Merge commit 'v2.6.27-rc6' into x86/pat
[linux-2.6] / 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                         /*
1034                          * Ignore pgoff.
1035                          */
1036                         pgoff = 0;
1037                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1038                         break;
1039                 case MAP_PRIVATE:
1040                         /*
1041                          * Set pgoff according to addr for anon_vma.
1042                          */
1043                         pgoff = addr >> PAGE_SHIFT;
1044                         break;
1045                 default:
1046                         return -EINVAL;
1047                 }
1048         }
1049
1050         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1051         if (error)
1052                 return error;
1053
1054         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1055                            accountable);
1056 }
1057 EXPORT_SYMBOL(do_mmap_pgoff);
1058
1059 /*
1060  * Some shared mappigns will want the pages marked read-only
1061  * to track write events. If so, we'll downgrade vm_page_prot
1062  * to the private version (using protection_map[] without the
1063  * VM_SHARED bit).
1064  */
1065 int vma_wants_writenotify(struct vm_area_struct *vma)
1066 {
1067         unsigned int vm_flags = vma->vm_flags;
1068
1069         /* If it was private or non-writable, the write bit is already clear */
1070         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1071                 return 0;
1072
1073         /* The backer wishes to know when pages are first written to? */
1074         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1075                 return 1;
1076
1077         /* The open routine did something to the protections already? */
1078         if (pgprot_val(vma->vm_page_prot) !=
1079             pgprot_val(vm_get_page_prot(vm_flags)))
1080                 return 0;
1081
1082         /* Specialty mapping? */
1083         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1084                 return 0;
1085
1086         /* Can the mapping track the dirty pages? */
1087         return vma->vm_file && vma->vm_file->f_mapping &&
1088                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1089 }
1090
1091 unsigned long mmap_region(struct file *file, unsigned long addr,
1092                           unsigned long len, unsigned long flags,
1093                           unsigned int vm_flags, unsigned long pgoff,
1094                           int accountable)
1095 {
1096         struct mm_struct *mm = current->mm;
1097         struct vm_area_struct *vma, *prev;
1098         int correct_wcount = 0;
1099         int error;
1100         struct rb_node **rb_link, *rb_parent;
1101         unsigned long charged = 0;
1102         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1103
1104         /* Clear old maps */
1105         error = -ENOMEM;
1106 munmap_back:
1107         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1108         if (vma && vma->vm_start < addr + len) {
1109                 if (do_munmap(mm, addr, len))
1110                         return -ENOMEM;
1111                 goto munmap_back;
1112         }
1113
1114         /* Check against address space limit. */
1115         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1116                 return -ENOMEM;
1117
1118         if (flags & MAP_NORESERVE)
1119                 vm_flags |= VM_NORESERVE;
1120
1121         if (accountable && (!(flags & MAP_NORESERVE) ||
1122                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1123                 if (vm_flags & VM_SHARED) {
1124                         /* Check memory availability in shmem_file_setup? */
1125                         vm_flags |= VM_ACCOUNT;
1126                 } else if (vm_flags & VM_WRITE) {
1127                         /*
1128                          * Private writable mapping: check memory availability
1129                          */
1130                         charged = len >> PAGE_SHIFT;
1131                         if (security_vm_enough_memory(charged))
1132                                 return -ENOMEM;
1133                         vm_flags |= VM_ACCOUNT;
1134                 }
1135         }
1136
1137         /*
1138          * Can we just expand an old private anonymous mapping?
1139          * The VM_SHARED test is necessary because shmem_zero_setup
1140          * will create the file object for a shared anonymous map below.
1141          */
1142         if (!file && !(vm_flags & VM_SHARED) &&
1143             vma_merge(mm, prev, addr, addr + len, vm_flags,
1144                                         NULL, NULL, pgoff, NULL))
1145                 goto out;
1146
1147         /*
1148          * Determine the object being mapped and call the appropriate
1149          * specific mapper. the address has already been validated, but
1150          * not unmapped, but the maps are removed from the list.
1151          */
1152         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1153         if (!vma) {
1154                 error = -ENOMEM;
1155                 goto unacct_error;
1156         }
1157
1158         vma->vm_mm = mm;
1159         vma->vm_start = addr;
1160         vma->vm_end = addr + len;
1161         vma->vm_flags = vm_flags;
1162         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1163         vma->vm_pgoff = pgoff;
1164
1165         if (file) {
1166                 error = -EINVAL;
1167                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1168                         goto free_vma;
1169                 if (vm_flags & VM_DENYWRITE) {
1170                         error = deny_write_access(file);
1171                         if (error)
1172                                 goto free_vma;
1173                         correct_wcount = 1;
1174                 }
1175                 vma->vm_file = file;
1176                 get_file(file);
1177                 error = file->f_op->mmap(file, vma);
1178                 if (error)
1179                         goto unmap_and_free_vma;
1180                 if (vm_flags & VM_EXECUTABLE)
1181                         added_exe_file_vma(mm);
1182         } else if (vm_flags & VM_SHARED) {
1183                 error = shmem_zero_setup(vma);
1184                 if (error)
1185                         goto free_vma;
1186         }
1187
1188         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1189          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1190          * that memory reservation must be checked; but that reservation
1191          * belongs to shared memory object, not to vma: so now clear it.
1192          */
1193         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1194                 vma->vm_flags &= ~VM_ACCOUNT;
1195
1196         /* Can addr have changed??
