ksm: clean up obsolete references
[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                 vm_flags |= VM_LOCKED;
970         }
971
972         /* mlock MCL_FUTURE? */
973         if (vm_flags & VM_LOCKED) {
974                 unsigned long locked, lock_limit;
975                 locked = len >> PAGE_SHIFT;
976                 locked += mm->locked_vm;
977                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
978                 lock_limit >>= PAGE_SHIFT;
979                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
980                         return -EAGAIN;
981         }
982
983         inode = file ? file->f_path.dentry->d_inode : NULL;
984
985         if (file) {
986                 switch (flags & MAP_TYPE) {
987                 case MAP_SHARED:
988                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
989                                 return -EACCES;
990
991                         /*
992                          * Make sure we don't allow writing to an append-only
993                          * file..
994                          */
995                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
996                                 return -EACCES;
997
998                         /*
999                          * Make sure there are no mandatory locks on the file.
1000                          */
1001                         if (locks_verify_locked(inode))
1002                                 return -EAGAIN;
1003
1004                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1005                         if (!(file->f_mode & FMODE_WRITE))
1006                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1007
1008                         /* fall through */
1009                 case MAP_PRIVATE:
1010                         if (!(file->f_mode & FMODE_READ))
1011                                 return -EACCES;
1012                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1013                                 if (vm_flags & VM_EXEC)
1014                                         return -EPERM;
1015                                 vm_flags &= ~VM_MAYEXEC;
1016                         }
1017
1018                         if (!file->f_op || !file->f_op->mmap)
1019                                 return -ENODEV;
1020                         break;
1021
1022                 default:
1023                         return -EINVAL;
1024                 }
1025         } else {
1026                 switch (flags & MAP_TYPE) {
1027                 case MAP_SHARED:
1028                         /*
1029                          * Ignore pgoff.
1030                          */
1031                         pgoff = 0;
1032                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1033                         break;
1034                 case MAP_PRIVATE:
1035                         /*
1036                          * Set pgoff according to addr for anon_vma.
1037                          */
1038                         pgoff = addr >> PAGE_SHIFT;
1039                         break;
1040                 default:
1041                         return -EINVAL;
1042                 }
1043         }
1044
1045         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1046         if (error)
1047                 return error;
1048         error = ima_file_mmap(file, prot);
1049         if (error)
1050                 return error;
1051
1052         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1053 }
1054 EXPORT_SYMBOL(do_mmap_pgoff);
1055
1056 /*
1057  * Some shared mappigns will want the pages marked read-only
1058  * to track write events. If so, we'll downgrade vm_page_prot
1059  * to the private version (using protection_map[] without the
1060  * VM_SHARED bit).
1061  */
1062 int vma_wants_writenotify(struct vm_area_struct *vma)
1063 {
1064         unsigned int vm_flags = vma->vm_flags;
1065
1066         /* If it was private or non-writable, the write bit is already clear */
1067         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1068                 return 0;
1069
1070         /* The backer wishes to know when pages are first written to? */
1071         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1072                 return 1;
1073
1074         /* The open routine did something to the protections already? */
1075         if (pgprot_val(vma->vm_page_prot) !=
1076             pgprot_val(vm_get_page_prot(vm_flags)))
1077                 return 0;
1078
1079         /* Specialty mapping? */
1080         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1081                 return 0;
1082
1083         /* Can the mapping track the dirty pages? */
1084         return vma->vm_file && vma->vm_file->f_mapping &&
1085                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1086 }
1087
1088 /*
1089  * We account for memory if it's a private writeable mapping,
1090  * not hugepages and VM_NORESERVE wasn't set.
1091  */
1092 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1093 {
1094         /*
1095          * hugetlb has its own accounting separate from the core VM
1096          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1097          */
1098         if (file && is_file_hugepages(file))
1099                 return 0;
1100
1101         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1102 }
1103
1104 unsigned long mmap_region(struct file *file, unsigned long addr,
1105                           unsigned long len, unsigned long flags,
1106                           unsigned int vm_flags, unsigned long pgoff)
1107 {
1108         struct mm_struct *mm = current->mm;
1109         struct vm_area_struct *vma, *prev;
1110         int correct_wcount = 0;
1111         int error;
1112         struct rb_node **rb_link, *rb_parent;
1113         unsigned long charged = 0;
1114         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1115
1116         /* Clear old maps */
1117         error = -ENOMEM;
1118 munmap_back:
1119         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1120         if (vma && vma->vm_start < addr + len) {
1121                 if (do_munmap(mm, addr, len))
1122                         return -ENOMEM;
1123                 goto munmap_back;
1124         }
1125
1126         /* Check against address space limit. */
1127         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1128                 return -ENOMEM;
1129
1130         /*
1131          * Set 'VM_NORESERVE' if we should not account for the
1132          * memory use of this mapping.
1133          */
1134         if ((flags & MAP_NORESERVE)) {
1135                 /* We honor MAP_NORESERVE if allowed to overcommit */
1136                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1137                         vm_flags |= VM_NORESERVE;
1138
1139                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1140                 if (file && is_file_hugepages(file))
1141                         vm_flags |= VM_NORESERVE;
1142         }
1143
1144         /*
1145          * Private writable mapping: check memory availability
1146          */
1147         if (accountable_mapping(file, vm_flags)) {
1148                 charged = len >> PAGE_SHIFT;
1149                 if (security_vm_enough_memory(charged))
1150                         return -ENOMEM;
1151                 vm_flags |= VM_ACCOUNT;
1152         }
1153
1154         /*
1155          * Can we just expand an old mapping?
1156          */
1157         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1158         if (vma)
1159                 goto out;
1160
1161         /*
1162          * Determine the object being mapped and call the appropriate
1163          * specific mapper. the address has already been validated, but
1164          * not unmapped, but the maps are removed from the list.
