seq_file: fix big-enough lseek() + read()
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/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         if (mm)
179                 allowed -= mm->total_vm / 32;
180
181         /*
182          * cast `allowed' as a signed long because vm_committed_space
183          * sometimes has a negative value
184          */
185         if (atomic_long_read(&vm_committed_space) < (long)allowed)
186                 return 0;
187 error:
188         vm_unacct_memory(pages);
189
190         return -ENOMEM;
191 }
192
193 /*
194  * Requires inode->i_mapping->i_mmap_lock
195  */
196 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
197                 struct file *file, struct address_space *mapping)
198 {
199         if (vma->vm_flags & VM_DENYWRITE)
200                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
201         if (vma->vm_flags & VM_SHARED)
202                 mapping->i_mmap_writable--;
203
204         flush_dcache_mmap_lock(mapping);
205         if (unlikely(vma->vm_flags & VM_NONLINEAR))
206                 list_del_init(&vma->shared.vm_set.list);
207         else
208                 vma_prio_tree_remove(vma, &mapping->i_mmap);
209         flush_dcache_mmap_unlock(mapping);
210 }
211
212 /*
213  * Unlink a file-based vm structure from its prio_tree, to hide
214  * vma from rmap and vmtruncate before freeing its page tables.
215  */
216 void unlink_file_vma(struct vm_area_struct *vma)
217 {
218         struct file *file = vma->vm_file;
219
220         if (file) {
221                 struct address_space *mapping = file->f_mapping;
222                 spin_lock(&mapping->i_mmap_lock);
223                 __remove_shared_vm_struct(vma, file, mapping);
224                 spin_unlock(&mapping->i_mmap_lock);
225         }
226 }
227
228 /*
229  * Close a vm structure and free it, returning the next.
230  */
231 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
232 {
233         struct vm_area_struct *next = vma->vm_next;
234
235         might_sleep();
236         if (vma->vm_ops && vma->vm_ops->close)
237                 vma->vm_ops->close(vma);
238         if (vma->vm_file) {
239                 fput(vma->vm_file);
240                 if (vma->vm_flags & VM_EXECUTABLE)
241                         removed_exe_file_vma(vma->vm_mm);
242         }
243         mpol_put(vma_policy(vma));
244         kmem_cache_free(vm_area_cachep, vma);
245         return next;
246 }
247
248 SYSCALL_DEFINE1(brk, unsigned long, brk)
249 {
250         unsigned long rlim, retval;
251         unsigned long newbrk, oldbrk;
252         struct mm_struct *mm = current->mm;
253         unsigned long min_brk;
254
255         down_write(&mm->mmap_sem);
256
257 #ifdef CONFIG_COMPAT_BRK
258         min_brk = mm->end_code;
259 #else
260         min_brk = mm->start_brk;
261 #endif
262         if (brk < min_brk)
263                 goto out;
264
265         /*
266          * Check against rlimit here. If this check is done later after the test
267          * of oldbrk with newbrk then it can escape the test and let the data
268          * segment grow beyond its set limit the in case where the limit is
269          * not page aligned -Ram Gupta
270          */
271         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
272         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
273                         (mm->end_data - mm->start_data) > rlim)
274                 goto out;
275
276         newbrk = PAGE_ALIGN(brk);
277         oldbrk = PAGE_ALIGN(mm->brk);
278         if (oldbrk == newbrk)
279                 goto set_brk;
280
281         /* Always allow shrinking brk. */
282         if (brk <= mm->brk) {
283                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
284                         goto set_brk;
285                 goto out;
286         }
287
288         /* Check against existing mmap mappings. */
289         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
290                 goto out;
291
292         /* Ok, looks good - let it rip. */
293         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
294                 goto out;
295 set_brk:
296         mm->brk = brk;
297 out:
298         retval = mm->brk;
299         up_write(&mm->mmap_sem);
300         return retval;
301 }
302
303 #ifdef DEBUG_MM_RB
304 static int browse_rb(struct rb_root *root)
305 {
306         int i = 0, j;
307         struct rb_node *nd, *pn = NULL;
308         unsigned long prev = 0, pend = 0;
309
310         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
311                 struct vm_area_struct *vma;
312                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
313                 if (vma->vm_start < prev)
314                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
315                 if (vma->vm_start < pend)
316                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
317                 if (vma->vm_start > vma->vm_end)
318                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
319                 i++;
320                 pn = nd;
321                 prev = vma->vm_start;
322                 pend = vma->vm_end;
323         }
324         j = 0;
325         for (nd = pn; nd; nd = rb_prev(nd)) {
326                 j++;
327         }
328         if (i != j)
329                 printk("backwards %d, forwards %d\n", j, i), i = 0;
330         return i;
331 }
332
333 void validate_mm(struct mm_struct *mm)
334 {
335         int bug = 0;
336         int i = 0;
337         struct vm_area_struct *tmp = mm->mmap;
338         while (tmp) {
339                 tmp = tmp->vm_next;
340                 i++;
341         }
342         if (i != mm->map_count)
343                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
344         i = browse_rb(&mm->mm_rb);
345         if (i != mm->map_count)
346                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
347         BUG_ON(bug);
348 }
349 #else
350 #define validate_mm(mm) do { } while (0)
351 #endif
352
353 static struct vm_area_struct *
354 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
355                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
356                 struct rb_node ** rb_parent)
357 {
358         struct vm_area_struct * vma;
359         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
360
361         __rb_link = &mm->mm_rb.rb_node;
362         rb_prev = __rb_parent = NULL;
363         vma = NULL;
364
365         while (*__rb_link) {
366                 struct vm_area_struct *vma_tmp;
367
368                 __rb_parent = *__rb_link;
369                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
370
371                 if (vma_tmp->vm_end > addr) {
372                         vma = vma_tmp;
373                         if (vma_tmp->vm_start <= addr)
374                                 break;
375                         __rb_link = &__rb_parent->rb_left;
376                 } else {
377                         rb_prev = __rb_parent;
378                         __rb_link = &__rb_parent->rb_right;
379                 }
380         }
381
382         *pprev = NULL;
383         if (rb_prev)
384                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
385         *rb_link = __rb_link;
386         *rb_parent = __rb_parent;
387         return vma;
388 }
389
390 static inline void
391 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
392                 struct vm_area_struct *prev, struct rb_node *rb_parent)
393 {
394         if (prev) {
395                 vma->vm_next = prev->vm_next;
396                 prev->vm_next = vma;
397         } else {
398                 mm->mmap = vma;
399                 if (rb_parent)
400                         vma->vm_next = rb_entry(rb_parent,
401                                         struct vm_area_struct, vm_rb);
402                 else
403                         vma->vm_next = NULL;
404         }
405 }
406
407 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
408                 struct rb_node **rb_link, struct rb_node *rb_parent)
409 {
410         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
411         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
412 }
413
414 static void __vma_link_file(struct vm_area_struct *vma)
415 {
416         struct file *file;
417
418         file = vma->vm_file;
419         if (file) {
420                 struct address_space *mapping = file->f_mapping;
421
422                 if (vma->vm_flags & VM_DENYWRITE)
423                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
424                 if (vma->vm_flags & VM_SHARED)
425                         mapping->i_mmap_writable++;
426
427                 flush_dcache_mmap_lock(mapping);
428                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
429                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
430                 else
431                         vma_prio_tree_insert(vma, &mapping->i_mmap);
432                 flush_dcache_mmap_unlock(mapping);
433         }
434 }
435
436 static void
437 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
438         struct vm_area_struct *prev, struct rb_node **rb_link,
439         struct rb_node *rb_parent)
440 {
441         __vma_link_list(mm, vma, prev, rb_parent);
442         __vma_link_rb(mm, vma, rb_link, rb_parent);
443         __anon_vma_link(vma);
444 }
445
446 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
447                         struct vm_area_struct *prev, struct rb_node **rb_link,
448                         struct rb_node *rb_parent)
449 {
450         struct address_space *mapping = NULL;
451
452         if (vma->vm_file)
453                 mapping = vma->vm_file->f_mapping;
454
455         if (mapping) {
456                 spin_lock(&mapping->i_mmap_lock);
457                 vma->vm_truncate_count = mapping->truncate_count;
458         }
459         anon_vma_lock(vma);
460
461         __vma_link(mm, vma, prev, rb_link, rb_parent);
462         __vma_link_file(vma);
463
464         anon_vma_unlock(vma);
465         if (mapping)
466                 spin_unlock(&mapping->i_mmap_lock);
467
468         mm->map_count++;
469         validate_mm(mm);
470 }
471
472 /*
473  * Helper for vma_adjust in the split_vma insert case:
474  * insert vm structure into list and rbtree and anon_vma,
475  * but it has already been inserted into prio_tree earlier.