1197          *
1198          * Answer: Yes, several device drivers can do it in their
1199          *         f_op->mmap method. -DaveM
1200          */
1201         addr = vma->vm_start;
1202         pgoff = vma->vm_pgoff;
1203         vm_flags = vma->vm_flags;
1204
1205         if (vma_wants_writenotify(vma))
1206                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1207
1208         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1209                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1210                 mpol_put(vma_policy(vma));
1211                 kmem_cache_free(vm_area_cachep, vma);
1212                 fput(file);
1213                 if (vm_flags & VM_EXECUTABLE)
1214                         removed_exe_file_vma(mm);
1215         } else {
1216                 vma_link(mm, vma, prev, rb_link, rb_parent);
1217                 file = vma->vm_file;
1218         }
1219
1220         /* Once vma denies write, undo our temporary denial count */
1221         if (correct_wcount)
1222                 atomic_inc(&inode->i_writecount);
1223 out:
1224         mm->total_vm += len >> PAGE_SHIFT;
1225         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1226         if (vm_flags & VM_LOCKED) {
1227                 mm->locked_vm += len >> PAGE_SHIFT;
1228                 make_pages_present(addr, addr + len);
1229         }
1230         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1231                 make_pages_present(addr, addr + len);
1232         return addr;
1233
1234 unmap_and_free_vma:
1235         if (correct_wcount)
1236                 atomic_inc(&inode->i_writecount);
1237         vma->vm_file = NULL;
1238         fput(file);
1239
1240         /* Undo any partial mapping done by a device driver. */
1241         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1242         charged = 0;
1243 free_vma:
1244         kmem_cache_free(vm_area_cachep, vma);
1245 unacct_error:
1246         if (charged)
1247                 vm_unacct_memory(charged);
1248         return error;
1249 }
1250
1251 /* Get an address range which is currently unmapped.
1252  * For shmat() with addr=0.
1253  *
1254  * Ugly calling convention alert:
1255  * Return value with the low bits set means error value,
1256  * ie
1257  *      if (ret & ~PAGE_MASK)
1258  *              error = ret;
1259  *
1260  * This function "knows" that -ENOMEM has the bits set.
1261  */
1262 #ifndef HAVE_ARCH_UNMAPPED_AREA
1263 unsigned long
1264 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1265                 unsigned long len, unsigned long pgoff, unsigned long flags)
1266 {
1267         struct mm_struct *mm = current->mm;
1268         struct vm_area_struct *vma;
1269         unsigned long start_addr;
1270
1271         if (len > TASK_SIZE)
1272                 return -ENOMEM;
1273
1274         if (flags & MAP_FIXED)
1275                 return addr;
1276
1277         if (addr) {
1278                 addr = PAGE_ALIGN(addr);
1279                 vma = find_vma(mm, addr);
1280                 if (TASK_SIZE - len >= addr &&
1281                     (!vma || addr + len <= vma->vm_start))
1282                         return addr;
1283         }
1284         if (len > mm->cached_hole_size) {
1285                 start_addr = addr = mm->free_area_cache;
1286         } else {
1287                 start_addr = addr = TASK_UNMAPPED_BASE;
1288                 mm->cached_hole_size = 0;
1289         }
1290
1291 full_search:
1292         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1293                 /* At this point:  (!vma || addr < vma->vm_end). */
1294                 if (TASK_SIZE - len < addr) {
1295                         /*
1296                          * Start a new search - just in case we missed
1297                          * some holes.
1298                          */
1299                         if (start_addr != TASK_UNMAPPED_BASE) {
1300                                 addr = TASK_UNMAPPED_BASE;
1301                                 start_addr = addr;
1302                                 mm->cached_hole_size = 0;
1303                                 goto full_search;
1304                         }
1305                         return -ENOMEM;
1306                 }
1307                 if (!vma || addr + len <= vma->vm_start) {
1308                         /*
1309                          * Remember the place where we stopped the search:
1310                          */
1311                         mm->free_area_cache = addr + len;
1312                         return addr;
1313                 }
1314                 if (addr + mm->cached_hole_size < vma->vm_start)
1315                         mm->cached_hole_size = vma->vm_start - addr;
1316                 addr = vma->vm_end;
1317         }
1318 }
1319 #endif  
1320
1321 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1322 {
1323         /*
1324          * Is this a new hole at the lowest possible address?
1325          */
1326         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1327                 mm->free_area_cache = addr;
1328                 mm->cached_hole_size = ~0UL;
1329         }
1330 }
1331
1332 /*
1333  * This mmap-allocator allocates new areas top-down from below the
1334  * stack's low limit (the base):
1335  */
1336 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1337 unsigned long
1338 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1339                           const unsigned long len, const unsigned long pgoff,
1340                           const unsigned long flags)
1341 {
1342         struct vm_area_struct *vma;
1343         struct mm_struct *mm = current->mm;
1344         unsigned long addr = addr0;
1345
1346         /* requested length too big for entire address space */
1347         if (len > TASK_SIZE)
1348                 return -ENOMEM;
1349
1350         if (flags & MAP_FIXED)
1351                 return addr;
1352
1353         /* requesting a specific address */
1354         if (addr) {
1355                 addr = PAGE_ALIGN(addr);
1356                 vma = find_vma(mm, addr);
1357                 if (TASK_SIZE - len >= addr &&
1358                                 (!vma || addr + len <= vma->vm_start))
1359                         return addr;
1360         }
1361
1362         /* check if free_area_cache is useful for us */
1363         if (len <= mm->cached_hole_size) {
1364                 mm->cached_hole_size = 0;
1365                 mm->free_area_cache = mm->mmap_base;
1366         }
1367
1368         /* either no address requested or can't fit in requested address hole */
1369         addr = mm->free_area_cache;
1370
1371         /* make sure it can fit in the remaining address space */
1372         if (addr > len) {
1373                 vma = find_vma(mm, addr-len);
1374                 if (!vma || addr <= vma->vm_start)
1375                         /* remember the address as a hint for next time */
1376                         return (mm->free_area_cache = addr-len);
1377         }
1378
1379         if (mm->mmap_base < len)
1380                 goto bottomup;
1381
1382         addr = mm->mmap_base-len;
1383
1384         do {
1385                 /*
1386                  * Lookup failure means no vma is above this address,
1387                  * else if new region fits below vma->vm_start,
1388                  * return with success:
1389                  */
1390                 vma = find_vma(mm, addr);
1391                 if (!vma || addr+len <= vma->vm_start)
1392                         /* remember the address as a hint for next time */
1393                         return (mm->free_area_cache = addr);
1394
1395                 /* remember the largest hole we saw so far */
1396                 if (addr + mm->cached_hole_size < vma->vm_start)
1397                         mm->cached_hole_size = vma->vm_start - addr;
1398
1399                 /* try just below the current vma->vm_start */
1400                 addr = vma->vm_start-len;
1401         } while (len < vma->vm_start);
1402
1403 bottomup:
1404         /*
1405          * A failed mmap() very likely causes application failure,
1406          * so fall back to the bottom-up function here. This scenario
1407          * can happen with large stack limits and large mmap()
1408          * allocations.
