Merge branch 'linus' into perfcounters/core
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/ima.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/module.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_counter.h>
32
33 #include <asm/uaccess.h>
34 #include <asm/cacheflush.h>
35 #include <asm/tlb.h>
36 #include <asm/mmu_context.h>
37
38 #include "internal.h"
39
40 #ifndef arch_mmap_check
41 #define arch_mmap_check(addr, len, flags)       (0)
42 #endif
43
44 #ifndef arch_rebalance_pgtables
45 #define arch_rebalance_pgtables(addr, len)              (addr)
46 #endif
47
48 static void unmap_region(struct mm_struct *mm,
49                 struct vm_area_struct *vma, struct vm_area_struct *prev,
50                 unsigned long start, unsigned long end);
51
52 /*
53  * WARNING: the debugging will use recursive algorithms so never enable this
54  * unless you know what you are doing.
55  */
56 #undef DEBUG_MM_RB
57
58 /* description of effects of mapping type and prot in current implementation.
59  * this is due to the limited x86 page protection hardware.  The expected
60  * behavior is in parens:
61  *
62  * map_type     prot
63  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
64  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
65  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
66  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
67  *              
68  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
69  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
70  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
71  *
72  */
73 pgprot_t protection_map[16] = {
74         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
75         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
76 };
77
78 pgprot_t vm_get_page_prot(unsigned long vm_flags)
79 {
80         return __pgprot(pgprot_val(protection_map[vm_flags &
81                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
82                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
83 }
84 EXPORT_SYMBOL(vm_get_page_prot);
85
86 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
87 int sysctl_overcommit_ratio = 50;       /* default is 50% */
88 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
89 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
90
91 /*
92  * Check that a process has enough memory to allocate a new virtual
93  * mapping. 0 means there is enough memory for the allocation to
94  * succeed and -ENOMEM implies there is not.
95  *
96  * We currently support three overcommit policies, which are set via the
97  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
98  *
99  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
100  * Additional code 2002 Jul 20 by Robert Love.
101  *
102  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
103  *
104  * Note this is a helper function intended to be used by LSMs which
105  * wish to use this logic.
106  */
107 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
108 {
109         unsigned long free, allowed;
110
111         vm_acct_memory(pages);
112
113         /*
114          * Sometimes we want to use more memory than we have
115          */
116         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
117                 return 0;
118
119         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
120                 unsigned long n;
121
122                 free = global_page_state(NR_FILE_PAGES);
123                 free += nr_swap_pages;
124
125                 /*
126                  * Any slabs which are created with the
127                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
128                  * which are reclaimable, under pressure.  The dentry
129                  * cache and most inode caches should fall into this
130                  */
131                 free += global_page_state(NR_SLAB_RECLAIMABLE);
132
133                 /*
134                  * Leave the last 3% for root
135                  */
136                 if (!cap_sys_admin)
137                         free -= free / 32;
138
139                 if (free > pages)
140                         return 0;
141
142                 /*
143                  * nr_free_pages() is very expensive on large systems,
144                  * only call if we're about to fail.
145                  */
146                 n = nr_free_pages();
147
148                 /*
149                  * Leave reserved pages. The pages are not for anonymous pages.
150                  */
151                 if (n <= totalreserve_pages)
152                         goto error;
153                 else
154                         n -= totalreserve_pages;
155
156                 /*
157                  * Leave the last 3% for root
158                  */
159                 if (!cap_sys_admin)
160                         n -= n / 32;
161                 free += n;
162
163                 if (free > pages)
164                         return 0;
165
166                 goto error;
167         }
168
169         allowed = (totalram_pages - hugetlb_total_pages())
170                 * sysctl_overcommit_ratio / 100;
171         /*
172          * Leave the last 3% for root
173          */
174         if (!cap_sys_admin)
175                 allowed -= allowed / 32;
176         allowed += total_swap_pages;
177
178         /* Don't let a single process grow too big:
179            leave 3% of the size of this process for other processes */
180         if (mm)
181                 allowed -= mm->total_vm / 32;
182
183         /*
184          * cast `allowed' as a signed long because vm_committed_space
185          * sometimes has a negative value
186          */
187         if (atomic_long_read(&vm_committed_space) < (long)allowed)
188                 return 0;
189 error:
190         vm_unacct_memory(pages);
191
192         return -ENOMEM;
193 }
194
195 /*
196  * Requires inode->i_mapping->i_mmap_lock
197  */
198 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
199                 struct file *file, struct address_space *mapping)
200 {
201         if (vma->vm_flags & VM_DENYWRITE)
202                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
203         if (vma->vm_flags & VM_SHARED)
204                 mapping->i_mmap_writable--;
205
206         flush_dcache_mmap_lock(mapping);
207         if (unlikely(vma->vm_flags & VM_NONLINEAR))
208                 list_del_init(&vma->shared.vm_set.list);
209         else
210                 vma_prio_tree_remove(vma, &mapping->i_mmap);
211         flush_dcache_mmap_unlock(mapping);
212 }
213
214 /*
215  * Unlink a file-based vm structure from its prio_tree, to hide
216  * vma from rmap and vmtruncate before freeing its page tables.
217  */
218 void unlink_file_vma(struct vm_area_struct *vma)
219 {
220         struct file *file = vma->vm_file;
221
222         if (file) {
223                 struct address_space *mapping = file->f_mapping;
224                 spin_lock(&mapping->i_mmap_lock);
225                 __remove_shared_vm_struct(vma, file, mapping);
226                 spin_unlock(&mapping->i_mmap_lock);
227         }
228 }
229
230 /*
231  * Close a vm structure and free it, returning the next.
232  */
233 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
234 {
235         struct vm_area_struct *next = vma->vm_next;
236
237         might_sleep();
238         if (vma->vm_ops && vma->vm_ops->close)
239                 vma->vm_ops->close(vma);
240         if (vma->vm_file) {
241                 fput(vma->vm_file);
242                 if (vma->vm_flags & VM_EXECUTABLE)
243                         removed_exe_file_vma(vma->vm_mm);
244         }
245         mpol_put(vma_policy(vma));
246         kmem_cache_free(vm_area_cachep, vma);
247         return next;
248 }
249
250 SYSCALL_DEFINE1(brk, unsigned long, brk)
251 {
252         unsigned long rlim, retval;
253         unsigned long newbrk, oldbrk;
254         struct mm_struct *mm = current->mm;
255         unsigned long min_brk;
256
257         down_write(&mm->mmap_sem);
258
259 #ifdef CONFIG_COMPAT_BRK
260         min_brk = mm->end_code;
261 #else
262         min_brk = mm->start_brk;
263 #endif
264         if (brk < min_brk)
265                 goto out;
266
267         /*
268          * Check against rlimit here. If this check is done later after the test
269          * of oldbrk with newbrk then it can escape the test and let the data
270          * segment grow beyond its set limit the in case where the limit is
271          * not page aligned -Ram Gupta
272          */
273         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
274         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
275                         (mm->end_data - mm->start_data) > rlim)
276                 goto out;
277
278         newbrk = PAGE_ALIGN(brk);
279         oldbrk = PAGE_ALIGN(mm->brk);
280         if (oldbrk == newbrk)
281                 goto set_brk;
282
283         /* Always allow shrinking brk. */
284         if (brk <= mm->brk) {
285                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
286                         goto set_brk;
287                 goto out;
288         }
289
290         /* Check against existing mmap mappings. */
291         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
292                 goto out;
293
294         /* Ok, looks good - let it rip. */
295         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
296                 goto out;
297 set_brk:
298         mm->brk = brk;
299 out:
300         retval = mm->brk;
301         up_write(&mm->mmap_sem);
302         return retval;
303 }
304
305 #ifdef DEBUG_MM_RB
306 static int browse_rb(struct rb_root *root)
307 {
308         int i = 0, j;
309         struct rb_node *nd, *pn = NULL;
310         unsigned long prev = 0, pend = 0;
311
312         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
313                 struct vm_area_struct *vma;
314                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
315                 if (vma->vm_start < prev)
316                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
317                 if (vma->vm_start < pend)
318                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
319                 if (vma->vm_start > vma->vm_end)
320                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
321                 i++;
322                 pn = nd;
323                 prev = vma->vm_start;
324                 pend = vma->vm_end;
325         }
326         j = 0;
327         for (nd = pn; nd; nd = rb_prev(nd)) {
328                 j++;
329         }
330         if (i != j)
331                 printk("backwards %d, forwards %d\n", j, i), i = 0;
332         return i;
333 }
334
335 void validate_mm(struct mm_struct *mm)
336 {
337         int bug = 0;
338         int i = 0;
339         struct vm_area_struct *tmp = mm->mmap;
340         while (tmp) {
341                 tmp = tmp->vm_next;
342                 i++;
343         }
344         if (i != mm->map_count)
345                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
346         i = browse_rb(&mm->mm_rb);
347         if (i != mm->map_count)
348                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
349         BUG_ON(bug);
350 }
351 #else
352 #define validate_mm(mm) do { } while (0)
353 #endif
354
355 static struct vm_area_struct *
356 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
357                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
358                 struct rb_node ** rb_parent)
359 {
360         struct vm_area_struct * vma;
361         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
362
363         __rb_link = &mm->mm_rb.rb_node;
364         rb_prev = __rb_parent = NULL;
365         vma = NULL;
366
367         while (*__rb_link) {
368                 struct vm_area_struct *vma_tmp;
