert) {
623                 /*
624                  * split_vma has split insert from vma, and needs
625                  * us to insert it before dropping the locks
626                  * (it may either follow vma or precede it).
627                  */
628                 __insert_vm_struct(mm, insert);
629         }
630
631         if (anon_vma)
632                 spin_unlock(&anon_vma->lock);
633         if (mapping)
634                 spin_unlock(&mapping->i_mmap_lock);
635
636         if (remove_next) {
637                 if (file) {
638                         fput(file);
639                         if (next->vm_flags & VM_EXECUTABLE)
640                                 removed_exe_file_vma(mm);
641                 }
642                 mm->map_count--;
643                 mpol_put(vma_policy(next));
644                 kmem_cache_free(vm_area_cachep, next);
645                 /*
646                  * In mprotect's case 6 (see comments on vma_merge),
647                  * we must remove another next too. It would clutter
648                  * up the code too much to do both in one go.
649                  */
650                 if (remove_next == 2) {
651                         next = vma->vm_next;
652                         goto again;
653                 }
654         }
655
656         validate_mm(mm);
657 }
658
659 /* Flags that can be inherited from an existing mapping when merging */
660 #define VM_MERGEABLE_FLAGS (VM_CAN_NONLINEAR)
661
662 /*
663  * If the vma has a ->close operation then the driver probably needs to release
664  * per-vma resources, so we don't attempt to merge those.
665  */
666 static inline int is_mergeable_vma(struct vm_area_struct *vma,
667                         struct file *file, unsigned long vm_flags)
668 {
669         if ((vma->vm_flags ^ vm_flags) & ~VM_MERGEABLE_FLAGS)
670                 return 0;
671         if (vma->vm_file != file)
672                 return 0;
673         if (vma->vm_ops && vma->vm_ops->close)
674                 return 0;
675         return 1;
676 }
677
678 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
679                                         struct anon_vma *anon_vma2)
680 {
681         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
682 }
683
684 /*
685  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
686  * in front of (at a lower virtual address and file offset than) the vma.
687  *
688  * We cannot merge two vmas if they have differently assigned (non-NULL)
689  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
690  *
691  * We don't check here for the merged mmap wrapping around the end of pagecache
692  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
693  * wrap, nor mmaps which cover the final page at index -1UL.
694  */
695 static int
696 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
697         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
698 {
699         if (is_mergeable_vma(vma, file, vm_flags) &&
700             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
701                 if (vma->vm_pgoff == vm_pgoff)
702                         return 1;
703         }
704         return 0;
705 }
706
707 /*
708  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
709  * beyond (at a higher virtual address and file offset than) the vma.
710  *
711  * We cannot merge two vmas if they have differently assigned (non-NULL)
712  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
713  */
714 static int
715 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
716         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
717 {
718         if (is_mergeable_vma(vma, file, vm_flags) &&
719             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
720                 pgoff_t vm_pglen;
721                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
722                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
723                         return 1;
724         }
725         return 0;
726 }
727
728 /*
729  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
730  * whether that can be merged with its predecessor or its successor.
731  * Or both (it neatly fills a hole).
732  *
733  * In most cases - when called for mmap, brk or mremap - [addr,end) is
734  * certain not to be mapped by the time vma_merge is called; but when
735  * called for mprotect, it is certain to be already mapped (either at
736  * an offset within prev, or at the start of next), and the flags of
737  * this area are about to be changed to vm_flags - and the no-change
738  * case has already been eliminated.
739  *
740  * The following mprotect cases have to be considered, where AAAA is
741  * the area passed down from mprotect_fixup, never extending beyond one
742  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
743  *
744  *     AAAA             AAAA                AAAA          AAAA
745  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
746  *    cannot merge    might become    might become    might become
747  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
748  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
749  *    mremap move:                                    PPPPNNNNNNNN 8
750  *        AAAA
751  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
752  *    might become    case 1 below    case 2 below    case 3 below
753  *
754  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
755  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
756  */
757 struct vm_area_struct *vma_merge(struct mm_struct *mm,
758                         struct vm_area_struct *prev, unsigned long addr,
759                         unsigned long end, unsigned long vm_flags,
760                         struct anon_vma *anon_vma, struct file *file,
761                         pgoff_t pgoff, struct mempolicy *policy)
762 {
763         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
764         struct vm_area_struct *area, *next;
765
766         /*
767          * We later require that vma->vm_flags == vm_flags,
768          * so this tests vma->vm_flags & VM_SPECIAL, too.
769          */
770         if (vm_flags & VM_SPECIAL)
771                 return NULL;
772
773         if (prev)
774                 next = prev->vm_next;
775         else
776                 next = mm->mmap;
777         area = next;
778         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
779                 next = next->vm_next;
780
781         /*
782          * Can it merge with the predecessor?
783          */
784         if (prev && prev->vm_end == addr &&
785                         mpol_equal(vma_policy(prev), policy) &&
786                         can_vma_merge_after(prev, vm_flags,
787                                                 anon_vma, file, pgoff)) {
788                 /*
789                  * OK, it can.  Can we now merge in the successor as well?
790                  */
791                 if (next && end == next->vm_start &&
792                                 mpol_equal(policy, vma_policy(next)) &&
793                                 can_vma_merge_before(next, vm_flags,
794                                         anon_vma, file, pgoff+pglen) &&
795                                 is_mergeable_anon_vma(prev->anon_vma,
796                                                       next->anon_vma)) {
797                                                         /* cases 1, 6 */
798                         vma_adjust(prev, prev->vm_start,
799                                 next->vm_end, prev->vm_pgoff, NULL);
800                 } else                                  /* cases 2, 5, 7 */
801                         vma_adjust(prev, prev->vm_start,
802                                 end, prev->vm_pgoff, NULL);
803                 return prev;
804         }
805
806         /*
807          * Can this new request be merged in front of next?
808          */
809         if (next && end == next->vm_start &&
810                         mpol_equal(policy, vma_policy(next)) &&
811                         can_vma_merge_before(next, vm_flags,
812                                         anon_vma, file, pgoff+pglen)) {
813                 if (prev && addr < prev->vm_end)        /* case 4 */
814                         vma_adjust(prev, prev->vm_start,
815                                 addr, prev->vm_pgoff, NULL);
816                 else                                    /* cases 3, 8 */
817                         vma_adjust(area, addr, next->vm_end,
818                                 next->vm_pgoff - pglen, NULL);
819                 return area;
820         }
821
822         return NULL;
823 }
824
825 /*
826  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
827  * neighbouring vmas for a suitable anon_vma, before it goes off
828  * to allocate a new anon_vma.  It checks because a repetitive
829  * sequence of mprotects and faults may otherwise lead to distinct
830  * anon_vmas being allocated, preventing vma merge in subsequent
831  * mprotect.
832  */
833 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
834 {
835         struct vm_area_struct *near;
836         unsigned long vm_flags;
837
838         near = vma->vm_next;
839         if (!near)
840                 goto try_prev;
841
842         /*
843          * Since only mprotect tries to remerge vmas, match flags
844          * which might be mprotected into each other later on.
845          * Neither mlock nor madvise tries to remerge at present,
846          * so leave their flags as obstructing a merge.