1165          */
1166         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1167         if (!vma) {
1168                 error = -ENOMEM;
1169                 goto unacct_error;
1170         }
1171
1172         vma->vm_mm = mm;
1173         vma->vm_start = addr;
1174         vma->vm_end = addr + len;
1175         vma->vm_flags = vm_flags;
1176         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1177         vma->vm_pgoff = pgoff;
1178
1179         if (file) {
1180                 error = -EINVAL;
1181                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1182                         goto free_vma;
1183                 if (vm_flags & VM_DENYWRITE) {
1184                         error = deny_write_access(file);
1185                         if (error)
1186                                 goto free_vma;
1187                         correct_wcount = 1;
1188                 }
1189                 vma->vm_file = file;
1190                 get_file(file);
1191                 error = file->f_op->mmap(file, vma);
1192                 if (error)
1193                         goto unmap_and_free_vma;
1194                 if (vm_flags & VM_EXECUTABLE)
1195                         added_exe_file_vma(mm);
1196         } else if (vm_flags & VM_SHARED) {
1197                 error = shmem_zero_setup(vma);
1198                 if (error)
1199                         goto free_vma;
1200         }
1201
1202         /* Can addr have changed??
1203          *
1204          * Answer: Yes, several device drivers can do it in their
1205          *         f_op->mmap method. -DaveM
1206          */
1207         addr = vma->vm_start;
1208         pgoff = vma->vm_pgoff;
1209         vm_flags = vma->vm_flags;
1210
1211         if (vma_wants_writenotify(vma))
1212                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1213
1214         vma_link(mm, vma, prev, rb_link, rb_parent);
1215         file = vma->vm_file;
1216
1217         /* Once vma denies write, undo our temporary denial count */
1218         if (correct_wcount)
1219                 atomic_inc(&inode->i_writecount);
1220 out:
1221         perf_event_mmap(vma);
1222
1223         mm->total_vm += len >> PAGE_SHIFT;
1224         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1225         if (vm_flags & VM_LOCKED) {
1226                 /*
1227                  * makes pages present; downgrades, drops, reacquires mmap_sem
1228                  */
1229                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1230                 if (nr_pages < 0)
1231                         return nr_pages;        /* vma gone! */
1232                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1233         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1234                 make_pages_present(addr, addr + len);
1235         return addr;
1236
1237 unmap_and_free_vma:
1238         if (correct_wcount)
1239                 atomic_inc(&inode->i_writecount);
1240         vma->vm_file = NULL;
1241         fput(file);
1242
1243         /* Undo any partial mapping done by a device driver. */
1244         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1245         charged = 0;
1246 free_vma:
1247         kmem_cache_free(vm_area_cachep, vma);
1248 unacct_error:
1249         if (charged)
1250                 vm_unacct_memory(charged);
1251         return error;
1252 }
1253
1254 /* Get an address range which is currently unmapped.
1255  * For shmat() with addr=0.
1256  *
1257  * Ugly calling convention alert:
1258  * Return value with the low bits set means error value,
1259  * ie
1260  *      if (ret & ~PAGE_MASK)
1261  *              error = ret;
1262  *
1263  * This function "knows" that -ENOMEM has the bits set.
1264  */
1265 #ifndef HAVE_ARCH_UNMAPPED_AREA
1266 unsigned long
1267 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1268                 unsigned long len, unsigned long pgoff, unsigned long flags)
1269 {
1270         struct mm_struct *mm = current->mm;
1271         struct vm_area_struct *vma;
1272         unsigned long start_addr;
1273
1274         if (len > TASK_SIZE)
1275                 return -ENOMEM;
1276
1277         if (flags & MAP_FIXED)
1278                 return addr;
1279
1280         if (addr) {
1281                 addr = PAGE_ALIGN(addr);
1282                 vma = find_vma(mm, addr);
1283                 if (TASK_SIZE - len >= addr &&
1284                     (!vma || addr + len <= vma->vm_start))
1285                         return addr;
1286         }
1287         if (len > mm->cached_hole_size) {
1288                 start_addr = addr = mm->free_area_cache;
1289         } else {
1290                 start_addr = addr = TASK_UNMAPPED_BASE;
1291                 mm->cached_hole_size = 0;
1292         }
1293
1294 full_search:
1295         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1296                 /* At this point:  (!vma || addr < vma->vm_end). */
1297                 if (TASK_SIZE - len < addr) {
1298                         /*
1299                          * Start a new search - just in case we missed
1300                          * some holes.
1301                          */
1302                         if (start_addr != TASK_UNMAPPED_BASE) {
1303                                 addr = TASK_UNMAPPED_BASE;
1304                                 start_addr = addr;
1305                                 mm->cached_hole_size = 0;
1306                                 goto full_search;
1307                         }
1308                         return -ENOMEM;
1309                 }
1310                 if (!vma || addr + len <= vma->vm_start) {
1311                         /*
1312                          * Remember the place where we stopped the search:
1313                          */
1314                         mm->free_area_cache = addr + len;
1315                         return addr;
1316                 }
1317                 if (addr + mm->cached_hole_size < vma->vm_start)
1318                         mm->cached_hole_size = vma->vm_start - addr;
1319                 addr = vma->vm_end;
1320         }
1321 }
1322 #endif  
1323
1324 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1325 {
1326         /*
1327          * Is this a new hole at the lowest possible address?
