mmap: remove unnecessary code
[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_event.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 /*
660  * If the vma has a ->close operation then the driver probably needs to release
661  * per-vma resources, so we don't attempt to merge those.
662  */
663 static inline int is_mergeable_vma(struct vm_area_struct *vma,
664                         struct file *file, unsigned long vm_flags)
665 {
666         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
667         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
668                 return 0;
669         if (vma->vm_file != file)
670                 return 0;
671         if (vma->vm_ops && vma->vm_ops->close)
672                 return 0;
673         return 1;
674 }
675
676 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
677                                         struct anon_vma *anon_vma2)
678 {
679         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
680 }
681
682 /*
683  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
684  * in front of (at a lower virtual address and file offset than) the vma.
685  *
686  * We cannot merge two vmas if they have differently assigned (non-NULL)
687  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
688  *
689  * We don't check here for the merged mmap wrapping around the end of pagecache
690  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
691  * wrap, nor mmaps which cover the final page at index -1UL.
692  */
693 static int
694 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
695         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
696 {
697         if (is_mergeable_vma(vma, file, vm_flags) &&
698             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
699                 if (vma->vm_pgoff == vm_pgoff)
700                         return 1;
701         }
702         return 0;
703 }
704
705 /*
706  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
707  * beyond (at a higher virtual address and file offset than) the vma.
708  *
709  * We cannot merge two vmas if they have differently assigned (non-NULL)
710  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
711  */
712 static int
713 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
714         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
715 {
716         if (is_mergeable_vma(vma, file, vm_flags) &&
717             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
718                 pgoff_t vm_pglen;
719                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
720                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
721                         return 1;
722         }
723         return 0;
724 }
725
726 /*
727  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
728  * whether that can be merged with its predecessor or its successor.
729  * Or both (it neatly fills a hole).
730  *
731  * In most cases - when called for mmap, brk or mremap - [addr,end) is
732  * certain not to be mapped by the time vma_merge is called; but when
733  * called for mprotect, it is certain to be already mapped (either at
734  * an offset within prev, or at the start of next), and the flags of
735  * this area are about to be changed to vm_flags - and the no-change
736  * case has already been eliminated.
737  *
738  * The following mprotect cases have to be considered, where AAAA is
739  * the area passed down from mprotect_fixup, never extending beyond one
740  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
741  *
742  *     AAAA             AAAA                AAAA          AAAA
743  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
744  *    cannot merge    might become    might become    might become
745  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
746  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
747  *    mremap move:                                    PPPPNNNNNNNN 8
748  *        AAAA
749  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
750  *    might become    case 1 below    case 2 below    case 3 below
751  *
752  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
753  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
754  */
755 struct vm_area_struct *vma_merge(struct mm_struct *mm,
756                         struct vm_area_struct *prev, unsigned long addr,
757                         unsigned long end, unsigned long vm_flags,
758                         struct anon_vma *anon_vma, struct file *file,
759                         pgoff_t pgoff, struct mempolicy *policy)
760 {
761         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
762         struct vm_area_struct *area, *next;
763
764         /*
765          * We later require that vma->vm_flags == vm_flags,
766          * so this tests vma->vm_flags & VM_SPECIAL, too.
767          */
768         if (vm_flags & VM_SPECIAL)
769                 return NULL;
770
771         if (prev)
772                 next = prev->vm_next;
773         else
774                 next = mm->mmap;
775         area = next;
776         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
777                 next = next->vm_next;
778
779         /*
780          * Can it merge with the predecessor?
781          */
782         if (prev && prev->vm_end == addr &&
783                         mpol_equal(vma_policy(prev), policy) &&
784                         can_vma_merge_after(prev, vm_flags,
785                                                 anon_vma, file, pgoff)) {
786                 /*
787                  * OK, it can.  Can we now merge in the successor as well?
788                  */
789                 if (next && end == next->vm_start &&
790                                 mpol_equal(policy, vma_policy(next)) &&
791                                 can_vma_merge_before(next, vm_flags,
792                                         anon_vma, file, pgoff+pglen) &&
793                                 is_mergeable_anon_vma(prev->anon_vma,
794                                                       next->anon_vma)) {
795                                                         /* cases 1, 6 */
796                         vma_adjust(prev, prev->vm_start,
797                                 next->vm_end, prev->vm_pgoff, NULL);
798                 } else                                  /* cases 2, 5, 7 */
799                         vma_adjust(prev, prev->vm_start,
800                                 end, prev->vm_pgoff, NULL);
801                 return prev;
802         }
803
804         /*
805          * Can this new request be merged in front of next?
806          */
807         if (next && end == next->vm_start &&
808                         mpol_equal(policy, vma_policy(next)) &&
809                         can_vma_merge_before(next, vm_flags,
810                                         anon_vma, file, pgoff+pglen)) {
811                 if (prev && addr < prev->vm_end)        /* case 4 */
812                         vma_adjust(prev, prev->vm_start,
813                                 addr, prev->vm_pgoff, NULL);
814                 else                                    /* cases 3, 8 */
815                         vma_adjust(area, addr, next->vm_end,
816                                 next->vm_pgoff - pglen, NULL);
817                 return area;
818         }
819
820         return NULL;
821 }
822
823 /*
824  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
825  * neighbouring vmas for a suitable anon_vma, before it goes off
826  * to allocate a new anon_vma.  It checks because a repetitive
827  * sequence of mprotects and faults may otherwise lead to distinct
828  * anon_vmas being allocated, preventing vma merge in subsequent
829  * mprotect.
830  */
831 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
832 {
833         struct vm_area_struct *near;
834         unsigned long vm_flags;
835
836         near = vma->vm_next;
837         if (!near)
838                 goto try_prev;
839
840         /*
841          * Since only mprotect tries to remerge vmas, match flags
842          * which might be mprotected into each other later on.
843          * Neither mlock nor madvise tries to remerge at present,
844          * so leave their flags as obstructing a merge.
845          */
846         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
847         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
848
849         if (near->anon_vma && vma->vm_end == near->vm_start &&
850                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
851                         can_vma_merge_before(near, vm_flags,
852                                 NULL, vma->vm_file, vma->vm_pgoff +
853                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
854                 return near->anon_vma;
855 try_prev:
856         /*
857          * It is potentially slow to have to call find_vma_prev here.
858          * But it's only on the first write fault on the vma, not
859          * every time, and we could devise a way to avoid it later
860          * (e.g. stash info in next's anon_vma_node when assigning
861          * an anon_vma, or when trying vma_merge).  Another time.
862          */
863         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
864         if (!near)
865                 goto none;
866
867         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
868         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
869
870         if (near->anon_vma && near->vm_end == vma->vm_start &&
871                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
872                         can_vma_merge_after(near, vm_flags,
873                                 NULL, vma->vm_file, vma->vm_pgoff))
874                 return near->anon_vma;
875 none:
876         /*
877          * There's no absolute need to look only at touching neighbours:
878          * we could search further afield for "compatible" anon_vmas.
879          * But it would probably just be a waste of time searching,
880          * or lead to too many vmas hanging off the same anon_vma.
