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