847          */
848         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
849         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
850
851         if (near->anon_vma && vma->vm_end == near->vm_start &&
852                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
853                         can_vma_merge_before(near, vm_flags,
854                                 NULL, vma->vm_file, vma->vm_pgoff +
855                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
856                 return near->anon_vma;
857 try_prev:
858         /*
859          * It is potentially slow to have to call find_vma_prev here.
860          * But it's only on the first write fault on the vma, not
861          * every time, and we could devise a way to avoid it later
862          * (e.g. stash info in next's anon_vma_node when assigning
863          * an anon_vma, or when trying vma_merge).  Another time.
864          */
865         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
866         if (!near)
867                 goto none;
868
869         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
870         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
871
872         if (near->anon_vma && near->vm_end == vma->vm_start &&
873                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
874                         can_vma_merge_after(near, vm_flags,
875                                 NULL, vma->vm_file, vma->vm_pgoff))
876                 return near->anon_vma;
877 none:
878         /*
879          * There's no absolute need to look only at touching neighbours:
880          * we could search further afield for "compatible" anon_vmas.
881          * But it would probably just be a waste of time searching,
882          * or lead to too many vmas hanging off the same anon_vma.
883          * We're trying to allow mprotect remerging later on,
884          * not trying to minimize memory used for anon_vmas.
885          */
886         return NULL;
887 }
888
889 #ifdef CONFIG_PROC_FS
890 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
891                                                 struct file *file, long pages)
892 {
893         const unsigned long stack_flags
894                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
895
896         if (file) {
897                 mm->shared_vm += pages;
898                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
899                         mm->exec_vm += pages;
900         } else if (flags & stack_flags)
901                 mm->stack_vm += pages;
902         if (flags & (VM_RESERVED|VM_IO))
903                 mm->reserved_vm += pages;
904 }
905 #endif /* CONFIG_PROC_FS */
906
907 /*
908  * The caller must hold down_write(current->mm->mmap_sem).
909  */
910
911 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
912                         unsigned long len, unsigned long prot,
913                         unsigned long flags, unsigned long pgoff)
914 {
915         struct mm_struct * mm = current->mm;
916         struct inode *inode;
917         unsigned int vm_flags;
918         int error;
919         unsigned long reqprot = prot;
920
921         /*
922          * Does the application expect PROT_READ to imply PROT_EXEC?
923          *
924          * (the exception is when the underlying filesystem is noexec
925          *  mounted, in which case we dont add PROT_EXEC.)
926          */
927         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
928                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
929                         prot |= PROT_EXEC;
930
931         if (!len)
932                 return -EINVAL;
933
934         if (!(flags & MAP_FIXED))
935                 addr = round_hint_to_min(addr);
936
937         error = arch_mmap_check(addr, len, flags);
938         if (error)
939                 return error;
940
941         /* Careful about overflows.. */
942         len = PAGE_ALIGN(len);
943         if (!len || len > TASK_SIZE)
944                 return -ENOMEM;
945
946         /* offset overflow? */
947         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
948                return -EOVERFLOW;
949
950         /* Too many mappings? */
951         if (mm->map_count > sysctl_max_map_count)
952                 return -ENOMEM;
953
954         /* Obtain the address to map to. we verify (or select) it and ensure
955          * that it represents a valid section of the address space.
956          */
957         addr = get_unmapped_area(file, addr, len, pgoff, flags);
958         if (addr & ~PAGE_MASK)
959                 return addr;
960
961         /* Do simple checking here so the lower-level routines won't have
962          * to. we assume access permissions have been handled by the open
963          * of the memory object, so we don't do any here.
964          */
965         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
966                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
967
968         if (flags & MAP_LOCKED) {
969                 if (!can_do_mlock())
970                         return -EPERM;
971                 vm_flags |= VM_LOCKED;
972         }
973
974         /* mlock MCL_FUTURE? */
975         if (vm_flags & VM_LOCKED) {
976                 unsigned long locked, lock_limit;
977                 locked = len >> PAGE_SHIFT;
978                 locked += mm->locked_vm;
979                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
980                 lock_limit >>= PAGE_SHIFT;
981                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
982                         return -EAGAIN;
983         }
984
985         inode = file ? file->f_path.dentry->d_inode : NULL;
986
987         if (file) {
988                 switch (flags & MAP_TYPE) {
989                 case MAP_SHARED:
990                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
991                                 return -EACCES;
992
993                         /*
994                          * Make sure we don't allow writing to an append-only
995                          * file..
996                          */
997                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
998                                 return -EACCES;
999
1000                         /*
1001                          * Make sure there are no mandatory locks on the file.
1002                          */
1003                         if (locks_verify_locked(inode))
1004                                 return -EAGAIN;
1005
1006                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1007                         if (!(file->f_mode & FMODE_WRITE))
1008                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1009
1010                         /* fall through */
1011                 case MAP_PRIVATE:
1012                         if (!(file->f_mode & FMODE_READ))
1013                                 return -EACCES;
1014                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1015                                 if (vm_flags & VM_EXEC)
1016                                         return -EPERM;
1017                                 vm_flags &= ~VM_MAYEXEC;
1018                         }
1019
1020                         if (!file->f_op || !file->f_op->mmap)
1021                                 return -ENODEV;
1022                         break;
1023
1024                 default:
1025                         return -EINVAL;
1026                 }
1027         } else {
1028                 switch (flags & MAP_TYPE) {
1029                 case MAP_SHARED:
1030                         /*
1031                          * Ignore pgoff.
1032                          */
1033                         pgoff = 0;
1034                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1035                         break;
1036                 case MAP_PRIVATE:
1037                         /*
1038                          * Set pgoff according to addr for anon_vma.
1039                          */
1040                         pgoff = addr >> PAGE_SHIFT;
1041                         break;
1042                 default:
1043                         return -EINVAL;
1044                 }
1045         }
1046
1047         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1048         if (error)
1049                 return error;
1050         error = ima_file_mmap(file, prot);
1051         if (error)
1052                 return error;
1053
1054         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1055 }
1056 EXPORT_SYMBOL(do_mmap_pgoff);
1057
1058 /*
1059  * Some shared mappigns will want the pages marked read-only
1060  * to track write events. If so, we'll downgrade vm_page_prot
1061  * to the private version (using protection_map[] without the
1062  * VM_SHARED bit).
1063  */
1064 int vma_wants_writenotify(struct vm_area_struct *vma)
1065 {
1066         unsigned int vm_flags = vma->vm_flags;
1067
1068         /* If it was private or non-writable, the write bit is already clear */
1069         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1070                 return 0;
1071
1072         /* The backer wishes to know when pages are first written to? */
1073         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1074                 return 1;
1075
1076         /* The open routine did something to the protections already? */
1077         if (pgprot_val(vma->vm_page_prot) !=
1078             pgprot_val(vm_get_page_prot(vm_flags)))
1079                 return 0;
1080
1081         /* Specialty mapping? */
1082         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1083                 return 0;
1084
1085         /* Can the mapping track the dirty pages? */
1086         return vma->vm_file && vma->vm_file->f_mapping &&
1087                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1088 }
1089
1090 /*
1091  * We account for memory if it's a private writeable mapping,
1092  * not hugepages and VM_NORESERVE wasn't set.