881          * We're trying to allow mprotect remerging later on,
882          * not trying to minimize memory used for anon_vmas.
883          */
884         return NULL;
885 }
886
887 #ifdef CONFIG_PROC_FS
888 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
889                                                 struct file *file, long pages)
890 {
891         const unsigned long stack_flags
892                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
893
894         if (file) {
895                 mm->shared_vm += pages;
896                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
897                         mm->exec_vm += pages;
898         } else if (flags & stack_flags)
899                 mm->stack_vm += pages;
900         if (flags & (VM_RESERVED|VM_IO))
901                 mm->reserved_vm += pages;
902 }
903 #endif /* CONFIG_PROC_FS */
904
905 /*
906  * The caller must hold down_write(&current->mm->mmap_sem).
907  */
908
909 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
910                         unsigned long len, unsigned long prot,
911                         unsigned long flags, unsigned long pgoff)
912 {
913         struct mm_struct * mm = current->mm;
914         struct inode *inode;
915         unsigned int vm_flags;
916         int error;
917         unsigned long reqprot = prot;
918
919         /*
920          * Does the application expect PROT_READ to imply PROT_EXEC?
921          *
922          * (the exception is when the underlying filesystem is noexec
923          *  mounted, in which case we dont add PROT_EXEC.)
924          */
925         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
926                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
927                         prot |= PROT_EXEC;
928
929         if (!len)
930                 return -EINVAL;
931
932         if (!(flags & MAP_FIXED))
933                 addr = round_hint_to_min(addr);
934
935         error = arch_mmap_check(addr, len, flags);
936         if (error)
937                 return error;
938
939         /* Careful about overflows.. */
940         len = PAGE_ALIGN(len);
941         if (!len || len > TASK_SIZE)
942                 return -ENOMEM;
943
944         /* offset overflow? */
945         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
946                return -EOVERFLOW;
947
948         /* Too many mappings? */
949         if (mm->map_count > sysctl_max_map_count)
950                 return -ENOMEM;
951
952         /* Obtain the address to map to. we verify (or select) it and ensure
953          * that it represents a valid section of the address space.
954          */
955         addr = get_unmapped_area(file, addr, len, pgoff, flags);
956         if (addr & ~PAGE_MASK)
957                 return addr;
958
959         /* Do simple checking here so the lower-level routines won't have
960          * to. we assume access permissions have been handled by the open
961          * of the memory object, so we don't do any here.
962          */
963         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
964                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
965
966         if (flags & MAP_LOCKED)
967                 if (!can_do_mlock())
968                         return -EPERM;
969
970         /* mlock MCL_FUTURE? */
971         if (vm_flags & VM_LOCKED) {
972                 unsigned long locked, lock_limit;
973                 locked = len >> PAGE_SHIFT;
974                 locked += mm->locked_vm;
975                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
976                 lock_limit >>= PAGE_SHIFT;
977                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
978                         return -EAGAIN;
979         }
980
981         inode = file ? file->f_path.dentry->d_inode : NULL;
982
983         if (file) {
984                 switch (flags & MAP_TYPE) {
985                 case MAP_SHARED:
986                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
987                                 return -EACCES;
988
989                         /*
990                          * Make sure we don't allow writing to an append-only
991                          * file..
992                          */
993                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
994                                 return -EACCES;
995
996                         /*
997                          * Make sure there are no mandatory locks on the file.
998                          */
999                         if (locks_verify_locked(inode))
1000                                 return -EAGAIN;
1001
1002                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1003                         if (!(file->f_mode & FMODE_WRITE))
1004                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1005
1006                         /* fall through */
1007                 case MAP_PRIVATE:
1008                         if (!(file->f_mode & FMODE_READ))
1009                                 return -EACCES;
1010                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1011                                 if (vm_flags & VM_EXEC)
1012                                         return -EPERM;
1013                                 vm_flags &= ~VM_MAYEXEC;
1014                         }
1015
1016                         if (!file->f_op || !file->f_op->mmap)
1017                                 return -ENODEV;
1018                         break;
1019
1020                 default:
1021                         return -EINVAL;
1022                 }
1023         } else {
1024                 switch (flags & MAP_TYPE) {
1025                 case MAP_SHARED:
1026                         /*
1027                          * Ignore pgoff.
1028                          */
1029                         pgoff = 0;
1030                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1031                         break;
1032                 case MAP_PRIVATE:
1033                         /*
1034                          * Set pgoff according to addr for anon_vma.
1035                          */
1036                         pgoff = addr >> PAGE_SHIFT;
1037                         break;
1038                 default:
1039                         return -EINVAL;
1040                 }
1041         }
1042
1043         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1044         if (error)
1045                 return error;
1046         error = ima_file_mmap(file, prot);
1047         if (error)
1048                 return error;
1049
1050         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1051 }
1052 EXPORT_SYMBOL(do_mmap_pgoff);
1053
1054 /*
1055  * Some shared mappigns will want the pages marked read-only
1056  * to track write events. If so, we'll downgrade vm_page_prot
1057  * to the private version (using protection_map[] without the
1058  * VM_SHARED bit).
1059  */
1060 int vma_wants_writenotify(struct vm_area_struct *vma)
1061 {
1062         unsigned int vm_flags = vma->vm_flags;
1063
1064         /* If it was private or non-writable, the write bit is already clear */
1065         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1066                 return 0;
1067
1068         /* The backer wishes to know when pages are first written to? */
1069         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1070                 return 1;
1071
1072         /* The open routine did something to the protections already? */
1073         if (pgprot_val(vma->vm_page_prot) !=
1074             pgprot_val(vm_get_page_prot(vm_flags)))
1075                 return 0;
1076
1077         /* Specialty mapping? */
1078         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1079                 return 0;
1080
1081         /* Can the mapping track the dirty pages? */
1082         return vma->vm_file && vma->vm_file->f_mapping &&
1083                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1084 }
1085
1086 /*
1087  * We account for memory if it's a private writeable mapping,
1088  * not hugepages and VM_NORESERVE wasn't set.
1089  */
1090 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1091 {
1092         /*
1093          * hugetlb has its own accounting separate from the core VM
1094          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1095          */
1096         if (file && is_file_hugepages(file))
1097                 return 0;
1098
1099         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1100 }
1101
1102 unsigned long mmap_region(struct file *file, unsigned long addr,
1103                           unsigned long len, unsigned long flags,
1104                           unsigned int vm_flags, unsigned long pgoff)
1105 {
1106         struct mm_struct *mm = current->mm;
1107         struct vm_area_struct *vma, *prev;
1108         int correct_wcount = 0;
1109         int error;
1110         struct rb_node **rb_link, *rb_parent;
1111         unsigned long charged = 0;
1112         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1113
1114         /* Clear old maps */
1115         error = -ENOMEM;
1116 munmap_back:
1117         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1118         if (vma && vma->vm_start < addr + len) {
1119                 if (do_munmap(mm, addr, len))
1120                         return -ENOMEM;
1121                 goto munmap_back;
1122         }
1123
1124         /* Check against address space limit. */
1125         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1126                 return -ENOMEM;
1127
1128         /*
1129          * Set 'VM_NORESERVE' if we should not account for the
1130          * memory use of this mapping.