1409          */
1410         mm->cached_hole_size = ~0UL;
1411         mm->free_area_cache = TASK_UNMAPPED_BASE;
1412         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1413         /*
1414          * Restore the topdown base:
1415          */
1416         mm->free_area_cache = mm->mmap_base;
1417         mm->cached_hole_size = ~0UL;
1418
1419         return addr;
1420 }
1421 #endif
1422
1423 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1424 {
1425         /*
1426          * Is this a new hole at the highest possible address?
1427          */
1428         if (addr > mm->free_area_cache)
1429                 mm->free_area_cache = addr;
1430
1431         /* dont allow allocations above current base */
1432         if (mm->free_area_cache > mm->mmap_base)
1433                 mm->free_area_cache = mm->mmap_base;
1434 }
1435
1436 unsigned long
1437 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1438                 unsigned long pgoff, unsigned long flags)
1439 {
1440         unsigned long (*get_area)(struct file *, unsigned long,
1441                                   unsigned long, unsigned long, unsigned long);
1442
1443         get_area = current->mm->get_unmapped_area;
1444         if (file && file->f_op && file->f_op->get_unmapped_area)
1445                 get_area = file->f_op->get_unmapped_area;
1446         addr = get_area(file, addr, len, pgoff, flags);
1447         if (IS_ERR_VALUE(addr))
1448                 return addr;
1449
1450         if (addr > TASK_SIZE - len)
1451                 return -ENOMEM;
1452         if (addr & ~PAGE_MASK)
1453                 return -EINVAL;
1454
1455         return arch_rebalance_pgtables(addr, len);
1456 }
1457
1458 EXPORT_SYMBOL(get_unmapped_area);
1459
1460 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1461 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1462 {
1463         struct vm_area_struct *vma = NULL;
1464
1465         if (mm) {
1466                 /* Check the cache first. */
1467                 /* (Cache hit rate is typically around 35%.) */
1468                 vma = mm->mmap_cache;
1469                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1470                         struct rb_node * rb_node;
1471
1472                         rb_node = mm->mm_rb.rb_node;
1473                         vma = NULL;
1474
1475                         while (rb_node) {
1476                                 struct vm_area_struct * vma_tmp;
1477
1478                                 vma_tmp = rb_entry(rb_node,
1479                                                 struct vm_area_struct, vm_rb);
1480
1481                                 if (vma_tmp->vm_end > addr) {
1482                                         vma = vma_tmp;
1483                                         if (vma_tmp->vm_start <= addr)
1484                                                 break;
1485                                         rb_node = rb_node->rb_left;
1486                                 } else
1487                                         rb_node = rb_node->rb_right;
1488                         }
1489                         if (vma)
1490                                 mm->mmap_cache = vma;
1491                 }
1492         }
1493         return vma;
1494 }
1495
1496 EXPORT_SYMBOL(find_vma);
1497
1498 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1499 struct vm_area_struct *
1500 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1501                         struct vm_area_struct **pprev)
1502 {
1503         struct vm_area_struct *vma = NULL, *prev = NULL;
1504         struct rb_node * rb_node;
1505         if (!mm)
1506                 goto out;
1507
1508         /* Guard against addr being lower than the first VMA */
1509         vma = mm->mmap;
1510
1511         /* Go through the RB tree quickly. */
1512         rb_node = mm->mm_rb.rb_node;
1513
1514         while (rb_node) {
1515                 struct vm_area_struct *vma_tmp;
1516                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1517
1518                 if (addr < vma_tmp->vm_end) {
1519                         rb_node = rb_node->rb_left;
1520                 } else {
1521                         prev = vma_tmp;
1522                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1523                                 break;
1524                         rb_node = rb_node->rb_right;
1525                 }
1526         }
1527
1528 out:
1529         *pprev = prev;
1530         return prev ? prev->vm_next : vma;
1531 }
1532
1533 /*
1534  * Verify that the stack growth is acceptable and
1535  * update accounting. This is shared with both the
1536  * grow-up and grow-down cases.