369
370                 __rb_parent = *__rb_link;
371                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
372
373                 if (vma_tmp->vm_end > addr) {
374                         vma = vma_tmp;
375                         if (vma_tmp->vm_start <= addr)
376                                 break;
377                         __rb_link = &__rb_parent->rb_left;
378                 } else {
379                         rb_prev = __rb_parent;
380                         __rb_link = &__rb_parent->rb_right;
381                 }
382         }
383
384         *pprev = NULL;
385         if (rb_prev)
386                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
387         *rb_link = __rb_link;
388         *rb_parent = __rb_parent;
389         return vma;
390 }
391
392 static inline void
393 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
394                 struct vm_area_struct *prev, struct rb_node *rb_parent)
395 {
396         if (prev) {
397                 vma->vm_next = prev->vm_next;
398                 prev->vm_next = vma;
399         } else {
400                 mm->mmap = vma;
401                 if (rb_parent)
402                         vma->vm_next = rb_entry(rb_parent,
403                                         struct vm_area_struct, vm_rb);
404                 else
405                         vma->vm_next = NULL;
406         }
407 }
408
409 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
410                 struct rb_node **rb_link, struct rb_node *rb_parent)
411 {
412         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
413         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
414 }
415
416 static void __vma_link_file(struct vm_area_struct *vma)
417 {
418         struct file *file;
419
420         file = vma->vm_file;
421         if (file) {
422                 struct address_space *mapping = file->f_mapping;
423
424                 if (vma->vm_flags & VM_DENYWRITE)
425                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
426                 if (vma->vm_flags & VM_SHARED)
427                         mapping->i_mmap_writable++;
428
429                 flush_dcache_mmap_lock(mapping);
430                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
431                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
432                 else
433                         vma_prio_tree_insert(vma, &mapping->i_mmap);
434                 flush_dcache_mmap_unlock(mapping);
435         }
436 }
437
438 static void
439 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
440         struct vm_area_struct *prev, struct rb_node **rb_link,
441         struct rb_node *rb_parent)
442 {
443         __vma_link_list(mm, vma, prev, rb_parent);
444         __vma_link_rb(mm, vma, rb_link, rb_parent);
445         __anon_vma_link(vma);
446 }
447
448 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
449                         struct vm_area_struct *prev, struct rb_node **rb_link,
450                         struct rb_node *rb_parent)
451 {
452         struct address_space *mapping = NULL;
453
454         if (vma->vm_file)
455                 mapping = vma->vm_file->f_mapping;
456
457         if (mapping) {
458                 spin_lock(&mapping->i_mmap_lock);
459                 vma->vm_truncate_count = mapping->truncate_count;
460         }
461         anon_vma_lock(vma);
462
463         __vma_link(mm, vma, prev, rb_link, rb_parent);
464         __vma_link_file(vma);
465
466         anon_vma_unlock(vma);
467         if (mapping)
468                 spin_unlock(&mapping->i_mmap_lock);
469
470         mm->map_count++;
471         validate_mm(mm);
472 }
473
474 /*
475  * Helper for vma_adjust in the split_vma insert case:
476  * insert vm structure into list and rbtree and anon_vma,
477  * but it has already been inserted into prio_tree earlier.
478  */
479 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
480 {
481         struct vm_area_struct *__vma, *prev;
482         struct rb_node **rb_link, *rb_parent;
483
484         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
485         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
486         __vma_link(mm, vma, prev, rb_link, rb_parent);
487         mm->map_count++;
488 }
489
490 static inline void
491 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
492                 struct vm_area_struct *prev)
493 {
494         prev->vm_next = vma->vm_next;
495         rb_erase(&vma->vm_rb, &mm->mm_rb);
496         if (mm->mmap_cache == vma)
497                 mm->mmap_cache = prev;
498 }
499
500 /*
501  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
502  * is already present in an i_mmap tree without adjusting the tree.
503  * The following helper function should be used when such adjustments
504  * are necessary.  The "insert" vma (if any) is to be inserted
505  * before we drop the necessary locks.
506  */
507 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
508         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
509 {
510         struct mm_struct *mm = vma->vm_mm;
511         struct vm_area_struct *next = vma->vm_next;
512         struct vm_area_struct *importer = NULL;
513         struct address_space *mapping = NULL;
514         struct prio_tree_root *root = NULL;
515         struct file *file = vma->vm_file;
516         struct anon_vma *anon_vma = NULL;
517         long adjust_next = 0;
518         int remove_next = 0;
519
520         if (next && !insert) {
521                 if (end >= next->vm_end) {
522                         /*
523                          * vma expands, overlapping all the next, and
524                          * perhaps the one after too (mprotect case 6).
525                          */
526 again:                  remove_next = 1 + (end > next->vm_end);
527                         end = next->vm_end;
528                         anon_vma = next->anon_vma;
529                         importer = vma;
530                 } else if (end > next->vm_start) {
531                         /*
532                          * vma expands, overlapping part of the next:
533                          * mprotect case 5 shifting the boundary up.
534                          */
535                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
536                         anon_vma = next->anon_vma;
537                         importer = vma;
538                 } else if (end < vma->vm_end) {
539                         /*
540                          * vma shrinks, and !insert tells it's not
541                          * split_vma inserting another: so it must be
542                          * mprotect case 4 shifting the boundary down.
543                          */
544                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
545                         anon_vma = next->anon_vma;
546                         importer = next;
547                 }
548         }
549
550         if (file) {
551                 mapping = file->f_mapping;
552                 if (!(vma->vm_flags & VM_NONLINEAR))
553                         root = &mapping->i_mmap;
554                 spin_lock(&mapping->i_mmap_lock);
555                 if (importer &&
556                     vma->vm_truncate_count != next->vm_truncate_count) {
557                         /*
558                          * unmap_mapping_range might be in progress:
559                          * ensure that the expanding vma is rescanned.
560                          */
561                         importer->vm_truncate_count = 0;
562                 }
563                 if (insert) {
564                         insert->vm_truncate_count = vma->vm_truncate_count;
565                         /*
566                          * Put into prio_tree now, so instantiated pages
567                          * are visible to arm/parisc __flush_dcache_page
568                          * throughout; but we cannot insert into address
569                          * space until vma start or end is updated.
570                          */
571                         __vma_link_file(insert);
572                 }
573         }
574
575         /*
576          * When changing only vma->vm_end, we don't really need
577          * anon_vma lock: but is that case worth optimizing out?
578          */
579         if (vma->anon_vma)
580                 anon_vma = vma->anon_vma;
581         if (anon_vma) {
582                 spin_lock(&anon_vma->lock);
583                 /*
584                  * Easily overlooked: when mprotect shifts the boundary,
585                  * make sure the expanding vma has anon_vma set if the
586                  * shrinking vma had, to cover any anon pages imported.
587                  */
588                 if (importer && !importer->anon_vma) {
589                         importer->anon_vma = anon_vma;
590                         __anon_vma_link(importer);
591                 }
592         }
593
594         if (root) {
595                 flush_dcache_mmap_lock(mapping);
596                 vma_prio_tree_remove(vma, root);
597                 if (adjust_next)
598                         vma_prio_tree_remove(next, root);
599         }
600
601         vma->vm_start = start;
602         vma->vm_end = end;
603         vma->vm_pgoff = pgoff;
604         if (adjust_next) {
605                 next->vm_start += adjust_next << PAGE_SHIFT;
606                 next->vm_pgoff += adjust_next;
607         }
608
609         if (root) {
610                 if (adjust_next)
611                         vma_prio_tree_insert(next, root);
612                 vma_prio_tree_insert(vma, root);
613                 flush_dcache_mmap_unlock(mapping);
614         }
615
616         if (remove_next) {
617                 /*
618                  * vma_merge has merged next into vma, and needs
619                  * us to remove next before dropping the locks.
620                  */
621                 __vma_unlink(mm, next, vma);
622                 if (file)
623                         __remove_shared_vm_struct(next, file, mapping);
624                 if (next->anon_vma)
625                         __anon_vma_merge(vma, next);
626         } else if (insert) {
627                 /*
628                  * split_vma has split insert from vma, and needs
629                  * us to insert it before dropping the locks
630                  * (it may either follow vma or precede it).
631                  */
632                 __insert_vm_struct(mm, insert);
633         }
634
635         if (anon_vma)
636                 spin_unlock(&anon_vma->lock);
637         if (mapping)
638                 spin_unlock(&mapping->i_mmap_lock);
639
640         if (remove_next) {
641                 if (file) {
642                         fput(file);
643                         if (next->vm_flags & VM_EXECUTABLE)
644                                 removed_exe_file_vma(mm);
645                 }
646                 mm->map_count--;
647                 mpol_put(vma_policy(next));
648                 kmem_cache_free(vm_area_cachep, next);
649                 /*
650                  * In mprotect's case 6 (see comments on vma_merge),
651                  * we must remove another next too. It would clutter
652                  * up the code too much to do both in one go.
653                  */
654                 if (remove_next == 2) {
655                         next = vma->vm_next;
656                         goto again;
657                 }
658         }
659
660         validate_mm(mm);
661 }
662
663 /* Flags that can be inherited from an existing mapping when merging */
664 #define VM_MERGEABLE_FLAGS (VM_CAN_NONLINEAR)
665
666 /*
667  * If the vma has a ->close operation then the driver probably needs to release
668  * per-vma resources, so we don't attempt to merge those.