1131          */
1132         if ((flags & MAP_NORESERVE)) {
1133                 /* We honor MAP_NORESERVE if allowed to overcommit */
1134                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1135                         vm_flags |= VM_NORESERVE;
1136
1137                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1138                 if (file && is_file_hugepages(file))
1139                         vm_flags |= VM_NORESERVE;
1140         }
1141
1142         /*
1143          * Private writable mapping: check memory availability
1144          */
1145         if (accountable_mapping(file, vm_flags)) {
1146                 charged = len >> PAGE_SHIFT;
1147                 if (security_vm_enough_memory(charged))
1148                         return -ENOMEM;
1149                 vm_flags |= VM_ACCOUNT;
1150         }
1151
1152         /*
1153          * Can we just expand an old mapping?
1154          */
1155         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1156         if (vma)
1157                 goto out;
1158
1159         /*
1160          * Determine the object being mapped and call the appropriate
1161          * specific mapper. the address has already been validated, but
1162          * not unmapped, but the maps are removed from the list.
1163          */
1164         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1165         if (!vma) {
1166                 error = -ENOMEM;
1167                 goto unacct_error;
1168         }
1169
1170         vma->vm_mm = mm;
1171         vma->vm_start = addr;
1172         vma->vm_end = addr + len;
1173         vma->vm_flags = vm_flags;
1174         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1175         vma->vm_pgoff = pgoff;
1176
1177         if (file) {
1178                 error = -EINVAL;
1179                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1180                         goto free_vma;
1181                 if (vm_flags & VM_DENYWRITE) {
1182                         error = deny_write_access(file);
1183                         if (error)
1184                                 goto free_vma;
1185                         correct_wcount = 1;
1186                 }
1187                 vma->vm_file = file;
1188                 get_file(file);
1189                 error = file->f_op->mmap(file, vma);
1190                 if (error)
1191                         goto unmap_and_free_vma;
1192                 if (vm_flags & VM_EXECUTABLE)
1193                         added_exe_file_vma(mm);
1194         } else if (vm_flags & VM_SHARED) {
1195                 error = shmem_zero_setup(vma);
1196                 if (error)
1197                         goto free_vma;
1198         }
1199
1200         /* Can addr have changed??
1201          *
1202          * Answer: Yes, several device drivers can do it in their
1203          *         f_op->mmap method. -DaveM
1204          */
1205         addr = vma->vm_start;
1206         pgoff = vma->vm_pgoff;
1207         vm_flags = vma->vm_flags;
1208
1209         if (vma_wants_writenotify(vma))
1210                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1211
1212         vma_link(mm, vma, prev, rb_link, rb_parent);
1213         file = vma->vm_file;
1214
1215         /* Once vma denies write, undo our temporary denial count */
1216         if (correct_wcount)
1217                 atomic_inc(&inode->i_writecount);
1218 out:
1219         perf_event_mmap(vma);
1220
1221         mm->total_vm += len >> PAGE_SHIFT;
1222         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1223         if (vm_flags & VM_LOCKED) {
1224                 /*
1225                  * makes pages present; downgrades, drops, reacquires mmap_sem
1226                  */
1227                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1228                 if (nr_pages < 0)
1229                         return nr_pages;        /* vma gone! */
1230                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1231         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1232                 make_pages_present(addr, addr + len);
1233         return addr;
1234
1235 unmap_and_free_vma:
1236         if (correct_wcount)
1237                 atomic_inc(&inode->i_writecount);
1238         vma->vm_file = NULL;
1239         fput(file);
1240
1241         /* Undo any partial mapping done by a device driver. */
1242         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1243         charged = 0;
1244 free_vma:
1245         kmem_cache_free(vm_area_cachep, vma);
1246 unacct_error:
1247         if (charged)
1248                 vm_unacct_memory(charged);
1249         return error;
1250 }
1251
1252 /* Get an address range which is currently unmapped.
1253  * For shmat() with addr=0.
1254  *
1255  * Ugly calling convention alert:
1256  * Return value with the low bits set means error value,
1257  * ie
1258  *      if (ret & ~PAGE_MASK)
1259  *              error = ret;
1260  *
1261  * This function "knows" that -ENOMEM has the bits set.
1262  */
1263 #ifndef HAVE_ARCH_UNMAPPED_AREA
1264 unsigned long
1265 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1266                 unsigned long len, unsigned long pgoff, unsigned long flags)
1267 {
1268         struct mm_struct *mm = current->mm;
1269         struct vm_area_struct *vma;
1270         unsigned long start_addr;
1271
1272         if (len > TASK_SIZE)
1273                 return -ENOMEM;
1274
1275         if (flags & MAP_FIXED)
1276                 return addr;
1277
1278         if (addr) {
1279                 addr = PAGE_ALIGN(addr);
1280                 vma = find_vma(mm, addr);
1281                 if (TASK_SIZE - len >= addr &&
1282                     (!vma || addr + len <= vma->vm_start))
1283                         return addr;
1284         }
1285         if (len > mm->cached_hole_size) {
1286                 start_addr = addr = mm->free_area_cache;
1287         } else {
1288                 start_addr = addr = TASK_UNMAPPED_BASE;
1289                 mm->cached_hole_size = 0;
1290         }
1291
1292 full_search:
1293         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1294                 /* At this point:  (!vma || addr < vma->vm_end). */
1295                 if (TASK_SIZE - len < addr) {
1296                         /*
1297                          * Start a new search - just in case we missed
1298                          * some holes.
1299                          */
1300                         if (start_addr != TASK_UNMAPPED_BASE) {
1301                                 addr = TASK_UNMAPPED_BASE;
1302                                 start_addr = addr;
1303                                 mm->cached_hole_size = 0;
1304                                 goto full_search;
1305                         }
1306                         return -ENOMEM;
1307                 }
1308                 if (!vma || addr + len <= vma->vm_start) {
1309                         /*
1310                          * Remember the place where we stopped the search:
1311                          */
1312                         mm->free_area_cache = addr + len;
1313                         return addr;
1314                 }
1315                 if (addr + mm->cached_hole_size < vma->vm_start)
1316                         mm->cached_hole_size = vma->vm_start - addr;
1317                 addr = vma->vm_end;
1318         }
1319 }
1320 #endif  
1321
1322 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1323 {
1324         /*
1325          * Is this a new hole at the lowest possible address?
