7bdfd2661f173a273d34eaeef7e93244d584990c
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@redhat.com>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30
31 #include <asm/uaccess.h>
32 #include <asm/cacheflush.h>
33 #include <asm/tlb.h>
34 #include <asm/mmu_context.h>
35
36 #include "internal.h"
37
38 #ifndef arch_mmap_check
39 #define arch_mmap_check(addr, len, flags)       (0)
40 #endif
41
42 #ifndef arch_rebalance_pgtables
43 #define arch_rebalance_pgtables(addr, len)              (addr)
44 #endif
45
46 static void unmap_region(struct mm_struct *mm,
47                 struct vm_area_struct *vma, struct vm_area_struct *prev,
48                 unsigned long start, unsigned long end);
49
50 /*
51  * WARNING: the debugging will use recursive algorithms so never enable this
52  * unless you know what you are doing.
53  */
54 #undef DEBUG_MM_RB
55
56 /* description of effects of mapping type and prot in current implementation.
57  * this is due to the limited x86 page protection hardware.  The expected
58  * behavior is in parens:
59  *
60  * map_type     prot
61  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
62  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
63  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
64  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
65  *              
66  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
67  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
68  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
69  *
70  */
71 pgprot_t protection_map[16] = {
72         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
73         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
74 };
75
76 pgprot_t vm_get_page_prot(unsigned long vm_flags)
77 {
78         return __pgprot(pgprot_val(protection_map[vm_flags &
79                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
80                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
81 }
82 EXPORT_SYMBOL(vm_get_page_prot);
83
84 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
85 int sysctl_overcommit_ratio = 50;       /* default is 50% */
86 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
87 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
88
89 /*
90  * Check that a process has enough memory to allocate a new virtual
91  * mapping. 0 means there is enough memory for the allocation to
92  * succeed and -ENOMEM implies there is not.
93  *
94  * We currently support three overcommit policies, which are set via the
95  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
96  *
97  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
98  * Additional code 2002 Jul 20 by Robert Love.
99  *
100  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
101  *
102  * Note this is a helper function intended to be used by LSMs which
103  * wish to use this logic.
104  */
105 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
106 {
107         unsigned long free, allowed;
108
109         vm_acct_memory(pages);
110
111         /*
112          * Sometimes we want to use more memory than we have
113          */
114         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
115                 return 0;
116
117         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
118                 unsigned long n;
119
120                 free = global_page_state(NR_FILE_PAGES);
121                 free += nr_swap_pages;
122
123                 /*
124                  * Any slabs which are created with the
125                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
126                  * which are reclaimable, under pressure.  The dentry
127                  * cache and most inode caches should fall into this
128                  */
129                 free += global_page_state(NR_SLAB_RECLAIMABLE);
130
131                 /*
132                  * Leave the last 3% for root
133                  */
134                 if (!cap_sys_admin)
135                         free -= free / 32;
136
137                 if (free > pages)
138                         return 0;
139
140                 /*
141                  * nr_free_pages() is very expensive on large systems,
142                  * only call if we're about to fail.
143                  */
144                 n = nr_free_pages();
145
146                 /*
147                  * Leave reserved pages. The pages are not for anonymous pages.
148                  */
149                 if (n <= totalreserve_pages)
150                         goto error;
151                 else
152                         n -= totalreserve_pages;
153
154                 /*
155                  * Leave the last 3% for root
156                  */
157                 if (!cap_sys_admin)
158                         n -= n / 32;
159                 free += n;
160
161                 if (free > pages)
162                         return 0;
163
164                 goto error;
165         }
166
167         allowed = (totalram_pages - hugetlb_total_pages())
168                 * sysctl_overcommit_ratio / 100;
169         /*
170          * Leave the last 3% for root
171          */
172         if (!cap_sys_admin)
173                 allowed -= allowed / 32;
174         allowed += total_swap_pages;
175
176         /* Don't let a single process grow too big:
177            leave 3% of the size of this process for other processes */
178         allowed -= mm->total_vm / 32;
179
180         /*
181          * cast `allowed' as a signed long because vm_committed_space
182          * sometimes has a negative value
183          */
184         if (atomic_long_read(&vm_committed_space) < (long)allowed)
185                 return 0;
186 error:
187         vm_unacct_memory(pages);
188
189         return -ENOMEM;
190 }
191
192 /*
193  * Requires inode->i_mapping->i_mmap_lock
194  */
195 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
196                 struct file *file, struct address_space *mapping)
197 {
198         if (vma->vm_flags & VM_DENYWRITE)
199                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
200         if (vma->vm_flags & VM_SHARED)
201                 mapping->i_mmap_writable--;
202
203         flush_dcache_mmap_lock(mapping);
204         if (unlikely(vma->vm_flags & VM_NONLINEAR))
205                 list_del_init(&vma->shared.vm_set.list);
206         else
207                 vma_prio_tree_remove(vma, &mapping->i_mmap);
208         flush_dcache_mmap_unlock(mapping);
209 }
210
211 /*
212  * Unlink a file-based vm structure from its prio_tree, to hide
213  * vma from rmap and vmtruncate before freeing its page tables.
214  */
215 void unlink_file_vma(struct vm_area_struct *vma)
216 {
217         struct file *file = vma->vm_file;
218
219         if (file) {
220                 struct address_space *mapping = file->f_mapping;
221                 spin_lock(&mapping->i_mmap_lock);
222                 __remove_shared_vm_struct(vma, file, mapping);
223                 spin_unlock(&mapping->i_mmap_lock);
224         }
225 }
226
227 /*
228  * Close a vm structure and free it, returning the next.
229  */
230 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
231 {
232         struct vm_area_struct *next = vma->vm_next;
233
234         might_sleep();
235         if (vma->vm_ops && vma->vm_ops->close)
236                 vma->vm_ops->close(vma);
237         if (vma->vm_file) {
238                 fput(vma->vm_file);
239                 if (vma->vm_flags & VM_EXECUTABLE)
240                         removed_exe_file_vma(vma->vm_mm);
241         }
242         mpol_put(vma_policy(vma));
243         kmem_cache_free(vm_area_cachep, vma);
244         return next;
245 }
246
247 asmlinkage unsigned long sys_brk(unsigned long brk)
248 {
249         unsigned long rlim, retval;
250         unsigned long newbrk, oldbrk;
251         struct mm_struct *mm = current->mm;
252         unsigned long min_brk;
253
254         down_write(&mm->mmap_sem);
255
256 #ifdef CONFIG_COMPAT_BRK
257         min_brk = mm->end_code;
258 #else
259         min_brk = mm->start_brk;
260 #endif
261         if (brk < min_brk)
262                 goto out;
263
264         /*
265          * Check against rlimit here. If this check is done later after the test
266          * of oldbrk with newbrk then it can escape the test and let the data
267          * segment grow beyond its set limit the in case where the limit is
268          * not page aligned -Ram Gupta
269          */
270         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
271         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
272                         (mm->end_data - mm->start_data) > rlim)
273                 goto out;
274
275         newbrk = PAGE_ALIGN(brk);
276         oldbrk = PAGE_ALIGN(mm->brk);
277         if (oldbrk == newbrk)
278                 goto set_brk;
279
280         /* Always allow shrinking brk. */
281         if (brk <= mm->brk) {
282                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
283                         goto set_brk;
284                 goto out;
285         }
286
287         /* Check against existing mmap mappings. */
288         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
289                 goto out;
290
291         /* Ok, looks good - let it rip. */
292         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
293                 goto out;
294 set_brk:
295         mm->brk = brk;
296 out:
297         retval = mm->brk;
298         up_write(&mm->mmap_sem);
299         return retval;
300 }
301
302 #ifdef DEBUG_MM_RB
303 static int browse_rb(struct rb_root *root)
304 {
305         int i = 0, j;
306         struct rb_node *nd, *pn = NULL;
307         unsigned long prev = 0, pend = 0;
308
309         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
310                 struct vm_area_struct *vma;
311                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
312                 if (vma->vm_start < prev)
313                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
314                 if (vma->vm_start < pend)
315                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
316                 if (vma->vm_start > vma->vm_end)
317                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
318                 i++;
319                 pn = nd;
320                 prev = vma->vm_start;
321                 pend = vma->vm_end;
322         }
323         j = 0;
324         for (nd = pn; nd; nd = rb_prev(nd)) {
325                 j++;
326         }
327         if (i != j)
328                 printk("backwards %d, forwards %d\n", j, i), i = 0;
329         return i;
330 }
331
332 void validate_mm(struct mm_struct *mm)
333 {
334         int bug = 0;
335         int i = 0;
336         struct vm_area_struct *tmp = mm->mmap;
337         while (tmp) {
338                 tmp = tmp->vm_next;
339                 i++;
340         }
341         if (i != mm->map_count)
342                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
343         i = browse_rb(&mm->mm_rb);
344         if (i != mm->map_count)
345                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
346         BUG_ON(bug);
347 }
348 #else
349 #define validate_mm(mm) do { } while (0)
350 #endif
351
352 static struct vm_area_struct *
353 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
354                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
355                 struct rb_node ** rb_parent)
356 {
357         struct vm_area_struct * vma;
358         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