1328          */
1329         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1330                 mm->free_area_cache = addr;
1331                 mm->cached_hole_size = ~0UL;
1332         }
1333 }
1334
1335 /*
1336  * This mmap-allocator allocates new areas top-down from below the
1337  * stack's low limit (the base):
1338  */
1339 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1340 unsigned long
1341 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1342                           const unsigned long len, const unsigned long pgoff,
1343                           const unsigned long flags)
1344 {
1345         struct vm_area_struct *vma;
1346         struct mm_struct *mm = current->mm;
1347         unsigned long addr = addr0;
1348
1349         /* requested length too big for entire address space */
1350         if (len > TASK_SIZE)
1351                 return -ENOMEM;
1352
1353         if (flags & MAP_FIXED)
1354                 return addr;
1355
1356         /* requesting a specific address */
1357         if (addr) {
1358                 addr = PAGE_ALIGN(addr);
1359                 vma = find_vma(mm, addr);
1360                 if (TASK_SIZE - len >= addr &&
1361                                 (!vma || addr + len <= vma->vm_start))
1362                         return addr;
1363         }
1364
1365         /* check if free_area_cache is useful for us */
1366         if (len <= mm->cached_hole_size) {
1367                 mm->cached_hole_size = 0;
1368                 mm->free_area_cache = mm->mmap_base;
1369         }
1370
1371         /* either no address requested or can't fit in requested address hole */
1372         addr = mm->free_area_cache;
1373
1374         /* make sure it can fit in the remaining address space */
1375         if (addr > len) {
1376                 vma = find_vma(mm, addr-len);
1377                 if (!vma || addr <= vma->vm_start)
1378                         /* remember the address as a hint for next time */
1379                         return (mm->free_area_cache = addr-len);
1380         }
1381
1382         if (mm->mmap_base < len)
1383                 goto bottomup;
1384
1385         addr = mm->mmap_base-len;
1386
1387         do {
1388                 /*
1389                  * Lookup failure means no vma is above this address,
1390                  * else if new region fits below vma->vm_start,
1391                  * return with success:
1392                  */
1393                 vma = find_vma(mm, addr);
1394                 if (!vma || addr+len <= vma->vm_start)
1395                         /* remember the address as a hint for next time */
1396                         return (mm->free_area_cache = addr);
1397
1398                 /* remember the largest hole we saw so far */
1399                 if (addr + mm->cached_hole_size < vma->vm_start)
1400                         mm->cached_hole_size = vma->vm_start - addr;
1401
1402                 /* try just below the current vma->vm_start */
1403                 addr = vma->vm_start-len;
1404         } while (len < vma->vm_start);
1405
1406 bottomup:
1407         /*
1408          * A failed mmap() very likely causes application failure,
1409          * so fall back to the bottom-up function here. This scenario
1410          * can happen with large stack limits and large mmap()
1411          * allocations.
1412          */
1413         mm->cached_hole_size = ~0UL;
1414         mm->free_area_cache = TASK_UNMAPPED_BASE;
1415         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1416         /*
1417          * Restore the topdown base:
1418          */
1419         mm->free_area_cache = mm->mmap_base;
1420         mm->cached_hole_size = ~0UL;
1421
1422         return addr;
1423 }
1424 #endif
1425
1426 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1427 {
1428         /*
1429          * Is this a new hole at the highest possible address?
1430          */
1431         if (addr > mm->free_area_cache)
1432                 mm->free_area_cache = addr;
1433
1434         /* dont allow allocations above current base */
1435         if (mm->free_area_cache > mm->mmap_base)
1436                 mm->free_area_cache = mm->mmap_base;
1437 }
1438
1439 unsigned long
1440 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1441                 unsigned long pgoff, unsigned long flags)
1442 {
1443         unsigned long (*get_area)(struct file *, unsigned long,
1444                                   unsigned long, unsigned long, unsigned long);
1445
1446         get_area = current->mm->get_unmapped_area;
1447         if (file && file->f_op && file->f_op->get_unmapped_area)
1448                 get_area = file->f_op->get_unmapped_area;
1449         addr = get_area(file, addr, len, pgoff, flags);
1450         if (IS_ERR_VALUE(addr))
1451                 return addr;
1452
1453         if (addr > TASK_SIZE - len)
1454                 return -ENOMEM;
1455         if (addr & ~PAGE_MASK)
1456                 return -EINVAL;
1457
1458         return arch_rebalance_pgtables(addr, len);
1459 }
1460
1461 EXPORT_SYMBOL(get_unmapped_area);
1462
1463 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1464 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1465 {
1466         struct vm_area_struct *vma = NULL;
1467
1468         if (mm) {
1469                 /* Check the cache first. */
1470                 /* (Cache hit rate is typically around 35%.) */
1471                 vma = mm->mmap_cache;
1472                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1473                         struct rb_node * rb_node;
1474
1475                         rb_node = mm->mm_rb.rb_node;
1476                         vma = NULL;
1477
1478                         while (rb_node) {
1479                                 struct vm_area_struct * vma_tmp;
1480
1481                                 vma_tmp = rb_entry(rb_node,
1482                                                 struct vm_area_struct, vm_rb);
1483
1484                                 if (vma_tmp->vm_end > addr) {
1485                                         vma = vma_tmp;
1486                                         if (vma_tmp->vm_start <= addr)
1487                                                 break;
1488                                         rb_node = rb_node->rb_left;
1489                                 } else
1490                                         rb_node = rb_node->rb_right;
1491                         }
1492                         if (vma)
1493                                 mm->mmap_cache = vma;
1494                 }
1495         }
1496         return vma;
1497 }
1498
1499 EXPORT_SYMBOL(find_vma);
1500
1501 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1502 struct vm_area_struct *
1503 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1504                         struct vm_area_struct **pprev)
1505 {
1506         struct vm_area_struct *vma = NULL, *prev = NULL;
1507         struct rb_node *rb_node;
1508         if (!mm)
1509                 goto out;
1510
1511         /* Guard against addr being lower than the first VMA */
1512         vma = mm->mmap;
1513
1514         /* Go through the RB tree quickly. */
1515         rb_node = mm->mm_rb.rb_node;
1516
1517         while (rb_node) {
1518                 struct vm_area_struct *vma_tmp;
1519                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1520
1521                 if (addr < vma_tmp->vm_end) {
1522                         rb_node = rb_node->rb_left;
1523                 } else {
1524                         prev = vma_tmp;
1525                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1526                                 break;
1527                         rb_node = rb_node->rb_right;
1528                 }
1529         }
1530
1531 out:
1532         *pprev = prev;
1533         return prev ? prev->vm_next : vma;
1534 }
1535
1536 /*
1537  * Verify that the stack growth is acceptable and
1538  * update accounting. This is shared with both the
1539  * grow-up and grow-down cases.