1326          */
1327         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1328                 mm->free_area_cache = addr;
1329                 mm->cached_hole_size = ~0UL;
1330         }
1331 }
1332
1333 /*
1334  * This mmap-allocator allocates new areas top-down from below the
1335  * stack's low limit (the base):
1336  */
1337 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1338 unsigned long
1339 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1340                           const unsigned long len, const unsigned long pgoff,
1341                           const unsigned long flags)
1342 {
1343         struct vm_area_struct *vma;
1344         struct mm_struct *mm = current->mm;
1345         unsigned long addr = addr0;
1346
1347         /* requested length too big for entire address space */
1348         if (len > TASK_SIZE)
1349                 return -ENOMEM;
1350
1351         if (flags & MAP_FIXED)
1352                 return addr;
1353
1354         /* requesting a specific address */
1355         if (addr) {
1356                 addr = PAGE_ALIGN(addr);
1357                 vma = find_vma(mm, addr);
1358                 if (TASK_SIZE - len >= addr &&
1359                                 (!vma || addr + len <= vma->vm_start))
1360                         return addr;
1361         }
1362
1363         /* check if free_area_cache is useful for us */
1364         if (len <= mm->cached_hole_size) {
1365                 mm->cached_hole_size = 0;
1366                 mm->free_area_cache = mm->mmap_base;
1367         }
1368
1369         /* either no address requested or can't fit in requested address hole */
1370         addr = mm->free_area_cache;
1371
1372         /* make sure it can fit in the remaining address space */
1373         if (addr > len) {
1374                 vma = find_vma(mm, addr-len);
1375                 if (!vma || addr <= vma->vm_start)
1376                         /* remember the address as a hint for next time */
1377                         return (mm->free_area_cache = addr-len);
1378         }
1379
1380         if (mm->mmap_base < len)
1381                 goto bottomup;
1382
1383         addr = mm->mmap_base-len;
1384
1385         do {
1386                 /*
1387                  * Lookup failure means no vma is above this address,
1388                  * else if new region fits below vma->vm_start,
1389                  * return with success:
1390                  */
1391                 vma = find_vma(mm, addr);
1392                 if (!vma || addr+len <= vma->vm_start)
1393                         /* remember the address as a hint for next time */
1394                         return (mm->free_area_cache = addr);
1395
1396                 /* remember the largest hole we saw so far */
1397                 if (addr + mm->cached_hole_size < vma->vm_start)
1398                         mm->cached_hole_size = vma->vm_start - addr;
1399
1400                 /* try just below the current vma->vm_start */
1401                 addr = vma->vm_start-len;
1402         } while (len < vma->vm_start);
1403
1404 bottomup:
1405         /*
1406          * A failed mmap() very likely causes application failure,
1407          * so fall back to the bottom-up function here. This scenario
1408          * can happen with large stack limits and large mmap()
1409          * allocations.
1410          */
1411         mm->cached_hole_size = ~0UL;
1412         mm->free_area_cache = TASK_UNMAPPED_BASE;
1413         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1414         /*
1415          * Restore the topdown base:
1416          */
1417         mm->free_area_cache = mm->mmap_base;
1418         mm->cached_hole_size = ~0UL;
1419
1420         return addr;
1421 }
1422 #endif
1423
1424 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1425 {
1426         /*
1427          * Is this a new hole at the highest possible address?
1428          */
1429         if (addr > mm->free_area_cache)
1430                 mm->free_area_cache = addr;
1431
1432         /* dont allow allocations above current base */
1433         if (mm->free_area_cache > mm->mmap_base)
1434                 mm->free_area_cache = mm->mmap_base;
1435 }
1436
1437 unsigned long
1438 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1439                 unsigned long pgoff, unsigned long flags)
1440 {
1441         unsigned long (*get_area)(struct file *, unsigned long,
1442                                   unsigned long, unsigned long, unsigned long);
1443
1444         get_area = current->mm->get_unmapped_area;
1445         if (file && file->f_op && file->f_op->get_unmapped_area)
1446                 get_area = file->f_op->get_unmapped_area;
1447         addr = get_area(file, addr, len, pgoff, flags);
1448         if (IS_ERR_VALUE(addr))
1449                 return addr;
1450
1451         if (addr > TASK_SIZE - len)
1452                 return -ENOMEM;
1453         if (addr & ~PAGE_MASK)
1454                 return -EINVAL;
1455
1456         return arch_rebalance_pgtables(addr, len);
1457 }
1458
1459 EXPORT_SYMBOL(get_unmapped_area);
1460
1461 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1462 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1463 {
1464         struct vm_area_struct *vma = NULL;
1465
1466         if (mm) {
1467                 /* Check the cache first. */
1468                 /* (Cache hit rate is typically around 35%.) */
1469                 vma = mm->mmap_cache;
1470                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1471                         struct rb_node * rb_node;
1472
1473                         rb_node = mm->mm_rb.rb_node;
1474                         vma = NULL;
1475
1476                         while (rb_node) {
1477                                 struct vm_area_struct * vma_tmp;
1478
1479                                 vma_tmp = rb_entry(rb_node,
1480                                                 struct vm_area_struct, vm_rb);
1481
1482                                 if (vma_tmp->vm_end > addr) {
1483                                         vma = vma_tmp;
1484                                         if (vma_tmp->vm_start <= addr)
1485                                                 break;
1486                                         rb_node = rb_node->rb_left;
1487                                 } else
1488                                         rb_node = rb_node->rb_right;
1489                         }
1490                         if (vma)
1491                                 mm->mmap_cache = vma;
1492                 }
1493         }
1494         return vma;
1495 }
1496
1497 EXPORT_SYMBOL(find_vma);
1498
1499 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1500 struct vm_area_struct *
1501 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1502                         struct vm_area_struct **pprev)
1503 {
1504         struct vm_area_struct *vma = NULL, *prev = NULL;
1505         struct rb_node *rb_node;
1506         if (!mm)
1507                 goto out;
1508
1509         /* Guard against addr being lower than the first VMA */
1510         vma = mm->mmap;
1511
1512         /* Go through the RB tree quickly. */
1513         rb_node = mm->mm_rb.rb_node;
1514
1515         while (rb_node) {
1516                 struct vm_area_struct *vma_tmp;
1517                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1518
1519                 if (addr < vma_tmp->vm_end) {
1520                         rb_node = rb_node->rb_left;
1521                 } else {
1522                         prev = vma_tmp;
1523                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1524                                 break;
1525                         rb_node = rb_node->rb_right;
1526                 }
1527         }
1528
1529 out:
1530         *pprev = prev;
1531         return prev ? prev->vm_next : vma;
1532 }
1533
1534 /*
1535  * Verify that the stack growth is acceptable and
1536  * update accounting. This is shared with both the
1537  * grow-up and grow-down cases.
1538  */
1539 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1540 {
1541         struct mm_struct *mm = vma->vm_mm;
1542         struct rlimit *rlim = current->signal->rlim;
1543         unsigned long new_start;
1544
1545         /* address space limit tests */
1546         if (!may_expand_vm(mm, grow))
1547                 return -ENOMEM;
1548
1549         /* Stack limit test */
1550         if (size > rlim[RLIMIT_STACK].rlim_cur)
1551                 return -ENOMEM;
1552
1553         /* mlock limit tests */
1554         if (vma->vm_flags & VM_LOCKED) {
1555                 unsigned long locked;
1556                 unsigned long limit;
1557                 locked = mm->locked_vm + grow;
1558                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1559                 if (locked > limit && !capable(CAP_IPC_LOCK))
1560                         return -ENOMEM;
1561         }
1562
1563         /* Check to ensure the stack will not grow into a hugetlb-only region */
1564         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1565                         vma->vm_end - size;
1566         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1567                 return -EFAULT;
1568
1569         /*
1570          * Overcommit..  This must be the final test, as it will
1571          * update security statistics.