1537  */
1538 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1539 {
1540         struct mm_struct *mm = vma->vm_mm;
1541         struct rlimit *rlim = current->signal->rlim;
1542         unsigned long new_start;
1543
1544         /* address space limit tests */
1545         if (!may_expand_vm(mm, grow))
1546                 return -ENOMEM;
1547
1548         /* Stack limit test */
1549         if (size > rlim[RLIMIT_STACK].rlim_cur)
1550                 return -ENOMEM;
1551
1552         /* mlock limit tests */
1553         if (vma->vm_flags & VM_LOCKED) {
1554                 unsigned long locked;
1555                 unsigned long limit;
1556                 locked = mm->locked_vm + grow;
1557                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1558                 if (locked > limit && !capable(CAP_IPC_LOCK))
1559                         return -ENOMEM;
1560         }
1561
1562         /* Check to ensure the stack will not grow into a hugetlb-only region */
1563         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1564                         vma->vm_end - size;
1565         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1566                 return -EFAULT;
1567
1568         /*
1569          * Overcommit..  This must be the final test, as it will
1570          * update security statistics.
1571          */
1572         if (security_vm_enough_memory(grow))
1573                 return -ENOMEM;
1574
1575         /* Ok, everything looks good - let it rip */
1576         mm->total_vm += grow;
1577         if (vma->vm_flags & VM_LOCKED)
1578                 mm->locked_vm += grow;
1579         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1580         return 0;
1581 }
1582
1583 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1584 /*
1585  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1586  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1587  */
1588 #ifndef CONFIG_IA64
1589 static inline
1590 #endif
1591 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1592 {
1593         int error;
1594
1595         if (!(vma->vm_flags & VM_GROWSUP))
1596                 return -EFAULT;
1597
1598         /*
1599          * We must make sure the anon_vma is allocated
1600          * so that the anon_vma locking is not a noop.
1601          */
1602         if (unlikely(anon_vma_prepare(vma)))
1603                 return -ENOMEM;
1604         anon_vma_lock(vma);
1605
1606         /*
1607          * vma->vm_start/vm_end cannot change under us because the caller
1608          * is required to hold the mmap_sem in read mode.  We need the
1609          * anon_vma lock to serialize against concurrent expand_stacks.
1610          * Also guard against wrapping around to address 0.
1611          */
1612         if (address < PAGE_ALIGN(address+4))
1613                 address = PAGE_ALIGN(address+4);
1614         else {
1615                 anon_vma_unlock(vma);
1616                 return -ENOMEM;
1617         }
1618         error = 0;
1619
1620         /* Somebody else might have raced and expanded it already */
1621         if (address > vma->vm_end) {
1622                 unsigned long size, grow;
1623
1624                 size = address - vma->vm_start;
1625                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1626
1627                 error = acct_stack_growth(vma, size, grow);
1628                 if (!error)
1629                         vma->vm_end = address;
1630         }
1631         anon_vma_unlock(vma);
1632         return error;
1633 }
1634 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1635
1636 /*
1637  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1638  */
1639 static inline int expand_downwards(struct vm_area_struct *vma,
1640                                    unsigned long address)
1641 {
1642         int error;
1643
1644         /*
1645          * We must make sure the anon_vma is allocated
1646          * so that the anon_vma locking is not a noop.
1647          */
1648         if (unlikely(anon_vma_prepare(vma)))
1649                 return -ENOMEM;
1650
1651         address &= PAGE_MASK;
1652         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1653         if (error)
1654                 return error;
1655
1656         anon_vma_lock(vma);
1657
1658         /*
1659          * vma->vm_start/vm_end cannot change under us because the caller
1660          * is required to hold the mmap_sem in read mode.  We need the
1661          * anon_vma lock to serialize against concurrent expand_stacks.
1662          */
1663
1664         /* Somebody else might have raced and expanded it already */
1665         if (address < vma->vm_start) {
1666                 unsigned long size, grow;
1667
1668                 size = vma->vm_end - address;
1669                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1670
1671                 error = acct_stack_growth(vma, size, grow);
1672                 if (!error) {
1673                         vma->vm_start = address;
1674                         vma->vm_pgoff -= grow;
1675                 }
1676         }
1677         anon_vma_unlock(vma);
1678         return error;
1679 }
1680
1681 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1682 {
1683         return expand_downwards(vma, address);
1684 }
1685
1686 #ifdef CONFIG_STACK_GROWSUP
1687 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1688 {
1689         return expand_upwards(vma, address);
1690 }
1691
1692 struct vm_area_struct *
1693 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1694 {
1695         struct vm_area_struct *vma, *prev;
1696
1697         addr &= PAGE_MASK;
1698         vma = find_vma_prev(mm, addr, &prev);
1699         if (vma && (vma->vm_start <= addr))
1700                 return vma;
1701         if (!prev || expand_stack(prev, addr))
1702                 return NULL;
1703         if (prev->vm_flags & VM_LOCKED)
1704                 make_pages_present(addr, prev->vm_end);
1705         return prev;
1706 }
1707 #else
1708 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1709 {
1710         return expand_downwards(vma, address);
1711 }
1712
1713 struct vm_area_struct *
1714 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1715 {
1716         struct vm_area_struct * vma;
1717         unsigned long start;
1718
1719         addr &= PAGE_MASK;
1720         vma = find_vma(mm,addr);
1721         if (!vma)
1722                 return NULL;
1723         if (vma->vm_start <= addr)
1724                 return vma;
1725         if (!(vma->vm_flags & VM_GROWSDOWN))
1726                 return NULL;
1727         start = vma->vm_start;
1728         if (expand_stack(vma, addr))
1729                 return NULL;
1730         if (vma->vm_flags & VM_LOCKED)
1731                 make_pages_present(addr, start);
1732         return vma;
1733 }
1734 #endif
1735
1736 /*
1737  * Ok - we have the memory areas we should free on the vma list,
1738  * so release them, and do the vma updates.