1540  */
1541 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1542 {
1543         struct mm_struct *mm = vma->vm_mm;
1544         struct rlimit *rlim = current->signal->rlim;
1545         unsigned long new_start;
1546
1547         /* address space limit tests */
1548         if (!may_expand_vm(mm, grow))
1549                 return -ENOMEM;
1550
1551         /* Stack limit test */
1552         if (size > rlim[RLIMIT_STACK].rlim_cur)
1553                 return -ENOMEM;
1554
1555         /* mlock limit tests */
1556         if (vma->vm_flags & VM_LOCKED) {
1557                 unsigned long locked;
1558                 unsigned long limit;
1559                 locked = mm->locked_vm + grow;
1560                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1561                 if (locked > limit && !capable(CAP_IPC_LOCK))
1562                         return -ENOMEM;
1563         }
1564
1565         /* Check to ensure the stack will not grow into a hugetlb-only region */
1566         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1567                         vma->vm_end - size;
1568         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1569                 return -EFAULT;
1570
1571         /*
1572          * Overcommit..  This must be the final test, as it will
1573          * update security statistics.
1574          */
1575         if (security_vm_enough_memory_mm(mm, grow))
1576                 return -ENOMEM;
1577
1578         /* Ok, everything looks good - let it rip */
1579         mm->total_vm += grow;
1580         if (vma->vm_flags & VM_LOCKED)
1581                 mm->locked_vm += grow;
1582         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1583         return 0;
1584 }
1585
1586 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1587 /*
1588  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1589  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1590  */
1591 #ifndef CONFIG_IA64
1592 static
1593 #endif
1594 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1595 {
1596         int error;
1597
1598         if (!(vma->vm_flags & VM_GROWSUP))
1599                 return -EFAULT;
1600
1601         /*
1602          * We must make sure the anon_vma is allocated
1603          * so that the anon_vma locking is not a noop.
1604          */
1605         if (unlikely(anon_vma_prepare(vma)))
1606                 return -ENOMEM;
1607         anon_vma_lock(vma);
1608
1609         /*
1610          * vma->vm_start/vm_end cannot change under us because the caller
1611          * is required to hold the mmap_sem in read mode.  We need the
1612          * anon_vma lock to serialize against concurrent expand_stacks.
1613          * Also guard against wrapping around to address 0.
1614          */
1615         if (address < PAGE_ALIGN(address+4))
1616                 address = PAGE_ALIGN(address+4);
1617         else {
1618                 anon_vma_unlock(vma);
1619                 return -ENOMEM;
1620         }
1621         error = 0;
1622
1623         /* Somebody else might have raced and expanded it already */
1624         if (address > vma->vm_end) {
1625                 unsigned long size, grow;
1626
1627                 size = address - vma->vm_start;
1628                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1629
1630                 error = acct_stack_growth(vma, size, grow);
1631                 if (!error)
1632                         vma->vm_end = address;
1633         }
1634         anon_vma_unlock(vma);
1635         return error;
1636 }
1637 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1638
1639 /*
1640  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1641  */
1642 static int expand_downwards(struct vm_area_struct *vma,
1643                                    unsigned long address)
1644 {
1645         int error;
1646
1647         /*
1648          * We must make sure the anon_vma is allocated
1649          * so that the anon_vma locking is not a noop.
1650          */
1651         if (unlikely(anon_vma_prepare(vma)))
1652                 return -ENOMEM;
1653
1654         address &= PAGE_MASK;
1655         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1656         if (error)
1657                 return error;
1658
1659         anon_vma_lock(vma);
1660
1661         /*
1662          * vma->vm_start/vm_end cannot change under us because the caller
1663          * is required to hold the mmap_sem in read mode.  We need the
1664          * anon_vma lock to serialize against concurrent expand_stacks.
1665          */
1666
1667         /* Somebody else might have raced and expanded it already */
1668         if (address < vma->vm_start) {
1669                 unsigned long size, grow;
1670
1671                 size = vma->vm_end - address;
1672                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1673
1674                 error = acct_stack_growth(vma, size, grow);
1675                 if (!error) {
1676                         vma->vm_start = address;
1677                         vma->vm_pgoff -= grow;
1678                 }
1679         }
1680         anon_vma_unlock(vma);
1681         return error;
1682 }
1683
1684 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1685 {
1686         return expand_downwards(vma, address);
1687 }
1688
1689 #ifdef CONFIG_STACK_GROWSUP
1690 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1691 {
1692         return expand_upwards(vma, address);
1693 }
1694
1695 struct vm_area_struct *
1696 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1697 {
1698         struct vm_area_struct *vma, *prev;
1699
1700         addr &= PAGE_MASK;
1701         vma = find_vma_prev(mm, addr, &prev);
1702         if (vma && (vma->vm_start <= addr))
1703                 return vma;
1704         if (!prev || expand_stack(prev, addr))
1705                 return NULL;
1706         if (prev->vm_flags & VM_LOCKED) {
1707                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1708                         return NULL;    /* vma gone! */
1709         }
1710         return prev;
1711 }
1712 #else
1713 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1714 {
1715         return expand_downwards(vma, address);
1716 }
1717
1718 struct vm_area_struct *
1719 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1720 {
1721         struct vm_area_struct * vma;
1722         unsigned long start;
1723
1724         addr &= PAGE_MASK;
1725         vma = find_vma(mm,addr);
1726         if (!vma)
1727                 return NULL;
1728         if (vma->vm_start <= addr)
1729                 return vma;
1730         if (!(vma->vm_flags & VM_GROWSDOWN))
1731                 return NULL;
1732         start = vma->vm_start;
1733         if (expand_stack(vma, addr))
1734                 return NULL;
1735         if (vma->vm_flags & VM_LOCKED) {
1736                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1737                         return NULL;    /* vma gone! */
1738         }
1739         return vma;
1740 }
1741 #endif
1742
1743 /*
1744  * Ok - we have the memory areas we should free on the vma list,
1745  * so release them, and do the vma updates.