1572          */
1573         if (security_vm_enough_memory_mm(mm, grow))
1574                 return -ENOMEM;
1575
1576         /* Ok, everything looks good - let it rip */
1577         mm->total_vm += grow;
1578         if (vma->vm_flags & VM_LOCKED)
1579                 mm->locked_vm += grow;
1580         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1581         return 0;
1582 }
1583
1584 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1585 /*
1586  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1587  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1588  */
1589 #ifndef CONFIG_IA64
1590 static
1591 #endif
1592 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1593 {
1594         int error;
1595
1596         if (!(vma->vm_flags & VM_GROWSUP))
1597                 return -EFAULT;
1598
1599         /*
1600          * We must make sure the anon_vma is allocated
1601          * so that the anon_vma locking is not a noop.
1602          */
1603         if (unlikely(anon_vma_prepare(vma)))
1604                 return -ENOMEM;
1605         anon_vma_lock(vma);
1606
1607         /*
1608          * vma->vm_start/vm_end cannot change under us because the caller
1609          * is required to hold the mmap_sem in read mode.  We need the
1610          * anon_vma lock to serialize against concurrent expand_stacks.
1611          * Also guard against wrapping around to address 0.
1612          */
1613         if (address < PAGE_ALIGN(address+4))
1614                 address = PAGE_ALIGN(address+4);
1615         else {
1616                 anon_vma_unlock(vma);
1617                 return -ENOMEM;
1618         }
1619         error = 0;
1620
1621         /* Somebody else might have raced and expanded it already */
1622         if (address > vma->vm_end) {
1623                 unsigned long size, grow;
1624
1625                 size = address - vma->vm_start;
1626                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1627
1628                 error = acct_stack_growth(vma, size, grow);
1629                 if (!error)
1630                         vma->vm_end = address;
1631         }
1632         anon_vma_unlock(vma);
1633         return error;
1634 }
1635 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1636
1637 /*
1638  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1639  */
1640 static int expand_downwards(struct vm_area_struct *vma,
1641                                    unsigned long address)
1642 {
1643         int error;
1644
1645         /*
1646          * We must make sure the anon_vma is allocated
1647          * so that the anon_vma locking is not a noop.
1648          */
1649         if (unlikely(anon_vma_prepare(vma)))
1650                 return -ENOMEM;
1651
1652         address &= PAGE_MASK;
1653         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1654         if (error)
1655                 return error;
1656
1657         anon_vma_lock(vma);
1658
1659         /*
1660          * vma->vm_start/vm_end cannot change under us because the caller
1661          * is required to hold the mmap_sem in read mode.  We need the
1662          * anon_vma lock to serialize against concurrent expand_stacks.
1663          */
1664
1665         /* Somebody else might have raced and expanded it already */
1666         if (address < vma->vm_start) {
1667                 unsigned long size, grow;
1668
1669                 size = vma->vm_end - address;
1670                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1671
1672                 error = acct_stack_growth(vma, size, grow);
1673                 if (!error) {
1674                         vma->vm_start = address;
1675                         vma->vm_pgoff -= grow;
1676                 }
1677         }
1678         anon_vma_unlock(vma);
1679         return error;
1680 }
1681
1682 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1683 {
1684         return expand_downwards(vma, address);
1685 }
1686
1687 #ifdef CONFIG_STACK_GROWSUP
1688 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1689 {
1690         return expand_upwards(vma, address);
1691 }
1692
1693 struct vm_area_struct *
1694 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1695 {
1696         struct vm_area_struct *vma, *prev;
1697
1698         addr &= PAGE_MASK;
1699         vma = find_vma_prev(mm, addr, &prev);
1700         if (vma && (vma->vm_start <= addr))
1701                 return vma;
1702         if (!prev || expand_stack(prev, addr))
1703                 return NULL;
1704         if (prev->vm_flags & VM_LOCKED) {
1705                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1706                         return NULL;    /* vma gone! */
1707         }
1708         return prev;
1709 }
1710 #else
1711 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1712 {
1713         return expand_downwards(vma, address);
1714 }
1715
1716 struct vm_area_struct *
1717 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1718 {
1719         struct vm_area_struct * vma;
1720         unsigned long start;
1721
1722         addr &= PAGE_MASK;
1723         vma = find_vma(mm,addr);
1724         if (!vma)
1725                 return NULL;
1726         if (vma->vm_start <= addr)
1727                 return vma;
1728         if (!(vma->vm_flags & VM_GROWSDOWN))
1729                 return NULL;
1730         start = vma->vm_start;
1731         if (expand_stack(vma, addr))
1732                 return NULL;
1733         if (vma->vm_flags & VM_LOCKED) {
1734                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1735                         return NULL;    /* vma gone! */
1736         }
1737         return vma;
1738 }
1739 #endif
1740
1741 /*
1742  * Ok - we have the memory areas we should free on the vma list,
1743  * so release them, and do the vma updates.
1744  *
1745  * Called with the mm semaphore held.
1746  */
1747 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1748 {
1749         /* Update high watermark before we lower total_vm */
1750         update_hiwater_vm(mm);
1751         do {
1752                 long nrpages = vma_pages(vma);
1753
1754                 mm->total_vm -= nrpages;
1755                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1756                 vma = remove_vma(vma);
1757         } while (vma);
1758         validate_mm(mm);
1759 }
1760
1761 /*
1762  * Get rid of page table information in the indicated region.
1763  *
1764  * Called with the mm semaphore held.
1765  */
1766 static void unmap_region(struct mm_struct *mm,
1767                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1768                 unsigned long start, unsigned long end)
1769 {
1770         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1771         struct mmu_gather *tlb;
1772         unsigned long nr_accounted = 0;
1773
1774         lru_add_drain();
1775         tlb = tlb_gather_mmu(mm, 0);
1776         update_hiwater_rss(mm);
1777         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1778         vm_unacct_memory(nr_accounted);
1779         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1780                                  next? next->vm_start: 0);
1781         tlb_finish_mmu(tlb, start, end);
1782 }
1783
1784 /*
1785  * Create a list of vma's touched by the unmap, removing them from the mm's
1786  * vma list as we go..
1787  */
1788 static void
1789 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1790         struct vm_area_struct *prev, unsigned long end)
1791 {
1792         struct vm_area_struct **insertion_point;
1793         struct vm_area_struct *tail_vma = NULL;
1794         unsigned long addr;
1795
1796         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1797         do {
1798                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1799                 mm->map_count--;
1800                 tail_vma = vma;
1801                 vma = vma->vm_next;
1802         } while (vma && vma->vm_start < end);
1803         *insertion_point = vma;
1804         tail_vma->vm_next = NULL;
1805         if (mm->unmap_area == arch_unmap_area)
1806                 addr = prev ? prev->vm_end : mm->mmap_base;
1807         else
1808                 addr = vma ?  vma->vm_start : mm->mmap_base;
1809         mm->unmap_area(mm, addr);
1810         mm->mmap_cache = NULL;          /* Kill the cache. */
1811 }
1812
1813 /*
1814  * Split a vma into two pieces at address 'addr', a new vma is allocated
1815  * either for the first part or the tail.