476  */
477 static void __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 /* Flags that can be inherited from an existing mapping when merging */
662 #define VM_MERGEABLE_FLAGS (VM_CAN_NONLINEAR)
663
664 /*
665  * If the vma has a ->close operation then the driver probably needs to release
666  * per-vma resources, so we don't attempt to merge those.
667  */
668 static inline int is_mergeable_vma(struct vm_area_struct *vma,
669                         struct file *file, unsigned long vm_flags)
670 {
671         if ((vma->vm_flags ^ vm_flags) & ~VM_MERGEABLE_FLAGS)
672                 return 0;
673         if (vma->vm_file != file)
674                 return 0;
675         if (vma->vm_ops && vma->vm_ops->close)
676                 return 0;
677         return 1;
678 }
679
680 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
681                                         struct anon_vma *anon_vma2)
682 {
683         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
684 }
685
686 /*
687  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
688  * in front of (at a lower virtual address and file offset than) the vma.
689  *
690  * We cannot merge two vmas if they have differently assigned (non-NULL)
691  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
692  *
693  * We don't check here for the merged mmap wrapping around the end of pagecache
694  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
695  * wrap, nor mmaps which cover the final page at index -1UL.
696  */
697 static int
698 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
699         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
700 {
701         if (is_mergeable_vma(vma, file, vm_flags) &&
702             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
703                 if (vma->vm_pgoff == vm_pgoff)
704                         return 1;
705         }
706         return 0;
707 }
708
709 /*
710  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
711  * beyond (at a higher virtual address and file offset than) the vma.
712  *
713  * We cannot merge two vmas if they have differently assigned (non-NULL)
714  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
715  */
716 static int
717 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
718         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
719 {
720         if (is_mergeable_vma(vma, file, vm_flags) &&
721             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
722                 pgoff_t vm_pglen;
723                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
724                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
725                         return 1;
726         }
727         return 0;
728 }
729
730 /*
731  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
732  * whether that can be merged with its predecessor or its successor.
733  * Or both (it neatly fills a hole).
734  *
735  * In most cases - when called for mmap, brk or mremap - [addr,end) is
736  * certain not to be mapped by the time vma_merge is called; but when
737  * called for mprotect, it is certain to be already mapped (either at
738  * an offset within prev, or at the start of next), and the flags of
739  * this area are about to be changed to vm_flags - and the no-change
740  * case has already been eliminated.
741  *
742  * The following mprotect cases have to be considered, where AAAA is
743  * the area passed down from mprotect_fixup, never extending beyond one
744  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
745  *
746  *     AAAA             AAAA                AAAA          AAAA
747  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
748  *    cannot merge    might become    might become    might become
749  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
750  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
751  *    mremap move:                                    PPPPNNNNNNNN 8
752  *        AAAA
753  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
754  *    might become    case 1 below    case 2 below    case 3 below
755  *
756  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
757  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
758  */
759 struct vm_area_struct *vma_merge(struct mm_struct *mm,
760                         struct vm_area_struct *prev, unsigned long addr,
761                         unsigned long end, unsigned long vm_flags,
762                         struct anon_vma *anon_vma, struct file *file,
763                         pgoff_t pgoff, struct mempolicy *policy)
764 {
765         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
766         struct vm_area_struct *area, *next;
767
768         /*
769          * We later require that vma->vm_flags == vm_flags,
770          * so this tests vma->vm_flags & VM_SPECIAL, too.
771          */
772         if (vm_flags & VM_SPECIAL)
773                 return NULL;
774
775         if (prev)
776                 next = prev->vm_next;
777         else
778                 next = mm->mmap;
779         area = next;
780         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
781                 next = next->vm_next;
782
783         /*
784          * Can it merge with the predecessor?
785          */
786         if (prev && prev->vm_end == addr &&
787                         mpol_equal(vma_policy(prev), policy) &&
788                         can_vma_merge_after(prev, vm_flags,
789                                                 anon_vma, file, pgoff)) {
790                 /*
791                  * OK, it can.  Can we now merge in the successor as well?
792                  */
793                 if (next && end == next->vm_start &&
794                                 mpol_equal(policy, vma_policy(next)) &&
795                                 can_vma_merge_before(next, vm_flags,
796                                         anon_vma, file, pgoff+pglen) &&
797                                 is_mergeable_anon_vma(prev->anon_vma,
798                                                       next->anon_vma)) {
799                                                         /* cases 1, 6 */
800                         vma_adjust(prev, prev->vm_start,
801                                 next->vm_end, prev->vm_pgoff, NULL);
802                 } else                                  /* cases 2, 5, 7 */
803                         vma_adjust(prev, prev->vm_start,
804                                 end, prev->vm_pgoff, NULL);
805                 return prev;
806         }
807
808         /*
809          * Can this new request be merged in front of next?
810          */
811         if (next && end == next->vm_start &&
812                         mpol_equal(policy, vma_policy(next)) &&
813                         can_vma_merge_before(next, vm_flags,
814                                         anon_vma, file, pgoff+pglen)) {
815                 if (prev && addr < prev->vm_end)        /* case 4 */
816                         vma_adjust(prev, prev->vm_start,
817                                 addr, prev->vm_pgoff, NULL);
818                 else                                    /* cases 3, 8 */
819                         vma_adjust(area, addr, next->vm_end,
820                                 next->vm_pgoff - pglen, NULL);
821                 return area;
822         }
823
824         return NULL;
825 }
826
827 /*
828  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
829  * neighbouring vmas for a suitable anon_vma, before it goes off
830  * to allocate a new anon_vma.  It checks because a repetitive
831  * sequence of mprotects and faults may otherwise lead to distinct
832  * anon_vmas being allocated, preventing vma merge in subsequent
833  * mprotect.
834  */
835 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
836 {
837         struct vm_area_struct *near;
838         unsigned long vm_flags;
839
840         near = vma->vm_next;
841         if (!near)
842                 goto try_prev;
843
844         /*
845          * Since only mprotect tries to remerge vmas, match flags
846          * which might be mprotected into each other later on.
847          * Neither mlock nor madvise tries to remerge at present,
848          * so leave their flags as obstructing a merge.
849          */
850         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
851         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
852
853         if (near->anon_vma && vma->vm_end == near->vm_start &&
854                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
855                         can_vma_merge_before(near, vm_flags,
856                                 NULL, vma->vm_file, vma->vm_pgoff +
857                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
858                 return near->anon_vma;
859 try_prev:
860         /*
861          * It is potentially slow to have to call find_vma_prev here.
862          * But it's only on the first write fault on the vma, not
863          * every time, and we could devise a way to avoid it later
864          * (e.g. stash info in next's anon_vma_node when assigning
865          * an anon_vma, or when trying vma_merge).  Another time.
866          */
867         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
868         if (!near)
869                 goto none;
870
871         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
872         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
873
874         if (near->anon_vma && near->vm_end == vma->vm_start &&
875                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
876                         can_vma_merge_after(near, vm_flags,
877                                 NULL, vma->vm_file, vma->vm_pgoff))
878                 return near->anon_vma;
879 none:
880         /*
881          * There's no absolute need to look only at touching neighbours:
882          * we could search further afield for "compatible" anon_vmas.
883          * But it would probably just be a waste of time searching,
884          * or lead to too many vmas hanging off the same anon_vma.
885          * We're trying to allow mprotect remerging later on,
886          * not trying to minimize memory used for anon_vmas.