359
360         __rb_link = &mm->mm_rb.rb_node;
361         rb_prev = __rb_parent = NULL;
362         vma = NULL;
363
364         while (*__rb_link) {
365                 struct vm_area_struct *vma_tmp;
366
367                 __rb_parent = *__rb_link;
368                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
369
370                 if (vma_tmp->vm_end > addr) {
371                         vma = vma_tmp;
372                         if (vma_tmp->vm_start <= addr)
373                                 break;
374                         __rb_link = &__rb_parent->rb_left;
375                 } else {
376                         rb_prev = __rb_parent;
377                         __rb_link = &__rb_parent->rb_right;
378                 }
379         }
380
381         *pprev = NULL;
382         if (rb_prev)
383                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
384         *rb_link = __rb_link;
385         *rb_parent = __rb_parent;
386         return vma;
387 }
388
389 static inline void
390 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
391                 struct vm_area_struct *prev, struct rb_node *rb_parent)
392 {
393         if (prev) {
394                 vma->vm_next = prev->vm_next;
395                 prev->vm_next = vma;
396         } else {
397                 mm->mmap = vma;
398                 if (rb_parent)
399                         vma->vm_next = rb_entry(rb_parent,
400                                         struct vm_area_struct, vm_rb);
401                 else
402                         vma->vm_next = NULL;
403         }
404 }
405
406 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
407                 struct rb_node **rb_link, struct rb_node *rb_parent)
408 {
409         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
410         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
411 }
412
413 static inline void __vma_link_file(struct vm_area_struct *vma)
414 {
415         struct file * file;
416
417         file = vma->vm_file;
418         if (file) {
419                 struct address_space *mapping = file->f_mapping;
420
421                 if (vma->vm_flags & VM_DENYWRITE)
422                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
423                 if (vma->vm_flags & VM_SHARED)
424                         mapping->i_mmap_writable++;
425
426                 flush_dcache_mmap_lock(mapping);
427                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
428                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
429                 else
430                         vma_prio_tree_insert(vma, &mapping->i_mmap);
431                 flush_dcache_mmap_unlock(mapping);
432         }
433 }
434
435 static void
436 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
437         struct vm_area_struct *prev, struct rb_node **rb_link,
438         struct rb_node *rb_parent)
439 {
440         __vma_link_list(mm, vma, prev, rb_parent);
441         __vma_link_rb(mm, vma, rb_link, rb_parent);
442         __anon_vma_link(vma);
443 }
444
445 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
446                         struct vm_area_struct *prev, struct rb_node **rb_link,
447                         struct rb_node *rb_parent)
448 {
449         struct address_space *mapping = NULL;
450
451         if (vma->vm_file)
452                 mapping = vma->vm_file->f_mapping;
453
454         if (mapping) {
455                 spin_lock(&mapping->i_mmap_lock);
456                 vma->vm_truncate_count = mapping->truncate_count;
457         }
458         anon_vma_lock(vma);
459
460         __vma_link(mm, vma, prev, rb_link, rb_parent);
461         __vma_link_file(vma);
462
463         anon_vma_unlock(vma);
464         if (mapping)
465                 spin_unlock(&mapping->i_mmap_lock);
466
467         mm->map_count++;
468         validate_mm(mm);
469 }
470
471 /*
472  * Helper for vma_adjust in the split_vma insert case:
473  * insert vm structure into list and rbtree and anon_vma,
474  * but it has already been inserted into prio_tree earlier.
475  */
476 static void
477 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
478 {
479         struct vm_area_struct * __vma, * prev;
480         struct rb_node ** rb_link, * rb_parent;
481
482         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
483         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
484         __vma_link(mm, vma, prev, rb_link, rb_parent);
485         mm->map_count++;
486 }
487
488 static inline void
489 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
490                 struct vm_area_struct *prev)
491 {
492         prev->vm_next = vma->vm_next;
493         rb_erase(&vma->vm_rb, &mm->mm_rb);
494         if (mm->mmap_cache == vma)
495                 mm->mmap_cache = prev;
496 }
497
498 /*
499  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
500  * is already present in an i_mmap tree without adjusting the tree.
501  * The following helper function should be used when such adjustments
502  * are necessary.  The "insert" vma (if any) is to be inserted
503  * before we drop the necessary locks.
504  */
505 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
506         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
507 {
508         struct mm_struct *mm = vma->vm_mm;
509         struct vm_area_struct *next = vma->vm_next;
510         struct vm_area_struct *importer = NULL;
511         struct address_space *mapping = NULL;
512         struct prio_tree_root *root = NULL;
513         struct file *file = vma->vm_file;
514         struct anon_vma *anon_vma = NULL;
515         long adjust_next = 0;
516         int remove_next = 0;
517
518         if (next && !insert) {
519                 if (end >= next->vm_end) {
520                         /*
521                          * vma expands, overlapping all the next, and
522                          * perhaps the one after too (mprotect case 6).
523                          */
524 again:                  remove_next = 1 + (end > next->vm_end);
525                         end = next->vm_end;
526                         anon_vma = next->anon_vma;
527                         importer = vma;
528                 } else if (end > next->vm_start) {
529                         /*
530                          * vma expands, overlapping part of the next:
531                          * mprotect case 5 shifting the boundary up.
532                          */
533                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
534                         anon_vma = next->anon_vma;
535                         importer = vma;
536                 } else if (end < vma->vm_end) {
537                         /*
538                          * vma shrinks, and !insert tells it's not
539                          * split_vma inserting another: so it must be
540                          * mprotect case 4 shifting the boundary down.
541                          */
542                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
543                         anon_vma = next->anon_vma;
544                         importer = next;
545                 }
546         }
547
548         if (file) {
549                 mapping = file->f_mapping;
550                 if (!(vma->vm_flags & VM_NONLINEAR))
551                         root = &mapping->i_mmap;
552                 spin_lock(&mapping->i_mmap_lock);
553                 if (importer &&
554                     vma->vm_truncate_count != next->vm_truncate_count) {
555                         /*
556                          * unmap_mapping_range might be in progress:
557                          * ensure that the expanding vma is rescanned.
558                          */
559                         importer->vm_truncate_count = 0;
560                 }
561                 if (insert) {
562                         insert->vm_truncate_count = vma->vm_truncate_count;
563                         /*
564                          * Put into prio_tree now, so instantiated pages
565                          * are visible to arm/parisc __flush_dcache_page
566                          * throughout; but we cannot insert into address
567                          * space until vma start or end is updated.
568                          */
569                         __vma_link_file(insert);
570                 }
571         }
572
573         /*
574          * When changing only vma->vm_end, we don't really need
575          * anon_vma lock: but is that case worth optimizing out?
576          */
577         if (vma->anon_vma)
578                 anon_vma = vma->anon_vma;
579         if (anon_vma) {
580                 spin_lock(&anon_vma->lock);
581                 /*
582                  * Easily overlooked: when mprotect shifts the boundary,
583                  * make sure the expanding vma has anon_vma set if the
584                  * shrinking vma had, to cover any anon pages imported.
585                  */
586                 if (importer && !importer->anon_vma) {
587                         importer->anon_vma = anon_vma;
588                         __anon_vma_link(importer);
589                 }
590         }
591
592         if (root) {
593                 flush_dcache_mmap_lock(mapping);
594                 vma_prio_tree_remove(vma, root);
595                 if (adjust_next)
596                         vma_prio_tree_remove(next, root);
597         }
598
599         vma->vm_start = start;
600         vma->vm_end = end;
601         vma->vm_pgoff = pgoff;
602         if (adjust_next) {
603                 next->vm_start += adjust_next << PAGE_SHIFT;
604                 next->vm_pgoff += adjust_next;
605         }
606
607         if (root) {
608                 if (adjust_next)
609                         vma_prio_tree_insert(next, root);
610                 vma_prio_tree_insert(vma, root);
611                 flush_dcache_mmap_unlock(mapping);
612         }
613
614         if (remove_next) {
615                 /*
616                  * vma_merge has merged next into vma, and needs
617                  * us to remove next before dropping the locks.
618                  */
619                 __vma_unlink(mm, next, vma);
620                 if (file)
621                         __remove_shared_vm_struct(next, file, mapping);
622                 if (next->anon_vma)
623                         __anon_vma_merge(vma, next);
624         } else if (insert) {
625                 /*
626                  * split_vma has split insert from vma, and needs
627                  * us to insert it before dropping the locks
628                  * (it may either follow vma or precede it).
629                  */
630                 __insert_vm_struct(mm, insert);
631         }
632
633         if (anon_vma)
634                 spin_unlock(&anon_vma->lock);
635         if (mapping)
636                 spin_unlock(&mapping->i_mmap_lock);
637
638         if (remove_next) {
639                 if (file) {
640                         fput(file);
641                         if (next->vm_flags & VM_EXECUTABLE)
642                                 removed_exe_file_vma(mm);
643                 }
644                 mm->map_count--;
645                 mpol_put(vma_policy(next));
646                 kmem_cache_free(vm_area_cachep, next);
647                 /*
648                  * In mprotect's case 6 (see comments on vma_merge),
649                  * we must remove another next too. It would clutter
650                  * up the code too much to do both in one go.