1746  *
1747  * Called with the mm semaphore held.
1748  */
1749 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1750 {
1751         /* Update high watermark before we lower total_vm */
1752         update_hiwater_vm(mm);
1753         do {
1754                 long nrpages = vma_pages(vma);
1755
1756                 mm->total_vm -= nrpages;
1757                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1758                 vma = remove_vma(vma);
1759         } while (vma);
1760         validate_mm(mm);
1761 }
1762
1763 /*
1764  * Get rid of page table information in the indicated region.
1765  *
1766  * Called with the mm semaphore held.
1767  */
1768 static void unmap_region(struct mm_struct *mm,
1769                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1770                 unsigned long start, unsigned long end)
1771 {
1772         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1773         struct mmu_gather *tlb;
1774         unsigned long nr_accounted = 0;
1775
1776         lru_add_drain();
1777         tlb = tlb_gather_mmu(mm, 0);
1778         update_hiwater_rss(mm);
1779         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1780         vm_unacct_memory(nr_accounted);
1781         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1782                                  next? next->vm_start: 0);
1783         tlb_finish_mmu(tlb, start, end);
1784 }
1785
1786 /*
1787  * Create a list of vma's touched by the unmap, removing them from the mm's
1788  * vma list as we go..
1789  */
1790 static void
1791 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1792         struct vm_area_struct *prev, unsigned long end)
1793 {
1794         struct vm_area_struct **insertion_point;
1795         struct vm_area_struct *tail_vma = NULL;
1796         unsigned long addr;
1797
1798         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1799         do {
1800                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1801                 mm->map_count--;
1802                 tail_vma = vma;
1803                 vma = vma->vm_next;
1804         } while (vma && vma->vm_start < end);
1805         *insertion_point = vma;
1806         tail_vma->vm_next = NULL;
1807         if (mm->unmap_area == arch_unmap_area)
1808                 addr = prev ? prev->vm_end : mm->mmap_base;
1809         else
1810                 addr = vma ?  vma->vm_start : mm->mmap_base;
1811         mm->unmap_area(mm, addr);
1812         mm->mmap_cache = NULL;          /* Kill the cache. */
1813 }
1814
1815 /*
1816  * Split a vma into two pieces at address 'addr', a new vma is allocated
1817  * either for the first part or the tail.
1818  */
1819 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1820               unsigned long addr, int new_below)
1821 {
1822         struct mempolicy *pol;
1823         struct vm_area_struct *new;
1824
1825         if (is_vm_hugetlb_page(vma) && (addr &
1826                                         ~(huge_page_mask(hstate_vma(vma)))))
1827                 return -EINVAL;
1828
1829         if (mm->map_count >= sysctl_max_map_count)
1830                 return -ENOMEM;
1831
1832         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1833         if (!new)
1834                 return -ENOMEM;
1835
1836         /* most fields are the same, copy all, and then fixup */
1837         *new = *vma;
1838
1839         if (new_below)
1840                 new->vm_end = addr;
1841         else {
1842                 new->vm_start = addr;
1843                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1844         }
1845
1846         pol = mpol_dup(vma_policy(vma));
1847         if (IS_ERR(pol)) {
1848                 kmem_cache_free(vm_area_cachep, new);
1849                 return PTR_ERR(pol);
1850         }
1851         vma_set_policy(new, pol);
1852
1853         if (new->vm_file) {
1854                 get_file(new->vm_file);
1855                 if (vma->vm_flags & VM_EXECUTABLE)
1856                         added_exe_file_vma(mm);
1857         }
1858
1859         if (new->vm_ops && new->vm_ops->open)
1860                 new->vm_ops->open(new);
1861
1862         if (new_below)
1863                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1864                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1865         else
1866                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1867
1868         return 0;
1869 }
1870
1871 /* Munmap is split into 2 main parts -- this part which finds
1872  * what needs doing, and the areas themselves, which do the
1873  * work.  This now handles partial unmappings.
1874  * Jeremy Fitzhardinge <jeremy@goop.org>
1875  */
1876 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1877 {
1878         unsigned long end;
1879         struct vm_area_struct *vma, *prev, *last;
1880
1881         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1882                 return -EINVAL;
1883
1884         if ((len = PAGE_ALIGN(len)) == 0)
1885                 return -EINVAL;
1886
1887         /* Find the first overlapping VMA */
1888         vma = find_vma_prev(mm, start, &prev);
1889         if (!vma)
1890                 return 0;
1891         /* we have  start < vma->vm_end  */
1892
1893         /* if it doesn't overlap, we have nothing.. */
1894         end = start + len;
1895         if (vma->vm_start >= end)
1896                 return 0;
1897
1898         /*
1899          * If we need to split any vma, do it now to save pain later.
1900          *
1901          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1902          * unmapped vm_area_struct will remain in use: so lower split_vma
1903          * places tmp vma above, and higher split_vma places tmp vma below.