669  */
670 static inline int is_mergeable_vma(struct vm_area_struct *vma,
671                         struct file *file, unsigned long vm_flags)
672 {
673         if ((vma->vm_flags ^ vm_flags) & ~VM_MERGEABLE_FLAGS)
674                 return 0;
675         if (vma->vm_file != file)
676                 return 0;
677         if (vma->vm_ops && vma->vm_ops->close)
678                 return 0;
679         return 1;
680 }
681
682 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
683                                         struct anon_vma *anon_vma2)
684 {
685         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
686 }
687
688 /*
689  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
690  * in front of (at a lower virtual address and file offset than) the vma.
691  *
692  * We cannot merge two vmas if they have differently assigned (non-NULL)
693  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
694  *
695  * We don't check here for the merged mmap wrapping around the end of pagecache
696  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
697  * wrap, nor mmaps which cover the final page at index -1UL.
698  */
699 static int
700 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
701         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
702 {
703         if (is_mergeable_vma(vma, file, vm_flags) &&
704             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
705                 if (vma->vm_pgoff == vm_pgoff)
706                         return 1;
707         }
708         return 0;
709 }
710
711 /*
712  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
713  * beyond (at a higher virtual address and file offset than) the vma.
714  *
715  * We cannot merge two vmas if they have differently assigned (non-NULL)
716  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
717  */
718 static int
719 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
720         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
721 {
722         if (is_mergeable_vma(vma, file, vm_flags) &&
723             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
724                 pgoff_t vm_pglen;
725                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
726                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
727                         return 1;
728         }
729         return 0;
730 }
731
732 /*
733  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
734  * whether that can be merged with its predecessor or its successor.
735  * Or both (it neatly fills a hole).
736  *
737  * In most cases - when called for mmap, brk or mremap - [addr,end) is
738  * certain not to be mapped by the time vma_merge is called; but when
739  * called for mprotect, it is certain to be already mapped (either at
740  * an offset within prev, or at the start of next), and the flags of
741  * this area are about to be changed to vm_flags - and the no-change
742  * case has already been eliminated.
743  *
744  * The following mprotect cases have to be considered, where AAAA is
745  * the area passed down from mprotect_fixup, never extending beyond one
746  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
747  *
748  *     AAAA             AAAA                AAAA          AAAA
749  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
750  *    cannot merge    might become    might become    might become
751  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
752  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
753  *    mremap move:                                    PPPPNNNNNNNN 8
754  *        AAAA
755  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
756  *    might become    case 1 below    case 2 below    case 3 below
757  *
758  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
759  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
760  */
761 struct vm_area_struct *vma_merge(struct mm_struct *mm,
762                         struct vm_area_struct *prev, unsigned long addr,
763                         unsigned long end, unsigned long vm_flags,
764                         struct anon_vma *anon_vma, struct file *file,
765                         pgoff_t pgoff, struct mempolicy *policy)
766 {
767         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
768         struct vm_area_struct *area, *next;
769
770         /*
771          * We later require that vma->vm_flags == vm_flags,
772          * so this tests vma->vm_flags & VM_SPECIAL, too.
773          */
774         if (vm_flags & VM_SPECIAL)
775                 return NULL;
776
777         if (prev)
778                 next = prev->vm_next;
779         else
780                 next = mm->mmap;
781         area = next;
782         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
783                 next = next->vm_next;
784
785         /*
786          * Can it merge with the predecessor?
787          */
788         if (prev && prev->vm_end == addr &&
789                         mpol_equal(vma_policy(prev), policy) &&
790                         can_vma_merge_after(prev, vm_flags,
791                                                 anon_vma, file, pgoff)) {
792                 /*
793                  * OK, it can.  Can we now merge in the successor as well?
794                  */
795                 if (next && end == next->vm_start &&
796                                 mpol_equal(policy, vma_policy(next)) &&
797                                 can_vma_merge_before(next, vm_flags,
798                                         anon_vma, file, pgoff+pglen) &&
799                                 is_mergeable_anon_vma(prev->anon_vma,
800                                                       next->anon_vma)) {
801                                                         /* cases 1, 6 */
802                         vma_adjust(prev, prev->vm_start,
803                                 next->vm_end, prev->vm_pgoff, NULL);
804                 } else                                  /* cases 2, 5, 7 */
805                         vma_adjust(prev, prev->vm_start,
806                                 end, prev->vm_pgoff, NULL);
807                 return prev;
808         }
809
810         /*
811          * Can this new request be merged in front of next?
812          */
813         if (next && end == next->vm_start &&
814                         mpol_equal(policy, vma_policy(next)) &&
815                         can_vma_merge_before(next, vm_flags,
816                                         anon_vma, file, pgoff+pglen)) {
817                 if (prev && addr < prev->vm_end)        /* case 4 */
818                         vma_adjust(prev, prev->vm_start,
819                                 addr, prev->vm_pgoff, NULL);
820                 else                                    /* cases 3, 8 */
821                         vma_adjust(area, addr, next->vm_end,
822                                 next->vm_pgoff - pglen, NULL);
823                 return area;
824         }
825
826         return NULL;
827 }
828
829 /*
830  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
831  * neighbouring vmas for a suitable anon_vma, before it goes off
832  * to allocate a new anon_vma.  It checks because a repetitive
833  * sequence of mprotects and faults may otherwise lead to distinct
834  * anon_vmas being allocated, preventing vma merge in subsequent
835  * mprotect.
836  */
837 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
838 {
839         struct vm_area_struct *near;
840         unsigned long vm_flags;
841
842         near = vma->vm_next;
843         if (!near)
844                 goto try_prev;
845
846         /*
847          * Since only mprotect tries to remerge vmas, match flags
848          * which might be mprotected into each other later on.
849          * Neither mlock nor madvise tries to remerge at present,
850          * so leave their flags as obstructing a merge.
851          */
852         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
853         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
854
855         if (near->anon_vma && vma->vm_end == near->vm_start &&
856                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
857                         can_vma_merge_before(near, vm_flags,
858                                 NULL, vma->vm_file, vma->vm_pgoff +
859                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
860                 return near->anon_vma;
861 try_prev:
862         /*
863          * It is potentially slow to have to call find_vma_prev here.
864          * But it's only on the first write fault on the vma, not
865          * every time, and we could devise a way to avoid it later
866          * (e.g. stash info in next's anon_vma_node when assigning
867          * an anon_vma, or when trying vma_merge).  Another time.
868          */
869         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
870         if (!near)
871                 goto none;
872
873         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
874         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
875
876         if (near->anon_vma && near->vm_end == vma->vm_start &&
877                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
878                         can_vma_merge_after(near, vm_flags,
879                                 NULL, vma->vm_file, vma->vm_pgoff))
880                 return near->anon_vma;
881 none:
882         /*
883          * There's no absolute need to look only at touching neighbours:
884          * we could search further afield for "compatible" anon_vmas.
885          * But it would probably just be a waste of time searching,
886          * or lead to too many vmas hanging off the same anon_vma.
887          * We're trying to allow mprotect remerging later on,
888          * not trying to minimize memory used for anon_vmas.
889          */
890         return NULL;
891 }
892
893 #ifdef CONFIG_PROC_FS
894 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
895                                                 struct file *file, long pages)
896 {
897         const unsigned long stack_flags
898                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
899
900         if (file) {
901                 mm->shared_vm += pages;
902                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
903                         mm->exec_vm += pages;
904         } else if (flags & stack_flags)
905                 mm->stack_vm += pages;
906         if (flags & (VM_RESERVED|VM_IO))
907                 mm->reserved_vm += pages;
908 }
909 #endif /* CONFIG_PROC_FS */
910
911 /*
912  * The caller must hold down_write(current->mm->mmap_sem).
913  */
914
915 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
916                         unsigned long len, unsigned long prot,
917                         unsigned long flags, unsigned long pgoff)
918 {
919         struct mm_struct * mm = current->mm;
920         struct inode *inode;
921         unsigned int vm_flags;
922         int error;
923         unsigned long reqprot = prot;
924
925         /*
926          * Does the application expect PROT_READ to imply PROT_EXEC?
927          *
928          * (the exception is when the underlying filesystem is noexec
929          *  mounted, in which case we dont add PROT_EXEC.)
930          */
931         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
932                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
933                         prot |= PROT_EXEC;
934
935         if (!len)
936                 return -EINVAL;
937
938         if (!(flags & MAP_FIXED))
939                 addr = round_hint_to_min(addr);
940
941         error = arch_mmap_check(addr, len, flags);
942         if (error)
943                 return error;
944
945         /* Careful about overflows.. */
946         len = PAGE_ALIGN(len);
947         if (!len || len > TASK_SIZE)
948                 return -ENOMEM;
949
950         /* offset overflow? */
951         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
952                return -EOVERFLOW;
953
954         /* Too many mappings? */
955         if (mm->map_count > sysctl_max_map_count)
956                 return -ENOMEM;
957
958         /* Obtain the address to map to. we verify (or select) it and ensure
959          * that it represents a valid section of the address space.
960          */
961         addr = get_unmapped_area(file, addr, len, pgoff, flags);
962         if (addr & ~PAGE_MASK)
963                 return addr;
964
965         /* Do simple checking here so the lower-level routines won't have
966          * to. we assume access permissions have been handled by the open
967          * of the memory object, so we don't do any here.