651                  */
652                 if (remove_next == 2) {
653                         next = vma->vm_next;
654                         goto again;
655                 }
656         }
657
658         validate_mm(mm);
659 }
660
661 /*
662  * If the vma has a ->close operation then the driver probably needs to release
663  * per-vma resources, so we don't attempt to merge those.
664  */
665 static inline int is_mergeable_vma(struct vm_area_struct *vma,
666                         struct file *file, unsigned long vm_flags)
667 {
668         if (vma->vm_flags != vm_flags)
669                 return 0;
670         if (vma->vm_file != file)
671                 return 0;
672         if (vma->vm_ops && vma->vm_ops->close)
673                 return 0;
674         return 1;
675 }
676
677 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
678                                         struct anon_vma *anon_vma2)
679 {
680         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
681 }
682
683 /*
684  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
685  * in front of (at a lower virtual address and file offset than) the vma.
686  *
687  * We cannot merge two vmas if they have differently assigned (non-NULL)
688  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
689  *
690  * We don't check here for the merged mmap wrapping around the end of pagecache
691  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
692  * wrap, nor mmaps which cover the final page at index -1UL.
693  */
694 static int
695 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
696         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
697 {
698         if (is_mergeable_vma(vma, file, vm_flags) &&
699             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
700                 if (vma->vm_pgoff == vm_pgoff)
701                         return 1;
702         }
703         return 0;
704 }
705
706 /*
707  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
708  * beyond (at a higher virtual address and file offset than) the vma.
709  *
710  * We cannot merge two vmas if they have differently assigned (non-NULL)
711  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
712  */
713 static int
714 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
715         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
716 {
717         if (is_mergeable_vma(vma, file, vm_flags) &&
718             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
719                 pgoff_t vm_pglen;
720                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
721                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
722                         return 1;
723         }
724         return 0;
725 }
726
727 /*
728  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
729  * whether that can be merged with its predecessor or its successor.
730  * Or both (it neatly fills a hole).
731  *
732  * In most cases - when called for mmap, brk or mremap - [addr,end) is
733  * certain not to be mapped by the time vma_merge is called; but when
734  * called for mprotect, it is certain to be already mapped (either at
735  * an offset within prev, or at the start of next), and the flags of
736  * this area are about to be changed to vm_flags - and the no-change
737  * case has already been eliminated.
738  *
739  * The following mprotect cases have to be considered, where AAAA is
740  * the area passed down from mprotect_fixup, never extending beyond one
741  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
742  *
743  *     AAAA             AAAA                AAAA          AAAA
744  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
745  *    cannot merge    might become    might become    might become
746  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
747  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
748  *    mremap move:                                    PPPPNNNNNNNN 8
749  *        AAAA
750  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
751  *    might become    case 1 below    case 2 below    case 3 below
752  *
753  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
754  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
755  */
756 struct vm_area_struct *vma_merge(struct mm_struct *mm,
757                         struct vm_area_struct *prev, unsigned long addr,
758                         unsigned long end, unsigned long vm_flags,
759                         struct anon_vma *anon_vma, struct file *file,
760                         pgoff_t pgoff, struct mempolicy *policy)
761 {
762         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
763         struct vm_area_struct *area, *next;
764
765         /*
766          * We later require that vma->vm_flags == vm_flags,
767          * so this tests vma->vm_flags & VM_SPECIAL, too.
768          */
769         if (vm_flags & VM_SPECIAL)
770                 return NULL;
771
772         if (prev)
773                 next = prev->vm_next;
774         else
775                 next = mm->mmap;
776         area = next;
777         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
778                 next = next->vm_next;
779
780         /*
781          * Can it merge with the predecessor?
782          */
783         if (prev && prev->vm_end == addr &&
784                         mpol_equal(vma_policy(prev), policy) &&
785                         can_vma_merge_after(prev, vm_flags,
786                                                 anon_vma, file, pgoff)) {
787                 /*
788                  * OK, it can.  Can we now merge in the successor as well?
789                  */
790                 if (next && end == next->vm_start &&
791                                 mpol_equal(policy, vma_policy(next)) &&
792                                 can_vma_merge_before(next, vm_flags,
793                                         anon_vma, file, pgoff+pglen) &&
794                                 is_mergeable_anon_vma(prev->anon_vma,
795                                                       next->anon_vma)) {
796                                                         /* cases 1, 6 */
797                         vma_adjust(prev, prev->vm_start,
798                                 next->vm_end, prev->vm_pgoff, NULL);
799                 } else                                  /* cases 2, 5, 7 */
800                         vma_adjust(prev, prev->vm_start,
801                                 end, prev->vm_pgoff, NULL);
802                 return prev;
803         }
804
805         /*
806          * Can this new request be merged in front of next?
807          */
808         if (next && end == next->vm_start &&
809                         mpol_equal(policy, vma_policy(next)) &&
810                         can_vma_merge_before(next, vm_flags,
811                                         anon_vma, file, pgoff+pglen)) {
812                 if (prev && addr < prev->vm_end)        /* case 4 */
813                         vma_adjust(prev, prev->vm_start,
814                                 addr, prev->vm_pgoff, NULL);
815                 else                                    /* cases 3, 8 */
816                         vma_adjust(area, addr, next->vm_end,
817                                 next->vm_pgoff - pglen, NULL);
818                 return area;
819         }
820
821         return NULL;
822 }
823
824 /*
825  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
826  * neighbouring vmas for a suitable anon_vma, before it goes off
827  * to allocate a new anon_vma.  It checks because a repetitive
828  * sequence of mprotects and faults may otherwise lead to distinct
829  * anon_vmas being allocated, preventing vma merge in subsequent
830  * mprotect.
831  */
832 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
833 {
834         struct vm_area_struct *near;
835         unsigned long vm_flags;
836
837         near = vma->vm_next;
838         if (!near)
839                 goto try_prev;
840
841         /*
842          * Since only mprotect tries to remerge vmas, match flags
843          * which might be mprotected into each other later on.
844          * Neither mlock nor madvise tries to remerge at present,
845          * so leave their flags as obstructing a merge.
846          */
847         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
848         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
849
850         if (near->anon_vma && vma->vm_end == near->vm_start &&
851                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
852                         can_vma_merge_before(near, vm_flags,
853                                 NULL, vma->vm_file, vma->vm_pgoff +
854                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
855                 return near->anon_vma;
856 try_prev:
857         /*
858          * It is potentially slow to have to call find_vma_prev here.
859          * But it's only on the first write fault on the vma, not
860          * every time, and we could devise a way to avoid it later
861          * (e.g. stash info in next's anon_vma_node when assigning
862          * an anon_vma, or when trying vma_merge).  Another time.
863          */
864         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
865         if (!near)
866                 goto none;
867
868         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
869         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
870
871         if (near->anon_vma && near->vm_end == vma->vm_start &&
872                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
873                         can_vma_merge_after(near, vm_flags,
874                                 NULL, vma->vm_file, vma->vm_pgoff))
875                 return near->anon_vma;
876 none:
877         /*
878          * There's no absolute need to look only at touching neighbours:
879          * we could search further afield for "compatible" anon_vmas.
880          * But it would probably just be a waste of time searching,
881          * or lead to too many vmas hanging off the same anon_vma.
882          * We're trying to allow mprotect remerging later on,
883          * not trying to minimize memory used for anon_vmas.
884          */
885         return NULL;
886 }
887
888 #ifdef CONFIG_PROC_FS
889 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
890                                                 struct file *file, long pages)
891 {
892         const unsigned long stack_flags
893                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
894
895         if (file) {
896                 mm->shared_vm += pages;
897                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
898                         mm->exec_vm += pages;
899         } else if (flags & stack_flags)
900                 mm->stack_vm += pages;
901         if (flags & (VM_RESERVED|VM_IO))
902                 mm->reserved_vm += pages;
903 }
904 #endif /* CONFIG_PROC_FS */
905
906 /*
907  * The caller must hold down_write(current->mm->mmap_sem).
908  */
909
910 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
911                         unsigned long len, unsigned long prot,
912                         unsigned long flags, unsigned long pgoff)
913 {
914         struct mm_struct * mm = current->mm;
915         struct inode *inode;
916         unsigned int vm_flags;
917         int error;
918         int accountable = 1;
919         unsigned long reqprot = prot;
920
921         /*
922          * Does the application expect PROT_READ to imply PROT_EXEC?
923          *
924          * (the exception is when the underlying filesystem is noexec
925          *  mounted, in which case we dont add PROT_EXEC.)
926          */
927         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
928                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
929                         prot |= PROT_EXEC;
930
931         if (!len)
932                 return -EINVAL;
933
934         if (!(flags & MAP_FIXED))
935                 addr = round_hint_to_min(addr);
936
937         error = arch_mmap_check(addr, len, flags);
938         if (error)
939                 return error;
940
941         /* Careful about overflows.. */
942         len = PAGE_ALIGN(len);
943         if (!len || len > TASK_SIZE)
944                 return -ENOMEM;
945
946         /* offset overflow? */
947         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
948                return -EOVERFLOW;
949
950         /* Too many mappings? */
951         if (mm->map_count > sysctl_max_map_count)
952                 return -ENOMEM;
953
954         /* Obtain the address to map to. we verify (or select) it and ensure
955          * that it represents a valid section of the address space.