1904          */
1905         if (start > vma->vm_start) {
1906                 int error = split_vma(mm, vma, start, 0);
1907                 if (error)
1908                         return error;
1909                 prev = vma;
1910         }
1911
1912         /* Does it split the last one? */
1913         last = find_vma(mm, end);
1914         if (last && end > last->vm_start) {
1915                 int error = split_vma(mm, last, end, 1);
1916                 if (error)
1917                         return error;
1918         }
1919         vma = prev? prev->vm_next: mm->mmap;
1920
1921         /*
1922          * unlock any mlock()ed ranges before detaching vmas
1923          */
1924         if (mm->locked_vm) {
1925                 struct vm_area_struct *tmp = vma;
1926                 while (tmp && tmp->vm_start < end) {
1927                         if (tmp->vm_flags & VM_LOCKED) {
1928                                 mm->locked_vm -= vma_pages(tmp);
1929                                 munlock_vma_pages_all(tmp);
1930                         }
1931                         tmp = tmp->vm_next;
1932                 }
1933         }
1934
1935         /*
1936          * Remove the vma's, and unmap the actual pages
1937          */
1938         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1939         unmap_region(mm, vma, prev, start, end);
1940
1941         /* Fix up all other VM information */
1942         remove_vma_list(mm, vma);
1943
1944         return 0;
1945 }
1946
1947 EXPORT_SYMBOL(do_munmap);
1948
1949 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1950 {
1951         int ret;
1952         struct mm_struct *mm = current->mm;
1953
1954         profile_munmap(addr);
1955
1956         down_write(&mm->mmap_sem);
1957         ret = do_munmap(mm, addr, len);
1958         up_write(&mm->mmap_sem);
1959         return ret;
1960 }
1961
1962 static inline void verify_mm_writelocked(struct mm_struct *mm)
1963 {
1964 #ifdef CONFIG_DEBUG_VM
1965         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1966                 WARN_ON(1);
1967                 up_read(&mm->mmap_sem);
1968         }
1969 #endif
1970 }
1971
1972 /*
1973  *  this is really a simplified "do_mmap".  it only handles
1974  *  anonymous maps.  eventually we may be able to do some
1975  *  brk-specific accounting here.
1976  */
1977 unsigned long do_brk(unsigned long addr, unsigned long len)
1978 {
1979         struct mm_struct * mm = current->mm;
1980         struct vm_area_struct * vma, * prev;
1981         unsigned long flags;
1982         struct rb_node ** rb_link, * rb_parent;
1983         pgoff_t pgoff = addr >> PAGE_SHIFT;
1984         int error;
1985
1986         len = PAGE_ALIGN(len);
1987         if (!len)
1988                 return addr;
1989
1990         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1991                 return -EINVAL;
1992
1993         if (is_hugepage_only_range(mm, addr, len))
1994                 return -EINVAL;
1995
1996         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1997         if (error)
1998                 return error;
1999
2000         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2001
2002         error = arch_mmap_check(addr, len, flags);
2003         if (error)
2004                 return error;
2005
2006         /*
2007          * mlock MCL_FUTURE?
2008          */
2009         if (mm->def_flags & VM_LOCKED) {
2010                 unsigned long locked, lock_limit;
2011                 locked = len >> PAGE_SHIFT;
2012                 locked += mm->locked_vm;
2013                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2014                 lock_limit >>= PAGE_SHIFT;
2015                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2016                         return -EAGAIN;
2017         }
2018
2019         /*
2020          * mm->mmap_sem is required to protect against another thread
2021          * changing the mappings in case we sleep.
2022          */
2023         verify_mm_writelocked(mm);
2024
2025         /*
2026          * Clear old maps.  this also does some error checking for us
2027          */
2028  munmap_back:
2029         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2030         if (vma && vma->vm_start < addr + len) {
2031                 if (do_munmap(mm, addr, len))
2032                         return -ENOMEM;
2033                 goto munmap_back;
2034         }
2035
2036         /* Check against address space limits *after* clearing old maps... */
2037         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2038                 return -ENOMEM;
2039
2040         if (mm->map_count > sysctl_max_map_count)
2041                 return -ENOMEM;
2042
2043         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2044                 return -ENOMEM;
2045
2046         /* Can we just expand an old private anonymous mapping? */
2047         vma = vma_merge(mm, prev, addr, addr + len, flags,
2048                                         NULL, NULL, pgoff, NULL);
2049         if (vma)
2050                 goto out;
2051
2052         /*
2053          * create a vma struct for an anonymous mapping
2054          */
2055         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2056         if (!vma) {
2057                 vm_unacct_memory(len >> PAGE_SHIFT);
2058                 return -ENOMEM;
2059         }
2060
2061         vma->vm_mm = mm;
2062         vma->vm_start = addr;
2063         vma->vm_end = addr + len;
2064         vma->vm_pgoff = pgoff;
2065         vma->vm_flags = flags;
2066         vma->vm_page_prot = vm_get_page_prot(flags);
2067         vma_link(mm, vma, prev, rb_link, rb_parent);
2068 out:
2069         mm->total_vm += len >> PAGE_SHIFT;
2070         if (flags & VM_LOCKED) {
2071                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2072                         mm->locked_vm += (len >> PAGE_SHIFT);
2073         }
2074         return addr;
2075 }
2076
2077 EXPORT_SYMBOL(do_brk);
2078
2079 /* Release all mmaps. */
2080 void exit_mmap(struct mm_struct *mm)
2081 {
2082         struct mmu_gather *tlb;
2083         struct vm_area_struct *vma;
2084         unsigned long nr_accounted = 0;
2085         unsigned long end;
2086
2087         /* mm's last user has gone, and its about to be pulled down */
2088         mmu_notifier_release(mm);
2089
2090         if (mm->locked_vm) {
2091                 vma = mm->mmap;
2092                 while (vma) {
2093                         if (vma->vm_flags & VM_LOCKED)
2094                                 munlock_vma_pages_all(vma);
2095                         vma = vma->vm_next;
2096                 }
2097         }
2098
2099         arch_exit_mmap(mm);
2100
2101         vma = mm->mmap;
2102         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2103                 return;
2104
2105         lru_add_drain();
2106         flush_cache_mm(mm);
2107         tlb = tlb_gather_mmu(mm, 1);
2108         /* update_hiwater_rss(mm) here? but nobody should be looking */
2109         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2110         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2111         vm_unacct_memory(nr_accounted);
2112
2113         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2114         tlb_finish_mmu(tlb, 0, end);
2115
2116         /*
2117          * Walk the list again, actually closing and freeing it,
2118          * with preemption enabled, without holding any MM locks.