1816  */
1817 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1818               unsigned long addr, int new_below)
1819 {
1820         struct mempolicy *pol;
1821         struct vm_area_struct *new;
1822
1823         if (is_vm_hugetlb_page(vma) && (addr &
1824                                         ~(huge_page_mask(hstate_vma(vma)))))
1825                 return -EINVAL;
1826
1827         if (mm->map_count >= sysctl_max_map_count)
1828                 return -ENOMEM;
1829
1830         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1831         if (!new)
1832                 return -ENOMEM;
1833
1834         /* most fields are the same, copy all, and then fixup */
1835         *new = *vma;
1836
1837         if (new_below)
1838                 new->vm_end = addr;
1839         else {
1840                 new->vm_start = addr;
1841                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1842         }
1843
1844         pol = mpol_dup(vma_policy(vma));
1845         if (IS_ERR(pol)) {
1846                 kmem_cache_free(vm_area_cachep, new);
1847                 return PTR_ERR(pol);
1848         }
1849         vma_set_policy(new, pol);
1850
1851         if (new->vm_file) {
1852                 get_file(new->vm_file);
1853                 if (vma->vm_flags & VM_EXECUTABLE)
1854                         added_exe_file_vma(mm);
1855         }
1856
1857         if (new->vm_ops && new->vm_ops->open)
1858                 new->vm_ops->open(new);
1859
1860         if (new_below)
1861                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1862                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1863         else
1864                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1865
1866         return 0;
1867 }
1868
1869 /* Munmap is split into 2 main parts -- this part which finds
1870  * what needs doing, and the areas themselves, which do the
1871  * work.  This now handles partial unmappings.
1872  * Jeremy Fitzhardinge <jeremy@goop.org>
1873  */
1874 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1875 {
1876         unsigned long end;
1877         struct vm_area_struct *vma, *prev, *last;
1878
1879         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1880                 return -EINVAL;
1881
1882         if ((len = PAGE_ALIGN(len)) == 0)
1883                 return -EINVAL;
1884
1885         /* Find the first overlapping VMA */
1886         vma = find_vma_prev(mm, start, &prev);
1887         if (!vma)
1888                 return 0;
1889         /* we have  start < vma->vm_end  */
1890
1891         /* if it doesn't overlap, we have nothing.. */
1892         end = start + len;
1893         if (vma->vm_start >= end)
1894                 return 0;
1895
1896         /*
1897          * If we need to split any vma, do it now to save pain later.
1898          *
1899          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1900          * unmapped vm_area_struct will remain in use: so lower split_vma
1901          * places tmp vma above, and higher split_vma places tmp vma below.
1902          */
1903         if (start > vma->vm_start) {
1904                 int error = split_vma(mm, vma, start, 0);
1905                 if (error)
1906                         return error;
1907                 prev = vma;
1908         }
1909
1910         /* Does it split the last one? */
1911         last = find_vma(mm, end);
1912         if (last && end > last->vm_start) {
1913                 int error = split_vma(mm, last, end, 1);
1914                 if (error)
1915                         return error;
1916         }
1917         vma = prev? prev->vm_next: mm->mmap;
1918
1919         /*
1920          * unlock any mlock()ed ranges before detaching vmas
1921          */
1922         if (mm->locked_vm) {
1923                 struct vm_area_struct *tmp = vma;
1924                 while (tmp && tmp->vm_start < end) {
1925                         if (tmp->vm_flags & VM_LOCKED) {
1926                                 mm->locked_vm -= vma_pages(tmp);
1927                                 munlock_vma_pages_all(tmp);
1928                         }
1929                         tmp = tmp->vm_next;
1930                 }
1931         }
1932
1933         /*
1934          * Remove the vma's, and unmap the actual pages
1935          */
1936         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1937         unmap_region(mm, vma, prev, start, end);
1938
1939         /* Fix up all other VM information */
1940         remove_vma_list(mm, vma);
1941
1942         return 0;
1943 }
1944
1945 EXPORT_SYMBOL(do_munmap);
1946
1947 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1948 {
1949         int ret;
1950         struct mm_struct *mm = current->mm;
1951
1952         profile_munmap(addr);
1953
1954         down_write(&mm->mmap_sem);
1955         ret = do_munmap(mm, addr, len);
1956         up_write(&mm->mmap_sem);
1957         return ret;
1958 }
1959
1960 static inline void verify_mm_writelocked(struct mm_struct *mm)
1961 {
1962 #ifdef CONFIG_DEBUG_VM
1963         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1964                 WARN_ON(1);
1965                 up_read(&mm->mmap_sem);
1966         }
1967 #endif
1968 }
1969
1970 /*
1971  *  this is really a simplified "do_mmap".  it only handles
1972  *  anonymous maps.  eventually we may be able to do some
1973  *  brk-specific accounting here.
1974  */
1975 unsigned long do_brk(unsigned long addr, unsigned long len)
1976 {
1977         struct mm_struct * mm = current->mm;
1978         struct vm_area_struct * vma, * prev;
1979         unsigned long flags;
1980         struct rb_node ** rb_link, * rb_parent;
1981         pgoff_t pgoff = addr >> PAGE_SHIFT;
1982         int error;
1983
1984         len = PAGE_ALIGN(len);
1985         if (!len)
1986                 return addr;
1987
1988         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1989                 return -EINVAL;
1990
1991         if (is_hugepage_only_range(mm, addr, len))
1992                 return -EINVAL;
1993
1994         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1995         if (error)
1996                 return error;
1997
1998         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1999
2000         error = arch_mmap_check(addr, len, flags);
2001         if (error)
2002                 return error;
2003
2004         /*
2005          * mlock MCL_FUTURE?
2006          */
2007         if (mm->def_flags & VM_LOCKED) {
2008                 unsigned long locked, lock_limit;
2009                 locked = len >> PAGE_SHIFT;
2010                 locked += mm->locked_vm;
2011                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2012                 lock_limit >>= PAGE_SHIFT;
2013                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2014                         return -EAGAIN;
2015         }
2016
2017         /*
2018          * mm->mmap_sem is required to protect against another thread
2019          * changing the mappings in case we sleep.