1093  */
1094 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1095 {
1096         /*
1097          * hugetlb has its own accounting separate from the core VM
1098          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1099          */
1100         if (file && is_file_hugepages(file))
1101                 return 0;
1102
1103         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1104 }
1105
1106 unsigned long mmap_region(struct file *file, unsigned long addr,
1107                           unsigned long len, unsigned long flags,
1108                           unsigned int vm_flags, unsigned long pgoff)
1109 {
1110         struct mm_struct *mm = current->mm;
1111         struct vm_area_struct *vma, *prev;
1112         int correct_wcount = 0;
1113         int error;
1114         struct rb_node **rb_link, *rb_parent;
1115         unsigned long charged = 0;
1116         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1117
1118         /* Clear old maps */
1119         error = -ENOMEM;
1120 munmap_back:
1121         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1122         if (vma && vma->vm_start < addr + len) {
1123                 if (do_munmap(mm, addr, len))
1124                         return -ENOMEM;
1125                 goto munmap_back;
1126         }
1127
1128         /* Check against address space limit. */
1129         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1130                 return -ENOMEM;
1131
1132         /*
1133          * Set 'VM_NORESERVE' if we should not account for the
1134          * memory use of this mapping.
1135          */
1136         if ((flags & MAP_NORESERVE)) {
1137                 /* We honor MAP_NORESERVE if allowed to overcommit */
1138                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1139                         vm_flags |= VM_NORESERVE;
1140
1141                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1142                 if (file && is_file_hugepages(file))
1143                         vm_flags |= VM_NORESERVE;
1144         }
1145
1146         /*
1147          * Private writable mapping: check memory availability
1148          */
1149         if (accountable_mapping(file, vm_flags)) {
1150                 charged = len >> PAGE_SHIFT;
1151                 if (security_vm_enough_memory(charged))
1152                         return -ENOMEM;
1153                 vm_flags |= VM_ACCOUNT;
1154         }
1155
1156         /*
1157          * Can we just expand an old mapping?
1158          */
1159         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1160         if (vma)
1161                 goto out;
1162
1163         /*
1164          * Determine the object being mapped and call the appropriate
1165          * specific mapper. the address has already been validated, but
1166          * not unmapped, but the maps are removed from the list.
1167          */
1168         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1169         if (!vma) {
1170                 error = -ENOMEM;
1171                 goto unacct_error;
1172         }
1173
1174         vma->vm_mm = mm;
1175         vma->vm_start = addr;
1176         vma->vm_end = addr + len;
1177         vma->vm_flags = vm_flags;
1178         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1179         vma->vm_pgoff = pgoff;
1180
1181         if (file) {
1182                 error = -EINVAL;
1183                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1184                         goto free_vma;
1185                 if (vm_flags & VM_DENYWRITE) {
1186                         error = deny_write_access(file);
1187                         if (error)
1188                                 goto free_vma;
1189                         correct_wcount = 1;
1190                 }
1191                 vma->vm_file = file;
1192                 get_file(file);
1193                 error = file->f_op->mmap(file, vma);
1194                 if (error)
1195                         goto unmap_and_free_vma;
1196                 if (vm_flags & VM_EXECUTABLE)
1197                         added_exe_file_vma(mm);
1198         } else if (vm_flags & VM_SHARED) {
1199                 error = shmem_zero_setup(vma);
1200                 if (error)
1201                         goto free_vma;
1202         }
1203
1204         /* Can addr have changed??
1205          *
1206          * Answer: Yes, several device drivers can do it in their
1207          *         f_op->mmap method. -DaveM
1208          */
1209         addr = vma->vm_start;
1210         pgoff = vma->vm_pgoff;
1211         vm_flags = vma->vm_flags;
1212
1213         if (vma_wants_writenotify(vma))
1214                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1215
1216         vma_link(mm, vma, prev, rb_link, rb_parent);
1217         file = vma->vm_file;
1218
1219         /* Once vma denies write, undo our temporary denial count */
1220         if (correct_wcount)
1221                 atomic_inc(&inode->i_writecount);
1222 out:
1223         if (vm_flags & VM_EXEC)
1224                 perf_counter_mmap(addr, len, pgoff, file);
1225
1226         mm->total_vm += len >> PAGE_SHIFT;
1227         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1228         if (vm_flags & VM_LOCKED) {
1229                 /*
1230                  * makes pages present; downgrades, drops, reacquires mmap_sem
1231                  */
1232                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1233                 if (nr_pages < 0)
1234                         return nr_pages;        /* vma gone! */
1235                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1236         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1237                 make_pages_present(addr, addr + len);
1238         return addr;
1239
1240 unmap_and_free_vma:
1241         if (correct_wcount)
1242                 atomic_inc(&inode->i_writecount);
1243         vma->vm_file = NULL;
1244         fput(file);
1245
1246         /* Undo any partial mapping done by a device driver. */
1247         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1248         charged = 0;
1249 free_vma:
1250         kmem_cache_free(vm_area_cachep, vma);
1251 unacct_error:
1252         if (charged)
1253                 vm_unacct_memory(charged);
1254         return error;
1255 }
1256
1257 /* Get an address range which is currently unmapped.
1258  * For shmat() with addr=0.
1259  *
1260  * Ugly calling convention alert:
1261  * Return value with the low bits set means error value,
1262  * ie
1263  *      if (ret & ~PAGE_MASK)
1264  *              error = ret;
1265  *
1266  * This function "knows" that -ENOMEM has the bits set.
1267  */
1268 #ifndef HAVE_ARCH_UNMAPPED_AREA
1269 unsigned long
1270 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1271                 unsigned long len, unsigned long pgoff, unsigned long flags)
1272 {
1273         struct mm_struct *mm = current->mm;
1274         struct vm_area_struct *vma;
1275         unsigned long start_addr;
1276
1277         if (len > TASK_SIZE)
1278                 return -ENOMEM;
1279
1280         if (flags & MAP_FIXED)
1281                 return addr;
1282
1283         if (addr) {
1284                 addr = PAGE_ALIGN(addr);
1285                 vma = find_vma(mm, addr);
1286                 if (TASK_SIZE - len >= addr &&
1287                     (!vma || addr + len <= vma->vm_start))
1288                         return addr;
1289         }
1290         if (len > mm->cached_hole_size) {
1291                 start_addr = addr = mm->free_area_cache;
1292         } else {
1293                 start_addr = addr = TASK_UNMAPPED_BASE;
1294                 mm->cached_hole_size = 0;
1295         }
1296
1297 full_search:
1298         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1299                 /* At this point:  (!vma || addr < vma->vm_end). */
1300                 if (TASK_SIZE - len < addr) {
1301                         /*
1302                          * Start a new search - just in case we missed
1303                          * some holes.
1304                          */
1305                         if (start_addr != TASK_UNMAPPED_BASE) {
1306                                 addr = TASK_UNMAPPED_BASE;
1307                                 start_addr = addr;
1308                                 mm->cached_hole_size = 0;
1309                                 goto full_search;
1310                         }
1311                         return -ENOMEM;
1312                 }
1313                 if (!vma || addr + len <= vma->vm_start) {
1314                         /*
1315                          * Remember the place where we stopped the search:
1316                          */
1317                         mm->free_area_cache = addr + len;
1318                         return addr;
1319                 }
1320                 if (addr + mm->cached_hole_size < vma->vm_start)
1321                         mm->cached_hole_size = vma->vm_start - addr;
1322                 addr = vma->vm_end;
1323         }
1324 }
1325 #endif  
1326
1327 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1328 {
1329         /*
1330          * Is this a new hole at the lowest possible address?