956          */
957         addr = get_unmapped_area(file, addr, len, pgoff, flags);
958         if (addr & ~PAGE_MASK)
959                 return addr;
960
961         /* Do simple checking here so the lower-level routines won't have
962          * to. we assume access permissions have been handled by the open
963          * of the memory object, so we don't do any here.
964          */
965         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
966                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
967
968         if (flags & MAP_LOCKED) {
969                 if (!can_do_mlock())
970                         return -EPERM;
971                 vm_flags |= VM_LOCKED;
972         }
973         /* mlock MCL_FUTURE? */
974         if (vm_flags & VM_LOCKED) {
975                 unsigned long locked, lock_limit;
976                 locked = len >> PAGE_SHIFT;
977                 locked += mm->locked_vm;
978                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
979                 lock_limit >>= PAGE_SHIFT;
980                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
981                         return -EAGAIN;
982         }
983
984         inode = file ? file->f_path.dentry->d_inode : NULL;
985
986         if (file) {
987                 switch (flags & MAP_TYPE) {
988                 case MAP_SHARED:
989                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
990                                 return -EACCES;
991
992                         /*
993                          * Make sure we don't allow writing to an append-only
994                          * file..
995                          */
996                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
997                                 return -EACCES;
998
999                         /*
1000                          * Make sure there are no mandatory locks on the file.
1001                          */
1002                         if (locks_verify_locked(inode))
1003                                 return -EAGAIN;
1004
1005                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1006                         if (!(file->f_mode & FMODE_WRITE))
1007                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1008
1009                         /* fall through */
1010                 case MAP_PRIVATE:
1011                         if (!(file->f_mode & FMODE_READ))
1012                                 return -EACCES;
1013                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1014                                 if (vm_flags & VM_EXEC)
1015                                         return -EPERM;
1016                                 vm_flags &= ~VM_MAYEXEC;
1017                         }
1018                         if (is_file_hugepages(file))
1019                                 accountable = 0;
1020
1021                         if (!file->f_op || !file->f_op->mmap)
1022                                 return -ENODEV;
1023                         break;
1024
1025                 default:
1026                         return -EINVAL;
1027                 }
1028         } else {
1029                 switch (flags & MAP_TYPE) {
1030                 case MAP_SHARED:
1031                         /*
1032                          * Ignore pgoff.
1033                          */
1034                         pgoff = 0;
1035                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1036                         break;
1037                 case MAP_PRIVATE:
1038                         /*
1039                          * Set pgoff according to addr for anon_vma.
1040                          */
1041                         pgoff = addr >> PAGE_SHIFT;
1042                         break;
1043                 default:
1044                         return -EINVAL;
1045                 }
1046         }
1047
1048         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1049         if (error)
1050                 return error;
1051
1052         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1053                            accountable);
1054 }
1055 EXPORT_SYMBOL(do_mmap_pgoff);
1056
1057 /*
1058  * Some shared mappigns will want the pages marked read-only
1059  * to track write events. If so, we'll downgrade vm_page_prot
1060  * to the private version (using protection_map[] without the
1061  * VM_SHARED bit).
1062  */
1063 int vma_wants_writenotify(struct vm_area_struct *vma)
1064 {
1065         unsigned int vm_flags = vma->vm_flags;
1066
1067         /* If it was private or non-writable, the write bit is already clear */
1068         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1069                 return 0;
1070
1071         /* The backer wishes to know when pages are first written to? */
1072         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1073                 return 1;
1074
1075         /* The open routine did something to the protections already? */
1076         if (pgprot_val(vma->vm_page_prot) !=
1077             pgprot_val(vm_get_page_prot(vm_flags)))
1078                 return 0;
1079
1080         /* Specialty mapping? */
1081         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1082                 return 0;
1083
1084         /* Can the mapping track the dirty pages? */
1085         return vma->vm_file && vma->vm_file->f_mapping &&
1086                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1087 }
1088
1089 unsigned long mmap_region(struct file *file, unsigned long addr,
1090                           unsigned long len, unsigned long flags,
1091                           unsigned int vm_flags, unsigned long pgoff,
1092                           int accountable)
1093 {
1094         struct mm_struct *mm = current->mm;
1095         struct vm_area_struct *vma, *prev;
1096         int correct_wcount = 0;
1097         int error;
1098         struct rb_node **rb_link, *rb_parent;
1099         unsigned long charged = 0;
1100         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1101
1102         /* Clear old maps */
1103         error = -ENOMEM;
1104 munmap_back:
1105         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1106         if (vma && vma->vm_start < addr + len) {
1107                 if (do_munmap(mm, addr, len))
1108                         return -ENOMEM;
1109                 goto munmap_back;
1110         }
1111
1112         /* Check against address space limit. */
1113         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1114                 return -ENOMEM;
1115
1116         if (flags & MAP_NORESERVE)
1117                 vm_flags |= VM_NORESERVE;
1118
1119         if (accountable && (!(flags & MAP_NORESERVE) ||
1120                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1121                 if (vm_flags & VM_SHARED) {
1122                         /* Check memory availability in shmem_file_setup? */
1123                         vm_flags |= VM_ACCOUNT;
1124                 } else if (vm_flags & VM_WRITE) {
1125                         /*
1126                          * Private writable mapping: check memory availability
1127                          */
1128                         charged = len >> PAGE_SHIFT;
1129                         if (security_vm_enough_memory(charged))
1130                                 return -ENOMEM;
1131                         vm_flags |= VM_ACCOUNT;
1132                 }
1133         }
1134
1135         /*
1136          * Can we just expand an old private anonymous mapping?
1137          * The VM_SHARED test is necessary because shmem_zero_setup
1138          * will create the file object for a shared anonymous map below.
1139          */
1140         if (!file && !(vm_flags & VM_SHARED) &&
1141             vma_merge(mm, prev, addr, addr + len, vm_flags,
1142                                         NULL, NULL, pgoff, NULL))
1143                 goto out;
1144
1145         /*
1146          * Determine the object being mapped and call the appropriate
1147          * specific mapper. the address has already been validated, but
1148          * not unmapped, but the maps are removed from the list.
1149          */
1150         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1151         if (!vma) {
1152                 error = -ENOMEM;
1153                 goto unacct_error;
1154         }
1155
1156         vma->vm_mm = mm;
1157         vma->vm_start = addr;
1158         vma->vm_end = addr + len;
1159         vma->vm_flags = vm_flags;
1160         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1161         vma->vm_pgoff = pgoff;
1162
1163         if (file) {
1164                 error = -EINVAL;
1165                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1166                         goto free_vma;
1167                 if (vm_flags & VM_DENYWRITE) {
1168                         error = deny_write_access(file);
1169                         if (error)
1170                                 goto free_vma;
1171                         correct_wcount = 1;
1172                 }
1173                 vma->vm_file = file;
1174                 get_file(file);
1175                 error = file->f_op->mmap(file, vma);
1176                 if (error)
1177                         goto unmap_and_free_vma;
1178                 if (vm_flags & VM_EXECUTABLE)
1179                         added_exe_file_vma(mm);
1180         } else if (vm_flags & VM_SHARED) {
1181                 error = shmem_zero_setup(vma);
1182                 if (error)
1183                         goto free_vma;
1184         }
1185
1186         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1187          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1188          * that memory reservation must be checked; but that reservation
1189          * belongs to shared memory object, not to vma: so now clear it.
1190          */
1191         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1192                 vma->vm_flags &= ~VM_ACCOUNT;
1193
1194         /* Can addr have changed??
1195          *
1196          * Answer: Yes, several device drivers can do it in their
1197          *         f_op->mmap method. -DaveM
1198          */
1199         addr = vma->vm_start;
1200         pgoff = vma->vm_pgoff;
1201         vm_flags = vma->vm_flags;
1202
1203         if (vma_wants_writenotify(vma))
1204                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1205
1206         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1207                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1208                 mpol_put(vma_policy(vma));
1209                 kmem_cache_free(vm_area_cachep, vma);
1210                 fput(file);
1211                 if (vm_flags & VM_EXECUTABLE)
1212                         removed_exe_file_vma(mm);
1213         } else {
1214                 vma_link(mm, vma, prev, rb_link, rb_parent);
1215                 file = vma->vm_file;
1216         }
1217
1218         /* Once vma denies write, undo our temporary denial count */
1219         if (correct_wcount)
1220                 atomic_inc(&inode->i_writecount);
1221 out:
1222         mm->total_vm += len >> PAGE_SHIFT;
1223         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1224         if (vm_flags & VM_LOCKED) {
1225                 mm->locked_vm += len >> PAGE_SHIFT;
1226                 make_pages_present(addr, addr + len);
1227         }
1228         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1229                 make_pages_present(addr, addr + len);
1230         return addr;
1231
1232 unmap_and_free_vma:
1233         if (correct_wcount)
1234                 atomic_inc(&inode->i_writecount);
1235         vma->vm_file = NULL;
1236         fput(file);
1237
1238         /* Undo any partial mapping done by a device driver. */
1239         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1240         charged = 0;
1241 free_vma:
1242         kmem_cache_free(vm_area_cachep, vma);
1243 unacct_error:
1244         if (charged)
1245                 vm_unacct_memory(charged);
1246         return error;
1247 }
1248
1249 /* Get an address range which is currently unmapped.
1250  * For shmat() with addr=0.