2020          */
2021         verify_mm_writelocked(mm);
2022
2023         /*
2024          * Clear old maps.  this also does some error checking for us
2025          */
2026  munmap_back:
2027         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2028         if (vma && vma->vm_start < addr + len) {
2029                 if (do_munmap(mm, addr, len))
2030                         return -ENOMEM;
2031                 goto munmap_back;
2032         }
2033
2034         /* Check against address space limits *after* clearing old maps... */
2035         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2036                 return -ENOMEM;
2037
2038         if (mm->map_count > sysctl_max_map_count)
2039                 return -ENOMEM;
2040
2041         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2042                 return -ENOMEM;
2043
2044         /* Can we just expand an old private anonymous mapping? */
2045         vma = vma_merge(mm, prev, addr, addr + len, flags,
2046                                         NULL, NULL, pgoff, NULL);
2047         if (vma)
2048                 goto out;
2049
2050         /*
2051          * create a vma struct for an anonymous mapping
2052          */
2053         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2054         if (!vma) {
2055                 vm_unacct_memory(len >> PAGE_SHIFT);
2056                 return -ENOMEM;
2057         }
2058
2059         vma->vm_mm = mm;
2060         vma->vm_start = addr;
2061         vma->vm_end = addr + len;
2062         vma->vm_pgoff = pgoff;
2063         vma->vm_flags = flags;
2064         vma->vm_page_prot = vm_get_page_prot(flags);
2065         vma_link(mm, vma, prev, rb_link, rb_parent);
2066 out:
2067         mm->total_vm += len >> PAGE_SHIFT;
2068         if (flags & VM_LOCKED) {
2069                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2070                         mm->locked_vm += (len >> PAGE_SHIFT);
2071         }
2072         return addr;
2073 }
2074
2075 EXPORT_SYMBOL(do_brk);
2076
2077 /* Release all mmaps. */
2078 void exit_mmap(struct mm_struct *mm)
2079 {
2080         struct mmu_gather *tlb;
2081         struct vm_area_struct *vma;
2082         unsigned long nr_accounted = 0;
2083         unsigned long end;
2084
2085         /* mm's last user has gone, and its about to be pulled down */
2086         mmu_notifier_release(mm);
2087
2088         if (mm->locked_vm) {
2089                 vma = mm->mmap;
2090                 while (vma) {
2091                         if (vma->vm_flags & VM_LOCKED)
2092                                 munlock_vma_pages_all(vma);
2093                         vma = vma->vm_next;
2094                 }
2095         }
2096
2097         arch_exit_mmap(mm);
2098
2099         vma = mm->mmap;
2100         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2101                 return;
2102
2103         lru_add_drain();
2104         flush_cache_mm(mm);
2105         tlb = tlb_gather_mmu(mm, 1);
2106         /* update_hiwater_rss(mm) here? but nobody should be looking */
2107         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2108         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2109         vm_unacct_memory(nr_accounted);
2110
2111         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2112         tlb_finish_mmu(tlb, 0, end);
2113
2114         /*
2115          * Walk the list again, actually closing and freeing it,
2116          * with preemption enabled, without holding any MM locks.
2117          */
2118         while (vma)
2119                 vma = remove_vma(vma);
2120
2121         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2122 }
2123
2124 /* Insert vm structure into process list sorted by address
2125  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2126  * then i_mmap_lock is taken here.
2127  */
2128 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2129 {
2130         struct vm_area_struct * __vma, * prev;
2131         struct rb_node ** rb_link, * rb_parent;
2132
2133         /*
2134          * The vm_pgoff of a purely anonymous vma should be irrelevant
2135          * until its first write fault, when page's anon_vma and index
2136          * are set.  But now set the vm_pgoff it will almost certainly
2137          * end up with (unless mremap moves it elsewhere before that
2138          * first wfault), so /proc/pid/maps tells a consistent story.
2139          *
2140          * By setting it to reflect the virtual start address of the
2141          * vma, merges and splits can happen in a seamless way, just
2142          * using the existing file pgoff checks and manipulations.
2143          * Similarly in do_mmap_pgoff and in do_brk.
2144          */
2145         if (!vma->vm_file) {
2146                 BUG_ON(vma->anon_vma);
2147                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2148         }
2149         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2150         if (__vma && __vma->vm_start < vma->vm_end)
2151                 return -ENOMEM;
2152         if ((vma->vm_flags & VM_ACCOUNT) &&
2153              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2154                 return -ENOMEM;
2155         vma_link(mm, vma, prev, rb_link, rb_parent);
2156         return 0;
2157 }
2158
2159 /*
2160  * Copy the vma structure to a new location in the same mm,
2161  * prior to moving page table entries, to effect an mremap move.
2162  */
2163 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2164         unsigned long addr, unsigned long len, pgoff_t pgoff)
2165 {
2166         struct vm_area_struct *vma = *vmap;
2167         unsigned long vma_start = vma->vm_start;
2168         struct mm_struct *mm = vma->vm_mm;
2169         struct vm_area_struct *new_vma, *prev;
2170         struct rb_node **rb_link, *rb_parent;
2171         struct mempolicy *pol;
2172
2173         /*
2174          * If anonymous vma has not yet been faulted, update new pgoff
2175          * to match new location, to increase its chance of merging.
2176          */
2177         if (!vma->vm_file && !vma->anon_vma)
2178                 pgoff = addr >> PAGE_SHIFT;
2179
2180         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2181         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2182                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2183         if (new_vma) {
2184                 /*
2185                  * Source vma may have been merged into new_vma
2186                  */
2187                 if (vma_start >= new_vma->vm_start &&
2188                     vma_start < new_vma->vm_end)
2189                         *vmap = new_vma;
2190         } else {
2191                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2192                 if (new_vma) {
2193                         *new_vma = *vma;
2194                         pol = mpol_dup(vma_policy(vma));
2195                         if (IS_ERR(pol)) {
2196                                 kmem_cache_free(vm_area_cachep, new_vma);
2197                                 return NULL;
2198                         }
2199                         vma_set_policy(new_vma, pol);
2200                         new_vma->vm_start = addr;
2201                         new_vma->vm_end = addr + len;
2202                         new_vma->vm_pgoff = pgoff;
2203                         if (new_vma->vm_file) {
2204                                 get_file(new_vma->vm_file);
2205                                 if (vma->vm_flags & VM_EXECUTABLE)
2206                                         added_exe_file_vma(mm);
2207                         }
2208                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2209                                 new_vma->vm_ops->open(new_vma);
2210                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2211                 }
2212         }
2213         return new_vma;
2214 }
2215
2216 /*
2217  * Return true if the calling process may expand its vm space by the passed
2218  * number of pages
2219  */
2220 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2221 {
2222         unsigned long cur = mm->total_vm;       /* pages */
2223         unsigned long lim;
2224
2225         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2226
2227         if (cur + npages > lim)
2228                 return 0;
2229         return 1;
2230 }
2231
2232
2233 static int special_mapping_fault(struct vm_area_struct *vma,
2234                                 struct vm_fault *vmf)
2235 {
2236         pgoff_t pgoff;
2237         struct page **pages;
2238
2239         /*
2240          * special mappings have no vm_file, and in that case, the mm
2241          * uses vm_pgoff internally. So we have to subtract it from here.
2242          * We are allowed to do this because we are the mm; do not copy
2243          * this code into drivers!
2244          */
2245         pgoff = vmf->pgoff - vma->vm_pgoff;
2246
2247         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2248                 pgoff--;
2249
2250         if (*pages) {
2251                 struct page *page = *pages;
2252                 get_page(page);
2253                 vmf->page = page;
2254                 return 0;
2255         }
2256
2257         return VM_FAULT_SIGBUS;
2258 }
2259
2260 /*
2261  * Having a close hook prevents vma merging regardless of flags.