887          */
888         return NULL;
889 }
890
891 #ifdef CONFIG_PROC_FS
892 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
893                                                 struct file *file, long pages)
894 {
895         const unsigned long stack_flags
896                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
897
898         if (file) {
899                 mm->shared_vm += pages;
900                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
901                         mm->exec_vm += pages;
902         } else if (flags & stack_flags)
903                 mm->stack_vm += pages;
904         if (flags & (VM_RESERVED|VM_IO))
905                 mm->reserved_vm += pages;
906 }
907 #endif /* CONFIG_PROC_FS */
908
909 /*
910  * The caller must hold down_write(current->mm->mmap_sem).
911  */
912
913 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
914                         unsigned long len, unsigned long prot,
915                         unsigned long flags, unsigned long pgoff)
916 {
917         struct mm_struct * mm = current->mm;
918         struct inode *inode;
919         unsigned int vm_flags;
920         int error;
921         int accountable = 1;
922         unsigned long reqprot = prot;
923
924         /*
925          * Does the application expect PROT_READ to imply PROT_EXEC?
926          *
927          * (the exception is when the underlying filesystem is noexec
928          *  mounted, in which case we dont add PROT_EXEC.)
929          */
930         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
931                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
932                         prot |= PROT_EXEC;
933
934         if (!len)
935                 return -EINVAL;
936
937         if (!(flags & MAP_FIXED))
938                 addr = round_hint_to_min(addr);
939
940         error = arch_mmap_check(addr, len, flags);
941         if (error)
942                 return error;
943
944         /* Careful about overflows.. */
945         len = PAGE_ALIGN(len);
946         if (!len || len > TASK_SIZE)
947                 return -ENOMEM;
948
949         /* offset overflow? */
950         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
951                return -EOVERFLOW;
952
953         /* Too many mappings? */
954         if (mm->map_count > sysctl_max_map_count)
955                 return -ENOMEM;
956
957         /* Obtain the address to map to. we verify (or select) it and ensure
958          * that it represents a valid section of the address space.
959          */
960         addr = get_unmapped_area(file, addr, len, pgoff, flags);
961         if (addr & ~PAGE_MASK)
962                 return addr;
963
964         /* Do simple checking here so the lower-level routines won't have
965          * to. we assume access permissions have been handled by the open
966          * of the memory object, so we don't do any here.
967          */
968         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
969                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
970
971         if (flags & MAP_LOCKED) {
972                 if (!can_do_mlock())
973                         return -EPERM;
974                 vm_flags |= VM_LOCKED;
975         }
976
977         /* mlock MCL_FUTURE? */
978         if (vm_flags & VM_LOCKED) {
979                 unsigned long locked, lock_limit;
980                 locked = len >> PAGE_SHIFT;
981                 locked += mm->locked_vm;
982                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
983                 lock_limit >>= PAGE_SHIFT;
984                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
985                         return -EAGAIN;
986         }
987
988         inode = file ? file->f_path.dentry->d_inode : NULL;
989
990         if (file) {
991                 switch (flags & MAP_TYPE) {
992                 case MAP_SHARED:
993                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
994                                 return -EACCES;
995
996                         /*
997                          * Make sure we don't allow writing to an append-only
998                          * file..
999                          */
1000                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1001                                 return -EACCES;
1002
1003                         /*
1004                          * Make sure there are no mandatory locks on the file.
1005                          */
1006                         if (locks_verify_locked(inode))
1007                                 return -EAGAIN;
1008
1009                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1010                         if (!(file->f_mode & FMODE_WRITE))
1011                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1012
1013                         /* fall through */
1014                 case MAP_PRIVATE:
1015                         if (!(file->f_mode & FMODE_READ))
1016                                 return -EACCES;
1017                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1018                                 if (vm_flags & VM_EXEC)
1019                                         return -EPERM;
1020                                 vm_flags &= ~VM_MAYEXEC;
1021                         }
1022                         if (is_file_hugepages(file))
1023                                 accountable = 0;
1024
1025                         if (!file->f_op || !file->f_op->mmap)
1026                                 return -ENODEV;
1027                         break;
1028
1029                 default:
1030                         return -EINVAL;
1031                 }
1032         } else {
1033                 switch (flags & MAP_TYPE) {
1034                 case MAP_SHARED:
1035                         /*
1036                          * Ignore pgoff.
1037                          */
1038                         pgoff = 0;
1039                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1040                         break;
1041                 case MAP_PRIVATE:
1042                         /*
1043                          * Set pgoff according to addr for anon_vma.
1044                          */
1045                         pgoff = addr >> PAGE_SHIFT;
1046                         break;
1047                 default:
1048                         return -EINVAL;
1049                 }
1050         }
1051
1052         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1053         if (error)
1054                 return error;
1055
1056         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1057                            accountable);
1058 }
1059 EXPORT_SYMBOL(do_mmap_pgoff);
1060
1061 /*
1062  * Some shared mappigns will want the pages marked read-only
1063  * to track write events. If so, we'll downgrade vm_page_prot
1064  * to the private version (using protection_map[] without the
1065  * VM_SHARED bit).
1066  */
1067 int vma_wants_writenotify(struct vm_area_struct *vma)
1068 {
1069         unsigned int vm_flags = vma->vm_flags;
1070
1071         /* If it was private or non-writable, the write bit is already clear */
1072         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1073                 return 0;
1074
1075         /* The backer wishes to know when pages are first written to? */
1076         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1077                 return 1;
1078
1079         /* The open routine did something to the protections already? */
1080         if (pgprot_val(vma->vm_page_prot) !=
1081             pgprot_val(vm_get_page_prot(vm_flags)))
1082                 return 0;
1083
1084         /* Specialty mapping? */
1085         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1086                 return 0;
1087
1088         /* Can the mapping track the dirty pages? */
1089         return vma->vm_file && vma->vm_file->f_mapping &&
1090                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1091 }
1092
1093 /*
1094  * We account for memory if it's a private writeable mapping,
1095  * and VM_NORESERVE wasn't set.
1096  */
1097 static inline int accountable_mapping(unsigned int vm_flags)
1098 {
1099         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1100 }
1101
1102 unsigned long mmap_region(struct file *file, unsigned long addr,
1103                           unsigned long len, unsigned long flags,
1104                           unsigned int vm_flags, unsigned long pgoff,
1105                           int accountable)
1106 {
1107         struct mm_struct *mm = current->mm;
1108         struct vm_area_struct *vma, *prev;
1109         int correct_wcount = 0;
1110         int error;
1111         struct rb_node **rb_link, *rb_parent;
1112         unsigned long charged = 0;
1113         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1114
1115         /* Clear old maps */
1116         error = -ENOMEM;
1117 munmap_back:
1118         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1119         if (vma && vma->vm_start < addr + len) {
1120                 if (do_munmap(mm, addr, len))
1121                         return -ENOMEM;
1122                 goto munmap_back;
1123         }
1124
1125         /* Check against address space limit. */
1126         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1127                 return -ENOMEM;
1128
1129         /*
1130          * Set 'VM_NORESERVE' if we should not account for the
1131          * memory use of this mapping. We only honor MAP_NORESERVE
1132          * if we're allowed to overcommit memory.
1133          */
1134         if ((flags & MAP_NORESERVE) && sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1135                 vm_flags |= VM_NORESERVE;
1136         if (!accountable)
1137                 vm_flags |= VM_NORESERVE;
1138
1139         /*
1140          * Private writable mapping: check memory availability
1141          */
1142         if (accountable_mapping(vm_flags)) {
1143                 charged = len >> PAGE_SHIFT;
1144                 if (security_vm_enough_memory(charged))
1145                         return -ENOMEM;
1146                 vm_flags |= VM_ACCOUNT;
1147         }
1148
1149         /*
1150          * Can we just expand an old mapping?
1151          */
1152         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1153         if (vma)
1154                 goto out;
1155
1156         /*
1157          * Determine the object being mapped and call the appropriate
1158          * specific mapper. the address has already been validated, but
1159          * not unmapped, but the maps are removed from the list.