1331          */
1332         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1333                 mm->free_area_cache = addr;
1334                 mm->cached_hole_size = ~0UL;
1335         }
1336 }
1337
1338 /*
1339  * This mmap-allocator allocates new areas top-down from below the
1340  * stack's low limit (the base):
1341  */
1342 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1343 unsigned long
1344 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1345                           const unsigned long len, const unsigned long pgoff,
1346                           const unsigned long flags)
1347 {
1348         struct vm_area_struct *vma;
1349         struct mm_struct *mm = current->mm;
1350         unsigned long addr = addr0;
1351
1352         /* requested length too big for entire address space */
1353         if (len > TASK_SIZE)
1354                 return -ENOMEM;
1355
1356         if (flags & MAP_FIXED)
1357                 return addr;
1358
1359         /* requesting a specific address */
1360         if (addr) {
1361                 addr = PAGE_ALIGN(addr);
1362                 vma = find_vma(mm, addr);
1363                 if (TASK_SIZE - len >= addr &&
1364                                 (!vma || addr + len <= vma->vm_start))
1365                         return addr;
1366         }
1367
1368         /* check if free_area_cache is useful for us */
1369         if (len <= mm->cached_hole_size) {
1370                 mm->cached_hole_size = 0;
1371                 mm->free_area_cache = mm->mmap_base;
1372         }
1373
1374         /* either no address requested or can't fit in requested address hole */
1375         addr = mm->free_area_cache;
1376
1377         /* make sure it can fit in the remaining address space */
1378         if (addr > len) {
1379                 vma = find_vma(mm, addr-len);
1380                 if (!vma || addr <= vma->vm_start)
1381                         /* remember the address as a hint for next time */
1382                         return (mm->free_area_cache = addr-len);
1383         }
1384
1385         if (mm->mmap_base < len)
1386                 goto bottomup;
1387
1388         addr = mm->mmap_base-len;
1389
1390         do {
1391                 /*
1392                  * Lookup failure means no vma is above this address,
1393                  * else if new region fits below vma->vm_start,
1394                  * return with success:
1395                  */
1396                 vma = find_vma(mm, addr);
1397                 if (!vma || addr+len <= vma->vm_start)
1398                         /* remember the address as a hint for next time */
1399                         return (mm->free_area_cache = addr);
1400
1401                 /* remember the largest hole we saw so far */
1402                 if (addr + mm->cached_hole_size < vma->vm_start)
1403                         mm->cached_hole_size = vma->vm_start - addr;
1404
1405                 /* try just below the current vma->vm_start */
1406                 addr = vma->vm_start-len;
1407         } while (len < vma->vm_start);
1408
1409 bottomup:
1410         /*
1411          * A failed mmap() very likely causes application failure,
1412          * so fall back to the bottom-up function here. This scenario
1413          * can happen with large stack limits and large mmap()
1414          * allocations.
1415          */
1416         mm->cached_hole_size = ~0UL;
1417         mm->free_area_cache = TASK_UNMAPPED_BASE;
1418         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1419         /*
1420          * Restore the topdown base:
1421          */
1422         mm->free_area_cache = mm->mmap_base;
1423         mm->cached_hole_size = ~0UL;
1424
1425         return addr;
1426 }
1427 #endif
1428
1429 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1430 {
1431         /*
1432          * Is this a new hole at the highest possible address?
1433          */
1434         if (addr > mm->free_area_cache)
1435                 mm->free_area_cache = addr;
1436
1437         /* dont allow allocations above current base */
1438         if (mm->free_area_cache > mm->mmap_base)
1439                 mm->free_area_cache = mm->mmap_base;
1440 }
1441
1442 unsigned long
1443 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1444                 unsigned long pgoff, unsigned long flags)
1445 {
1446         unsigned long (*get_area)(struct file *, unsigned long,
1447                                   unsigned long, unsigned long, unsigned long);
1448
1449         get_area = current->mm->get_unmapped_area;
1450         if (file && file->f_op && file->f_op->get_unmapped_area)
1451                 get_area = file->f_op->get_unmapped_area;
1452         addr = get_area(file, addr, len, pgoff, flags);
1453         if (IS_ERR_VALUE(addr))
1454                 return addr;
1455
1456         if (addr > TASK_SIZE - len)
1457                 return -ENOMEM;
1458         if (addr & ~PAGE_MASK)
1459                 return -EINVAL;
1460
1461         return arch_rebalance_pgtables(addr, len);
1462 }
1463
1464 EXPORT_SYMBOL(get_unmapped_area);
1465
1466 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1467 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1468 {
1469         struct vm_area_struct *vma = NULL;
1470
1471         if (mm) {
1472                 /* Check the cache first. */
1473                 /* (Cache hit rate is typically around 35%.) */
1474                 vma = mm->mmap_cache;
1475                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1476                         struct rb_node * rb_node;
1477
1478                         rb_node = mm->mm_rb.rb_node;
1479                         vma = NULL;
1480
1481                         while (rb_node) {
1482                                 struct vm_area_struct * vma_tmp;
1483
1484                                 vma_tmp = rb_entry(rb_node,
1485                                                 struct vm_area_struct, vm_rb);
1486
1487                                 if (vma_tmp->vm_end > addr) {
1488                                         vma = vma_tmp;
1489                                         if (vma_tmp->vm_start <= addr)
1490                                                 break;
1491                                         rb_node = rb_node->rb_left;
1492                                 } else
1493                                         rb_node = rb_node->rb_right;
1494                         }
1495                         if (vma)
1496                                 mm->mmap_cache = vma;
1497                 }
1498         }
1499         return vma;
1500 }
1501
1502 EXPORT_SYMBOL(find_vma);
1503
1504 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1505 struct vm_area_struct *
1506 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1507                         struct vm_area_struct **pprev)
1508 {
1509         struct vm_area_struct *vma = NULL, *prev = NULL;
1510         struct rb_node *rb_node;
1511         if (!mm)
1512                 goto out;
1513
1514         /* Guard against addr being lower than the first VMA */
1515         vma = mm->mmap;
1516
1517         /* Go through the RB tree quickly. */
1518         rb_node = mm->mm_rb.rb_node;
1519
1520         while (rb_node) {
1521                 struct vm_area_struct *vma_tmp;
1522                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1523
1524                 if (addr < vma_tmp->vm_end) {
1525                         rb_node = rb_node->rb_left;
1526                 } else {
1527                         prev = vma_tmp;
1528                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1529                                 break;
1530                         rb_node = rb_node->rb_right;
1531                 }
1532         }
1533
1534 out:
1535         *pprev = prev;
1536         return prev ? prev->vm_next : vma;
1537 }
1538
1539 /*
1540  * Verify that the stack growth is acceptable and
1541  * update accounting. This is shared with both the
1542  * grow-up and grow-down cases.
