rmap: move exclusively owned pages to own anon_vma in do_wp_page()
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.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 struct percpu_counter vm_committed_as;
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         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
183                 return 0;
184 error:
185         vm_unacct_memory(pages);
186
187         return -ENOMEM;
188 }
189
190 /*
191  * Requires inode->i_mapping->i_mmap_lock
192  */
193 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
194                 struct file *file, struct address_space *mapping)
195 {
196         if (vma->vm_flags & VM_DENYWRITE)
197                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
198         if (vma->vm_flags & VM_SHARED)
199                 mapping->i_mmap_writable--;
200
201         flush_dcache_mmap_lock(mapping);
202         if (unlikely(vma->vm_flags & VM_NONLINEAR))
203                 list_del_init(&vma->shared.vm_set.list);
204         else
205                 vma_prio_tree_remove(vma, &mapping->i_mmap);
206         flush_dcache_mmap_unlock(mapping);
207 }
208
209 /*
210  * Unlink a file-based vm structure from its prio_tree, to hide
211  * vma from rmap and vmtruncate before freeing its page tables.
212  */
213 void unlink_file_vma(struct vm_area_struct *vma)
214 {
215         struct file *file = vma->vm_file;
216
217         if (file) {
218                 struct address_space *mapping = file->f_mapping;
219                 spin_lock(&mapping->i_mmap_lock);
220                 __remove_shared_vm_struct(vma, file, mapping);
221                 spin_unlock(&mapping->i_mmap_lock);
222         }
223 }
224
225 /*
226  * Close a vm structure and free it, returning the next.
227  */
228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
229 {
230         struct vm_area_struct *next = vma->vm_next;
231
232         might_sleep();
233         if (vma->vm_ops && vma->vm_ops->close)
234                 vma->vm_ops->close(vma);
235         if (vma->vm_file) {
236                 fput(vma->vm_file);
237                 if (vma->vm_flags & VM_EXECUTABLE)
238                         removed_exe_file_vma(vma->vm_mm);
239         }
240         mpol_put(vma_policy(vma));
241         kmem_cache_free(vm_area_cachep, vma);
242         return next;
243 }
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         min_brk = mm->end_code;
256 #else
257         min_brk = mm->start_brk;
258 #endif
259         if (brk < min_brk)
260                 goto out;
261
262         /*
263          * Check against rlimit here. If this check is done later after the test
264          * of oldbrk with newbrk then it can escape the test and let the data
265          * segment grow beyond its set limit the in case where the limit is
266          * not page aligned -Ram Gupta
267          */
268         rlim = rlimit(RLIMIT_DATA);
269         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
270                         (mm->end_data - mm->start_data) > rlim)
271                 goto out;
272
273         newbrk = PAGE_ALIGN(brk);
274         oldbrk = PAGE_ALIGN(mm->brk);
275         if (oldbrk == newbrk)
276                 goto set_brk;
277
278         /* Always allow shrinking brk. */
279         if (brk <= mm->brk) {
280                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
281                         goto set_brk;
282                 goto out;
283         }
284
285         /* Check against existing mmap mappings. */
286         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
287                 goto out;
288
289         /* Ok, looks good - let it rip. */
290         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
291                 goto out;
292 set_brk:
293         mm->brk = brk;
294 out:
295         retval = mm->brk;
296         up_write(&mm->mmap_sem);
297         return retval;
298 }
299
300 #ifdef DEBUG_MM_RB
301 static int browse_rb(struct rb_root *root)
302 {
303         int i = 0, j;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev)
311                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
312                 if (vma->vm_start < pend)
313                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
314                 if (vma->vm_start > vma->vm_end)
315                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
316                 i++;
317                 pn = nd;
318                 prev = vma->vm_start;
319                 pend = vma->vm_end;
320         }
321         j = 0;
322         for (nd = pn; nd; nd = rb_prev(nd)) {
323                 j++;
324         }
325         if (i != j)
326                 printk("backwards %d, forwards %d\n", j, i), i = 0;
327         return i;
328 }
329
330 void validate_mm(struct mm_struct *mm)
331 {
332         int bug = 0;
333         int i = 0;