968          */
969         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
970                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
971
972         if (flags & MAP_LOCKED) {
973                 if (!can_do_mlock())
974                         return -EPERM;
975                 vm_flags |= VM_LOCKED;
976         }
977
978         /* mlock MCL_FUTURE? */
979         if (vm_flags & VM_LOCKED) {
980                 unsigned long locked, lock_limit;
981                 locked = len >> PAGE_SHIFT;
982                 locked += mm->locked_vm;
983                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
984                 lock_limit >>= PAGE_SHIFT;
985                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
986                         return -EAGAIN;
987         }
988
989         inode = file ? file->f_path.dentry->d_inode : NULL;
990
991         if (file) {
992                 switch (flags & MAP_TYPE) {
993                 case MAP_SHARED:
994                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
995                                 return -EACCES;
996
997                         /*
998                          * Make sure we don't allow writing to an append-only
999                          * file..
1000                          */
1001                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1002                                 return -EACCES;
1003
1004                         /*
1005                          * Make sure there are no mandatory locks on the file.
1006                          */
1007                         if (locks_verify_locked(inode))
1008                                 return -EAGAIN;
1009
1010                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1011                         if (!(file->f_mode & FMODE_WRITE))
1012                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1013
1014                         /* fall through */
1015                 case MAP_PRIVATE:
1016                         if (!(file->f_mode & FMODE_READ))
1017                                 return -EACCES;
1018                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1019                                 if (vm_flags & VM_EXEC)
1020                                         return -EPERM;
1021                                 vm_flags &= ~VM_MAYEXEC;
1022                         }
1023
1024                         if (!file->f_op || !file->f_op->mmap)
1025                                 return -ENODEV;
1026                         break;
1027
1028                 default:
1029                         return -EINVAL;
1030                 }
1031         } else {
1032                 switch (flags & MAP_TYPE) {
1033                 case MAP_SHARED:
1034                         /*
1035                          * Ignore pgoff.
1036                          */
1037                         pgoff = 0;
1038                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1039                         break;
1040                 case MAP_PRIVATE:
1041                         /*
1042                          * Set pgoff according to addr for anon_vma.
1043                          */
1044                         pgoff = addr >> PAGE_SHIFT;
1045                         break;
1046                 default:
1047                         return -EINVAL;
1048                 }
1049         }
1050
1051         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1052         if (error)
1053                 return error;
1054         error = ima_file_mmap(file, prot);
1055         if (error)
1056                 return error;
1057
1058         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1059 }
1060 EXPORT_SYMBOL(do_mmap_pgoff);
1061
1062 /*
1063  * Some shared mappigns will want the pages marked read-only
1064  * to track write events. If so, we'll downgrade vm_page_prot
1065  * to the private version (using protection_map[] without the
1066  * VM_SHARED bit).
1067  */
1068 int vma_wants_writenotify(struct vm_area_struct *vma)
1069 {
1070         unsigned int vm_flags = vma->vm_flags;
1071
1072         /* If it was private or non-writable, the write bit is already clear */
1073         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1074                 return 0;
1075
1076         /* The backer wishes to know when pages are first written to? */
1077         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1078                 return 1;
1079
1080         /* The open routine did something to the protections already? */
1081         if (pgprot_val(vma->vm_page_prot) !=
1082             pgprot_val(vm_get_page_prot(vm_flags)))
1083                 return 0;
1084
1085         /* Specialty mapping? */
1086         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1087                 return 0;
1088
1089         /* Can the mapping track the dirty pages? */
1090         return vma->vm_file && vma->vm_file->f_mapping &&
1091                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1092 }
1093
1094 /*
1095  * We account for memory if it's a private writeable mapping,
1096  * not hugepages and VM_NORESERVE wasn't set.
1097  */
1098 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1099 {
1100         /*
1101          * hugetlb has its own accounting separate from the core VM
1102          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1103          */
1104         if (file && is_file_hugepages(file))
1105                 return 0;
1106
1107         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1108 }
1109
1110 unsigned long mmap_region(struct file *file, unsigned long addr,
1111                           unsigned long len, unsigned long flags,
1112                           unsigned int vm_flags, unsigned long pgoff)
1113 {
1114         struct mm_struct *mm = current->mm;
1115         struct vm_area_struct *vma, *prev;
1116         int correct_wcount = 0;
1117         int error;
1118         struct rb_node **rb_link, *rb_parent;
1119         unsigned long charged = 0;
1120         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1121
1122         /* Clear old maps */
1123         error = -ENOMEM;
1124 munmap_back:
1125         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1126         if (vma && vma->vm_start < addr + len) {
1127                 if (do_munmap(mm, addr, len))
1128                         return -ENOMEM;
1129                 goto munmap_back;
1130         }
1131
1132         /* Check against address space limit. */
1133         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1134                 return -ENOMEM;
1135
1136         /*
1137          * Set 'VM_NORESERVE' if we should not account for the
1138          * memory use of this mapping.
1139          */
1140         if ((flags & MAP_NORESERVE)) {
1141                 /* We honor MAP_NORESERVE if allowed to overcommit */
1142                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1143                         vm_flags |= VM_NORESERVE;
1144
1145                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1146                 if (file && is_file_hugepages(file))
1147                         vm_flags |= VM_NORESERVE;
1148         }
1149
1150         /*
1151          * Private writable mapping: check memory availability
1152          */
1153         if (accountable_mapping(file, vm_flags)) {
1154                 charged = len >> PAGE_SHIFT;
1155                 if (security_vm_enough_memory(charged))
1156                         return -ENOMEM;
1157                 vm_flags |= VM_ACCOUNT;
1158         }
1159
1160         /*
1161          * Can we just expand an old mapping?
1162          */
1163         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1164         if (vma)
1165                 goto out;
1166
1167         /*
1168          * Determine the object being mapped and call the appropriate
1169          * specific mapper. the address has already been validated, but
1170          * not unmapped, but the maps are removed from the list.
1171          */
1172         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1173         if (!vma) {
1174                 error = -ENOMEM;
1175                 goto unacct_error;
1176         }
1177
1178         vma->vm_mm = mm;
1179         vma->vm_start = addr;
1180         vma->vm_end = addr + len;
1181         vma->vm_flags = vm_flags;
1182         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1183         vma->vm_pgoff = pgoff;
1184
1185         if (file) {
1186                 error = -EINVAL;
1187                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1188                         goto free_vma;
1189                 if (vm_flags & VM_DENYWRITE) {
1190                         error = deny_write_access(file);
1191                         if (error)
1192                                 goto free_vma;
1193                         correct_wcount = 1;
1194                 }
1195                 vma->vm_file = file;
1196                 get_file(file);
1197                 error = file->f_op->mmap(file, vma);
1198                 if (error)
1199                         goto unmap_and_free_vma;
1200                 if (vm_flags & VM_EXECUTABLE)
1201                         added_exe_file_vma(mm);
1202         } else if (vm_flags & VM_SHARED) {
1203                 error = shmem_zero_setup(vma);
1204                 if (error)
1205                         goto free_vma;
1206         }
1207
1208         /* Can addr have changed??
1209          *
1210          * Answer: Yes, several device drivers can do it in their
1211          *         f_op->mmap method. -DaveM
1212          */
1213         addr = vma->vm_start;
1214         pgoff = vma->vm_pgoff;
1215         vm_flags = vma->vm_flags;
1216
1217         if (vma_wants_writenotify(vma))
1218                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1219
1220         vma_link(mm, vma, prev, rb_link, rb_parent);
1221         file = vma->vm_file;
1222
1223         /* Once vma denies write, undo our temporary denial count */
1224         if (correct_wcount)
1225                 atomic_inc(&inode->i_writecount);
1226 out:
1227         if (vm_flags & VM_EXEC)
1228                 perf_counter_mmap(addr, len, pgoff, file);
1229
1230         mm->total_vm += len >> PAGE_SHIFT;
1231         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1232         if (vm_flags & VM_LOCKED) {
1233                 /*
1234                  * makes pages present; downgrades, drops, reacquires mmap_sem
1235                  */
1236                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1237                 if (nr_pages < 0)
1238                         return nr_pages;        /* vma gone! */
1239                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1240         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1241                 make_pages_present(addr, addr + len);
1242         return addr;
1243
1244 unmap_and_free_vma:
1245         if (correct_wcount)
1246                 atomic_inc(&inode->i_writecount);
1247         vma->vm_file = NULL;
1248         fput(file);
1249
1250         /* Undo any partial mapping done by a device driver. */
1251         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1252         charged = 0;
1253 free_vma:
1254         kmem_cache_free(vm_area_cachep, vma);
1255 unacct_error:
1256         if (charged)
1257                 vm_unacct_memory(charged);
1258         return error;
1259 }
1260
1261 /* Get an address range which is currently unmapped.
1262  * For shmat() with addr=0.
1263  *
1264  * Ugly calling convention alert:
1265  * Return value with the low bits set means error value,
1266  * ie
1267  *      if (ret & ~PAGE_MASK)
1268  *              error = ret;
1269  *
1270  * This function "knows" that -ENOMEM has the bits set.