1160          */
1161         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1162         if (!vma) {
1163                 error = -ENOMEM;
1164                 goto unacct_error;
1165         }
1166
1167         vma->vm_mm = mm;
1168         vma->vm_start = addr;
1169         vma->vm_end = addr + len;
1170         vma->vm_flags = vm_flags;
1171         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1172         vma->vm_pgoff = pgoff;
1173
1174         if (file) {
1175                 error = -EINVAL;
1176                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1177                         goto free_vma;
1178                 if (vm_flags & VM_DENYWRITE) {
1179                         error = deny_write_access(file);
1180                         if (error)
1181                                 goto free_vma;
1182                         correct_wcount = 1;
1183                 }
1184                 vma->vm_file = file;
1185                 get_file(file);
1186                 error = file->f_op->mmap(file, vma);
1187                 if (error)
1188                         goto unmap_and_free_vma;
1189                 if (vm_flags & VM_EXECUTABLE)
1190                         added_exe_file_vma(mm);
1191         } else if (vm_flags & VM_SHARED) {
1192                 error = shmem_zero_setup(vma);
1193                 if (error)
1194                         goto free_vma;
1195         }
1196
1197         /* Can addr have changed??
1198          *
1199          * Answer: Yes, several device drivers can do it in their
1200          *         f_op->mmap method. -DaveM
1201          */
1202         addr = vma->vm_start;
1203         pgoff = vma->vm_pgoff;
1204         vm_flags = vma->vm_flags;
1205
1206         if (vma_wants_writenotify(vma))
1207                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1208
1209         vma_link(mm, vma, prev, rb_link, rb_parent);
1210         file = vma->vm_file;
1211
1212         /* Once vma denies write, undo our temporary denial count */
1213         if (correct_wcount)
1214                 atomic_inc(&inode->i_writecount);
1215 out:
1216         mm->total_vm += len >> PAGE_SHIFT;
1217         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1218         if (vm_flags & VM_LOCKED) {
1219                 /*
1220                  * makes pages present; downgrades, drops, reacquires mmap_sem
1221                  */
1222                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1223                 if (nr_pages < 0)
1224                         return nr_pages;        /* vma gone! */
1225                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1226         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1227                 make_pages_present(addr, addr + len);
1228         return addr;
1229
1230 unmap_and_free_vma:
1231         if (correct_wcount)
1232                 atomic_inc(&inode->i_writecount);
1233         vma->vm_file = NULL;
1234         fput(file);
1235
1236         /* Undo any partial mapping done by a device driver. */
1237         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1238         charged = 0;
1239 free_vma:
1240         kmem_cache_free(vm_area_cachep, vma);
1241 unacct_error:
1242         if (charged)
1243                 vm_unacct_memory(charged);
1244         return error;
1245 }
1246
1247 /* Get an address range which is currently unmapped.
1248  * For shmat() with addr=0.
1249  *
1250  * Ugly calling convention alert:
1251  * Return value with the low bits set means error value,
1252  * ie
1253  *      if (ret & ~PAGE_MASK)
1254  *              error = ret;
1255  *
1256  * This function "knows" that -ENOMEM has the bits set.
1257  */
1258 #ifndef HAVE_ARCH_UNMAPPED_AREA
1259 unsigned long
1260 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1261                 unsigned long len, unsigned long pgoff, unsigned long flags)
1262 {
1263         struct mm_struct *mm = current->mm;
1264         struct vm_area_struct *vma;
1265         unsigned long start_addr;
1266
1267         if (len > TASK_SIZE)
1268                 return -ENOMEM;
1269
1270         if (flags & MAP_FIXED)
1271                 return addr;
1272
1273         if (addr) {
1274                 addr = PAGE_ALIGN(addr);
1275                 vma = find_vma(mm, addr);
1276                 if (TASK_SIZE - len >= addr &&
1277                     (!vma || addr + len <= vma->vm_start))
1278                         return addr;
1279         }
1280         if (len > mm->cached_hole_size) {
1281                 start_addr = addr = mm->free_area_cache;
1282         } else {
1283                 start_addr = addr = TASK_UNMAPPED_BASE;
1284                 mm->cached_hole_size = 0;
1285         }
1286
1287 full_search:
1288         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1289                 /* At this point:  (!vma || addr < vma->vm_end). */
1290                 if (TASK_SIZE - len < addr) {
1291                         /*
1292                          * Start a new search - just in case we missed
1293                          * some holes.
1294                          */
1295                         if (start_addr != TASK_UNMAPPED_BASE) {
1296                                 addr = TASK_UNMAPPED_BASE;
1297                                 start_addr = addr;
1298                                 mm->cached_hole_size = 0;
1299                                 goto full_search;
1300                         }
1301                         return -ENOMEM;
1302                 }
1303                 if (!vma || addr + len <= vma->vm_start) {
1304                         /*
1305                          * Remember the place where we stopped the search:
1306                          */
1307                         mm->free_area_cache = addr + len;
1308                         return addr;
1309                 }
1310                 if (addr + mm->cached_hole_size < vma->vm_start)
1311                         mm->cached_hole_size = vma->vm_start - addr;
1312                 addr = vma->vm_end;
1313         }
1314 }
1315 #endif  
1316
1317 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1318 {
1319         /*
1320          * Is this a new hole at the lowest possible address?
1321          */
1322         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1323                 mm->free_area_cache = addr;
1324                 mm->cached_hole_size = ~0UL;
1325         }
1326 }
1327
1328 /*
1329  * This mmap-allocator allocates new areas top-down from below the
1330  * stack's low limit (the base):
1331  */
1332 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1333 unsigned long
1334 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1335                           const unsigned long len, const unsigned long pgoff,
1336                           const unsigned long flags)
1337 {
1338         struct vm_area_struct *vma;
1339         struct mm_struct *mm = current->mm;
1340         unsigned long addr = addr0;
1341
1342         /* requested length too big for entire address space */
1343         if (len > TASK_SIZE)
1344                 return -ENOMEM;
1345
1346         if (flags & MAP_FIXED)
1347                 return addr;
1348
1349         /* requesting a specific address */
1350         if (addr) {
1351                 addr = PAGE_ALIGN(addr);
1352                 vma = find_vma(mm, addr);
1353                 if (TASK_SIZE - len >= addr &&
1354                                 (!vma || addr + len <= vma->vm_start))
1355                         return addr;
1356         }
1357
1358         /* check if free_area_cache is useful for us */
1359         if (len <= mm->cached_hole_size) {
1360                 mm->cached_hole_size = 0;
1361                 mm->free_area_cache = mm->mmap_base;
1362         }
1363
1364         /* either no address requested or can't fit in requested address hole */
1365         addr = mm->free_area_cache;
1366
1367         /* make sure it can fit in the remaining address space */
1368         if (addr > len) {
1369                 vma = find_vma(mm, addr-len);
1370                 if (!vma || addr <= vma->vm_start)
1371                         /* remember the address as a hint for next time */
1372                         return (mm->free_area_cache = addr-len);
1373         }
1374
1375         if (mm->mmap_base < len)
1376                 goto bottomup;
1377
1378         addr = mm->mmap_base-len;
1379
1380         do {
1381                 /*
1382                  * Lookup failure means no vma is above this address,
1383                  * else if new region fits below vma->vm_start,
1384                  * return with success:
1385                  */
1386                 vma = find_vma(mm, addr);
1387                 if (!vma || addr+len <= vma->vm_start)
1388                         /* remember the address as a hint for next time */
1389                         return (mm->free_area_cache = addr);
1390
1391                 /* remember the largest hole we saw so far */
1392                 if (addr + mm->cached_hole_size < vma->vm_start)
1393                         mm->cached_hole_size = vma->vm_start - addr;
1394
1395                 /* try just below the current vma->vm_start */
1396                 addr = vma->vm_start-len;
1397         } while (len < vma->vm_start);
1398
1399 bottomup:
1400         /*
1401          * A failed mmap() very likely causes application failure,
1402          * so fall back to the bottom-up function here. This scenario
1403          * can happen with large stack limits and large mmap()
1404          * allocations.