1251  *
1252  * Ugly calling convention alert:
1253  * Return value with the low bits set means error value,
1254  * ie
1255  *      if (ret & ~PAGE_MASK)
1256  *              error = ret;
1257  *
1258  * This function "knows" that -ENOMEM has the bits set.
1259  */
1260 #ifndef HAVE_ARCH_UNMAPPED_AREA
1261 unsigned long
1262 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1263                 unsigned long len, unsigned long pgoff, unsigned long flags)
1264 {
1265         struct mm_struct *mm = current->mm;
1266         struct vm_area_struct *vma;
1267         unsigned long start_addr;
1268
1269         if (len > TASK_SIZE)
1270                 return -ENOMEM;
1271
1272         if (flags & MAP_FIXED)
1273                 return addr;
1274
1275         if (addr) {
1276                 addr = PAGE_ALIGN(addr);
1277                 vma = find_vma(mm, addr);
1278                 if (TASK_SIZE - len >= addr &&
1279                     (!vma || addr + len <= vma->vm_start))
1280                         return addr;
1281         }
1282         if (len > mm->cached_hole_size) {
1283                 start_addr = addr = mm->free_area_cache;
1284         } else {
1285                 start_addr = addr = TASK_UNMAPPED_BASE;
1286                 mm->cached_hole_size = 0;
1287         }
1288
1289 full_search:
1290         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1291                 /* At this point:  (!vma || addr < vma->vm_end). */
1292                 if (TASK_SIZE - len < addr) {
1293                         /*
1294                          * Start a new search - just in case we missed
1295                          * some holes.
1296                          */
1297                         if (start_addr != TASK_UNMAPPED_BASE) {
1298                                 addr = TASK_UNMAPPED_BASE;
1299                                 start_addr = addr;
1300                                 mm->cached_hole_size = 0;
1301                                 goto full_search;
1302                         }
1303                         return -ENOMEM;
1304                 }
1305                 if (!vma || addr + len <= vma->vm_start) {
1306                         /*
1307                          * Remember the place where we stopped the search:
1308                          */
1309                         mm->free_area_cache = addr + len;
1310                         return addr;
1311                 }
1312                 if (addr + mm->cached_hole_size < vma->vm_start)
1313                         mm->cached_hole_size = vma->vm_start - addr;
1314                 addr = vma->vm_end;
1315         }
1316 }
1317 #endif  
1318
1319 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1320 {
1321         /*
1322          * Is this a new hole at the lowest possible address?
1323          */
1324         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1325                 mm->free_area_cache = addr;
1326                 mm->cached_hole_size = ~0UL;
1327         }
1328 }
1329
1330 /*
1331  * This mmap-allocator allocates new areas top-down from below the
1332  * stack's low limit (the base):
1333  */
1334 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1335 unsigned long
1336 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1337                           const unsigned long len, const unsigned long pgoff,
1338                           const unsigned long flags)
1339 {
1340         struct vm_area_struct *vma;
1341         struct mm_struct *mm = current->mm;
1342         unsigned long addr = addr0;
1343
1344         /* requested length too big for entire address space */
1345         if (len > TASK_SIZE)
1346                 return -ENOMEM;
1347
1348         if (flags & MAP_FIXED)
1349                 return addr;
1350
1351         /* requesting a specific address */
1352         if (addr) {
1353                 addr = PAGE_ALIGN(addr);
1354                 vma = find_vma(mm, addr);
1355                 if (TASK_SIZE - len >= addr &&
1356                                 (!vma || addr + len <= vma->vm_start))
1357                         return addr;
1358         }
1359
1360         /* check if free_area_cache is useful for us */
1361         if (len <= mm->cached_hole_size) {
1362                 mm->cached_hole_size = 0;
1363                 mm->free_area_cache = mm->mmap_base;
1364         }
1365
1366         /* either no address requested or can't fit in requested address hole */
1367         addr = mm->free_area_cache;
1368
1369         /* make sure it can fit in the remaining address space */
1370         if (addr > len) {
1371                 vma = find_vma(mm, addr-len);
1372                 if (!vma || addr <= vma->vm_start)
1373                         /* remember the address as a hint for next time */
1374                         return (mm->free_area_cache = addr-len);
1375         }
1376
1377         if (mm->mmap_base < len)
1378                 goto bottomup;
1379
1380         addr = mm->mmap_base-len;
1381
1382         do {
1383                 /*
1384                  * Lookup failure means no vma is above this address,
1385                  * else if new region fits below vma->vm_start,
1386                  * return with success:
1387                  */
1388                 vma = find_vma(mm, addr);
1389                 if (!vma || addr+len <= vma->vm_start)
1390                         /* remember the address as a hint for next time */
1391                         return (mm->free_area_cache = addr);
1392
1393                 /* remember the largest hole we saw so far */
1394                 if (addr + mm->cached_hole_size < vma->vm_start)
1395                         mm->cached_hole_size = vma->vm_start - addr;
1396
1397                 /* try just below the current vma->vm_start */
1398                 addr = vma->vm_start-len;
1399         } while (len < vma->vm_start);
1400
1401 bottomup:
1402         /*
1403          * A failed mmap() very likely causes application failure,
1404          * so fall back to the bottom-up function here. This scenario
1405          * can happen with large stack limits and large mmap()
1406          * allocations.
1407          */
1408         mm->cached_hole_size = ~0UL;
1409         mm->free_area_cache = TASK_UNMAPPED_BASE;
1410         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1411         /*
1412          * Restore the topdown base:
1413          */
1414         mm->free_area_cache = mm->mmap_base;
1415         mm->cached_hole_size = ~0UL;
1416
1417         return addr;
1418 }
1419 #endif
1420
1421 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1422 {
1423         /*
1424          * Is this a new hole at the highest possible address?
1425          */
1426         if (addr > mm->free_area_cache)
1427                 mm->free_area_cache = addr;
1428
1429         /* dont allow allocations above current base */
1430         if (mm->free_area_cache > mm->mmap_base)
1431                 mm->free_area_cache = mm->mmap_base;
1432 }
1433
1434 unsigned long
1435 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1436                 unsigned long pgoff, unsigned long flags)
1437 {
1438         unsigned long (*get_area)(struct file *, unsigned long,
1439                                   unsigned long, unsigned long, unsigned long);
1440
1441         get_area = current->mm->get_unmapped_area;
1442         if (file && file->f_op && file->f_op->get_unmapped_area)
1443                 get_area = file->f_op->get_unmapped_area;
1444         addr = get_area(file, addr, len, pgoff, flags);
1445         if (IS_ERR_VALUE(addr))
1446                 return addr;
1447
1448         if (addr > TASK_SIZE - len)
1449                 return -ENOMEM;
1450         if (addr & ~PAGE_MASK)
1451                 return -EINVAL;
1452
1453         return arch_rebalance_pgtables(addr, len);
1454 }
1455
1456 EXPORT_SYMBOL(get_unmapped_area);
1457
1458 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1459 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1460 {
1461         struct vm_area_struct *vma = NULL;
1462
1463         if (mm) {
1464                 /* Check the cache first. */
1465                 /* (Cache hit rate is typically around 35%.) */
1466                 vma = mm->mmap_cache;
1467                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1468                         struct rb_node * rb_node;
1469
1470                         rb_node = mm->mm_rb.rb_node;
1471                         vma = NULL;
1472
1473                         while (rb_node) {
1474                                 struct vm_area_struct * vma_tmp;
1475
1476                                 vma_tmp = rb_entry(rb_node,
1477                                                 struct vm_area_struct, vm_rb);
1478
1479                                 if (vma_tmp->vm_end > addr) {
1480                                         vma = vma_tmp;
1481                                         if (vma_tmp->vm_start <= addr)
1482                                                 break;
1483                                         rb_node = rb_node->rb_left;
1484                                 } else
1485                                         rb_node = rb_node->rb_right;
1486                         }
1487                         if (vma)
1488                                 mm->mmap_cache = vma;
1489                 }
1490         }
1491         return vma;
1492 }
1493
1494 EXPORT_SYMBOL(find_vma);
1495
1496 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1497 struct vm_area_struct *
1498 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1499                         struct vm_area_struct **pprev)
1500 {
1501         struct vm_area_struct *vma = NULL, *prev = NULL;
1502         struct rb_node * rb_node;
1503         if (!mm)
1504                 goto out;
1505
1506         /* Guard against addr being lower than the first VMA */
1507         vma = mm->mmap;
1508
1509         /* Go through the RB tree quickly. */
1510         rb_node = mm->mm_rb.rb_node;
1511
1512         while (rb_node) {
1513                 struct vm_area_struct *vma_tmp;
1514                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1515
1516                 if (addr < vma_tmp->vm_end) {
1517                         rb_node = rb_node->rb_left;
1518                 } else {
1519                         prev = vma_tmp;
1520                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1521                                 break;
1522                         rb_node = rb_node->rb_right;
1523                 }
1524         }
1525
1526 out:
1527         *pprev = prev;
1528         return prev ? prev->vm_next : vma;
1529 }
1530
1531 /*
1532  * Verify that the stack growth is acceptable and
1533  * update accounting. This is shared with both the
1534  * grow-up and grow-down cases.