1271  */
1272 #ifndef HAVE_ARCH_UNMAPPED_AREA
1273 unsigned long
1274 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1275                 unsigned long len, unsigned long pgoff, unsigned long flags)
1276 {
1277         struct mm_struct *mm = current->mm;
1278         struct vm_area_struct *vma;
1279         unsigned long start_addr;
1280
1281         if (len > TASK_SIZE)
1282                 return -ENOMEM;
1283
1284         if (flags & MAP_FIXED)
1285                 return addr;
1286
1287         if (addr) {
1288                 addr = PAGE_ALIGN(addr);
1289                 vma = find_vma(mm, addr);
1290                 if (TASK_SIZE - len >= addr &&
1291                     (!vma || addr + len <= vma->vm_start))
1292                         return addr;
1293         }
1294         if (len > mm->cached_hole_size) {
1295                 start_addr = addr = mm->free_area_cache;
1296         } else {
1297                 start_addr = addr = TASK_UNMAPPED_BASE;
1298                 mm->cached_hole_size = 0;
1299         }
1300
1301 full_search:
1302         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1303                 /* At this point:  (!vma || addr < vma->vm_end). */
1304                 if (TASK_SIZE - len < addr) {
1305                         /*
1306                          * Start a new search - just in case we missed
1307                          * some holes.
1308                          */
1309                         if (start_addr != TASK_UNMAPPED_BASE) {
1310                                 addr = TASK_UNMAPPED_BASE;
1311                                 start_addr = addr;
1312                                 mm->cached_hole_size = 0;
1313                                 goto full_search;
1314                         }
1315                         return -ENOMEM;
1316                 }
1317                 if (!vma || addr + len <= vma->vm_start) {
1318                         /*
1319                          * Remember the place where we stopped the search:
1320                          */
1321                         mm->free_area_cache = addr + len;
1322                         return addr;
1323                 }
1324                 if (addr + mm->cached_hole_size < vma->vm_start)
1325                         mm->cached_hole_size = vma->vm_start - addr;
1326                 addr = vma->vm_end;
1327         }
1328 }
1329 #endif  
1330
1331 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1332 {
1333         /*
1334          * Is this a new hole at the lowest possible address?
1335          */
1336         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1337                 mm->free_area_cache = addr;
1338                 mm->cached_hole_size = ~0UL;
1339         }
1340 }
1341
1342 /*
1343  * This mmap-allocator allocates new areas top-down from below the
1344  * stack's low limit (the base):
1345  */
1346 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1347 unsigned long
1348 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1349                           const unsigned long len, const unsigned long pgoff,
1350                           const unsigned long flags)
1351 {
1352         struct vm_area_struct *vma;
1353         struct mm_struct *mm = current->mm;
1354         unsigned long addr = addr0;
1355
1356         /* requested length too big for entire address space */
1357         if (len > TASK_SIZE)
1358                 return -ENOMEM;
1359
1360         if (flags & MAP_FIXED)
1361                 return addr;
1362
1363         /* requesting a specific address */
1364         if (addr) {
1365                 addr = PAGE_ALIGN(addr);
1366                 vma = find_vma(mm, addr);
1367                 if (TASK_SIZE - len >= addr &&
1368                                 (!vma || addr + len <= vma->vm_start))
1369                         return addr;
1370         }
1371
1372         /* check if free_area_cache is useful for us */
1373         if (len <= mm->cached_hole_size) {
1374                 mm->cached_hole_size = 0;
1375                 mm->free_area_cache = mm->mmap_base;
1376         }
1377
1378         /* either no address requested or can't fit in requested address hole */
1379         addr = mm->free_area_cache;
1380
1381         /* make sure it can fit in the remaining address space */
1382         if (addr > len) {
1383                 vma = find_vma(mm, addr-len);
1384                 if (!vma || addr <= vma->vm_start)
1385                         /* remember the address as a hint for next time */
1386                         return (mm->free_area_cache = addr-len);
1387         }
1388
1389         if (mm->mmap_base < len)
1390                 goto bottomup;
1391
1392         addr = mm->mmap_base-len;
1393
1394         do {
1395                 /*
1396                  * Lookup failure means no vma is above this address,
1397                  * else if new region fits below vma->vm_start,
1398                  * return with success:
1399                  */
1400                 vma = find_vma(mm, addr);
1401                 if (!vma || addr+len <= vma->vm_start)
1402                         /* remember the address as a hint for next time */
1403                         return (mm->free_area_cache = addr);
1404
1405                 /* remember the largest hole we saw so far */
1406                 if (addr + mm->cached_hole_size < vma->vm_start)
1407                         mm->cached_hole_size = vma->vm_start - addr;
1408
1409                 /* try just below the current vma->vm_start */
1410                 addr = vma->vm_start-len;
1411         } while (len < vma->vm_start);
1412
1413 bottomup:
1414         /*
1415          * A failed mmap() very likely causes application failure,
1416          * so fall back to the bottom-up function here. This scenario
1417          * can happen with large stack limits and large mmap()
1418          * allocations.
1419          */
1420         mm->cached_hole_size = ~0UL;
1421         mm->free_area_cache = TASK_UNMAPPED_BASE;
1422         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1423         /*
1424          * Restore the topdown base:
1425          */
1426         mm->free_area_cache = mm->mmap_base;
1427         mm->cached_hole_size = ~0UL;
1428
1429         return addr;
1430 }
1431 #endif
1432
1433 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1434 {
1435         /*
1436          * Is this a new hole at the highest possible address?
1437          */
1438         if (addr > mm->free_area_cache)
1439                 mm->free_area_cache = addr;
1440
1441         /* dont allow allocations above current base */
1442         if (mm->free_area_cache > mm->mmap_base)
1443                 mm->free_area_cache = mm->mmap_base;
1444 }
1445
1446 unsigned long
1447 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1448                 unsigned long pgoff, unsigned long flags)
1449 {
1450         unsigned long (*get_area)(struct file *, unsigned long,
1451                                   unsigned long, unsigned long, unsigned long);
1452
1453         get_area = current->mm->get_unmapped_area;
1454         if (file && file->f_op && file->f_op->get_unmapped_area)
1455                 get_area = file->f_op->get_unmapped_area;
1456         addr = get_area(file, addr, len, pgoff, flags);
1457         if (IS_ERR_VALUE(addr))
1458                 return addr;
1459
1460         if (addr > TASK_SIZE - len)
1461                 return -ENOMEM;
1462         if (addr & ~PAGE_MASK)
1463                 return -EINVAL;
1464
1465         return arch_rebalance_pgtables(addr, len);
1466 }
1467
1468 EXPORT_SYMBOL(get_unmapped_area);
1469
1470 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1471 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1472 {
1473         struct vm_area_struct *vma = NULL;
1474
1475         if (mm) {
1476                 /* Check the cache first. */
1477                 /* (Cache hit rate is typically around 35%.) */
1478                 vma = mm->mmap_cache;
1479                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1480                         struct rb_node * rb_node;
1481
1482                         rb_node = mm->mm_rb.rb_node;
1483                         vma = NULL;
1484
1485                         while (rb_node) {
1486                                 struct vm_area_struct * vma_tmp;
1487
1488                                 vma_tmp = rb_entry(rb_node,
1489                                                 struct vm_area_struct, vm_rb);
1490
1491                                 if (vma_tmp->vm_end > addr) {
1492                                         vma = vma_tmp;
1493                                         if (vma_tmp->vm_start <= addr)
1494                                                 break;
1495                                         rb_node = rb_node->rb_left;
1496                                 } else
1497                                         rb_node = rb_node->rb_right;
1498                         }
1499                         if (vma)
1500                                 mm->mmap_cache = vma;
1501                 }
1502         }
1503         return vma;
1504 }
1505
1506 EXPORT_SYMBOL(find_vma);
1507
1508 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1509 struct vm_area_struct *
1510 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1511                         struct vm_area_struct **pprev)
1512 {
1513         struct vm_area_struct *vma = NULL, *prev = NULL;
1514         struct rb_node *rb_node;
1515         if (!mm)
1516                 goto out;
1517
1518         /* Guard against addr being lower than the first VMA */
1519         vma = mm->mmap;
1520
1521         /* Go through the RB tree quickly. */
1522         rb_node = mm->mm_rb.rb_node;
1523
1524         while (rb_node) {
1525                 struct vm_area_struct *vma_tmp;
1526                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1527
1528                 if (addr < vma_tmp->vm_end) {
1529                         rb_node = rb_node->rb_left;
1530                 } else {
1531                         prev = vma_tmp;
1532                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1533                                 break;
1534                         rb_node = rb_node->rb_right;
1535                 }
1536         }
1537
1538 out:
1539         *pprev = prev;
1540         return prev ? prev->vm_next : vma;
1541 }
1542
1543 /*
1544  * Verify that the stack growth is acceptable and
1545  * update accounting. This is shared with both the
1546  * grow-up and grow-down cases.
1547  */
1548 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1549 {
1550         struct mm_struct *mm = vma->vm_mm;
1551         struct rlimit *rlim = current->signal->rlim;
1552         unsigned long new_start;
1553
1554         /* address space limit tests */
1555         if (!may_expand_vm(mm, grow))
1556                 return -ENOMEM;
1557
1558         /* Stack limit test */
1559         if (size > rlim[RLIMIT_STACK].rlim_cur)
1560                 return -ENOMEM;
1561
1562         /* mlock limit tests */
1563         if (vma->vm_flags & VM_LOCKED) {
1564                 unsigned long locked;
1565                 unsigned long limit;
1566                 locked = mm->locked_vm + grow;
1567                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1568                 if (locked > limit && !capable(CAP_IPC_LOCK))
1569                         return -ENOMEM;
1570         }
1571
1572         /* Check to ensure the stack will not grow into a hugetlb-only region */
1573         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1574                         vma->vm_end - size;
1575         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1576                 return -EFAULT;
1577
1578         /*
1579          * Overcommit..  This must be the final test, as it will
1580          * update security statistics.