1405          */
1406         mm->cached_hole_size = ~0UL;
1407         mm->free_area_cache = TASK_UNMAPPED_BASE;
1408         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1409         /*
1410          * Restore the topdown base:
1411          */
1412         mm->free_area_cache = mm->mmap_base;
1413         mm->cached_hole_size = ~0UL;
1414
1415         return addr;
1416 }
1417 #endif
1418
1419 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1420 {
1421         /*
1422          * Is this a new hole at the highest possible address?
1423          */
1424         if (addr > mm->free_area_cache)
1425                 mm->free_area_cache = addr;
1426
1427         /* dont allow allocations above current base */
1428         if (mm->free_area_cache > mm->mmap_base)
1429                 mm->free_area_cache = mm->mmap_base;
1430 }
1431
1432 unsigned long
1433 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1434                 unsigned long pgoff, unsigned long flags)
1435 {
1436         unsigned long (*get_area)(struct file *, unsigned long,
1437                                   unsigned long, unsigned long, unsigned long);
1438
1439         get_area = current->mm->get_unmapped_area;
1440         if (file && file->f_op && file->f_op->get_unmapped_area)
1441                 get_area = file->f_op->get_unmapped_area;
1442         addr = get_area(file, addr, len, pgoff, flags);
1443         if (IS_ERR_VALUE(addr))
1444                 return addr;
1445
1446         if (addr > TASK_SIZE - len)
1447                 return -ENOMEM;
1448         if (addr & ~PAGE_MASK)
1449                 return -EINVAL;
1450
1451         return arch_rebalance_pgtables(addr, len);
1452 }
1453
1454 EXPORT_SYMBOL(get_unmapped_area);
1455
1456 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1457 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1458 {
1459         struct vm_area_struct *vma = NULL;
1460
1461         if (mm) {
1462                 /* Check the cache first. */
1463                 /* (Cache hit rate is typically around 35%.) */
1464                 vma = mm->mmap_cache;
1465                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1466                         struct rb_node * rb_node;
1467
1468                         rb_node = mm->mm_rb.rb_node;
1469                         vma = NULL;
1470
1471                         while (rb_node) {
1472                                 struct vm_area_struct * vma_tmp;
1473
1474                                 vma_tmp = rb_entry(rb_node,
1475                                                 struct vm_area_struct, vm_rb);
1476
1477                                 if (vma_tmp->vm_end > addr) {
1478                                         vma = vma_tmp;
1479                                         if (vma_tmp->vm_start <= addr)
1480                                                 break;
1481                                         rb_node = rb_node->rb_left;
1482                                 } else
1483                                         rb_node = rb_node->rb_right;
1484                         }
1485                         if (vma)
1486                                 mm->mmap_cache = vma;
1487                 }
1488         }
1489         return vma;
1490 }
1491
1492 EXPORT_SYMBOL(find_vma);
1493
1494 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1495 struct vm_area_struct *
1496 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1497                         struct vm_area_struct **pprev)
1498 {
1499         struct vm_area_struct *vma = NULL, *prev = NULL;
1500         struct rb_node *rb_node;
1501         if (!mm)
1502                 goto out;
1503
1504         /* Guard against addr being lower than the first VMA */
1505         vma = mm->mmap;
1506
1507         /* Go through the RB tree quickly. */
1508         rb_node = mm->mm_rb.rb_node;
1509
1510         while (rb_node) {
1511                 struct vm_area_struct *vma_tmp;
1512                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1513
1514                 if (addr < vma_tmp->vm_end) {
1515                         rb_node = rb_node->rb_left;
1516                 } else {
1517                         prev = vma_tmp;
1518                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1519                                 break;
1520                         rb_node = rb_node->rb_right;
1521                 }
1522         }
1523
1524 out:
1525         *pprev = prev;
1526         return prev ? prev->vm_next : vma;
1527 }
1528
1529 /*
1530  * Verify that the stack growth is acceptable and
1531  * update accounting. This is shared with both the
1532  * grow-up and grow-down cases.
1533  */
1534 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1535 {
1536         struct mm_struct *mm = vma->vm_mm;
1537         struct rlimit *rlim = current->signal->rlim;
1538         unsigned long new_start;
1539
1540         /* address space limit tests */
1541         if (!may_expand_vm(mm, grow))
1542                 return -ENOMEM;
1543
1544         /* Stack limit test */
1545         if (size > rlim[RLIMIT_STACK].rlim_cur)
1546                 return -ENOMEM;
1547
1548         /* mlock limit tests */
1549         if (vma->vm_flags & VM_LOCKED) {
1550                 unsigned long locked;
1551                 unsigned long limit;
1552                 locked = mm->locked_vm + grow;
1553                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1554                 if (locked > limit && !capable(CAP_IPC_LOCK))
1555                         return -ENOMEM;
1556         }
1557
1558         /* Check to ensure the stack will not grow into a hugetlb-only region */
1559         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1560                         vma->vm_end - size;
1561         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1562                 return -EFAULT;
1563
1564         /*
1565          * Overcommit..  This must be the final test, as it will
1566          * update security statistics.
1567          */
1568         if (security_vm_enough_memory(grow))
1569                 return -ENOMEM;
1570
1571         /* Ok, everything looks good - let it rip */
1572         mm->total_vm += grow;
1573         if (vma->vm_flags & VM_LOCKED)
1574                 mm->locked_vm += grow;
1575         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1576         return 0;
1577 }
1578
1579 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1580 /*
1581  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1582  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1583  */
1584 #ifndef CONFIG_IA64
1585 static
1586 #endif
1587 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1588 {
1589         int error;
1590
1591         if (!(vma->vm_flags & VM_GROWSUP))
1592                 return -EFAULT;
1593
1594         /*
1595          * We must make sure the anon_vma is allocated
1596          * so that the anon_vma locking is not a noop.
1597          */
1598         if (unlikely(anon_vma_prepare(vma)))
1599                 return -ENOMEM;
1600         anon_vma_lock(vma);
1601
1602         /*
1603          * vma->vm_start/vm_end cannot change under us because the caller
1604          * is required to hold the mmap_sem in read mode.  We need the
1605          * anon_vma lock to serialize against concurrent expand_stacks.
1606          * Also guard against wrapping around to address 0.
1607          */
1608         if (address < PAGE_ALIGN(address+4))
1609                 address = PAGE_ALIGN(address+4);
1610         else {
1611                 anon_vma_unlock(vma);
1612                 return -ENOMEM;
1613         }
1614         error = 0;
1615
1616         /* Somebody else might have raced and expanded it already */
1617         if (address > vma->vm_end) {
1618                 unsigned long size, grow;
1619
1620                 size = address - vma->vm_start;
1621                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1622
1623                 error = acct_stack_growth(vma, size, grow);
1624                 if (!error)
1625                         vma->vm_end = address;
1626         }
1627         anon_vma_unlock(vma);
1628         return error;
1629 }
1630 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1631
1632 /*
1633  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1634  */
1635 static int expand_downwards(struct vm_area_struct *vma,
1636                                    unsigned long address)
1637 {
1638         int error;
1639
1640         /*
1641          * We must make sure the anon_vma is allocated
1642          * so that the anon_vma locking is not a noop.
1643          */
1644         if (unlikely(anon_vma_prepare(vma)))
1645                 return -ENOMEM;
1646
1647         address &= PAGE_MASK;
1648         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1649         if (error)
1650                 return error;
1651
1652         anon_vma_lock(vma);
1653
1654         /*
1655          * vma->vm_start/vm_end cannot change under us because the caller
1656          * is required to hold the mmap_sem in read mode.  We need the
1657          * anon_vma lock to serialize against concurrent expand_stacks.