1535  */
1536 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1537 {
1538         struct mm_struct *mm = vma->vm_mm;
1539         struct rlimit *rlim = current->signal->rlim;
1540         unsigned long new_start;
1541
1542         /* address space limit tests */
1543         if (!may_expand_vm(mm, grow))
1544                 return -ENOMEM;
1545
1546         /* Stack limit test */
1547         if (size > rlim[RLIMIT_STACK].rlim_cur)
1548                 return -ENOMEM;
1549
1550         /* mlock limit tests */
1551         if (vma->vm_flags & VM_LOCKED) {
1552                 unsigned long locked;
1553                 unsigned long limit;
1554                 locked = mm->locked_vm + grow;
1555                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1556                 if (locked > limit && !capable(CAP_IPC_LOCK))
1557                         return -ENOMEM;
1558         }
1559
1560         /* Check to ensure the stack will not grow into a hugetlb-only region */
1561         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1562                         vma->vm_end - size;
1563         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1564                 return -EFAULT;
1565
1566         /*
1567          * Overcommit..  This must be the final test, as it will
1568          * update security statistics.
1569          */
1570         if (security_vm_enough_memory(grow))
1571                 return -ENOMEM;
1572
1573         /* Ok, everything looks good - let it rip */
1574         mm->total_vm += grow;
1575         if (vma->vm_flags & VM_LOCKED)
1576                 mm->locked_vm += grow;
1577         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1578         return 0;
1579 }
1580
1581 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1582 /*
1583  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1584  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1585  */
1586 #ifndef CONFIG_IA64
1587 static inline
1588 #endif
1589 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1590 {
1591         int error;
1592
1593         if (!(vma->vm_flags & VM_GROWSUP))
1594                 return -EFAULT;
1595
1596         /*
1597          * We must make sure the anon_vma is allocated
1598          * so that the anon_vma locking is not a noop.
1599          */
1600         if (unlikely(anon_vma_prepare(vma)))
1601                 return -ENOMEM;
1602         anon_vma_lock(vma);
1603
1604         /*
1605          * vma->vm_start/vm_end cannot change under us because the caller
1606          * is required to hold the mmap_sem in read mode.  We need the
1607          * anon_vma lock to serialize against concurrent expand_stacks.
1608          * Also guard against wrapping around to address 0.
1609          */
1610         if (address < PAGE_ALIGN(address+4))
1611                 address = PAGE_ALIGN(address+4);
1612         else {
1613                 anon_vma_unlock(vma);
1614                 return -ENOMEM;
1615         }
1616         error = 0;
1617
1618         /* Somebody else might have raced and expanded it already */
1619         if (address > vma->vm_end) {
1620                 unsigned long size, grow;
1621
1622                 size = address - vma->vm_start;
1623                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1624
1625                 error = acct_stack_growth(vma, size, grow);
1626                 if (!error)
1627                         vma->vm_end = address;
1628         }
1629         anon_vma_unlock(vma);
1630         return error;
1631 }
1632 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1633
1634 /*
1635  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1636  */
1637 static inline int expand_downwards(struct vm_area_struct *vma,
1638                                    unsigned long address)
1639 {
1640         int error;
1641
1642         /*
1643          * We must make sure the anon_vma is allocated
1644          * so that the anon_vma locking is not a noop.
1645          */
1646         if (unlikely(anon_vma_prepare(vma)))
1647                 return -ENOMEM;
1648
1649         address &= PAGE_MASK;
1650         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1651         if (error)
1652                 return error;
1653
1654         anon_vma_lock(vma);
1655
1656         /*
1657          * vma->vm_start/vm_end cannot change under us because the caller
1658          * is required to hold the mmap_sem in read mode.  We need the
1659          * anon_vma lock to serialize against concurrent expand_stacks.
1660          */
1661
1662         /* Somebody else might have raced and expanded it already */
1663         if (address < vma->vm_start) {
1664                 unsigned long size, grow;
1665
1666                 size = vma->vm_end - address;
1667                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1668
1669                 error = acct_stack_growth(vma, size, grow);
1670                 if (!error) {
1671                         vma->vm_start = address;
1672                         vma->vm_pgoff -= grow;
1673                 }
1674         }
1675         anon_vma_unlock(vma);
1676         return error;
1677 }
1678
1679 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1680 {
1681         return expand_downwards(vma, address);
1682 }
1683
1684 #ifdef CONFIG_STACK_GROWSUP
1685 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1686 {
1687         return expand_upwards(vma, address);
1688 }
1689
1690 struct vm_area_struct *
1691 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1692 {
1693         struct vm_area_struct *vma, *prev;
1694
1695         addr &= PAGE_MASK;
1696         vma = find_vma_prev(mm, addr, &prev);
1697         if (vma && (vma->vm_start <= addr))
1698                 return vma;
1699         if (!prev || expand_stack(prev, addr))
1700                 return NULL;
1701         if (prev->vm_flags & VM_LOCKED)
1702                 make_pages_present(addr, prev->vm_end);
1703         return prev;
1704 }
1705 #else
1706 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1707 {
1708         return expand_downwards(vma, address);
1709 }
1710
1711 struct vm_area_struct *
1712 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1713 {
1714         struct vm_area_struct * vma;
1715         unsigned long start;
1716
1717         addr &= PAGE_MASK;
1718         vma = find_vma(mm,addr);
1719         if (!vma)
1720                 return NULL;
1721         if (vma->vm_start <= addr)
1722                 return vma;
1723         if (!(vma->vm_flags & VM_GROWSDOWN))
1724                 return NULL;
1725         start = vma->vm_start;
1726         if (expand_stack(vma, addr))
1727                 return NULL;
1728         if (vma->vm_flags & VM_LOCKED)
1729                 make_pages_present(addr, start);
1730         return vma;
1731 }
1732 #endif
1733
1734 /*
1735  * Ok - we have the memory areas we should free on the vma list,
1736  * so release them, and do the vma updates.
1737  *
1738  * Called with the mm semaphore held.
1739  */
1740 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1741 {
1742         /* Update high watermark before we lower total_vm */
1743         update_hiwater_vm(mm);
1744         do {
1745                 long nrpages = vma_pages(vma);
1746
1747                 mm->total_vm -= nrpages;
1748                 if (vma->vm_flags & VM_LOCKED)
1749                         mm->locked_vm -= nrpages;
1750                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1751                 vma = remove_vma(vma);
1752         } while (vma);
1753         validate_mm(mm);
1754 }
1755
1756 /*
1757  * Get rid of page table information in the indicated region.
1758  *
1759  * Called with the mm semaphore held.
1760  */
1761 static void unmap_region(struct mm_struct *mm,
1762                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1763                 unsigned long start, unsigned long end)
1764 {
1765         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1766         struct mmu_gather *tlb;
1767         unsigned long nr_accounted = 0;
1768
1769         lru_add_drain();
1770         tlb = tlb_gather_mmu(mm, 0);
1771         update_hiwater_rss(mm);
1772         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1773         vm_unacct_memory(nr_accounted);
1774         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1775                                  next? next->vm_start: 0);
1776         tlb_finish_mmu(tlb, start, end);
1777 }
1778
1779 /*
1780  * Create a list of vma's touched by the unmap, removing them from the mm's
1781  * vma list as we go..
1782  */
1783 static void
1784 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1785         struct vm_area_struct *prev, unsigned long end)
1786 {
1787         struct vm_area_struct **insertion_point;
1788         struct vm_area_struct *tail_vma = NULL;
1789         unsigned long addr;
1790
1791         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1792         do {
1793                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1794                 mm->map_count--;
1795                 tail_vma = vma;
1796                 vma = vma->vm_next;
1797         } while (vma && vma->vm_start < end);
1798         *insertion_point = vma;
1799         tail_vma->vm_next = NULL;
1800         if (mm->unmap_area == arch_unmap_area)
1801                 addr = prev ? prev->vm_end : mm->mmap_base;
1802         else
1803                 addr = vma ?  vma->vm_start : mm->mmap_base;
1804         mm->unmap_area(mm, addr);
1805         mm->mmap_cache = NULL;          /* Kill the cache. */
1806 }
1807
1808 /*
1809  * Split a vma into two pieces at address 'addr', a new vma is allocated
1810  * either for the first part or the tail.
1811  */
1812 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1813               unsigned long addr, int new_below)
1814 {
1815         struct mempolicy *pol;
1816         struct vm_area_struct *new;
1817
1818         if (is_vm_hugetlb_page(vma) && (addr &
1819                                         ~(huge_page_mask(hstate_vma(vma)))))
1820                 return -EINVAL;
1821
1822         if (mm->map_count >= sysctl_max_map_count)
1823                 return -ENOMEM;
1824
1825         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1826         if (!new)
1827                 return -ENOMEM;
1828
1829         /* most fields are the same, copy all, and then fixup */
1830         *new = *vma;
1831
1832         if (new_below)
1833                 new->vm_end = addr;
1834         else {
1835                 new->vm_start = addr;
1836                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1837         }
1838
1839         pol = mpol_dup(vma_policy(vma));
1840         if (IS_ERR(pol)) {
1841                 kmem_cache_free(vm_area_cachep, new);
1842                 return PTR_ERR(pol);
1843         }
1844         vma_set_policy(new, pol);
1845
1846         if (new->vm_file) {
1847                 get_file(new->vm_file);
1848                 if (vma->vm_flags & VM_EXECUTABLE)
1849                         added_exe_file_vma(mm);
1850         }
1851
1852         if (new->vm_ops && new->vm_ops->open)
1853                 new->vm_ops->open(new);
1854
1855         if (new_below)
1856                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1857                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1858         else
1859                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1860
1861         return 0;
1862 }
1863
1864 /* Munmap is split into 2 main parts -- this part which finds
1865  * what needs doing, and the areas themselves, which do the
1866  * work.  This now handles partial unmappings.