1581          */
1582         if (security_vm_enough_memory_mm(mm, grow))
1583                 return -ENOMEM;
1584
1585         /* Ok, everything looks good - let it rip */
1586         mm->total_vm += grow;
1587         if (vma->vm_flags & VM_LOCKED)
1588                 mm->locked_vm += grow;
1589         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1590         return 0;
1591 }
1592
1593 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1594 /*
1595  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1596  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1597  */
1598 #ifndef CONFIG_IA64
1599 static
1600 #endif
1601 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1602 {
1603         int error;
1604
1605         if (!(vma->vm_flags & VM_GROWSUP))
1606                 return -EFAULT;
1607
1608         /*
1609          * We must make sure the anon_vma is allocated
1610          * so that the anon_vma locking is not a noop.
1611          */
1612         if (unlikely(anon_vma_prepare(vma)))
1613                 return -ENOMEM;
1614         anon_vma_lock(vma);
1615
1616         /*
1617          * vma->vm_start/vm_end cannot change under us because the caller
1618          * is required to hold the mmap_sem in read mode.  We need the
1619          * anon_vma lock to serialize against concurrent expand_stacks.
1620          * Also guard against wrapping around to address 0.
1621          */
1622         if (address < PAGE_ALIGN(address+4))
1623                 address = PAGE_ALIGN(address+4);
1624         else {
1625                 anon_vma_unlock(vma);
1626                 return -ENOMEM;
1627         }
1628         error = 0;
1629
1630         /* Somebody else might have raced and expanded it already */
1631         if (address > vma->vm_end) {
1632                 unsigned long size, grow;
1633
1634                 size = address - vma->vm_start;
1635                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1636
1637                 error = acct_stack_growth(vma, size, grow);
1638                 if (!error)
1639                         vma->vm_end = address;
1640         }
1641         anon_vma_unlock(vma);
1642         return error;
1643 }
1644 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1645
1646 /*
1647  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1648  */
1649 static int expand_downwards(struct vm_area_struct *vma,
1650                                    unsigned long address)
1651 {
1652         int error;
1653
1654         /*
1655          * We must make sure the anon_vma is allocated
1656          * so that the anon_vma locking is not a noop.
1657          */
1658         if (unlikely(anon_vma_prepare(vma)))
1659                 return -ENOMEM;
1660
1661         address &= PAGE_MASK;
1662         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1663         if (error)
1664                 return error;
1665
1666         anon_vma_lock(vma);
1667
1668         /*
1669          * vma->vm_start/vm_end cannot change under us because the caller
1670          * is required to hold the mmap_sem in read mode.  We need the
1671          * anon_vma lock to serialize against concurrent expand_stacks.
1672          */
1673
1674         /* Somebody else might have raced and expanded it already */
1675         if (address < vma->vm_start) {
1676                 unsigned long size, grow;
1677
1678                 size = vma->vm_end - address;
1679                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1680
1681                 error = acct_stack_growth(vma, size, grow);
1682                 if (!error) {
1683                         vma->vm_start = address;
1684                         vma->vm_pgoff -= grow;
1685                 }
1686         }
1687         anon_vma_unlock(vma);
1688         return error;
1689 }
1690
1691 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1692 {
1693         return expand_downwards(vma, address);
1694 }
1695
1696 #ifdef CONFIG_STACK_GROWSUP
1697 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1698 {
1699         return expand_upwards(vma, address);
1700 }
1701
1702 struct vm_area_struct *
1703 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1704 {
1705         struct vm_area_struct *vma, *prev;
1706
1707         addr &= PAGE_MASK;
1708         vma = find_vma_prev(mm, addr, &prev);
1709         if (vma && (vma->vm_start <= addr))
1710                 return vma;
1711         if (!prev || expand_stack(prev, addr))
1712                 return NULL;
1713         if (prev->vm_flags & VM_LOCKED) {
1714                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1715                         return NULL;    /* vma gone! */
1716         }
1717         return prev;
1718 }
1719 #else
1720 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1721 {
1722         return expand_downwards(vma, address);
1723 }
1724
1725 struct vm_area_struct *
1726 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1727 {
1728         struct vm_area_struct * vma;
1729         unsigned long start;
1730
1731         addr &= PAGE_MASK;
1732         vma = find_vma(mm,addr);
1733         if (!vma)
1734                 return NULL;
1735         if (vma->vm_start <= addr)
1736                 return vma;
1737         if (!(vma->vm_flags & VM_GROWSDOWN))
1738                 return NULL;
1739         start = vma->vm_start;
1740         if (expand_stack(vma, addr))
1741                 return NULL;
1742         if (vma->vm_flags & VM_LOCKED) {
1743                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1744                         return NULL;    /* vma gone! */
1745         }
1746         return vma;
1747 }
1748 #endif
1749
1750 /*
1751  * Ok - we have the memory areas we should free on the vma list,
1752  * so release them, and do the vma updates.
1753  *
1754  * Called with the mm semaphore held.
1755  */
1756 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1757 {
1758         /* Update high watermark before we lower total_vm */
1759         update_hiwater_vm(mm);
1760         do {
1761                 long nrpages = vma_pages(vma);
1762
1763                 if (vma->vm_flags & VM_EXEC) {
1764                         perf_counter_munmap(vma->vm_start,
1765                                         nrpages << PAGE_SHIFT,
1766                                         vma->vm_pgoff, vma->vm_file);
1767                 }
1768
1769                 mm->total_vm -= nrpages;
1770                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1771                 vma = remove_vma(vma);
1772         } while (vma);
1773         validate_mm(mm);
1774 }
1775
1776 /*
1777  * Get rid of page table information in the indicated region.
1778  *
1779  * Called with the mm semaphore held.
1780  */
1781 static void unmap_region(struct mm_struct *mm,
1782                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1783                 unsigned long start, unsigned long end)
1784 {
1785         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1786         struct mmu_gather *tlb;
1787         unsigned long nr_accounted = 0;
1788
1789         lru_add_drain();
1790         tlb = tlb_gather_mmu(mm, 0);
1791         update_hiwater_rss(mm);
1792         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1793         vm_unacct_memory(nr_accounted);
1794         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1795                                  next? next->vm_start: 0);
1796         tlb_finish_mmu(tlb, start, end);
1797 }
1798
1799 /*
1800  * Create a list of vma's touched by the unmap, removing them from the mm's
1801  * vma list as we go..
1802  */
1803 static void
1804 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1805         struct vm_area_struct *prev, unsigned long end)
1806 {
1807         struct vm_area_struct **insertion_point;
1808         struct vm_area_struct *tail_vma = NULL;
1809         unsigned long addr;
1810
1811         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1812         do {
1813                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1814                 mm->map_count--;
1815                 tail_vma = vma;
1816                 vma = vma->vm_next;
1817         } while (vma && vma->vm_start < end);
1818         *insertion_point = vma;
1819         tail_vma->vm_next = NULL;
1820         if (mm->unmap_area == arch_unmap_area)
1821                 addr = prev ? prev->vm_end : mm->mmap_base;
1822         else
1823                 addr = vma ?  vma->vm_start : mm->mmap_base;
1824         mm->unmap_area(mm, addr);
1825         mm->mmap_cache = NULL;          /* Kill the cache. */
1826 }
1827
1828 /*
1829  * Split a vma into two pieces at address 'addr', a new vma is allocated
1830  * either for the first part or the tail.
1831  */
1832 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1833               unsigned long addr, int new_below)
1834 {
1835         struct mempolicy *pol;
1836         struct vm_area_struct *new;
1837
1838         if (is_vm_hugetlb_page(vma) && (addr &
1839                                         ~(huge_page_mask(hstate_vma(vma)))))
1840                 return -EINVAL;
1841
1842         if (mm->map_count >= sysctl_max_map_count)
1843                 return -ENOMEM;
1844
1845         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1846         if (!new)
1847                 return -ENOMEM;
1848
1849         /* most fields are the same, copy all, and then fixup */
1850         *new = *vma;
1851
1852         if (new_below)
1853                 new->vm_end = addr;
1854         else {
1855                 new->vm_start = addr;
1856                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1857         }
1858
1859         pol = mpol_dup(vma_policy(vma));
1860         if (IS_ERR(pol)) {
1861                 kmem_cache_free(vm_area_cachep, new);
1862                 return PTR_ERR(pol);
1863         }
1864         vma_set_policy(new, pol);
1865
1866         if (new->vm_file) {
1867                 get_file(new->vm_file);
1868                 if (vma->vm_flags & VM_EXECUTABLE)
1869                         added_exe_file_vma(mm);
1870         }
1871
1872         if (new->vm_ops && new->vm_ops->open)
1873                 new->vm_ops->open(new);
1874
1875         if (new_below)
1876                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1877                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1878         else
1879                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1880
1881         return 0;
1882 }
1883
1884 /* Munmap is split into 2 main parts -- this part which finds
1885  * what needs doing, and the areas themselves, which do the
1886  * work.  This now handles partial unmappings.