2119          */
2120         while (vma)
2121                 vma = remove_vma(vma);
2122
2123         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2124 }
2125
2126 /* Insert vm structure into process list sorted by address
2127  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2128  * then i_mmap_lock is taken here.
2129  */
2130 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2131 {
2132         struct vm_area_struct * __vma, * prev;
2133         struct rb_node ** rb_link, * rb_parent;
2134
2135         /*
2136          * The vm_pgoff of a purely anonymous vma should be irrelevant
2137          * until its first write fault, when page's anon_vma and index
2138          * are set.  But now set the vm_pgoff it will almost certainly
2139          * end up with (unless mremap moves it elsewhere before that
2140          * first wfault), so /proc/pid/maps tells a consistent story.
2141          *
2142          * By setting it to reflect the virtual start address of the
2143          * vma, merges and splits can happen in a seamless way, just
2144          * using the existing file pgoff checks and manipulations.
2145          * Similarly in do_mmap_pgoff and in do_brk.
2146          */
2147         if (!vma->vm_file) {
2148                 BUG_ON(vma->anon_vma);
2149                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2150         }
2151         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2152         if (__vma && __vma->vm_start < vma->vm_end)
2153                 return -ENOMEM;
2154         if ((vma->vm_flags & VM_ACCOUNT) &&
2155              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2156                 return -ENOMEM;
2157         vma_link(mm, vma, prev, rb_link, rb_parent);
2158         return 0;
2159 }
2160
2161 /*
2162  * Copy the vma structure to a new location in the same mm,
2163  * prior to moving page table entries, to effect an mremap move.
2164  */
2165 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2166         unsigned long addr, unsigned long len, pgoff_t pgoff)
2167 {
2168         struct vm_area_struct *vma = *vmap;
2169         unsigned long vma_start = vma->vm_start;
2170         struct mm_struct *mm = vma->vm_mm;
2171         struct vm_area_struct *new_vma, *prev;
2172         struct rb_node **rb_link, *rb_parent;
2173         struct mempolicy *pol;
2174
2175         /*
2176          * If anonymous vma has not yet been faulted, update new pgoff
2177          * to match new location, to increase its chance of merging.
2178          */
2179         if (!vma->vm_file && !vma->anon_vma)
2180                 pgoff = addr >> PAGE_SHIFT;
2181
2182         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2183         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2184                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2185         if (new_vma) {
2186                 /*
2187                  * Source vma may have been merged into new_vma
2188                  */
2189                 if (vma_start >= new_vma->vm_start &&
2190                     vma_start < new_vma->vm_end)
2191                         *vmap = new_vma;
2192         } else {
2193                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2194                 if (new_vma) {
2195                         *new_vma = *vma;
2196                         pol = mpol_dup(vma_policy(vma));
2197                         if (IS_ERR(pol)) {
2198                                 kmem_cache_free(vm_area_cachep, new_vma);
2199                                 return NULL;
2200                         }
2201                         vma_set_policy(new_vma, pol);
2202                         new_vma->vm_start = addr;
2203                         new_vma->vm_end = addr + len;
2204                         new_vma->vm_pgoff = pgoff;
2205                         if (new_vma->vm_file) {
2206                                 get_file(new_vma->vm_file);
2207                                 if (vma->vm_flags & VM_EXECUTABLE)
2208                                         added_exe_file_vma(mm);
2209                         }
2210                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2211                                 new_vma->vm_ops->open(new_vma);
2212                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2213                 }
2214         }
2215         return new_vma;
2216 }
2217
2218 /*
2219  * Return true if the calling process may expand its vm space by the passed
2220  * number of pages
2221  */
2222 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2223 {
2224         unsigned long cur = mm->total_vm;       /* pages */
2225         unsigned long lim;
2226
2227         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2228
2229         if (cur + npages > lim)
2230                 return 0;
2231         return 1;
2232 }
2233
2234
2235 static int special_mapping_fault(struct vm_area_struct *vma,
2236                                 struct vm_fault *vmf)
2237 {
2238         pgoff_t pgoff;
2239         struct page **pages;
2240
2241         /*
2242          * special mappings have no vm_file, and in that case, the mm
2243          * uses vm_pgoff internally. So we have to subtract it from here.
2244          * We are allowed to do this because we are the mm; do not copy
2245          * this code into drivers!
2246          */
2247         pgoff = vmf->pgoff - vma->vm_pgoff;
2248
2249         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2250                 pgoff--;
2251
2252         if (*pages) {
2253                 struct page *page = *pages;
2254                 get_page(page);
2255                 vmf->page = page;
2256                 return 0;
2257         }
2258
2259         return VM_FAULT_SIGBUS;
2260 }
2261
2262 /*
2263  * Having a close hook prevents vma merging regardless of flags.
2264  */
2265 static void special_mapping_close(struct vm_area_struct *vma)
2266 {
2267 }
2268
2269 static struct vm_operations_struct special_mapping_vmops = {
2270         .close = special_mapping_close,
2271         .fault = special_mapping_fault,
2272 };
2273
2274 /*
2275  * Called with mm->mmap_sem held for writing.
2276  * Insert a new vma covering the given region, with the given flags.
2277  * Its pages are supplied by the given array of struct page *.
2278  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2279  * The region past the last page supplied will always produce SIGBUS.
2280  * The array pointer and the pages it points to are assumed to stay alive
2281  * for as long as this mapping might exist.