334         struct vm_area_struct *tmp = mm->mmap;
335         while (tmp) {
336                 tmp = tmp->vm_next;
337                 i++;
338         }
339         if (i != mm->map_count)
340                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
341         i = browse_rb(&mm->mm_rb);
342         if (i != mm->map_count)
343                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
344         BUG_ON(bug);
345 }
346 #else
347 #define validate_mm(mm) do { } while (0)
348 #endif
349
350 static struct vm_area_struct *
351 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
352                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
353                 struct rb_node ** rb_parent)
354 {
355         struct vm_area_struct * vma;
356         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
357
358         __rb_link = &mm->mm_rb.rb_node;
359         rb_prev = __rb_parent = NULL;
360         vma = NULL;
361
362         while (*__rb_link) {
363                 struct vm_area_struct *vma_tmp;
364
365                 __rb_parent = *__rb_link;
366                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
367
368                 if (vma_tmp->vm_end > addr) {
369                         vma = vma_tmp;
370                         if (vma_tmp->vm_start <= addr)
371                                 break;
372                         __rb_link = &__rb_parent->rb_left;
373                 } else {
374                         rb_prev = __rb_parent;
375                         __rb_link = &__rb_parent->rb_right;
376                 }
377         }
378
379         *pprev = NULL;
380         if (rb_prev)
381                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
382         *rb_link = __rb_link;
383         *rb_parent = __rb_parent;
384         return vma;
385 }
386
387 static inline void
388 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
389                 struct vm_area_struct *prev, struct rb_node *rb_parent)
390 {
391         if (prev) {
392                 vma->vm_next = prev->vm_next;
393                 prev->vm_next = vma;
394         } else {
395                 mm->mmap = vma;
396                 if (rb_parent)
397                         vma->vm_next = rb_entry(rb_parent,
398                                         struct vm_area_struct, vm_rb);
399                 else
400                         vma->vm_next = NULL;
401         }
402 }
403
404 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
405                 struct rb_node **rb_link, struct rb_node *rb_parent)
406 {
407         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
408         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
409 }
410
411 static void __vma_link_file(struct vm_area_struct *vma)
412 {
413         struct file *file;
414
415         file = vma->vm_file;
416         if (file) {
417                 struct address_space *mapping = file->f_mapping;
418
419                 if (vma->vm_flags & VM_DENYWRITE)
420                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
421                 if (vma->vm_flags & VM_SHARED)
422                         mapping->i_mmap_writable++;
423
424                 flush_dcache_mmap_lock(mapping);
425                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
426                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
427                 else
428                         vma_prio_tree_insert(vma, &mapping->i_mmap);
429                 flush_dcache_mmap_unlock(mapping);
430         }
431 }
432
433 static void
434 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435         struct vm_area_struct *prev, struct rb_node **rb_link,
436         struct rb_node *rb_parent)
437 {
438         __vma_link_list(mm, vma, prev, rb_parent);
439         __vma_link_rb(mm, vma, rb_link, rb_parent);
440 }
441
442 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
443                         struct vm_area_struct *prev, struct rb_node **rb_link,
444                         struct rb_node *rb_parent)
445 {
446         struct address_space *mapping = NULL;
447
448         if (vma->vm_file)
449                 mapping = vma->vm_file->f_mapping;
450
451         if (mapping) {
452                 spin_lock(&mapping->i_mmap_lock);
453                 vma->vm_truncate_count = mapping->truncate_count;
454         }
455         anon_vma_lock(vma);
456
457         __vma_link(mm, vma, prev, rb_link, rb_parent);
458         __vma_link_file(vma);
459
460         anon_vma_unlock(vma);
461         if (mapping)
462                 spin_unlock(&mapping->i_mmap_lock);
463
464         mm->map_count++;
465         validate_mm(mm);
466 }
467
468 /*
469  * Helper for vma_adjust in the split_vma insert case:
470  * insert vm structure into list and rbtree and anon_vma,
471  * but it has already been inserted into prio_tree earlier.