1739  *
1740  * Called with the mm semaphore held.
1741  */
1742 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1743 {
1744         /* Update high watermark before we lower total_vm */
1745         update_hiwater_vm(mm);
1746         do {
1747                 long nrpages = vma_pages(vma);
1748
1749                 mm->total_vm -= nrpages;
1750                 if (vma->vm_flags & VM_LOCKED)
1751                         mm->locked_vm -= nrpages;
1752                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1753                 vma = remove_vma(vma);
1754         } while (vma);
1755         validate_mm(mm);
1756 }
1757
1758 /*
1759  * Get rid of page table information in the indicated region.
1760  *
1761  * Called with the mm semaphore held.
1762  */
1763 static void unmap_region(struct mm_struct *mm,
1764                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1765                 unsigned long start, unsigned long end)
1766 {
1767         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1768         struct mmu_gather *tlb;
1769         unsigned long nr_accounted = 0;
1770
1771         lru_add_drain();
1772         tlb = tlb_gather_mmu(mm, 0);
1773         update_hiwater_rss(mm);
1774         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1775         vm_unacct_memory(nr_accounted);
1776         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1777                                  next? next->vm_start: 0);
1778         tlb_finish_mmu(tlb, start, end);
1779 }
1780
1781 /*
1782  * Create a list of vma's touched by the unmap, removing them from the mm's
1783  * vma list as we go..
1784  */
1785 static void
1786 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1787         struct vm_area_struct *prev, unsigned long end)
1788 {
1789         struct vm_area_struct **insertion_point;
1790         struct vm_area_struct *tail_vma = NULL;
1791         unsigned long addr;
1792
1793         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1794         do {
1795                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1796                 mm->map_count--;
1797                 tail_vma = vma;
1798                 vma = vma->vm_next;
1799         } while (vma && vma->vm_start < end);
1800         *insertion_point = vma;
1801         tail_vma->vm_next = NULL;
1802         if (mm->unmap_area == arch_unmap_area)
1803                 addr = prev ? prev->vm_end : mm->mmap_base;
1804         else
1805                 addr = vma ?  vma->vm_start : mm->mmap_base;
1806         mm->unmap_area(mm, addr);
1807         mm->mmap_cache = NULL;          /* Kill the cache. */
1808 }
1809
1810 /*
1811  * Split a vma into two pieces at address 'addr', a new vma is allocated
1812  * either for the first part or the tail.
1813  */
1814 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1815               unsigned long addr, int new_below)
1816 {
1817         struct mempolicy *pol;
1818         struct vm_area_struct *new;
1819
1820         if (is_vm_hugetlb_page(vma) && (addr &
1821                                         ~(huge_page_mask(hstate_vma(vma)))))
1822                 return -EINVAL;
1823
1824         if (mm->map_count >= sysctl_max_map_count)
1825                 return -ENOMEM;
1826
1827         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1828         if (!new)
1829                 return -ENOMEM;
1830
1831         /* most fields are the same, copy all, and then fixup */
1832         *new = *vma;
1833
1834         if (new_below)
1835                 new->vm_end = addr;
1836         else {
1837                 new->vm_start = addr;
1838                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1839         }
1840
1841         pol = mpol_dup(vma_policy(vma));
1842         if (IS_ERR(pol)) {
1843                 kmem_cache_free(vm_area_cachep, new);
1844                 return PTR_ERR(pol);
1845         }
1846         vma_set_policy(new, pol);
1847
1848         if (new->vm_file) {
1849                 get_file(new->vm_file);
1850                 if (vma->vm_flags & VM_EXECUTABLE)
1851                         added_exe_file_vma(mm);
1852         }
1853
1854         if (new->vm_ops && new->vm_ops->open)
1855                 new->vm_ops->open(new);
1856
1857         if (new_below)
1858                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1859                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1860         else
1861                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1862
1863         return 0;
1864 }
1865
1866 /* Munmap is split into 2 main parts -- this part which finds
1867  * what needs doing, and the areas themselves, which do the
1868  * work.  This now handles partial unmappings.
1869  * Jeremy Fitzhardinge <jeremy@goop.org>
1870  */
1871 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1872 {
1873         unsigned long end;
1874         struct vm_area_struct *vma, *prev, *last;
1875
1876         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1877                 return -EINVAL;
1878
1879         if ((len = PAGE_ALIGN(len)) == 0)
1880                 return -EINVAL;
1881
1882         /* Find the first overlapping VMA */
1883         vma = find_vma_prev(mm, start, &prev);
1884         if (!vma)
1885                 return 0;
1886         /* we have  start < vma->vm_end  */
1887
1888         /* if it doesn't overlap, we have nothing.. */
1889         end = start + len;
1890         if (vma->vm_start >= end)
1891                 return 0;
1892
1893         /*
1894          * If we need to split any vma, do it now to save pain later.
1895          *
1896          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1897          * unmapped vm_area_struct will remain in use: so lower split_vma
1898          * places tmp vma above, and higher split_vma places tmp vma below.
