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