1658          */
1659
1660         /* Somebody else might have raced and expanded it already */
1661         if (address < vma->vm_start) {
1662                 unsigned long size, grow;
1663
1664                 size = vma->vm_end - address;
1665                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1666
1667                 error = acct_stack_growth(vma, size, grow);
1668                 if (!error) {
1669                         vma->vm_start = address;
1670                         vma->vm_pgoff -= grow;
1671                 }
1672         }
1673         anon_vma_unlock(vma);
1674         return error;
1675 }
1676
1677 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1678 {
1679         return expand_downwards(vma, address);
1680 }
1681
1682 #ifdef CONFIG_STACK_GROWSUP
1683 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1684 {
1685         return expand_upwards(vma, address);
1686 }
1687
1688 struct vm_area_struct *
1689 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1690 {
1691         struct vm_area_struct *vma, *prev;
1692
1693         addr &= PAGE_MASK;
1694         vma = find_vma_prev(mm, addr, &prev);
1695         if (vma && (vma->vm_start <= addr))
1696                 return vma;
1697         if (!prev || expand_stack(prev, addr))
1698                 return NULL;
1699         if (prev->vm_flags & VM_LOCKED) {
1700                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1701                         return NULL;    /* vma gone! */
1702         }
1703         return prev;
1704 }
1705 #else
1706 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1707 {
1708         return expand_downwards(vma, address);
1709 }
1710
1711 struct vm_area_struct *
1712 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1713 {
1714         struct vm_area_struct * vma;
1715         unsigned long start;
1716
1717         addr &= PAGE_MASK;
1718         vma = find_vma(mm,addr);
1719         if (!vma)
1720                 return NULL;
1721         if (vma->vm_start <= addr)
1722                 return vma;
1723         if (!(vma->vm_flags & VM_GROWSDOWN))
1724                 return NULL;
1725         start = vma->vm_start;
1726         if (expand_stack(vma, addr))
1727                 return NULL;
1728         if (vma->vm_flags & VM_LOCKED) {
1729                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1730                         return NULL;    /* vma gone! */
1731         }
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                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1751                 vma = remove_vma(vma);
1752         } while (vma);
1753         validate_mm(mm);
1754 }
1755
1756 /*
1757  * Get rid of page table information in the indicated region.
1758  *
1759  * Called with the mm semaphore held.
1760  */
1761 static void unmap_region(struct mm_struct *mm,
1762                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1763                 unsigned long start, unsigned long end)
1764 {
1765         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1766         struct mmu_gather *tlb;
1767         unsigned long nr_accounted = 0;
1768
1769         lru_add_drain();
1770         tlb = tlb_gather_mmu(mm, 0);
1771         update_hiwater_rss(mm);
1772         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1773         vm_unacct_memory(nr_accounted);
1774         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1775                                  next? next->vm_start: 0);
1776         tlb_finish_mmu(tlb, start, end);
1777 }
1778
1779 /*
1780  * Create a list of vma's touched by the unmap, removing them from the mm's
1781  * vma list as we go..
1782  */
1783 static void
1784 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1785         struct vm_area_struct *prev, unsigned long end)
1786 {
1787         struct vm_area_struct **insertion_point;
1788         struct vm_area_struct *tail_vma = NULL;
1789         unsigned long addr;
1790
1791         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1792         do {
1793                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1794                 mm->map_count--;
1795                 tail_vma = vma;
1796                 vma = vma->vm_next;
1797         } while (vma && vma->vm_start < end);
1798         *insertion_point = vma;
1799         tail_vma->vm_next = NULL;
1800         if (mm->unmap_area == arch_unmap_area)
1801                 addr = prev ? prev->vm_end : mm->mmap_base;
1802         else
1803                 addr = vma ?  vma->vm_start : mm->mmap_base;
1804         mm->unmap_area(mm, addr);
1805         mm->mmap_cache = NULL;          /* Kill the cache. */
1806 }
1807
1808 /*
1809  * Split a vma into two pieces at address 'addr', a new vma is allocated
1810  * either for the first part or the tail.
1811  */
1812 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1813               unsigned long addr, int new_below)
1814 {
1815         struct mempolicy *pol;
1816         struct vm_area_struct *new;
1817
1818         if (is_vm_hugetlb_page(vma) && (addr &
1819                                         ~(huge_page_mask(hstate_vma(vma)))))
1820                 return -EINVAL;
1821
1822         if (mm->map_count >= sysctl_max_map_count)
1823                 return -ENOMEM;
1824
1825         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1826         if (!new)
1827                 return -ENOMEM;
1828
1829         /* most fields are the same, copy all, and then fixup */
1830         *new = *vma;
1831
1832         if (new_below)
1833                 new->vm_end = addr;
1834         else {
1835                 new->vm_start = addr;
1836                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1837         }
1838
1839         pol = mpol_dup(vma_policy(vma));
1840         if (IS_ERR(pol)) {
1841                 kmem_cache_free(vm_area_cachep, new);
1842                 return PTR_ERR(pol);
1843         }
1844         vma_set_policy(new, pol);
1845
1846         if (new->vm_file) {
1847                 get_file(new->vm_file);
1848                 if (vma->vm_flags & VM_EXECUTABLE)
1849                         added_exe_file_vma(mm);
1850         }
1851
1852         if (new->vm_ops && new->vm_ops->open)
1853                 new->vm_ops->open(new);
1854
1855         if (new_below)
1856                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1857                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1858         else
1859                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1860
1861         return 0;
1862 }
1863
1864 /* Munmap is split into 2 main parts -- this part which finds
1865  * what needs doing, and the areas themselves, which do the
1866  * work.  This now handles partial unmappings.
1867  * Jeremy Fitzhardinge <jeremy@goop.org>
1868  */
1869 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1870 {
1871         unsigned long end;
1872         struct vm_area_struct *vma, *prev, *last;
1873
1874         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1875                 return -EINVAL;
1876
1877         if ((len = PAGE_ALIGN(len)) == 0)
1878                 return -EINVAL;
1879
1880         /* Find the first overlapping VMA */
1881         vma = find_vma_prev(mm, start, &prev);
1882         if (!vma)
1883                 return 0;
1884         /* we have  start < vma->vm_end  */
1885
1886         /* if it doesn't overlap, we have nothing.. */
1887         end = start + len;
1888         if (vma->vm_start >= end)
1889                 return 0;
1890
1891         /*
1892          * If we need to split any vma, do it now to save pain later.
1893          *
1894          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1895          * unmapped vm_area_struct will remain in use: so lower split_vma
1896          * places tmp vma above, and higher split_vma places tmp vma below.
1897          */
1898         if (start > vma->vm_start) {
1899                 int error = split_vma(mm, vma, start, 0);
1900                 if (error)
1901                         return error;
1902                 prev = vma;
1903         }
1904
1905         /* Does it split the last one? */
1906         last = find_vma(mm, end);
1907         if (last && end > last->vm_start) {
1908                 int error = split_vma(mm, last, end, 1);
1909                 if (error)
1910                         return error;
1911         }
1912         vma = prev? prev->vm_next: mm->mmap;
1913
1914         /*
1915          * unlock any mlock()ed ranges before detaching vmas
1916          */
1917         if (mm->locked_vm) {
1918                 struct vm_area_struct *tmp = vma;
1919                 while (tmp && tmp->vm_start < end) {
1920                         if (tmp->vm_flags & VM_LOCKED) {
1921                                 mm->locked_vm -= vma_pages(tmp);
1922                                 munlock_vma_pages_all(tmp);
1923                         }
1924                         tmp = tmp->vm_next;
1925                 }
1926         }
1927
1928         /*
1929          * Remove the vma's, and unmap the actual pages
1930          */
1931         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1932         unmap_region(mm, vma, prev, start, end);
1933
1934         /* Fix up all other VM information */
1935         remove_vma_list(mm, vma);
1936
1937         return 0;
1938 }
1939
1940 EXPORT_SYMBOL(do_munmap);
1941
1942 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1943 {
1944         int ret;
1945         struct mm_struct *mm = current->mm;
1946
1947         profile_munmap(addr);
1948
1949         down_write(&mm->mmap_sem);
1950         ret = do_munmap(mm, addr, len);
1951         up_write(&mm->mmap_sem);
1952         return ret;
1953 }
1954
1955 static inline void verify_mm_writelocked(struct mm_struct *mm)
1956 {
1957 #ifdef CONFIG_DEBUG_VM
1958         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1959                 WARN_ON(1);
1960                 up_read(&mm->mmap_sem);
1961         }
1962 #endif
1963 }
1964
1965 /*
1966  *  this is really a simplified "do_mmap".  it only handles
1967  *  anonymous maps.  eventually we may be able to do some
1968  *  brk-specific accounting here.