2262  */
2263 static void special_mapping_close(struct vm_area_struct *vma)
2264 {
2265 }
2266
2267 static struct vm_operations_struct special_mapping_vmops = {
2268         .close = special_mapping_close,
2269         .fault = special_mapping_fault,
2270 };
2271
2272 /*
2273  * Called with mm->mmap_sem held for writing.
2274  * Insert a new vma covering the given region, with the given flags.
2275  * Its pages are supplied by the given array of struct page *.
2276  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2277  * The region past the last page supplied will always produce SIGBUS.
2278  * The array pointer and the pages it points to are assumed to stay alive
2279  * for as long as this mapping might exist.
2280  */
2281 int install_special_mapping(struct mm_struct *mm,
2282                             unsigned long addr, unsigned long len,
2283                             unsigned long vm_flags, struct page **pages)
2284 {
2285         struct vm_area_struct *vma;
2286
2287         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2288         if (unlikely(vma == NULL))
2289                 return -ENOMEM;
2290
2291         vma->vm_mm = mm;
2292         vma->vm_start = addr;
2293         vma->vm_end = addr + len;
2294
2295         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2296         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2297
2298         vma->vm_ops = &special_mapping_vmops;
2299         vma->vm_private_data = pages;
2300
2301         if (unlikely(insert_vm_struct(mm, vma))) {
2302                 kmem_cache_free(vm_area_cachep, vma);
2303                 return -ENOMEM;
2304         }
2305
2306         mm->total_vm += len >> PAGE_SHIFT;
2307
2308         perf_event_mmap(vma);
2309
2310         return 0;
2311 }
2312
2313 static DEFINE_MUTEX(mm_all_locks_mutex);
2314
2315 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2316 {
2317         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2318                 /*
2319                  * The LSB of head.next can't change from under us
2320                  * because we hold the mm_all_locks_mutex.
2321                  */
2322                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2323                 /*
2324                  * We can safely modify head.next after taking the
2325                  * anon_vma->lock. If some other vma in this mm shares
2326                  * the same anon_vma we won't take it again.
2327                  *
2328                  * No need of atomic instructions here, head.next
2329                  * can't change from under us thanks to the
2330                  * anon_vma->lock.
2331                  */
2332                 if (__test_and_set_bit(0, (unsigned long *)
2333                                        &anon_vma->head.next))
2334                         BUG();
2335         }
2336 }
2337
2338 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2339 {
2340         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2341                 /*
2342                  * AS_MM_ALL_LOCKS can't change from under us because
2343                  * we hold the mm_all_locks_mutex.
2344                  *
2345                  * Operations on ->flags have to be atomic because
2346                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2347                  * mm_all_locks_mutex, there may be other cpus
2348                  * changing other bitflags in parallel to us.
2349                  */
2350                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2351                         BUG();
2352                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2353         }
2354 }
2355
2356 /*
2357  * This operation locks against the VM for all pte/vma/mm related
2358  * operations that could ever happen on a certain mm. This includes
2359  * vmtruncate, try_to_unmap, and all page faults.
2360  *
2361  * The caller must take the mmap_sem in write mode before calling
2362  * mm_take_all_locks(). The caller isn't allowed to release the
2363  * mmap_sem until mm_drop_all_locks() returns.
2364  *
2365  * mmap_sem in write mode is required in order to block all operations
2366  * that could modify pagetables and free pages without need of
2367  * altering the vma layout (for example populate_range() with
2368  * nonlinear vmas). It's also needed in write mode to avoid new
2369  * anon_vmas to be associated with existing vmas.
2370  *
2371  * A single task can't take more than one mm_take_all_locks() in a row
2372  * or it would deadlock.
2373  *
2374  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2375  * mapping->flags avoid to take the same lock twice, if more than one
2376  * vma in this mm is backed by the same anon_vma or address_space.
2377  *
2378  * We can take all the locks in random order because the VM code
2379  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2380  * takes more than one of them in a row. Secondly we're protected
2381  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2382  *
2383  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2384  * that may have to take thousand of locks.
2385  *
2386  * mm_take_all_locks() can fail if it's interrupted by signals.
2387  */
2388 int mm_take_all_locks(struct mm_struct *mm)
2389 {
2390         struct vm_area_struct *vma;
2391         int ret = -EINTR;
2392
2393         BUG_ON(down_read_trylock(&mm->mmap_sem));
2394
2395         mutex_lock(&mm_all_locks_mutex);
2396
2397         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2398                 if (signal_pending(current))
2399                         goto out_unlock;
2400                 if (vma->vm_file && vma->vm_file->f_mapping)
2401                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2402         }
2403
2404         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2405                 if (signal_pending(current))
2406                         goto out_unlock;
2407                 if (vma->anon_vma)
2408                         vm_lock_anon_vma(mm, vma->anon_vma);
2409         }
2410
2411         ret = 0;
2412
2413 out_unlock:
2414         if (ret)
2415                 mm_drop_all_locks(mm);
2416
2417         return ret;
2418 }
2419
2420 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2421 {
2422         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2423                 /*
2424                  * The LSB of head.next can't change to 0 from under
2425                  * us because we hold the mm_all_locks_mutex.
2426                  *
2427                  * We must however clear the bitflag before unlocking
2428                  * the vma so the users using the anon_vma->head will
2429                  * never see our bitflag.
2430                  *
2431                  * No need of atomic instructions here, head.next
2432                  * can't change from under us until we release the
2433                  * anon_vma->lock.
2434                  */
2435                 if (!__test_and_clear_bit(0, (unsigned long *)
2436                                           &anon_vma->head.next))
2437                         BUG();
2438                 spin_unlock(&anon_vma->lock);
2439         }
2440 }
2441
2442 static void vm_unlock_mapping(struct address_space *mapping)
2443 {
2444         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2445                 /*
2446                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2447                  * because we hold the mm_all_locks_mutex.
2448                  */
2449                 spin_unlock(&mapping->i_mmap_lock);
2450                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2451                                         &mapping->flags))
2452                         BUG();
2453         }
2454 }
2455
2456 /*
2457  * The mmap_sem cannot be released by the caller until
2458  * mm_drop_all_locks() returns.
2459  */
2460 void mm_drop_all_locks(struct mm_struct *mm)
2461 {
2462         struct vm_area_struct *vma;
2463
2464         BUG_ON(down_read_trylock(&mm->mmap_sem));
2465         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2466
2467         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2468                 if (vma->anon_vma)
2469                         vm_unlock_anon_vma(vma->anon_vma);
2470                 if (vma->vm_file && vma->vm_file->f_mapping)
2471                         vm_unlock_mapping(vma->vm_file->f_mapping);
2472         }
2473
2474         mutex_unlock(&mm_all_locks_mutex);
2475 }
2476
2477 /*
2478  * initialise the VMA slab
2479  */
2480 void __init mmap_init(void)
2481 {
2482         int ret;
2483
2484         ret = percpu_counter_init(&vm_committed_as, 0);
2485         VM_BUG_ON(ret);
2486 }