1543  */
1544 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1545 {
1546         struct mm_struct *mm = vma->vm_mm;
1547         struct rlimit *rlim = current->signal->rlim;
1548         unsigned long new_start;
1549
1550         /* address space limit tests */
1551         if (!may_expand_vm(mm, grow))
1552                 return -ENOMEM;
1553
1554         /* Stack limit test */
1555         if (size > rlim[RLIMIT_STACK].rlim_cur)
1556                 return -ENOMEM;
1557
1558         /* mlock limit tests */
1559         if (vma->vm_flags & VM_LOCKED) {
1560                 unsigned long locked;
1561                 unsigned long limit;
1562                 locked = mm->locked_vm + grow;
1563                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1564                 if (locked > limit && !capable(CAP_IPC_LOCK))
1565                         return -ENOMEM;
1566         }
1567
1568         /* Check to ensure the stack will not grow into a hugetlb-only region */
1569         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1570                         vma->vm_end - size;
1571         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1572                 return -EFAULT;
1573
1574         /*
1575          * Overcommit..  This must be the final test, as it will
1576          * update security statistics.
1577          */
1578         if (security_vm_enough_memory_mm(mm, grow))
1579                 return -ENOMEM;
1580
1581         /* Ok, everything looks good - let it rip */
1582         mm->total_vm += grow;
1583         if (vma->vm_flags & VM_LOCKED)
1584                 mm->locked_vm += grow;
1585         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1586         return 0;
1587 }
1588
1589 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1590 /*
1591  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1592  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1593  */
1594 #ifndef CONFIG_IA64
1595 static
1596 #endif
1597 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1598 {
1599         int error;
1600
1601         if (!(vma->vm_flags & VM_GROWSUP))
1602                 return -EFAULT;
1603
1604         /*
1605          * We must make sure the anon_vma is allocated
1606          * so that the anon_vma locking is not a noop.
1607          */
1608         if (unlikely(anon_vma_prepare(vma)))
1609                 return -ENOMEM;
1610         anon_vma_lock(vma);
1611
1612         /*
1613          * vma->vm_start/vm_end cannot change under us because the caller
1614          * is required to hold the mmap_sem in read mode.  We need the
1615          * anon_vma lock to serialize against concurrent expand_stacks.
1616          * Also guard against wrapping around to address 0.
1617          */
1618         if (address < PAGE_ALIGN(address+4))
1619                 address = PAGE_ALIGN(address+4);
1620         else {
1621                 anon_vma_unlock(vma);
1622                 return -ENOMEM;
1623         }
1624         error = 0;
1625
1626         /* Somebody else might have raced and expanded it already */
1627         if (address > vma->vm_end) {
1628                 unsigned long size, grow;
1629
1630                 size = address - vma->vm_start;
1631                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1632
1633                 error = acct_stack_growth(vma, size, grow);
1634                 if (!error)
1635                         vma->vm_end = address;
1636         }
1637         anon_vma_unlock(vma);
1638         return error;
1639 }
1640 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1641
1642 /*
1643  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1644  */
1645 static int expand_downwards(struct vm_area_struct *vma,
1646                                    unsigned long address)
1647 {
1648         int error;
1649
1650         /*
1651          * We must make sure the anon_vma is allocated
1652          * so that the anon_vma locking is not a noop.
1653          */
1654         if (unlikely(anon_vma_prepare(vma)))
1655                 return -ENOMEM;
1656
1657         address &= PAGE_MASK;
1658         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1659         if (error)
1660                 return error;
1661
1662         anon_vma_lock(vma);
1663
1664         /*
1665          * vma->vm_start/vm_end cannot change under us because the caller
1666          * is required to hold the mmap_sem in read mode.  We need the
1667          * anon_vma lock to serialize against concurrent expand_stacks.
1668          */
1669
1670         /* Somebody else might have raced and expanded it already */
1671         if (address < vma->vm_start) {
1672                 unsigned long size, grow;
1673
1674                 size = vma->vm_end - address;
1675                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1676
1677                 error = acct_stack_growth(vma, size, grow);
1678                 if (!error) {
1679                         vma->vm_start = address;
1680                         vma->vm_pgoff -= grow;
1681                 }
1682         }
1683         anon_vma_unlock(vma);
1684         return error;
1685 }
1686
1687 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1688 {
1689         return expand_downwards(vma, address);
1690 }
1691
1692 #ifdef CONFIG_STACK_GROWSUP
1693 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1694 {
1695         return expand_upwards(vma, address);
1696 }
1697
1698 struct vm_area_struct *
1699 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1700 {
1701         struct vm_area_struct *vma, *prev;
1702
1703         addr &= PAGE_MASK;
1704         vma = find_vma_prev(mm, addr, &prev);
1705         if (vma && (vma->vm_start <= addr))
1706                 return vma;
1707         if (!prev || expand_stack(prev, addr))
1708                 return NULL;
1709         if (prev->vm_flags & VM_LOCKED) {
1710                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1711                         return NULL;    /* vma gone! */
1712         }
1713         return prev;
1714 }
1715 #else
1716 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1717 {
1718         return expand_downwards(vma, address);
1719 }
1720
1721 struct vm_area_struct *
1722 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1723 {
1724         struct vm_area_struct * vma;
1725         unsigned long start;
1726
1727         addr &= PAGE_MASK;
1728         vma = find_vma(mm,addr);
1729         if (!vma)
1730                 return NULL;
1731         if (vma->vm_start <= addr)
1732                 return vma;
1733         if (!(vma->vm_flags & VM_GROWSDOWN))
1734                 return NULL;
1735         start = vma->vm_start;
1736         if (expand_stack(vma, addr))
1737                 return NULL;
1738         if (vma->vm_flags & VM_LOCKED) {
1739                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1740                         return NULL;    /* vma gone! */
1741         }
1742         return vma;
1743 }
1744 #endif
1745
1746 /*
1747  * Ok - we have the memory areas we should free on the vma list,
1748  * so release them, and do the vma updates.
1749  *
1750  * Called with the mm semaphore held.
1751  */
1752 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1753 {
1754         /* Update high watermark before we lower total_vm */
1755         update_hiwater_vm(mm);
1756         do {
1757                 long nrpages = vma_pages(vma);
1758
1759                 if (vma->vm_flags & VM_EXEC) {
1760                         perf_counter_munmap(vma->vm_start,
1761                                         nrpages << PAGE_SHIFT,
1762                                         vma->vm_pgoff, vma->vm_file);
1763                 }
1764
1765                 mm->total_vm -= nrpages;
1766                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1767                 vma = remove_vma(vma);
1768         } while (vma);
1769         validate_mm(mm);
1770 }
1771
1772 /*
1773  * Get rid of page table information in the indicated region.
1774  *
1775  * Called with the mm semaphore held.
1776  */
1777 static void unmap_region(struct mm_struct *mm,
1778                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1779                 unsigned long start, unsigned long end)
1780 {
1781         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1782         struct mmu_gather *tlb;
1783         unsigned long nr_accounted = 0;
1784
1785         lru_add_drain();
1786         tlb = tlb_gather_mmu(mm, 0);
1787         update_hiwater_rss(mm);
1788         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1789         vm_unacct_memory(nr_accounted);
1790         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1791                                  next? next->vm_start: 0);
1792         tlb_finish_mmu(tlb, start, end);
1793 }
1794
1795 /*
1796  * Create a list of vma's touched by the unmap, removing them from the mm's
1797  * vma list as we go..