1867  * Jeremy Fitzhardinge <jeremy@goop.org>
1868  */
1869 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1870 {
1871         unsigned long end;
1872         struct vm_area_struct *vma, *prev, *last;
1873
1874         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1875                 return -EINVAL;
1876
1877         if ((len = PAGE_ALIGN(len)) == 0)
1878                 return -EINVAL;
1879
1880         /* Find the first overlapping VMA */
1881         vma = find_vma_prev(mm, start, &prev);
1882         if (!vma)
1883                 return 0;
1884         /* we have  start < vma->vm_end  */
1885
1886         /* if it doesn't overlap, we have nothing.. */
1887         end = start + len;
1888         if (vma->vm_start >= end)
1889                 return 0;
1890
1891         /*
1892          * If we need to split any vma, do it now to save pain later.
1893          *
1894          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1895          * unmapped vm_area_struct will remain in use: so lower split_vma
1896          * places tmp vma above, and higher split_vma places tmp vma below.
1897          */
1898         if (start > vma->vm_start) {
1899                 int error = split_vma(mm, vma, start, 0);
1900                 if (error)
1901                         return error;
1902                 prev = vma;
1903         }
1904
1905         /* Does it split the last one? */
1906         last = find_vma(mm, end);
1907         if (last && end > last->vm_start) {
1908                 int error = split_vma(mm, last, end, 1);
1909                 if (error)
1910                         return error;
1911         }
1912         vma = prev? prev->vm_next: mm->mmap;
1913
1914         /*
1915          * Remove the vma's, and unmap the actual pages
1916          */
1917         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1918         unmap_region(mm, vma, prev, start, end);
1919
1920         /* Fix up all other VM information */
1921         remove_vma_list(mm, vma);
1922
1923         return 0;
1924 }
1925
1926 EXPORT_SYMBOL(do_munmap);
1927
1928 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1929 {
1930         int ret;
1931         struct mm_struct *mm = current->mm;
1932
1933         profile_munmap(addr);
1934
1935         down_write(&mm->mmap_sem);
1936         ret = do_munmap(mm, addr, len);
1937         up_write(&mm->mmap_sem);
1938         return ret;
1939 }
1940
1941 static inline void verify_mm_writelocked(struct mm_struct *mm)
1942 {
1943 #ifdef CONFIG_DEBUG_VM
1944         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1945                 WARN_ON(1);
1946                 up_read(&mm->mmap_sem);
1947         }
1948 #endif
1949 }
1950
1951 /*
1952  *  this is really a simplified "do_mmap".  it only handles
1953  *  anonymous maps.  eventually we may be able to do some
1954  *  brk-specific accounting here.
1955  */
1956 unsigned long do_brk(unsigned long addr, unsigned long len)
1957 {
1958         struct mm_struct * mm = current->mm;
1959         struct vm_area_struct * vma, * prev;
1960         unsigned long flags;
1961         struct rb_node ** rb_link, * rb_parent;
1962         pgoff_t pgoff = addr >> PAGE_SHIFT;
1963         int error;
1964
1965         len = PAGE_ALIGN(len);
1966         if (!len)
1967                 return addr;
1968
1969         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1970                 return -EINVAL;
1971
1972         if (is_hugepage_only_range(mm, addr, len))
1973                 return -EINVAL;
1974
1975         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1976         if (error)
1977                 return error;
1978
1979         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1980
1981         error = arch_mmap_check(addr, len, flags);
1982         if (error)
1983                 return error;
1984
1985         /*
1986          * mlock MCL_FUTURE?
1987          */
1988         if (mm->def_flags & VM_LOCKED) {
1989                 unsigned long locked, lock_limit;
1990                 locked = len >> PAGE_SHIFT;
1991                 locked += mm->locked_vm;
1992                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1993                 lock_limit >>= PAGE_SHIFT;
1994                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1995                         return -EAGAIN;
1996         }
1997
1998         /*
1999          * mm->mmap_sem is required to protect against another thread
2000          * changing the mappings in case we sleep.
2001          */
2002         verify_mm_writelocked(mm);
2003
2004         /*
2005          * Clear old maps.  this also does some error checking for us
2006          */
2007  munmap_back:
2008         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2009         if (vma && vma->vm_start < addr + len) {
2010                 if (do_munmap(mm, addr, len))
2011                         return -ENOMEM;
2012                 goto munmap_back;
2013         }
2014
2015         /* Check against address space limits *after* clearing old maps... */
2016         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2017                 return -ENOMEM;
2018
2019         if (mm->map_count > sysctl_max_map_count)
2020                 return -ENOMEM;
2021
2022         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2023                 return -ENOMEM;
2024
2025         /* Can we just expand an old private anonymous mapping? */
2026         if (vma_merge(mm, prev, addr, addr + len, flags,
2027                                         NULL, NULL, pgoff, NULL))
2028                 goto out;
2029
2030         /*
2031          * create a vma struct for an anonymous mapping
2032          */
2033         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2034         if (!vma) {
2035                 vm_unacct_memory(len >> PAGE_SHIFT);
2036                 return -ENOMEM;
2037         }
2038
2039         vma->vm_mm = mm;
2040         vma->vm_start = addr;
2041         vma->vm_end = addr + len;
2042         vma->vm_pgoff = pgoff;
2043         vma->vm_flags = flags;
2044         vma->vm_page_prot = vm_get_page_prot(flags);
2045         vma_link(mm, vma, prev, rb_link, rb_parent);
2046 out:
2047         mm->total_vm += len >> PAGE_SHIFT;
2048         if (flags & VM_LOCKED) {
2049                 mm->locked_vm += len >> PAGE_SHIFT;
2050                 make_pages_present(addr, addr + len);
2051         }
2052         return addr;
2053 }
2054
2055 EXPORT_SYMBOL(do_brk);
2056
2057 /* Release all mmaps. */
2058 void exit_mmap(struct mm_struct *mm)
2059 {
2060         struct mmu_gather *tlb;
2061         struct vm_area_struct *vma = mm->mmap;
2062         unsigned long nr_accounted = 0;
2063         unsigned long end;
2064
2065         /* mm's last user has gone, and its about to be pulled down */
2066         arch_exit_mmap(mm);
2067         mmu_notifier_release(mm);
2068
2069         lru_add_drain();
2070         flush_cache_mm(mm);
2071         tlb = tlb_gather_mmu(mm, 1);
2072         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2073         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2074         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2075         vm_unacct_memory(nr_accounted);
2076         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2077         tlb_finish_mmu(tlb, 0, end);
2078
2079         /*
2080          * Walk the list again, actually closing and freeing it,
2081          * with preemption enabled, without holding any MM locks.
2082          */
2083         while (vma)
2084                 vma = remove_vma(vma);
2085
2086         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2087 }
2088
2089 /* Insert vm structure into process list sorted by address
2090  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2091  * then i_mmap_lock is taken here.
2092  */
2093 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2094 {
2095         struct vm_area_struct * __vma, * prev;
2096         struct rb_node ** rb_link, * rb_parent;
2097
2098         /*
2099          * The vm_pgoff of a purely anonymous vma should be irrelevant
2100          * until its first write fault, when page's anon_vma and index
2101          * are set.  But now set the vm_pgoff it will almost certainly
2102          * end up with (unless mremap moves it elsewhere before that
2103          * first wfault), so /proc/pid/maps tells a consistent story.
2104          *
2105          * By setting it to reflect the virtual start address of the
2106          * vma, merges and splits can happen in a seamless way, just
2107          * using the existing file pgoff checks and manipulations.
2108          * Similarly in do_mmap_pgoff and in do_brk.
2109          */
2110         if (!vma->vm_file) {
2111                 BUG_ON(vma->anon_vma);
2112                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2113         }
2114         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2115         if (__vma && __vma->vm_start < vma->vm_end)
2116                 return -ENOMEM;
2117         if ((vma->vm_flags & VM_ACCOUNT) &&
2118              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2119                 return -ENOMEM;
2120         vma_link(mm, vma, prev, rb_link, rb_parent);
2121         return 0;
2122 }
2123
2124 /*
2125  * Copy the vma structure to a new location in the same mm,
2126  * prior to moving page table entries, to effect an mremap move.
2127  */
2128 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2129         unsigned long addr, unsigned long len, pgoff_t pgoff)
2130 {
2131         struct vm_area_struct *vma = *vmap;
2132         unsigned long vma_start = vma->vm_start;
2133         struct mm_struct *mm = vma->vm_mm;
2134         struct vm_area_struct *new_vma, *prev;
2135         struct rb_node **rb_link, *rb_parent;
2136         struct mempolicy *pol;
2137
2138         /*
2139          * If anonymous vma has not yet been faulted, update new pgoff
2140          * to match new location, to increase its chance of merging.