1887  * Jeremy Fitzhardinge <jeremy@goop.org>
1888  */
1889 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1890 {
1891         unsigned long end;
1892         struct vm_area_struct *vma, *prev, *last;
1893
1894         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1895                 return -EINVAL;
1896
1897         if ((len = PAGE_ALIGN(len)) == 0)
1898                 return -EINVAL;
1899
1900         /* Find the first overlapping VMA */
1901         vma = find_vma_prev(mm, start, &prev);
1902         if (!vma)
1903                 return 0;
1904         /* we have  start < vma->vm_end  */
1905
1906         /* if it doesn't overlap, we have nothing.. */
1907         end = start + len;
1908         if (vma->vm_start >= end)
1909                 return 0;
1910
1911         /*
1912          * If we need to split any vma, do it now to save pain later.
1913          *
1914          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1915          * unmapped vm_area_struct will remain in use: so lower split_vma
1916          * places tmp vma above, and higher split_vma places tmp vma below.
1917          */
1918         if (start > vma->vm_start) {
1919                 int error = split_vma(mm, vma, start, 0);
1920                 if (error)
1921                         return error;
1922                 prev = vma;
1923         }
1924
1925         /* Does it split the last one? */
1926         last = find_vma(mm, end);
1927         if (last && end > last->vm_start) {
1928                 int error = split_vma(mm, last, end, 1);
1929                 if (error)
1930                         return error;
1931         }
1932         vma = prev? prev->vm_next: mm->mmap;
1933
1934         /*
1935          * unlock any mlock()ed ranges before detaching vmas
1936          */
1937         if (mm->locked_vm) {
1938                 struct vm_area_struct *tmp = vma;
1939                 while (tmp && tmp->vm_start < end) {
1940                         if (tmp->vm_flags & VM_LOCKED) {
1941                                 mm->locked_vm -= vma_pages(tmp);
1942                                 munlock_vma_pages_all(tmp);
1943                         }
1944                         tmp = tmp->vm_next;
1945                 }
1946         }
1947
1948         /*
1949          * Remove the vma's, and unmap the actual pages
1950          */
1951         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1952         unmap_region(mm, vma, prev, start, end);
1953
1954         /* Fix up all other VM information */
1955         remove_vma_list(mm, vma);
1956
1957         return 0;
1958 }
1959
1960 EXPORT_SYMBOL(do_munmap);
1961
1962 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1963 {
1964         int ret;
1965         struct mm_struct *mm = current->mm;
1966
1967         profile_munmap(addr);
1968
1969         down_write(&mm->mmap_sem);
1970         ret = do_munmap(mm, addr, len);
1971         up_write(&mm->mmap_sem);
1972         return ret;
1973 }
1974
1975 static inline void verify_mm_writelocked(struct mm_struct *mm)
1976 {
1977 #ifdef CONFIG_DEBUG_VM
1978         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1979                 WARN_ON(1);
1980                 up_read(&mm->mmap_sem);
1981         }
1982 #endif
1983 }
1984
1985 /*
1986  *  this is really a simplified "do_mmap".  it only handles
1987  *  anonymous maps.  eventually we may be able to do some
1988  *  brk-specific accounting here.
1989  */
1990 unsigned long do_brk(unsigned long addr, unsigned long len)
1991 {
1992         struct mm_struct * mm = current->mm;
1993         struct vm_area_struct * vma, * prev;
1994         unsigned long flags;
1995         struct rb_node ** rb_link, * rb_parent;
1996         pgoff_t pgoff = addr >> PAGE_SHIFT;
1997         int error;
1998
1999         len = PAGE_ALIGN(len);
2000         if (!len)
2001                 return addr;
2002
2003         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
2004                 return -EINVAL;
2005
2006         if (is_hugepage_only_range(mm, addr, len))
2007                 return -EINVAL;
2008
2009         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2010         if (error)
2011                 return error;
2012
2013         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2014
2015         error = arch_mmap_check(addr, len, flags);
2016         if (error)
2017                 return error;
2018
2019         /*
2020          * mlock MCL_FUTURE?
2021          */
2022         if (mm->def_flags & VM_LOCKED) {
2023                 unsigned long locked, lock_limit;
2024                 locked = len >> PAGE_SHIFT;
2025                 locked += mm->locked_vm;
2026                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2027                 lock_limit >>= PAGE_SHIFT;
2028                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2029                         return -EAGAIN;
2030         }
2031
2032         /*
2033          * mm->mmap_sem is required to protect against another thread
2034          * changing the mappings in case we sleep.
2035          */
2036         verify_mm_writelocked(mm);
2037
2038         /*
2039          * Clear old maps.  this also does some error checking for us
2040          */
2041  munmap_back:
2042         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2043         if (vma && vma->vm_start < addr + len) {
2044                 if (do_munmap(mm, addr, len))
2045                         return -ENOMEM;
2046                 goto munmap_back;
2047         }
2048
2049         /* Check against address space limits *after* clearing old maps... */
2050         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2051                 return -ENOMEM;
2052
2053         if (mm->map_count > sysctl_max_map_count)
2054                 return -ENOMEM;
2055
2056         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2057                 return -ENOMEM;
2058
2059         /* Can we just expand an old private anonymous mapping? */
2060         vma = vma_merge(mm, prev, addr, addr + len, flags,
2061                                         NULL, NULL, pgoff, NULL);
2062         if (vma)
2063                 goto out;
2064
2065         /*
2066          * create a vma struct for an anonymous mapping
2067          */
2068         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2069         if (!vma) {
2070                 vm_unacct_memory(len >> PAGE_SHIFT);
2071                 return -ENOMEM;
2072         }
2073
2074         vma->vm_mm = mm;
2075         vma->vm_start = addr;
2076         vma->vm_end = addr + len;
2077         vma->vm_pgoff = pgoff;
2078         vma->vm_flags = flags;
2079         vma->vm_page_prot = vm_get_page_prot(flags);
2080         vma_link(mm, vma, prev, rb_link, rb_parent);
2081 out:
2082         mm->total_vm += len >> PAGE_SHIFT;
2083         if (flags & VM_LOCKED) {
2084                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2085                         mm->locked_vm += (len >> PAGE_SHIFT);
2086         }
2087         return addr;
2088 }
2089
2090 EXPORT_SYMBOL(do_brk);
2091
2092 /* Release all mmaps. */
2093 void exit_mmap(struct mm_struct *mm)
2094 {
2095         struct mmu_gather *tlb;
2096         struct vm_area_struct *vma;
2097         unsigned long nr_accounted = 0;
2098         unsigned long end;
2099
2100         /* mm's last user has gone, and its about to be pulled down */
2101         mmu_notifier_release(mm);
2102
2103         if (mm->locked_vm) {
2104                 vma = mm->mmap;
2105                 while (vma) {
2106                         if (vma->vm_flags & VM_LOCKED)
2107                                 munlock_vma_pages_all(vma);
2108                         vma = vma->vm_next;
2109                 }
2110         }
2111
2112         arch_exit_mmap(mm);
2113
2114         vma = mm->mmap;
2115         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2116                 return;
2117
2118         lru_add_drain();
2119         flush_cache_mm(mm);
2120         tlb = tlb_gather_mmu(mm, 1);
2121         /* update_hiwater_rss(mm) here? but nobody should be looking */
2122         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2123         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2124         vm_unacct_memory(nr_accounted);
2125         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2126         tlb_finish_mmu(tlb, 0, end);
2127
2128         /*
2129          * Walk the list again, actually closing and freeing it,
2130          * with preemption enabled, without holding any MM locks.
2131          */
2132         while (vma)
2133                 vma = remove_vma(vma);
2134
2135         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2136 }
2137
2138 /* Insert vm structure into process list sorted by address
2139  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2140  * then i_mmap_lock is taken here.
2141  */
2142 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2143 {
2144         struct vm_area_struct * __vma, * prev;
2145         struct rb_node ** rb_link, * rb_parent;
2146
2147         /*
2148          * The vm_pgoff of a purely anonymous vma should be irrelevant
2149          * until its first write fault, when page's anon_vma and index
2150          * are set.  But now set the vm_pgoff it will almost certainly
2151          * end up with (unless mremap moves it elsewhere before that
2152          * first wfault), so /proc/pid/maps tells a consistent story.
2153          *
2154          * By setting it to reflect the virtual start address of the
2155          * vma, merges and splits can happen in a seamless way, just
2156          * using the existing file pgoff checks and manipulations.
2157          * Similarly in do_mmap_pgoff and in do_brk.
2158          */
2159         if (!vma->vm_file) {
2160                 BUG_ON(vma->anon_vma);
2161                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2162         }
2163         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2164         if (__vma && __vma->vm_start < vma->vm_end)
2165                 return -ENOMEM;
2166         if ((vma->vm_flags & VM_ACCOUNT) &&
2167              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2168                 return -ENOMEM;
2169         vma_link(mm, vma, prev, rb_link, rb_parent);
2170         return 0;
2171 }
2172
2173 /*
2174  * Copy the vma structure to a new location in the same mm,
2175  * prior to moving page table entries, to effect an mremap move.
2176  */
2177 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2178         unsigned long addr, unsigned long len, pgoff_t pgoff)
2179 {
2180         struct vm_area_struct *vma = *vmap;
2181         unsigned long vma_start = vma->vm_start;
2182         struct mm_struct *mm = vma->vm_mm;
2183         struct vm_area_struct *new_vma, *prev;
2184         struct rb_node **rb_link, *rb_parent;
2185         struct mempolicy *pol;
2186
2187         /*
2188          * If anonymous vma has not yet been faulted, update new pgoff
2189          * to match new location, to increase its chance of merging.