1899          */
1900         if (start > vma->vm_start) {
1901                 int error = split_vma(mm, vma, start, 0);
1902                 if (error)
1903                         return error;
1904                 prev = vma;
1905         }
1906
1907         /* Does it split the last one? */
1908         last = find_vma(mm, end);
1909         if (last && end > last->vm_start) {
1910                 int error = split_vma(mm, last, end, 1);
1911                 if (error)
1912                         return error;
1913         }
1914         vma = prev? prev->vm_next: mm->mmap;
1915
1916         /*
1917          * Remove the vma's, and unmap the actual pages
1918          */
1919         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1920         unmap_region(mm, vma, prev, start, end);
1921
1922         /* Fix up all other VM information */
1923         remove_vma_list(mm, vma);
1924
1925         return 0;
1926 }
1927
1928 EXPORT_SYMBOL(do_munmap);
1929
1930 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1931 {
1932         int ret;
1933         struct mm_struct *mm = current->mm;
1934
1935         profile_munmap(addr);
1936
1937         down_write(&mm->mmap_sem);
1938         ret = do_munmap(mm, addr, len);
1939         up_write(&mm->mmap_sem);
1940         return ret;
1941 }
1942
1943 static inline void verify_mm_writelocked(struct mm_struct *mm)
1944 {
1945 #ifdef CONFIG_DEBUG_VM
1946         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1947                 WARN_ON(1);
1948                 up_read(&mm->mmap_sem);
1949         }
1950 #endif
1951 }
1952
1953 /*
1954  *  this is really a simplified "do_mmap".  it only handles
1955  *  anonymous maps.  eventually we may be able to do some
1956  *  brk-specific accounting here.
1957  */
1958 unsigned long do_brk(unsigned long addr, unsigned long len)
1959 {
1960         struct mm_struct * mm = current->mm;
1961         struct vm_area_struct * vma, * prev;
1962         unsigned long flags;
1963         struct rb_node ** rb_link, * rb_parent;
1964         pgoff_t pgoff = addr >> PAGE_SHIFT;
1965         int error;
1966
1967         len = PAGE_ALIGN(len);
1968         if (!len)
1969                 return addr;
1970
1971         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1972                 return -EINVAL;
1973
1974         if (is_hugepage_only_range(mm, addr, len))
1975                 return -EINVAL;
1976
1977         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1978         if (error)
1979                 return error;
1980
1981         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1982
1983         error = arch_mmap_check(addr, len, flags);
1984         if (error)
1985                 return error;
1986
1987         /*
1988          * mlock MCL_FUTURE?
1989          */
1990         if (mm->def_flags & VM_LOCKED) {
1991                 unsigned long locked, lock_limit;
1992                 locked = len >> PAGE_SHIFT;
1993                 locked += mm->locked_vm;
1994                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1995                 lock_limit >>= PAGE_SHIFT;
1996                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1997                         return -EAGAIN;
1998         }
1999
2000         /*
2001          * mm->mmap_sem is required to protect against another thread
2002          * changing the mappings in case we sleep.
2003          */
2004         verify_mm_writelocked(mm);
2005
2006         /*
2007          * Clear old maps.  this also does some error checking for us
2008          */
2009  munmap_back:
2010         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2011         if (vma && vma->vm_start < addr + len) {
2012                 if (do_munmap(mm, addr, len))
2013                         return -ENOMEM;
2014                 goto munmap_back;
2015         }
2016
2017         /* Check against address space limits *after* clearing old maps... */
2018         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2019                 return -ENOMEM;
2020
2021         if (mm->map_count > sysctl_max_map_count)
2022                 return -ENOMEM;
2023
2024         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2025                 return -ENOMEM;
2026
2027         /* Can we just expand an old private anonymous mapping? */
2028         if (vma_merge(mm, prev, addr, addr + len, flags,
2029                                         NULL, NULL, pgoff, NULL))
2030                 goto out;
2031
2032         /*
2033          * create a vma struct for an anonymous mapping
2034          */
2035         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2036         if (!vma) {
2037                 vm_unacct_memory(len >> PAGE_SHIFT);
2038                 return -ENOMEM;
2039         }
2040
2041         vma->vm_mm = mm;
2042         vma->vm_start = addr;
2043         vma->vm_end = addr + len;
2044         vma->vm_pgoff = pgoff;
2045         vma->vm_flags = flags;
2046         vma->vm_page_prot = vm_get_page_prot(flags);
2047         vma_link(mm, vma, prev, rb_link, rb_parent);
2048 out:
2049         mm->total_vm += len >> PAGE_SHIFT;
2050         if (flags & VM_LOCKED) {
2051                 mm->locked_vm += len >> PAGE_SHIFT;
2052                 make_pages_present(addr, addr + len);
2053         }
2054         return addr;
2055 }
2056
2057 EXPORT_SYMBOL(do_brk);
2058
2059 /* Release all mmaps. */
2060 void exit_mmap(struct mm_struct *mm)
2061 {
2062         struct mmu_gather *tlb;
2063         struct vm_area_struct *vma = mm->mmap;
2064         unsigned long nr_accounted = 0;
2065         unsigned long end;
2066
2067         /* mm's last user has gone, and its about to be pulled down */
2068         arch_exit_mmap(mm);
2069         mmu_notifier_release(mm);
2070
2071         lru_add_drain();
2072         flush_cache_mm(mm);
2073         tlb = tlb_gather_mmu(mm, 1);
2074         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2075         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2076         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2077         vm_unacct_memory(nr_accounted);
2078         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2079         tlb_finish_mmu(tlb, 0, end);
2080
2081         /*
2082          * Walk the list again, actually closing and freeing it,
2083          * with preemption enabled, without holding any MM locks.
2084          */
2085         while (vma)
2086                 vma = remove_vma(vma);
2087
2088         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2089 }
2090
2091 /* Insert vm structure into process list sorted by address
2092  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2093  * then i_mmap_lock is taken here.