2141          */
2142         if (!vma->vm_file && !vma->anon_vma)
2143                 pgoff = addr >> PAGE_SHIFT;
2144
2145         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2146         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2147                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2148         if (new_vma) {
2149                 /*
2150                  * Source vma may have been merged into new_vma
2151                  */
2152                 if (vma_start >= new_vma->vm_start &&
2153                     vma_start < new_vma->vm_end)
2154                         *vmap = new_vma;
2155         } else {
2156                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2157                 if (new_vma) {
2158                         *new_vma = *vma;
2159                         pol = mpol_dup(vma_policy(vma));
2160                         if (IS_ERR(pol)) {
2161                                 kmem_cache_free(vm_area_cachep, new_vma);
2162                                 return NULL;
2163                         }
2164                         vma_set_policy(new_vma, pol);
2165                         new_vma->vm_start = addr;
2166                         new_vma->vm_end = addr + len;
2167                         new_vma->vm_pgoff = pgoff;
2168                         if (new_vma->vm_file) {
2169                                 get_file(new_vma->vm_file);
2170                                 if (vma->vm_flags & VM_EXECUTABLE)
2171                                         added_exe_file_vma(mm);
2172                         }
2173                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2174                                 new_vma->vm_ops->open(new_vma);
2175                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2176                 }
2177         }
2178         return new_vma;
2179 }
2180
2181 /*
2182  * Return true if the calling process may expand its vm space by the passed
2183  * number of pages
2184  */
2185 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2186 {
2187         unsigned long cur = mm->total_vm;       /* pages */
2188         unsigned long lim;
2189
2190         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2191
2192         if (cur + npages > lim)
2193                 return 0;
2194         return 1;
2195 }
2196
2197
2198 static int special_mapping_fault(struct vm_area_struct *vma,
2199                                 struct vm_fault *vmf)
2200 {
2201         pgoff_t pgoff;
2202         struct page **pages;
2203
2204         /*
2205          * special mappings have no vm_file, and in that case, the mm
2206          * uses vm_pgoff internally. So we have to subtract it from here.
2207          * We are allowed to do this because we are the mm; do not copy
2208          * this code into drivers!
2209          */
2210         pgoff = vmf->pgoff - vma->vm_pgoff;
2211
2212         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2213                 pgoff--;
2214
2215         if (*pages) {
2216                 struct page *page = *pages;
2217                 get_page(page);
2218                 vmf->page = page;
2219                 return 0;
2220         }
2221
2222         return VM_FAULT_SIGBUS;
2223 }
2224
2225 /*
2226  * Having a close hook prevents vma merging regardless of flags.
2227  */
2228 static void special_mapping_close(struct vm_area_struct *vma)
2229 {
2230 }
2231
2232 static struct vm_operations_struct special_mapping_vmops = {
2233         .close = special_mapping_close,
2234         .fault = special_mapping_fault,
2235 };
2236
2237 /*
2238  * Called with mm->mmap_sem held for writing.
2239  * Insert a new vma covering the given region, with the given flags.
2240  * Its pages are supplied by the given array of struct page *.
2241  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2242  * The region past the last page supplied will always produce SIGBUS.
2243  * The array pointer and the pages it points to are assumed to stay alive
2244  * for as long as this mapping might exist.
2245  */
2246 int install_special_mapping(struct mm_struct *mm,
2247                             unsigned long addr, unsigned long len,
2248                             unsigned long vm_flags, struct page **pages)
2249 {
2250         struct vm_area_struct *vma;
2251
2252         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2253         if (unlikely(vma == NULL))
2254                 return -ENOMEM;
2255
2256         vma->vm_mm = mm;
2257         vma->vm_start = addr;
2258         vma->vm_end = addr + len;
2259
2260         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2261         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2262
2263         vma->vm_ops = &special_mapping_vmops;
2264         vma->vm_private_data = pages;
2265
2266         if (unlikely(insert_vm_struct(mm, vma))) {
2267                 kmem_cache_free(vm_area_cachep, vma);
2268                 return -ENOMEM;
2269         }
2270
2271         mm->total_vm += len >> PAGE_SHIFT;
2272
2273         return 0;
2274 }
2275
2276 static DEFINE_MUTEX(mm_all_locks_mutex);
2277
2278 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2279 {
2280         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2281                 /*
2282                  * The LSB of head.next can't change from under us
2283                  * because we hold the mm_all_locks_mutex.
2284                  */
2285                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2286                 /*
2287                  * We can safely modify head.next after taking the
2288                  * anon_vma->lock. If some other vma in this mm shares
2289                  * the same anon_vma we won't take it again.
2290                  *
2291                  * No need of atomic instructions here, head.next
2292                  * can't change from under us thanks to the
2293                  * anon_vma->lock.
2294                  */
2295                 if (__test_and_set_bit(0, (unsigned long *)
2296                                        &anon_vma->head.next))
2297                         BUG();
2298         }
2299 }
2300
2301 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2302 {
2303         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2304                 /*
2305                  * AS_MM_ALL_LOCKS can't change from under us because
2306                  * we hold the mm_all_locks_mutex.
2307                  *
2308                  * Operations on ->flags have to be atomic because
2309                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2310                  * mm_all_locks_mutex, there may be other cpus
2311                  * changing other bitflags in parallel to us.
2312                  */
2313                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2314                         BUG();
2315                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2316         }
2317 }
2318
2319 /*
2320  * This operation locks against the VM for all pte/vma/mm related
2321  * operations that could ever happen on a certain mm. This includes
2322  * vmtruncate, try_to_unmap, and all page faults.
2323  *
2324  * The caller must take the mmap_sem in write mode before calling
2325  * mm_take_all_locks(). The caller isn't allowed to release the
2326  * mmap_sem until mm_drop_all_locks() returns.
2327  *
2328  * mmap_sem in write mode is required in order to block all operations
2329  * that could modify pagetables and free pages without need of
2330  * altering the vma layout (for example populate_range() with
2331  * nonlinear vmas). It's also needed in write mode to avoid new
2332  * anon_vmas to be associated with existing vmas.
2333  *
2334  * A single task can't take more than one mm_take_all_locks() in a row
2335  * or it would deadlock.
2336  *
2337  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2338  * mapping->flags avoid to take the same lock twice, if more than one
2339  * vma in this mm is backed by the same anon_vma or address_space.
2340  *
2341  * We can take all the locks in random order because the VM code
2342  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2343  * takes more than one of them in a row. Secondly we're protected
2344  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2345  *
2346  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2347  * that may have to take thousand of locks.
2348  *
2349  * mm_take_all_locks() can fail if it's interrupted by signals.
2350  */
2351 int mm_take_all_locks(struct mm_struct *mm)
2352 {
2353         struct vm_area_struct *vma;
2354         int ret = -EINTR;
2355
2356         BUG_ON(down_read_trylock(&mm->mmap_sem));
2357
2358         mutex_lock(&mm_all_locks_mutex);
2359
2360         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2361                 if (signal_pending(current))
2362                         goto out_unlock;
2363                 if (vma->vm_file && vma->vm_file->f_mapping)
2364                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2365         }
2366
2367         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2368                 if (signal_pending(current))
2369                         goto out_unlock;
2370                 if (vma->anon_vma)
2371                         vm_lock_anon_vma(mm, vma->anon_vma);
2372         }
2373
2374         ret = 0;
2375
2376 out_unlock:
2377         if (ret)
2378                 mm_drop_all_locks(mm);
2379
2380         return ret;
2381 }
2382
2383 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2384 {
2385         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2386                 /*
2387                  * The LSB of head.next can't change to 0 from under
2388                  * us because we hold the mm_all_locks_mutex.
2389                  *
2390                  * We must however clear the bitflag before unlocking
2391                  * the vma so the users using the anon_vma->head will
2392                  * never see our bitflag.
2393                  *
2394                  * No need of atomic instructions here, head.next
2395                  * can't change from under us until we release the
2396                  * anon_vma->lock.
2397                  */
2398                 if (!__test_and_clear_bit(0, (unsigned long *)
2399                                           &anon_vma->head.next))
2400                         BUG();
2401                 spin_unlock(&anon_vma->lock);
2402         }
2403 }
2404
2405 static void vm_unlock_mapping(struct address_space *mapping)
2406 {
2407         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2408                 /*
2409                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2410                  * because we hold the mm_all_locks_mutex.
2411                  */
2412                 spin_unlock(&mapping->i_mmap_lock);
2413                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2414                                         &mapping->flags))
2415                         BUG();
2416         }
2417 }
2418
2419 /*
2420  * The mmap_sem cannot be released by the caller until
2421  * mm_drop_all_locks() returns.
2422  */
2423 void mm_drop_all_locks(struct mm_struct *mm)
2424 {
2425         struct vm_area_struct *vma;
2426
2427         BUG_ON(down_read_trylock(&mm->mmap_sem));
2428         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2429
2430         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2431                 if (vma->anon_vma)
2432                         vm_unlock_anon_vma(vma->anon_vma);
2433                 if (vma->vm_file && vma->vm_file->f_mapping)
2434                         vm_unlock_mapping(vma->vm_file->f_mapping);
2435         }
2436
2437         mutex_unlock(&mm_all_locks_mutex);
2438 }