1798  */
1799 static void
1800 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1801         struct vm_area_struct *prev, unsigned long end)
1802 {
1803         struct vm_area_struct **insertion_point;
1804         struct vm_area_struct *tail_vma = NULL;
1805         unsigned long addr;
1806
1807         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1808         do {
1809                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1810                 mm->map_count--;
1811                 tail_vma = vma;
1812                 vma = vma->vm_next;
1813         } while (vma && vma->vm_start < end);
1814         *insertion_point = vma;
1815         tail_vma->vm_next = NULL;
1816         if (mm->unmap_area == arch_unmap_area)
1817                 addr = prev ? prev->vm_end : mm->mmap_base;
1818         else
1819                 addr = vma ?  vma->vm_start : mm->mmap_base;
1820         mm->unmap_area(mm, addr);
1821         mm->mmap_cache = NULL;          /* Kill the cache. */
1822 }
1823
1824 /*
1825  * Split a vma into two pieces at address 'addr', a new vma is allocated
1826  * either for the first part or the tail.
1827  */
1828 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1829               unsigned long addr, int new_below)
1830 {
1831         struct mempolicy *pol;
1832         struct vm_area_struct *new;
1833
1834         if (is_vm_hugetlb_page(vma) && (addr &
1835                                         ~(huge_page_mask(hstate_vma(vma)))))
1836                 return -EINVAL;
1837
1838         if (mm->map_count >= sysctl_max_map_count)
1839                 return -ENOMEM;
1840
1841         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1842         if (!new)
1843                 return -ENOMEM;
1844
1845         /* most fields are the same, copy all, and then fixup */
1846         *new = *vma;
1847
1848         if (new_below)
1849                 new->vm_end = addr;
1850         else {
1851                 new->vm_start = addr;
1852                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1853         }
1854
1855         pol = mpol_dup(vma_policy(vma));
1856         if (IS_ERR(pol)) {
1857                 kmem_cache_free(vm_area_cachep, new);
1858                 return PTR_ERR(pol);
1859         }
1860         vma_set_policy(new, pol);
1861
1862         if (new->vm_file) {
1863                 get_file(new->vm_file);
1864                 if (vma->vm_flags & VM_EXECUTABLE)
1865                         added_exe_file_vma(mm);
1866         }
1867
1868         if (new->vm_ops && new->vm_ops->open)
1869                 new->vm_ops->open(new);
1870
1871         if (new_below)
1872                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1873                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1874         else
1875                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1876
1877         return 0;
1878 }
1879
1880 /* Munmap is split into 2 main parts -- this part which finds
1881  * what needs doing, and the areas themselves, which do the
1882  * work.  This now handles partial unmappings.
1883  * Jeremy Fitzhardinge <jeremy@goop.org>
1884  */
1885 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1886 {
1887         unsigned long end;
1888         struct vm_area_struct *vma, *prev, *last;
1889
1890         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1891                 return -EINVAL;
1892
1893         if ((len = PAGE_ALIGN(len)) == 0)
1894                 return -EINVAL;
1895
1896         /* Find the first overlapping VMA */
1897         vma = find_vma_prev(mm, start, &prev);
1898         if (!vma)
1899                 return 0;
1900         /* we have  start < vma->vm_end  */
1901
1902         /* if it doesn't overlap, we have nothing.. */
1903         end = start + len;
1904         if (vma->vm_start >= end)
1905                 return 0;
1906
1907         /*
1908          * If we need to split any vma, do it now to save pain later.
1909          *
1910          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1911          * unmapped vm_area_struct will remain in use: so lower split_vma
1912          * places tmp vma above, and higher split_vma places tmp vma below.
1913          */
1914         if (start > vma->vm_start) {
1915                 int error = split_vma(mm, vma, start, 0);
1916                 if (error)
1917                         return error;
1918                 prev = vma;
1919         }
1920
1921         /* Does it split the last one? */
1922         last = find_vma(mm, end);
1923         if (last && end > last->vm_start) {
1924                 int error = split_vma(mm, last, end, 1);
1925                 if (error)
1926                         return error;
1927         }
1928         vma = prev? prev->vm_next: mm->mmap;
1929
1930         /*
1931          * unlock any mlock()ed ranges before detaching vmas
1932          */
1933         if (mm->locked_vm) {
1934                 struct vm_area_struct *tmp = vma;
1935                 while (tmp && tmp->vm_start < end) {
1936                         if (tmp->vm_flags & VM_LOCKED) {
1937                                 mm->locked_vm -= vma_pages(tmp);
1938                                 munlock_vma_pages_all(tmp);
1939                         }
1940                         tmp = tmp->vm_next;
1941                 }
1942         }
1943
1944         /*
1945          * Remove the vma's, and unmap the actual pages
1946          */
1947         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1948         unmap_region(mm, vma, prev, start, end);
1949
1950         /* Fix up all other VM information */
1951         remove_vma_list(mm, vma);
1952
1953         return 0;
1954 }
1955
1956 EXPORT_SYMBOL(do_munmap);
1957
1958 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1959 {
1960         int ret;
1961         struct mm_struct *mm = current->mm;
1962
1963         profile_munmap(addr);
1964
1965         down_write(&mm->mmap_sem);
1966         ret = do_munmap(mm, addr, len);
1967         up_write(&mm->mmap_sem);
1968         return ret;
1969 }
1970
1971 static inline void verify_mm_writelocked(struct mm_struct *mm)
1972 {
1973 #ifdef CONFIG_DEBUG_VM
1974         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1975                 WARN_ON(1);
1976                 up_read(&mm->mmap_sem);
1977         }
1978 #endif
1979 }
1980
1981 /*
1982  *  this is really a simplified "do_mmap".  it only handles
1983  *  anonymous maps.  eventually we may be able to do some
1984  *  brk-specific accounting here.
1985  */
1986 unsigned long do_brk(unsigned long addr, unsigned long len)
1987 {
1988         struct mm_struct * mm = current->mm;
1989         struct vm_area_struct * vma, * prev;
1990         unsigned long flags;
1991         struct rb_node ** rb_link, * rb_parent;
1992         pgoff_t pgoff = addr >> PAGE_SHIFT;
1993         int error;
1994
1995         len = PAGE_ALIGN(len);
1996         if (!len)
1997                 return addr;
1998
1999         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
2000                 return -EINVAL;
2001
2002         if (is_hugepage_only_range(mm, addr, len))
2003                 return -EINVAL;
2004
2005         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2006         if (error)
2007                 return error;
2008
2009         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2010
2011         error = arch_mmap_check(addr, len, flags);
2012         if (error)
2013                 return error;
2014
2015         /*
2016          * mlock MCL_FUTURE?
2017          */
2018         if (mm->def_flags & VM_LOCKED) {
2019                 unsigned long locked, lock_limit;
2020                 locked = len >> PAGE_SHIFT;
2021                 locked += mm->locked_vm;
2022                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2023                 lock_limit >>= PAGE_SHIFT;
2024                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2025                         return -EAGAIN;
2026         }
2027
2028         /*
2029          * mm->mmap_sem is required to protect against another thread
2030          * changing the mappings in case we sleep.