1969  */
1970 unsigned long do_brk(unsigned long addr, unsigned long len)
1971 {
1972         struct mm_struct * mm = current->mm;
1973         struct vm_area_struct * vma, * prev;
1974         unsigned long flags;
1975         struct rb_node ** rb_link, * rb_parent;
1976         pgoff_t pgoff = addr >> PAGE_SHIFT;
1977         int error;
1978
1979         len = PAGE_ALIGN(len);
1980         if (!len)
1981                 return addr;
1982
1983         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1984                 return -EINVAL;
1985
1986         if (is_hugepage_only_range(mm, addr, len))
1987                 return -EINVAL;
1988
1989         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1990         if (error)
1991                 return error;
1992
1993         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1994
1995         error = arch_mmap_check(addr, len, flags);
1996         if (error)
1997                 return error;
1998
1999         /*
2000          * mlock MCL_FUTURE?
2001          */
2002         if (mm->def_flags & VM_LOCKED) {
2003                 unsigned long locked, lock_limit;
2004                 locked = len >> PAGE_SHIFT;
2005                 locked += mm->locked_vm;
2006                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2007                 lock_limit >>= PAGE_SHIFT;
2008                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2009                         return -EAGAIN;
2010         }
2011
2012         /*
2013          * mm->mmap_sem is required to protect against another thread
2014          * changing the mappings in case we sleep.
2015          */
2016         verify_mm_writelocked(mm);
2017
2018         /*
2019          * Clear old maps.  this also does some error checking for us
2020          */
2021  munmap_back:
2022         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2023         if (vma && vma->vm_start < addr + len) {
2024                 if (do_munmap(mm, addr, len))
2025                         return -ENOMEM;
2026                 goto munmap_back;
2027         }
2028
2029         /* Check against address space limits *after* clearing old maps... */
2030         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2031                 return -ENOMEM;
2032
2033         if (mm->map_count > sysctl_max_map_count)
2034                 return -ENOMEM;
2035
2036         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2037                 return -ENOMEM;
2038
2039         /* Can we just expand an old private anonymous mapping? */
2040         vma = vma_merge(mm, prev, addr, addr + len, flags,
2041                                         NULL, NULL, pgoff, NULL);
2042         if (vma)
2043                 goto out;
2044
2045         /*
2046          * create a vma struct for an anonymous mapping
2047          */
2048         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2049         if (!vma) {
2050                 vm_unacct_memory(len >> PAGE_SHIFT);
2051                 return -ENOMEM;
2052         }
2053
2054         vma->vm_mm = mm;
2055         vma->vm_start = addr;
2056         vma->vm_end = addr + len;
2057         vma->vm_pgoff = pgoff;
2058         vma->vm_flags = flags;
2059         vma->vm_page_prot = vm_get_page_prot(flags);
2060         vma_link(mm, vma, prev, rb_link, rb_parent);
2061 out:
2062         mm->total_vm += len >> PAGE_SHIFT;
2063         if (flags & VM_LOCKED) {
2064                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2065                         mm->locked_vm += (len >> PAGE_SHIFT);
2066         }
2067         return addr;
2068 }
2069
2070 EXPORT_SYMBOL(do_brk);
2071
2072 /* Release all mmaps. */
2073 void exit_mmap(struct mm_struct *mm)
2074 {
2075         struct mmu_gather *tlb;
2076         struct vm_area_struct *vma;
2077         unsigned long nr_accounted = 0;
2078         unsigned long end;
2079
2080         /* mm's last user has gone, and its about to be pulled down */
2081         arch_exit_mmap(mm);
2082         mmu_notifier_release(mm);
2083
2084         if (!mm->mmap)  /* Can happen if dup_mmap() received an OOM */
2085                 return;
2086
2087         if (mm->locked_vm) {
2088                 vma = mm->mmap;
2089                 while (vma) {
2090                         if (vma->vm_flags & VM_LOCKED)
2091                                 munlock_vma_pages_all(vma);
2092                         vma = vma->vm_next;
2093                 }
2094         }
2095         vma = mm->mmap;
2096         lru_add_drain();
2097         flush_cache_mm(mm);
2098         tlb = tlb_gather_mmu(mm, 1);
2099         /* update_hiwater_rss(mm) here? but nobody should be looking */
2100         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2101         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2102         vm_unacct_memory(nr_accounted);
2103         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2104         tlb_finish_mmu(tlb, 0, end);
2105
2106         /*
2107          * Walk the list again, actually closing and freeing it,
2108          * with preemption enabled, without holding any MM locks.
2109          */
2110         while (vma)
2111                 vma = remove_vma(vma);
2112
2113         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2114 }
2115
2116 /* Insert vm structure into process list sorted by address
2117  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2118  * then i_mmap_lock is taken here.
2119  */
2120 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2121 {
2122         struct vm_area_struct * __vma, * prev;
2123         struct rb_node ** rb_link, * rb_parent;
2124
2125         /*
2126          * The vm_pgoff of a purely anonymous vma should be irrelevant
2127          * until its first write fault, when page's anon_vma and index
2128          * are set.  But now set the vm_pgoff it will almost certainly
2129          * end up with (unless mremap moves it elsewhere before that
2130          * first wfault), so /proc/pid/maps tells a consistent story.
2131          *
2132          * By setting it to reflect the virtual start address of the
2133          * vma, merges and splits can happen in a seamless way, just
2134          * using the existing file pgoff checks and manipulations.
2135          * Similarly in do_mmap_pgoff and in do_brk.
2136          */
2137         if (!vma->vm_file) {
2138                 BUG_ON(vma->anon_vma);
2139                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2140         }
2141         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2142         if (__vma && __vma->vm_start < vma->vm_end)
2143                 return -ENOMEM;
2144         if ((vma->vm_flags & VM_ACCOUNT) &&
2145              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2146                 return -ENOMEM;
2147         vma_link(mm, vma, prev, rb_link, rb_parent);
2148         return 0;
2149 }
2150
2151 /*
2152  * Copy the vma structure to a new location in the same mm,
2153  * prior to moving page table entries, to effect an mremap move.
2154  */
2155 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2156         unsigned long addr, unsigned long len, pgoff_t pgoff)
2157 {
2158         struct vm_area_struct *vma = *vmap;
2159         unsigned long vma_start = vma->vm_start;
2160         struct mm_struct *mm = vma->vm_mm;
2161         struct vm_area_struct *new_vma, *prev;
2162         struct rb_node **rb_link, *rb_parent;
2163         struct mempolicy *pol;
2164
2165         /*
2166          * If anonymous vma has not yet been faulted, update new pgoff
2167          * to match new location, to increase its chance of merging.
2168          */
2169         if (!vma->vm_file && !vma->anon_vma)
2170                 pgoff = addr >> PAGE_SHIFT;
2171
2172         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2173         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2174                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2175         if (new_vma) {
2176                 /*
2177                  * Source vma may have been merged into new_vma
2178                  */
2179                 if (vma_start >= new_vma->vm_start &&
2180                     vma_start < new_vma->vm_end)
2181                         *vmap = new_vma;
2182         } else {
2183                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2184                 if (new_vma) {
2185                         *new_vma = *vma;
2186                         pol = mpol_dup(vma_policy(vma));
2187                         if (IS_ERR(pol)) {
2188                                 kmem_cache_free(vm_area_cachep, new_vma);
2189                                 return NULL;
2190                         }
2191                         vma_set_policy(new_vma, pol);
2192                         new_vma->vm_start = addr;
2193                         new_vma->vm_end = addr + len;
2194                         new_vma->vm_pgoff = pgoff;
2195                         if (new_vma->vm_file) {
2196                                 get_file(new_vma->vm_file);
2197                                 if (vma->vm_flags & VM_EXECUTABLE)
2198                                         added_exe_file_vma(mm);
2199                         }
2200                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2201                                 new_vma->vm_ops->open(new_vma);
2202                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2203                 }
2204         }
2205         return new_vma;
2206 }
2207
2208 /*
2209  * Return true if the calling process may expand its vm space by the passed
2210  * number of pages
2211  */
2212 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2213 {
2214         unsigned long cur = mm->total_vm;       /* pages */
2215         unsigned long lim;
2216
2217         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2218
2219         if (cur + npages > lim)
2220                 return 0;
2221         return 1;
2222 }
2223
2224
2225 static int special_mapping_fault(struct vm_area_struct *vma,
2226                                 struct vm_fault *vmf)
2227 {
2228         pgoff_t pgoff;
2229         struct page **pages;
2230
2231         /*
2232          * special mappings have no vm_file, and in that case, the mm
2233          * uses vm_pgoff internally. So we have to subtract it from here.