472  */
473 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
474 {
475         struct vm_area_struct *__vma, *prev;
476         struct rb_node **rb_link, *rb_parent;
477
478         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
479         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
480         __vma_link(mm, vma, prev, rb_link, rb_parent);
481         mm->map_count++;
482 }
483
484 static inline void
485 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
486                 struct vm_area_struct *prev)
487 {
488         prev->vm_next = vma->vm_next;
489         rb_erase(&vma->vm_rb, &mm->mm_rb);
490         if (mm->mmap_cache == vma)
491                 mm->mmap_cache = prev;
492 }
493
494 /*
495  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
496  * is already present in an i_mmap tree without adjusting the tree.
497  * The following helper function should be used when such adjustments
498  * are necessary.  The "insert" vma (if any) is to be inserted
499  * before we drop the necessary locks.
500  */
501 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
502         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
503 {
504         struct mm_struct *mm = vma->vm_mm;
505         struct vm_area_struct *next = vma->vm_next;
506         struct vm_area_struct *importer = NULL;
507         struct address_space *mapping = NULL;
508         struct prio_tree_root *root = NULL;
509         struct file *file = vma->vm_file;
510         struct anon_vma *anon_vma = NULL;
511         long adjust_next = 0;
512         int remove_next = 0;
513
514         if (next && !insert) {
515                 if (end >= next->vm_end) {
516                         /*
517                          * vma expands, overlapping all the next, and
518                          * perhaps the one after too (mprotect case 6).
519                          */
520 again:                  remove_next = 1 + (end > next->vm_end);
521                         end = next->vm_end;
522                         anon_vma = next->anon_vma;
523                         importer = vma;
524                 } else if (end > next->vm_start) {
525                         /*
526                          * vma expands, overlapping part of the next:
527                          * mprotect case 5 shifting the boundary up.
528                          */
529                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
530                         anon_vma = next->anon_vma;
531                         importer = vma;
532                 } else if (end < vma->vm_end) {
533                         /*
534                          * vma shrinks, and !insert tells it's not
535                          * split_vma inserting another: so it must be
536                          * mprotect case 4 shifting the boundary down.
537                          */
538                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
539                         anon_vma = next->anon_vma;
540                         importer = next;
541                 }
542         }
543
544         /*
545          * When changing only vma->vm_end, we don't really need anon_vma lock.
546          */
547         if (vma->anon_vma && (insert || importer || start != vma->vm_start))
548                 anon_vma = vma->anon_vma;
549         if (anon_vma) {
550                 /*
551                  * Easily overlooked: when mprotect shifts the boundary,
552                  * make sure the expanding vma has anon_vma set if the
553                  * shrinking vma had, to cover any anon pages imported.
554                  */
555                 if (importer && !importer->anon_vma) {
556                         /* Block reverse map lookups until things are set up. */
557                         importer->vm_flags |= VM_LOCK_RMAP;
558                         if (anon_vma_clone(importer, vma)) {
559                                 importer->vm_flags &= ~VM_LOCK_RMAP;
560                                 return -ENOMEM;
561                         }
562                         importer->anon_vma = anon_vma;
563                 }
564         }
565
566         if (file) {
567                 mapping = file->f_mapping;
568                 if (!(vma->vm_flags & VM_NONLINEAR))
569                         root = &mapping->i_mmap;
570                 spin_lock(&mapping->i_mmap_lock);
571                 if (importer &&
572                     vma->vm_truncate_count != next->vm_truncate_count) {
573                         /*
574                          * unmap_mapping_range might be in progress:
575                          * ensure that the expanding vma is rescanned.
576                          */
577                         importer->vm_truncate_count = 0;
578                 }
579                 if (insert) {
580                         insert->vm_truncate_count = vma->vm_truncate_count;
581                         /*
582                          * Put into prio_tree now, so instantiated pages
583                          * are visible to arm/parisc __flush_dcache_page
584                          * throughout; but we cannot insert into address
585                          * space until vma start or end is updated.