2282  */
2283 int install_special_mapping(struct mm_struct *mm,
2284                             unsigned long addr, unsigned long len,
2285                             unsigned long vm_flags, struct page **pages)
2286 {
2287         struct vm_area_struct *vma;
2288
2289         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2290         if (unlikely(vma == NULL))
2291                 return -ENOMEM;
2292
2293         vma->vm_mm = mm;
2294         vma->vm_start = addr;
2295         vma->vm_end = addr + len;
2296
2297         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2298         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2299
2300         vma->vm_ops = &special_mapping_vmops;
2301         vma->vm_private_data = pages;
2302
2303         if (unlikely(insert_vm_struct(mm, vma))) {
2304                 kmem_cache_free(vm_area_cachep, vma);
2305                 return -ENOMEM;
2306         }
2307
2308         mm->total_vm += len >> PAGE_SHIFT;
2309
2310         perf_event_mmap(vma);
2311
2312         return 0;
2313 }
2314
2315 static DEFINE_MUTEX(mm_all_locks_mutex);
2316
2317 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2318 {
2319         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2320                 /*
2321                  * The LSB of head.next can't change from under us
2322                  * because we hold the mm_all_locks_mutex.
2323                  */
2324                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2325                 /*
2326                  * We can safely modify head.next after taking the
2327                  * anon_vma->lock. If some other vma in this mm shares
2328                  * the same anon_vma we won't take it again.
2329                  *
2330                  * No need of atomic instructions here, head.next
2331                  * can't change from under us thanks to the
2332                  * anon_vma->lock.
2333                  */
2334                 if (__test_and_set_bit(0, (unsigned long *)
2335                                        &anon_vma->head.next))
2336                         BUG();
2337         }
2338 }
2339
2340 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2341 {
2342         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2343                 /*
2344                  * AS_MM_ALL_LOCKS can't change from under us because
2345                  * we hold the mm_all_locks_mutex.
2346                  *
2347                  * Operations on ->flags have to be atomic because
2348                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2349                  * mm_all_locks_mutex, there may be other cpus
2350                  * changing other bitflags in parallel to us.
2351                  */
2352                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2353                         BUG();
2354                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2355         }
2356 }
2357
2358 /*
2359  * This operation locks against the VM for all pte/vma/mm related
2360  * operations that could ever happen on a certain mm. This includes
2361  * vmtruncate, try_to_unmap, and all page faults.
2362  *
2363  * The caller must take the mmap_sem in write mode before calling
2364  * mm_take_all_locks(). The caller isn't allowed to release the
2365  * mmap_sem until mm_drop_all_locks() returns.
2366  *
2367  * mmap_sem in write mode is required in order to block all operations
2368  * that could modify pagetables and free pages without need of
2369  * altering the vma layout (for example populate_range() with
2370  * nonlinear vmas). It's also needed in write mode to avoid new
2371  * anon_vmas to be associated with existing vmas.
2372  *
2373  * A single task can't take more than one mm_take_all_locks() in a row
2374  * or it would deadlock.
2375  *
2376  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2377  * mapping->flags avoid to take the same lock twice, if more than one
2378  * vma in this mm is backed by the same anon_vma or address_space.
2379  *
2380  * We can take all the locks in random order because the VM code
2381  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2382  * takes more than one of them in a row. Secondly we're protected
2383  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2384  *
2385  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2386  * that may have to take thousand of locks.
2387  *
2388  * mm_take_all_locks() can fail if it's interrupted by signals.
2389  */
2390 int mm_take_all_locks(struct mm_struct *mm)
2391 {
2392         struct vm_area_struct *vma;
2393         int ret = -EINTR;
2394
2395         BUG_ON(down_read_trylock(&mm->mmap_sem));
2396
2397         mutex_lock(&mm_all_locks_mutex);
2398
2399         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2400                 if (signal_pending(current))
2401                         goto out_unlock;
2402                 if (vma->vm_file && vma->vm_file->f_mapping)
2403                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2404         }
2405
2406         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2407                 if (signal_pending(current))
2408                         goto out_unlock;
2409                 if (vma->anon_vma)
2410                         vm_lock_anon_vma(mm, vma->anon_vma);
2411         }
2412
2413         ret = 0;
2414
2415 out_unlock:
2416         if (ret)
2417                 mm_drop_all_locks(mm);
2418
2419         return ret;
2420 }
2421
2422 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2423 {
2424         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2425                 /*
2426                  * The LSB of head.next can't change to 0 from under
2427                  * us because we hold the mm_all_locks_mutex.
2428                  *
2429                  * We must however clear the bitflag before unlocking
2430                  * the vma so the users using the anon_vma->head will
2431                  * never see our bitflag.
2432                  *
2433                  * No need of atomic instructions here, head.next
2434                  * can't change from under us until we release the
2435                  * anon_vma->lock.
2436                  */
2437                 if (!__test_and_clear_bit(0, (unsigned long *)
2438                                           &anon_vma->head.next))
2439                         BUG();
2440                 spin_unlock(&anon_vma->lock);
2441         }
2442 }
2443
2444 static void vm_unlock_mapping(struct address_space *mapping)
2445 {
2446         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2447                 /*
2448                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2449                  * because we hold the mm_all_locks_mutex.
2450                  */
2451                 spin_unlock(&mapping->i_mmap_lock);
2452                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2453                                         &mapping->flags))
2454                         BUG();
2455         }
2456 }
2457
2458 /*
2459  * The mmap_sem cannot be released by the caller until
2460  * mm_drop_all_locks() returns.
2461  */
2462 void mm_drop_all_locks(struct mm_struct *mm)
2463 {
2464         struct vm_area_struct *vma;
2465
2466         BUG_ON(down_read_trylock(&mm->mmap_sem));
2467         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2468
2469         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2470                 if (vma->anon_vma)
2471                         vm_unlock_anon_vma(vma->anon_vma);
2472                 if (vma->vm_file && vma->vm_file->f_mapping)
2473                         vm_unlock_mapping(vma->vm_file->f_mapping);
2474         }
2475
2476         mutex_unlock(&mm_all_locks_mutex);
2477 }
2478
2479 /*
2480  * initialise the VMA slab
2481  */
2482 void __init mmap_init(void)
2483 {
2484         int ret;
2485
2486         ret = percpu_counter_init(&vm_committed_as, 0);
2487         VM_BUG_ON(ret);
2488 }