2031          */
2032         verify_mm_writelocked(mm);
2033
2034         /*
2035          * Clear old maps.  this also does some error checking for us
2036          */
2037  munmap_back:
2038         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2039         if (vma && vma->vm_start < addr + len) {
2040                 if (do_munmap(mm, addr, len))
2041                         return -ENOMEM;
2042                 goto munmap_back;
2043         }
2044
2045         /* Check against address space limits *after* clearing old maps... */
2046         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2047                 return -ENOMEM;
2048
2049         if (mm->map_count > sysctl_max_map_count)
2050                 return -ENOMEM;
2051
2052         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2053                 return -ENOMEM;
2054
2055         /* Can we just expand an old private anonymous mapping? */
2056         vma = vma_merge(mm, prev, addr, addr + len, flags,
2057                                         NULL, NULL, pgoff, NULL);
2058         if (vma)
2059                 goto out;
2060
2061         /*
2062          * create a vma struct for an anonymous mapping
2063          */
2064         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2065         if (!vma) {
2066                 vm_unacct_memory(len >> PAGE_SHIFT);
2067                 return -ENOMEM;
2068         }
2069
2070         vma->vm_mm = mm;
2071         vma->vm_start = addr;
2072         vma->vm_end = addr + len;
2073         vma->vm_pgoff = pgoff;
2074         vma->vm_flags = flags;
2075         vma->vm_page_prot = vm_get_page_prot(flags);
2076         vma_link(mm, vma, prev, rb_link, rb_parent);
2077 out:
2078         mm->total_vm += len >> PAGE_SHIFT;
2079         if (flags & VM_LOCKED) {
2080                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2081                         mm->locked_vm += (len >> PAGE_SHIFT);
2082         }
2083         return addr;
2084 }
2085
2086 EXPORT_SYMBOL(do_brk);
2087
2088 /* Release all mmaps. */
2089 void exit_mmap(struct mm_struct *mm)
2090 {
2091         struct mmu_gather *tlb;
2092         struct vm_area_struct *vma;
2093         unsigned long nr_accounted = 0;
2094         unsigned long end;
2095
2096         /* mm's last user has gone, and its about to be pulled down */
2097         mmu_notifier_release(mm);
2098
2099         if (mm->locked_vm) {
2100                 vma = mm->mmap;
2101                 while (vma) {
2102                         if (vma->vm_flags & VM_LOCKED)
2103                                 munlock_vma_pages_all(vma);
2104                         vma = vma->vm_next;
2105                 }
2106         }
2107
2108         arch_exit_mmap(mm);
2109
2110         vma = mm->mmap;
2111         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2112                 return;
2113
2114         lru_add_drain();
2115         flush_cache_mm(mm);
2116         tlb = tlb_gather_mmu(mm, 1);
2117         /* update_hiwater_rss(mm) here? but nobody should be looking */
2118         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2119         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2120         vm_unacct_memory(nr_accounted);
2121         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2122         tlb_finish_mmu(tlb, 0, end);
2123
2124         /*
2125          * Walk the list again, actually closing and freeing it,
2126          * with preemption enabled, without holding any MM locks.
2127          */
2128         while (vma)
2129                 vma = remove_vma(vma);
2130
2131         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2132 }
2133
2134 /* Insert vm structure into process list sorted by address
2135  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2136  * then i_mmap_lock is taken here.
2137  */
2138 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2139 {
2140         struct vm_area_struct * __vma, * prev;
2141         struct rb_node ** rb_link, * rb_parent;
2142
2143         /*
2144          * The vm_pgoff of a purely anonymous vma should be irrelevant
2145          * until its first write fault, when page's anon_vma and index
2146          * are set.  But now set the vm_pgoff it will almost certainly
2147          * end up with (unless mremap moves it elsewhere before that
2148          * first wfault), so /proc/pid/maps tells a consistent story.
2149          *
2150          * By setting it to reflect the virtual start address of the
2151          * vma, merges and splits can happen in a seamless way, just
2152          * using the existing file pgoff checks and manipulations.
2153          * Similarly in do_mmap_pgoff and in do_brk.
2154          */
2155         if (!vma->vm_file) {
2156                 BUG_ON(vma->anon_vma);
2157                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2158         }
2159         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2160         if (__vma && __vma->vm_start < vma->vm_end)
2161                 return -ENOMEM;
2162         if ((vma->vm_flags & VM_ACCOUNT) &&
2163              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2164                 return -ENOMEM;
2165         vma_link(mm, vma, prev, rb_link, rb_parent);
2166         return 0;
2167 }
2168
2169 /*
2170  * Copy the vma structure to a new location in the same mm,
2171  * prior to moving page table entries, to effect an mremap move.
2172  */
2173 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2174         unsigned long addr, unsigned long len, pgoff_t pgoff)
2175 {
2176         struct vm_area_struct *vma = *vmap;
2177         unsigned long vma_start = vma->vm_start;
2178         struct mm_struct *mm = vma->vm_mm;
2179         struct vm_area_struct *new_vma, *prev;
2180         struct rb_node **rb_link, *rb_parent;
2181         struct mempolicy *pol;
2182
2183         /*
2184          * If anonymous vma has not yet been faulted, update new pgoff
2185          * to match new location, to increase its chance of merging.
2186          */
2187         if (!vma->vm_file && !vma->anon_vma)
2188                 pgoff = addr >> PAGE_SHIFT;
2189
2190         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2191         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2192                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2193         if (new_vma) {
2194                 /*
2195                  * Source vma may have been merged into new_vma
2196                  */
2197                 if (vma_start >= new_vma->vm_start &&
2198                     vma_start < new_vma->vm_end)
2199                         *vmap = new_vma;
2200         } else {
2201                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2202                 if (new_vma) {
2203                         *new_vma = *vma;
2204                         pol = mpol_dup(vma_policy(vma));
2205                         if (IS_ERR(pol)) {
2206                                 kmem_cache_free(vm_area_cachep, new_vma);
2207                                 return NULL;
2208                         }
2209                         vma_set_policy(new_vma, pol);
2210                         new_vma->vm_start = addr;
2211                         new_vma->vm_end = addr + len;
2212                         new_vma->vm_pgoff = pgoff;
2213                         if (new_vma->vm_file) {
2214                                 get_file(new_vma->vm_file);
2215                                 if (vma->vm_flags & VM_EXECUTABLE)
2216                                         added_exe_file_vma(mm);
2217                         }
2218                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2219                                 new_vma->vm_ops->open(new_vma);
2220                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2221                 }
2222         }
2223         return new_vma;
2224 }
2225
2226 /*
2227  * Return true if the calling process may expand its vm space by the passed
2228  * number of pages
2229  */
2230 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2231 {
2232         unsigned long cur = mm->total_vm;       /* pages */
2233         unsigned long lim;
2234
2235         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2236
2237         if (cur + npages > lim)
2238                 return 0;
2239         return 1;
2240 }
2241
2242
2243 static int special_mapping_fault(struct vm_area_struct *vma,
2244                                 struct vm_fault *vmf)
2245 {
2246         pgoff_t pgoff;
2247         struct page **pages;
2248
2249         /*
2250          * special mappings have no vm_file, and in that case, the mm
2251          * uses vm_pgoff internally. So we have to subtract it from here.
2252          * We are allowed to do this because we are the mm; do not copy
2253          * this code into drivers!