2094  */
2095 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2096 {
2097         struct vm_area_struct * __vma, * prev;
2098         struct rb_node ** rb_link, * rb_parent;
2099
2100         /*
2101          * The vm_pgoff of a purely anonymous vma should be irrelevant
2102          * until its first write fault, when page's anon_vma and index
2103          * are set.  But now set the vm_pgoff it will almost certainly
2104          * end up with (unless mremap moves it elsewhere before that
2105          * first wfault), so /proc/pid/maps tells a consistent story.
2106          *
2107          * By setting it to reflect the virtual start address of the
2108          * vma, merges and splits can happen in a seamless way, just
2109          * using the existing file pgoff checks and manipulations.
2110          * Similarly in do_mmap_pgoff and in do_brk.
2111          */
2112         if (!vma->vm_file) {
2113                 BUG_ON(vma->anon_vma);
2114                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2115         }
2116         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2117         if (__vma && __vma->vm_start < vma->vm_end)
2118                 return -ENOMEM;
2119         if ((vma->vm_flags & VM_ACCOUNT) &&
2120              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2121                 return -ENOMEM;
2122         vma_link(mm, vma, prev, rb_link, rb_parent);
2123         return 0;
2124 }
2125
2126 /*
2127  * Copy the vma structure to a new location in the same mm,
2128  * prior to moving page table entries, to effect an mremap move.
2129  */
2130 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2131         unsigned long addr, unsigned long len, pgoff_t pgoff)
2132 {
2133         struct vm_area_struct *vma = *vmap;
2134         unsigned long vma_start = vma->vm_start;
2135         struct mm_struct *mm = vma->vm_mm;
2136         struct vm_area_struct *new_vma, *prev;
2137         struct rb_node **rb_link, *rb_parent;
2138         struct mempolicy *pol;
2139
2140         /*
2141          * If anonymous vma has not yet been faulted, update new pgoff
2142          * to match new location, to increase its chance of merging.
2143          */
2144         if (!vma->vm_file && !vma->anon_vma)
2145                 pgoff = addr >> PAGE_SHIFT;
2146
2147         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2148         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2149                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2150         if (new_vma) {
2151                 /*
2152                  * Source vma may have been merged into new_vma
2153                  */
2154                 if (vma_start >= new_vma->vm_start &&
2155                     vma_start < new_vma->vm_end)
2156                         *vmap = new_vma;
2157         } else {
2158                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2159                 if (new_vma) {
2160                         *new_vma = *vma;
2161                         pol = mpol_dup(vma_policy(vma));
2162                         if (IS_ERR(pol)) {
2163                                 kmem_cache_free(vm_area_cachep, new_vma);
2164                                 return NULL;
2165                         }
2166                         vma_set_policy(new_vma, pol);
2167                         new_vma->vm_start = addr;
2168                         new_vma->vm_end = addr + len;
2169                         new_vma->vm_pgoff = pgoff;
2170                         if (new_vma->vm_file) {
2171                                 get_file(new_vma->vm_file);
2172                                 if (vma->vm_flags & VM_EXECUTABLE)
2173                                         added_exe_file_vma(mm);
2174                         }
2175                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2176                                 new_vma->vm_ops->open(new_vma);
2177                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2178                 }
2179         }
2180         return new_vma;
2181 }
2182
2183 /*
2184  * Return true if the calling process may expand its vm space by the passed
2185  * number of pages
2186  */
2187 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2188 {
2189         unsigned long cur = mm->total_vm;       /* pages */
2190         unsigned long lim;
2191
2192         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2193
2194         if (cur + npages > lim)
2195                 return 0;
2196         return 1;
2197 }
2198
2199
2200 static int special_mapping_fault(struct vm_area_struct *vma,
2201                                 struct vm_fault *vmf)
2202 {
2203         pgoff_t pgoff;
2204         struct page **pages;
2205
2206         /*
2207          * special mappings have no vm_file, and in that case, the mm
2208          * uses vm_pgoff internally. So we have to subtract it from here.
2209          * We are allowed to do this because we are the mm; do not copy
2210          * this code into drivers!
2211          */
2212         pgoff = vmf->pgoff - vma->vm_pgoff;
2213
2214         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2215                 pgoff--;
2216
2217         if (*pages) {
2218                 struct page *page = *pages;
2219                 get_page(page);
2220                 vmf->page = page;
2221                 return 0;
2222         }
2223
2224         return VM_FAULT_SIGBUS;
2225 }
2226
2227 /*
2228  * Having a close hook prevents vma merging regardless of flags.
2229  */
2230 static void special_mapping_close(struct vm_area_struct *vma)
2231 {
2232 }
2233
2234 static struct vm_operations_struct special_mapping_vmops = {
2235         .close = special_mapping_close,
2236         .fault = special_mapping_fault,
2237 };
2238
2239 /*
2240  * Called with mm->mmap_sem held for writing.
2241  * Insert a new vma covering the given region, with the given flags.
2242  * Its pages are supplied by the given array of struct page *.
2243  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2244  * The region past the last page supplied will always produce SIGBUS.
2245  * The array pointer and the pages it points to are assumed to stay alive
2246  * for as long as this mapping might exist.
