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