2234          * We are allowed to do this because we are the mm; do not copy
2235          * this code into drivers!
2236          */
2237         pgoff = vmf->pgoff - vma->vm_pgoff;
2238
2239         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2240                 pgoff--;
2241
2242         if (*pages) {
2243                 struct page *page = *pages;
2244                 get_page(page);
2245                 vmf->page = page;
2246                 return 0;
2247         }
2248
2249         return VM_FAULT_SIGBUS;
2250 }
2251
2252 /*
2253  * Having a close hook prevents vma merging regardless of flags.
2254  */
2255 static void special_mapping_close(struct vm_area_struct *vma)
2256 {
2257 }
2258
2259 static struct vm_operations_struct special_mapping_vmops = {
2260         .close = special_mapping_close,
2261         .fault = special_mapping_fault,
2262 };
2263
2264 /*
2265  * Called with mm->mmap_sem held for writing.
2266  * Insert a new vma covering the given region, with the given flags.
2267  * Its pages are supplied by the given array of struct page *.
2268  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2269  * The region past the last page supplied will always produce SIGBUS.
2270  * The array pointer and the pages it points to are assumed to stay alive
2271  * for as long as this mapping might exist.
2272  */
2273 int install_special_mapping(struct mm_struct *mm,
2274                             unsigned long addr, unsigned long len,
2275                             unsigned long vm_flags, struct page **pages)
2276 {
2277         struct vm_area_struct *vma;
2278
2279         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2280         if (unlikely(vma == NULL))
2281                 return -ENOMEM;
2282
2283         vma->vm_mm = mm;
2284         vma->vm_start = addr;
2285         vma->vm_end = addr + len;
2286
2287         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2288         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2289
2290         vma->vm_ops = &special_mapping_vmops;
2291         vma->vm_private_data = pages;
2292
2293         if (unlikely(insert_vm_struct(mm, vma))) {
2294                 kmem_cache_free(vm_area_cachep, vma);
2295                 return -ENOMEM;
2296         }
2297
2298         mm->total_vm += len >> PAGE_SHIFT;
2299
2300         return 0;
2301 }
2302
2303 static DEFINE_MUTEX(mm_all_locks_mutex);
2304
2305 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2306 {
2307         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2308                 /*
2309                  * The LSB of head.next can't change from under us
2310                  * because we hold the mm_all_locks_mutex.
2311                  */
2312                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2313                 /*
2314                  * We can safely modify head.next after taking the
2315                  * anon_vma->lock. If some other vma in this mm shares
2316                  * the same anon_vma we won't take it again.
2317                  *
2318                  * No need of atomic instructions here, head.next
2319                  * can't change from under us thanks to the
2320                  * anon_vma->lock.
2321                  */
2322                 if (__test_and_set_bit(0, (unsigned long *)
2323                                        &anon_vma->head.next))
2324                         BUG();
2325         }
2326 }
2327
2328 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2329 {
2330         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2331                 /*
2332                  * AS_MM_ALL_LOCKS can't change from under us because
2333                  * we hold the mm_all_locks_mutex.
2334                  *
2335                  * Operations on ->flags have to be atomic because
2336                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2337                  * mm_all_locks_mutex, there may be other cpus
2338                  * changing other bitflags in parallel to us.
2339                  */
2340                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2341                         BUG();
2342                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2343         }
2344 }
2345
2346 /*
2347  * This operation locks against the VM for all pte/vma/mm related
2348  * operations that could ever happen on a certain mm. This includes
2349  * vmtruncate, try_to_unmap, and all page faults.
2350  *
2351  * The caller must take the mmap_sem in write mode before calling
2352  * mm_take_all_locks(). The caller isn't allowed to release the
2353  * mmap_sem until mm_drop_all_locks() returns.
2354  *
2355  * mmap_sem in write mode is required in order to block all operations
2356  * that could modify pagetables and free pages without need of
2357  * altering the vma layout (for example populate_range() with
2358  * nonlinear vmas). It's also needed in write mode to avoid new
2359  * anon_vmas to be associated with existing vmas.
2360  *
2361  * A single task can't take more than one mm_take_all_locks() in a row
2362  * or it would deadlock.
2363  *
2364  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2365  * mapping->flags avoid to take the same lock twice, if more than one
2366  * vma in this mm is backed by the same anon_vma or address_space.
2367  *
2368  * We can take all the locks in random order because the VM code
2369  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2370  * takes more than one of them in a row. Secondly we're protected
2371  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2372  *
2373  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2374  * that may have to take thousand of locks.
2375  *
2376  * mm_take_all_locks() can fail if it's interrupted by signals.
2377  */
2378 int mm_take_all_locks(struct mm_struct *mm)
2379 {
2380         struct vm_area_struct *vma;
2381         int ret = -EINTR;
2382
2383         BUG_ON(down_read_trylock(&mm->mmap_sem));
2384
2385         mutex_lock(&mm_all_locks_mutex);
2386
2387         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2388                 if (signal_pending(current))
2389                         goto out_unlock;
2390                 if (vma->vm_file && vma->vm_file->f_mapping)
2391                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2392         }
2393
2394         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2395                 if (signal_pending(current))
2396                         goto out_unlock;
2397                 if (vma->anon_vma)
2398                         vm_lock_anon_vma(mm, vma->anon_vma);
2399         }
2400
2401         ret = 0;
2402
2403 out_unlock:
2404         if (ret)
2405                 mm_drop_all_locks(mm);
2406
2407         return ret;
2408 }
2409
2410 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2411 {
2412         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2413                 /*
2414                  * The LSB of head.next can't change to 0 from under
2415                  * us because we hold the mm_all_locks_mutex.
2416                  *
2417                  * We must however clear the bitflag before unlocking
2418                  * the vma so the users using the anon_vma->head will
2419                  * never see our bitflag.
2420                  *
2421                  * No need of atomic instructions here, head.next
2422                  * can't change from under us until we release the
2423                  * anon_vma->lock.
2424                  */
2425                 if (!__test_and_clear_bit(0, (unsigned long *)
2426                                           &anon_vma->head.next))
2427                         BUG();
2428                 spin_unlock(&anon_vma->lock);
2429         }
2430 }
2431
2432 static void vm_unlock_mapping(struct address_space *mapping)
2433 {
2434         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2435                 /*
2436                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2437                  * because we hold the mm_all_locks_mutex.
2438                  */
2439                 spin_unlock(&mapping->i_mmap_lock);
2440                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2441                                         &mapping->flags))
2442                         BUG();
2443         }
2444 }
2445
2446 /*
2447  * The mmap_sem cannot be released by the caller until
2448  * mm_drop_all_locks() returns.
2449  */
2450 void mm_drop_all_locks(struct mm_struct *mm)
2451 {
2452         struct vm_area_struct *vma;
2453
2454         BUG_ON(down_read_trylock(&mm->mmap_sem));
2455         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2456
2457         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2458                 if (vma->anon_vma)
2459                         vm_unlock_anon_vma(vma->anon_vma);
2460                 if (vma->vm_file && vma->vm_file->f_mapping)
2461                         vm_unlock_mapping(vma->vm_file->f_mapping);
2462         }
2463
2464         mutex_unlock(&mm_all_locks_mutex);
2465 }
2466
2467 /*
2468  * initialise the VMA slab
2469  */
2470 void __init mmap_init(void)
2471 {
2472         vm_area_cachep = kmem_cache_create("vm_area_struct",
2473                         sizeof(struct vm_area_struct), 0,
2474                         SLAB_PANIC, NULL);
2475 }