2247  */
2248 int install_special_mapping(struct mm_struct *mm,
2249                             unsigned long addr, unsigned long len,
2250                             unsigned long vm_flags, struct page **pages)
2251 {
2252         struct vm_area_struct *vma;
2253
2254         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2255         if (unlikely(vma == NULL))
2256                 return -ENOMEM;
2257
2258         vma->vm_mm = mm;
2259         vma->vm_start = addr;
2260         vma->vm_end = addr + len;
2261
2262         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2263         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2264
2265         vma->vm_ops = &special_mapping_vmops;
2266         vma->vm_private_data = pages;
2267
2268         if (unlikely(insert_vm_struct(mm, vma))) {
2269                 kmem_cache_free(vm_area_cachep, vma);
2270                 return -ENOMEM;
2271         }
2272
2273         mm->total_vm += len >> PAGE_SHIFT;
2274
2275         return 0;
2276 }
2277
2278 static DEFINE_MUTEX(mm_all_locks_mutex);
2279
2280 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2281 {
2282         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2283                 /*
2284                  * The LSB of head.next can't change from under us
2285                  * because we hold the mm_all_locks_mutex.
2286                  */
2287                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2288                 /*
2289                  * We can safely modify head.next after taking the
2290                  * anon_vma->lock. If some other vma in this mm shares
2291                  * the same anon_vma we won't take it again.
2292                  *
2293                  * No need of atomic instructions here, head.next
2294                  * can't change from under us thanks to the
2295                  * anon_vma->lock.
2296                  */
2297                 if (__test_and_set_bit(0, (unsigned long *)
2298                                        &anon_vma->head.next))
2299                         BUG();
2300         }
2301 }
2302
2303 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2304 {
2305         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2306                 /*
2307                  * AS_MM_ALL_LOCKS can't change from under us because
2308                  * we hold the mm_all_locks_mutex.
2309                  *
2310                  * Operations on ->flags have to be atomic because
2311                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2312                  * mm_all_locks_mutex, there may be other cpus
2313                  * changing other bitflags in parallel to us.
2314                  */
2315                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2316                         BUG();
2317                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2318         }
2319 }
2320
2321 /*
2322  * This operation locks against the VM for all pte/vma/mm related
2323  * operations that could ever happen on a certain mm. This includes
2324  * vmtruncate, try_to_unmap, and all page faults.
2325  *
2326  * The caller must take the mmap_sem in write mode before calling
2327  * mm_take_all_locks(). The caller isn't allowed to release the
2328  * mmap_sem until mm_drop_all_locks() returns.
2329  *
2330  * mmap_sem in write mode is required in order to block all operations
2331  * that could modify pagetables and free pages without need of
2332  * altering the vma layout (for example populate_range() with
2333  * nonlinear vmas). It's also needed in write mode to avoid new
2334  * anon_vmas to be associated with existing vmas.
2335  *
2336  * A single task can't take more than one mm_take_all_locks() in a row
2337  * or it would deadlock.
2338  *
2339  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2340  * mapping->flags avoid to take the same lock twice, if more than one
2341  * vma in this mm is backed by the same anon_vma or address_space.
2342  *
2343  * We can take all the locks in random order because the VM code
2344  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2345  * takes more than one of them in a row. Secondly we're protected
2346  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2347  *
2348  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2349  * that may have to take thousand of locks.
2350  *
2351  * mm_take_all_locks() can fail if it's interrupted by signals.
2352  */
2353 int mm_take_all_locks(struct mm_struct *mm)
2354 {
2355         struct vm_area_struct *vma;
2356         int ret = -EINTR;
2357
2358         BUG_ON(down_read_trylock(&mm->mmap_sem));
2359
2360         mutex_lock(&mm_all_locks_mutex);
2361
2362         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2363                 if (signal_pending(current))
2364                         goto out_unlock;
2365                 if (vma->vm_file && vma->vm_file->f_mapping)
2366                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2367         }
2368
2369         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2370                 if (signal_pending(current))
2371                         goto out_unlock;
2372                 if (vma->anon_vma)
2373                         vm_lock_anon_vma(mm, vma->anon_vma);
2374         }
2375
2376         ret = 0;
2377
2378 out_unlock:
2379         if (ret)
2380                 mm_drop_all_locks(mm);
2381
2382         return ret;
2383 }
2384
2385 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2386 {
2387         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2388                 /*
2389                  * The LSB of head.next can't change to 0 from under
2390                  * us because we hold the mm_all_locks_mutex.
2391                  *
2392                  * We must however clear the bitflag before unlocking
2393                  * the vma so the users using the anon_vma->head will
2394                  * never see our bitflag.
2395                  *
2396                  * No need of atomic instructions here, head.next
2397                  * can't change from under us until we release the
2398                  * anon_vma->lock.
2399                  */
2400                 if (!__test_and_clear_bit(0, (unsigned long *)
2401                                           &anon_vma->head.next))
2402                         BUG();
2403                 spin_unlock(&anon_vma->lock);
2404         }
2405 }
2406
2407 static void vm_unlock_mapping(struct address_space *mapping)
2408 {
2409         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2410                 /*
2411                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2412                  * because we hold the mm_all_locks_mutex.
2413                  */
2414                 spin_unlock(&mapping->i_mmap_lock);
2415                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2416                                         &mapping->flags))
2417                         BUG();
2418         }
2419 }
2420
2421 /*
2422  * The mmap_sem cannot be released by the caller until
2423  * mm_drop_all_locks() returns.
2424  */
2425 void mm_drop_all_locks(struct mm_struct *mm)
2426 {
2427         struct vm_area_struct *vma;
2428
2429         BUG_ON(down_read_trylock(&mm->mmap_sem));
2430         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2431
2432         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2433                 if (vma->anon_vma)
2434                         vm_unlock_anon_vma(vma->anon_vma);
2435                 if (vma->vm_file && vma->vm_file->f_mapping)
2436                         vm_unlock_mapping(vma->vm_file->f_mapping);
2437         }
2438
2439         mutex_unlock(&mm_all_locks_mutex);
2440 }