2190          */
2191         if (!vma->vm_file && !vma->anon_vma)
2192                 pgoff = addr >> PAGE_SHIFT;
2193
2194         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2195         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2196                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2197         if (new_vma) {
2198                 /*
2199                  * Source vma may have been merged into new_vma
2200                  */
2201                 if (vma_start >= new_vma->vm_start &&
2202                     vma_start < new_vma->vm_end)
2203                         *vmap = new_vma;
2204         } else {
2205                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2206                 if (new_vma) {
2207                         *new_vma = *vma;
2208                         pol = mpol_dup(vma_policy(vma));
2209                         if (IS_ERR(pol)) {
2210                                 kmem_cache_free(vm_area_cachep, new_vma);
2211                                 return NULL;
2212                         }
2213                         vma_set_policy(new_vma, pol);
2214                         new_vma->vm_start = addr;
2215                         new_vma->vm_end = addr + len;
2216                         new_vma->vm_pgoff = pgoff;
2217                         if (new_vma->vm_file) {
2218                                 get_file(new_vma->vm_file);
2219                                 if (vma->vm_flags & VM_EXECUTABLE)
2220                                         added_exe_file_vma(mm);
2221                         }
2222                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2223                                 new_vma->vm_ops->open(new_vma);
2224                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2225                 }
2226         }
2227         return new_vma;
2228 }
2229
2230 /*
2231  * Return true if the calling process may expand its vm space by the passed
2232  * number of pages
2233  */
2234 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2235 {
2236         unsigned long cur = mm->total_vm;       /* pages */
2237         unsigned long lim;
2238
2239         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2240
2241         if (cur + npages > lim)
2242                 return 0;
2243         return 1;
2244 }
2245
2246
2247 static int special_mapping_fault(struct vm_area_struct *vma,
2248                                 struct vm_fault *vmf)
2249 {
2250         pgoff_t pgoff;
2251         struct page **pages;
2252
2253         /*
2254          * special mappings have no vm_file, and in that case, the mm
2255          * uses vm_pgoff internally. So we have to subtract it from here.
2256          * We are allowed to do this because we are the mm; do not copy
2257          * this code into drivers!
2258          */
2259         pgoff = vmf->pgoff - vma->vm_pgoff;
2260
2261         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2262                 pgoff--;
2263
2264         if (*pages) {
2265                 struct page *page = *pages;
2266                 get_page(page);
2267                 vmf->page = page;
2268                 return 0;
2269         }
2270
2271         return VM_FAULT_SIGBUS;
2272 }
2273
2274 /*
2275  * Having a close hook prevents vma merging regardless of flags.
2276  */
2277 static void special_mapping_close(struct vm_area_struct *vma)
2278 {
2279 }
2280
2281 static struct vm_operations_struct special_mapping_vmops = {
2282         .close = special_mapping_close,
2283         .fault = special_mapping_fault,
2284 };
2285
2286 /*
2287  * Called with mm->mmap_sem held for writing.
2288  * Insert a new vma covering the given region, with the given flags.
2289  * Its pages are supplied by the given array of struct page *.
2290  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2291  * The region past the last page supplied will always produce SIGBUS.
2292  * The array pointer and the pages it points to are assumed to stay alive
2293  * for as long as this mapping might exist.
2294  */
2295 int install_special_mapping(struct mm_struct *mm,
2296                             unsigned long addr, unsigned long len,
2297                             unsigned long vm_flags, struct page **pages)
2298 {
2299         struct vm_area_struct *vma;
2300
2301         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2302         if (unlikely(vma == NULL))
2303                 return -ENOMEM;
2304
2305         vma->vm_mm = mm;
2306         vma->vm_start = addr;
2307         vma->vm_end = addr + len;
2308
2309         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2310         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2311
2312         vma->vm_ops = &special_mapping_vmops;
2313         vma->vm_private_data = pages;
2314
2315         if (unlikely(insert_vm_struct(mm, vma))) {
2316                 kmem_cache_free(vm_area_cachep, vma);
2317                 return -ENOMEM;
2318         }
2319
2320         mm->total_vm += len >> PAGE_SHIFT;
2321
2322         return 0;
2323 }
2324
2325 static DEFINE_MUTEX(mm_all_locks_mutex);
2326
2327 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2328 {
2329         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2330                 /*
2331                  * The LSB of head.next can't change from under us
2332                  * because we hold the mm_all_locks_mutex.
2333                  */
2334                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2335                 /*
2336                  * We can safely modify head.next after taking the
2337                  * anon_vma->lock. If some other vma in this mm shares
2338                  * the same anon_vma we won't take it again.
2339                  *
2340                  * No need of atomic instructions here, head.next
2341                  * can't change from under us thanks to the
2342                  * anon_vma->lock.
2343                  */
2344                 if (__test_and_set_bit(0, (unsigned long *)
2345                                        &anon_vma->head.next))
2346                         BUG();
2347         }
2348 }
2349
2350 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2351 {
2352         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2353                 /*
2354                  * AS_MM_ALL_LOCKS can't change from under us because
2355                  * we hold the mm_all_locks_mutex.
2356                  *
2357                  * Operations on ->flags have to be atomic because
2358                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2359                  * mm_all_locks_mutex, there may be other cpus
2360                  * changing other bitflags in parallel to us.
2361                  */
2362                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2363                         BUG();
2364                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2365         }
2366 }
2367
2368 /*
2369  * This operation locks against the VM for all pte/vma/mm related
2370  * operations that could ever happen on a certain mm. This includes
2371  * vmtruncate, try_to_unmap, and all page faults.
2372  *
2373  * The caller must take the mmap_sem in write mode before calling
2374  * mm_take_all_locks(). The caller isn't allowed to release the
2375  * mmap_sem until mm_drop_all_locks() returns.
2376  *
2377  * mmap_sem in write mode is required in order to block all operations
2378  * that could modify pagetables and free pages without need of
2379  * altering the vma layout (for example populate_range() with
2380  * nonlinear vmas). It's also needed in write mode to avoid new
2381  * anon_vmas to be associated with existing vmas.
2382  *
2383  * A single task can't take more than one mm_take_all_locks() in a row
2384  * or it would deadlock.
2385  *
2386  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2387  * mapping->flags avoid to take the same lock twice, if more than one
2388  * vma in this mm is backed by the same anon_vma or address_space.
2389  *
2390  * We can take all the locks in random order because the VM code
2391  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2392  * takes more than one of them in a row. Secondly we're protected
2393  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2394  *
2395  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2396  * that may have to take thousand of locks.
2397  *
2398  * mm_take_all_locks() can fail if it's interrupted by signals.
2399  */
2400 int mm_take_all_locks(struct mm_struct *mm)
2401 {
2402         struct vm_area_struct *vma;
2403         int ret = -EINTR;
2404
2405         BUG_ON(down_read_trylock(&mm->mmap_sem));
2406
2407         mutex_lock(&mm_all_locks_mutex);
2408
2409         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2410                 if (signal_pending(current))
2411                         goto out_unlock;
2412                 if (vma->vm_file && vma->vm_file->f_mapping)
2413                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2414         }
2415
2416         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2417                 if (signal_pending(current))
2418                         goto out_unlock;
2419                 if (vma->anon_vma)
2420                         vm_lock_anon_vma(mm, vma->anon_vma);
2421         }
2422
2423         ret = 0;
2424
2425 out_unlock:
2426         if (ret)
2427                 mm_drop_all_locks(mm);
2428
2429         return ret;
2430 }
2431
2432 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2433 {
2434         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2435                 /*
2436                  * The LSB of head.next can't change to 0 from under
2437                  * us because we hold the mm_all_locks_mutex.
2438                  *
2439                  * We must however clear the bitflag before unlocking
2440                  * the vma so the users using the anon_vma->head will
2441                  * never see our bitflag.
2442                  *
2443                  * No need of atomic instructions here, head.next
2444                  * can't change from under us until we release the
2445                  * anon_vma->lock.
2446                  */
2447                 if (!__test_and_clear_bit(0, (unsigned long *)
2448                                           &anon_vma->head.next))
2449                         BUG();
2450                 spin_unlock(&anon_vma->lock);
2451         }
2452 }
2453
2454 static void vm_unlock_mapping(struct address_space *mapping)
2455 {
2456         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2457                 /*
2458                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2459                  * because we hold the mm_all_locks_mutex.
2460                  */
2461                 spin_unlock(&mapping->i_mmap_lock);
2462                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2463                                         &mapping->flags))
2464                         BUG();
2465         }
2466 }
2467
2468 /*
2469  * The mmap_sem cannot be released by the caller until
2470  * mm_drop_all_locks() returns.
2471  */
2472 void mm_drop_all_locks(struct mm_struct *mm)
2473 {
2474         struct vm_area_struct *vma;
2475
2476         BUG_ON(down_read_trylock(&mm->mmap_sem));
2477         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2478
2479         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2480                 if (vma->anon_vma)
2481                         vm_unlock_anon_vma(vma->anon_vma);
2482                 if (vma->vm_file && vma->vm_file->f_mapping)
2483                         vm_unlock_mapping(vma->vm_file->f_mapping);
2484         }
2485
2486         mutex_unlock(&mm_all_locks_mutex);
2487 }
2488
2489 /*
2490  * initialise the VMA slab
2491  */
2492 void __init mmap_init(void)
2493 {
2494 }