2254          */
2255         pgoff = vmf->pgoff - vma->vm_pgoff;
2256
2257         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2258                 pgoff--;
2259
2260         if (*pages) {
2261                 struct page *page = *pages;
2262                 get_page(page);
2263                 vmf->page = page;
2264                 return 0;
2265         }
2266
2267         return VM_FAULT_SIGBUS;
2268 }
2269
2270 /*
2271  * Having a close hook prevents vma merging regardless of flags.
2272  */
2273 static void special_mapping_close(struct vm_area_struct *vma)
2274 {
2275 }
2276
2277 static struct vm_operations_struct special_mapping_vmops = {
2278         .close = special_mapping_close,
2279         .fault = special_mapping_fault,
2280 };
2281
2282 /*
2283  * Called with mm->mmap_sem held for writing.
2284  * Insert a new vma covering the given region, with the given flags.
2285  * Its pages are supplied by the given array of struct page *.
2286  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2287  * The region past the last page supplied will always produce SIGBUS.
2288  * The array pointer and the pages it points to are assumed to stay alive
2289  * for as long as this mapping might exist.
2290  */
2291 int install_special_mapping(struct mm_struct *mm,
2292                             unsigned long addr, unsigned long len,
2293                             unsigned long vm_flags, struct page **pages)
2294 {
2295         struct vm_area_struct *vma;
2296
2297         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2298         if (unlikely(vma == NULL))
2299                 return -ENOMEM;
2300
2301         vma->vm_mm = mm;
2302         vma->vm_start = addr;
2303         vma->vm_end = addr + len;
2304
2305         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2306         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2307
2308         vma->vm_ops = &special_mapping_vmops;
2309         vma->vm_private_data = pages;
2310
2311         if (unlikely(insert_vm_struct(mm, vma))) {
2312                 kmem_cache_free(vm_area_cachep, vma);
2313                 return -ENOMEM;
2314         }
2315
2316         mm->total_vm += len >> PAGE_SHIFT;
2317
2318         return 0;
2319 }
2320
2321 static DEFINE_MUTEX(mm_all_locks_mutex);
2322
2323 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2324 {
2325         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2326                 /*
2327                  * The LSB of head.next can't change from under us
2328                  * because we hold the mm_all_locks_mutex.
2329                  */
2330                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2331                 /*
2332                  * We can safely modify head.next after taking the
2333                  * anon_vma->lock. If some other vma in this mm shares
2334                  * the same anon_vma we won't take it again.
2335                  *
2336                  * No need of atomic instructions here, head.next
2337                  * can't change from under us thanks to the
2338                  * anon_vma->lock.
2339                  */
2340                 if (__test_and_set_bit(0, (unsigned long *)
2341                                        &anon_vma->head.next))
2342                         BUG();
2343         }
2344 }
2345
2346 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2347 {
2348         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2349                 /*
2350                  * AS_MM_ALL_LOCKS can't change from under us because
2351                  * we hold the mm_all_locks_mutex.
2352                  *
2353                  * Operations on ->flags have to be atomic because
2354                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2355                  * mm_all_locks_mutex, there may be other cpus
2356                  * changing other bitflags in parallel to us.
2357                  */
2358                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2359                         BUG();
2360                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2361         }
2362 }
2363
2364 /*
2365  * This operation locks against the VM for all pte/vma/mm related
2366  * operations that could ever happen on a certain mm. This includes
2367  * vmtruncate, try_to_unmap, and all page faults.
2368  *
2369  * The caller must take the mmap_sem in write mode before calling
2370  * mm_take_all_locks(). The caller isn't allowed to release the
2371  * mmap_sem until mm_drop_all_locks() returns.
2372  *
2373  * mmap_sem in write mode is required in order to block all operations
2374  * that could modify pagetables and free pages without need of
2375  * altering the vma layout (for example populate_range() with
2376  * nonlinear vmas). It's also needed in write mode to avoid new
2377  * anon_vmas to be associated with existing vmas.
2378  *
2379  * A single task can't take more than one mm_take_all_locks() in a row
2380  * or it would deadlock.
2381  *
2382  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2383  * mapping->flags avoid to take the same lock twice, if more than one
2384  * vma in this mm is backed by the same anon_vma or address_space.
2385  *
2386  * We can take all the locks in random order because the VM code
2387  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2388  * takes more than one of them in a row. Secondly we're protected
2389  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2390  *
2391  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2392  * that may have to take thousand of locks.
2393  *
2394  * mm_take_all_locks() can fail if it's interrupted by signals.
2395  */
2396 int mm_take_all_locks(struct mm_struct *mm)
2397 {
2398         struct vm_area_struct *vma;
2399         int ret = -EINTR;
2400
2401         BUG_ON(down_read_trylock(&mm->mmap_sem));
2402
2403         mutex_lock(&mm_all_locks_mutex);
2404
2405         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2406                 if (signal_pending(current))
2407                         goto out_unlock;
2408                 if (vma->vm_file && vma->vm_file->f_mapping)
2409                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2410         }
2411
2412         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2413                 if (signal_pending(current))
2414                         goto out_unlock;
2415                 if (vma->anon_vma)
2416                         vm_lock_anon_vma(mm, vma->anon_vma);
2417         }
2418
2419         ret = 0;
2420
2421 out_unlock:
2422         if (ret)
2423                 mm_drop_all_locks(mm);
2424
2425         return ret;
2426 }
2427
2428 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2429 {
2430         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2431                 /*
2432                  * The LSB of head.next can't change to 0 from under
2433                  * us because we hold the mm_all_locks_mutex.
2434                  *
2435                  * We must however clear the bitflag before unlocking
2436                  * the vma so the users using the anon_vma->head will
2437                  * never see our bitflag.
2438                  *
2439                  * No need of atomic instructions here, head.next
2440                  * can't change from under us until we release the
2441                  * anon_vma->lock.
2442                  */
2443                 if (!__test_and_clear_bit(0, (unsigned long *)
2444                                           &anon_vma->head.next))
2445                         BUG();
2446                 spin_unlock(&anon_vma->lock);
2447         }
2448 }
2449
2450 static void vm_unlock_mapping(struct address_space *mapping)
2451 {
2452         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2453                 /*
2454                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2455                  * because we hold the mm_all_locks_mutex.
2456                  */
2457                 spin_unlock(&mapping->i_mmap_lock);
2458                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2459                                         &mapping->flags))
2460                         BUG();
2461         }
2462 }
2463
2464 /*
2465  * The mmap_sem cannot be released by the caller until
2466  * mm_drop_all_locks() returns.
2467  */
2468 void mm_drop_all_locks(struct mm_struct *mm)
2469 {
2470         struct vm_area_struct *vma;
2471
2472         BUG_ON(down_read_trylock(&mm->mmap_sem));
2473         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2474
2475         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2476                 if (vma->anon_vma)
2477                         vm_unlock_anon_vma(vma->anon_vma);
2478                 if (vma->vm_file && vma->vm_file->f_mapping)
2479                         vm_unlock_mapping(vma->vm_file->f_mapping);
2480         }
2481
2482         mutex_unlock(&mm_all_locks_mutex);
2483 }
2484
2485 /*
2486  * initialise the VMA slab
2487  */
2488 void __init mmap_init(void)
2489 {
2490         int ret;
2491
2492         ret = percpu_counter_init(&vm_committed_as, 0);
2493         VM_BUG_ON(ret);
2494 }