586                          */
587                         __vma_link_file(insert);
588                 }
589         }
590
591         if (root) {
592                 flush_dcache_mmap_lock(mapping);
593                 vma_prio_tree_remove(vma, root);
594                 if (adjust_next)
595                         vma_prio_tree_remove(next, root);
596         }
597
598         vma->vm_start = start;
599         vma->vm_end = end;
600         vma->vm_pgoff = pgoff;
601         if (adjust_next) {
602                 next->vm_start += adjust_next << PAGE_SHIFT;
603                 next->vm_pgoff += adjust_next;
604         }
605
606         if (root) {
607                 if (adjust_next)
608                         vma_prio_tree_insert(next, root);
609                 vma_prio_tree_insert(vma, root);
610                 flush_dcache_mmap_unlock(mapping);
611         }
612
613         if (remove_next) {
614                 /*
615                  * vma_merge has merged next into vma, and needs
616                  * us to remove next before dropping the locks.
617                  */
618                 __vma_unlink(mm, next, vma);
619                 if (file)
620                         __remove_shared_vm_struct(next, file, mapping);
621                 /*
622                  * This VMA is now dead, no need for rmap to follow it.
623                  * Call anon_vma_merge below, outside of i_mmap_lock.
624                  */
625                 next->vm_flags |= VM_LOCK_RMAP;
626         } else if (insert) {
627                 /*
628                  * split_vma has split insert from vma, and needs
629                  * us to insert it before dropping the locks
630                  * (it may either follow vma or precede it).
631                  */
632                 __insert_vm_struct(mm, insert);
633         }
634
635         if (mapping)
636                 spin_unlock(&mapping->i_mmap_lock);
637
638         /*
639          * The current VMA has been set up. It is now safe for the
640          * rmap code to get from the pages to the ptes.
641          */
642         if (anon_vma && importer)
643                 importer->vm_flags &= ~VM_LOCK_RMAP;
644
645         if (remove_next) {
646                 if (file) {
647                         fput(file);
648                         if (next->vm_flags & VM_EXECUTABLE)
649                                 removed_exe_file_vma(mm);
650                 }
651                 /* Protected by mmap_sem and VM_LOCK_RMAP. */
652                 if (next->anon_vma)
653                         anon_vma_merge(vma, next);
654                 mm->map_count--;
655                 mpol_put(vma_policy(next));
656                 kmem_cache_free(vm_area_cachep, next);
657                 /*
658                  * In mprotect's case 6 (see comments on vma_merge),
659                  * we must remove another next too. It would clutter
660                  * up the code too much to do both in one go.
661                  */
662                 if (remove_next == 2) {
663                         next = vma->vm_next;
664                         goto again;
665                 }
666         }
667
668         validate_mm(mm);
669
670         return 0;
671 }
672
673 /*
674  * If the vma has a ->close operation then the driver probably needs to release
675  * per-vma resources, so we don't attempt to merge those.
676  */
677 static inline int is_mergeable_vma(struct vm_area_struct *vma,
678                         struct file *file, unsigned long vm_flags)
679 {
680         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
681         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
682                 return 0;
683         if (vma->vm_file != file)
684                 return 0;
685         if (vma->vm_ops && vma->vm_ops->close)
686                 return 0;
687         return 1;
688 }
689
690 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
691                                         struct anon_vma *anon_vma2)
692 {
693         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
694 }
695
696 /*
697  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
698  * in front of (at a lower virtual address and file offset than) the vma.
699  *
700  * We cannot merge two vmas if they have differently assigned (non-NULL)
701  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
702  *
703  * We don't check here for the merged mmap wrapping around the end of pagecache
704  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
705  * wrap, nor mmaps which cover the final page at index -1UL.
706  */
707 static int
708 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
709         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
710 {
711         if (is_mergeable_vma(vma, file, vm_flags) &&
712             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
713                 if (vma->vm_pgoff == vm_pgoff)
714                         return 1;
715         }
716         return 0;
717 }
718
719 /*
720  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
721  * beyond (at a higher virtual address and file offset than) the vma.
722  *
723  * We cannot merge two vmas if they have differently assigned (non-NULL)
724  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
725  */
726 static int
727 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
728         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
729 {
730         if (is_mergeable_vma(vma, file, vm_flags) &&
731             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
732                 pgoff_t vm_pglen;
733                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
734                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
735                         return 1;
736         }
737         return 0;
738 }
739
740 /*
741  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
742  * whether that can be merged with its predecessor or its successor.
743  * Or both (it neatly fills a hole).
744  *
745  * In most cases - when called for mmap, brk or mremap - [addr,end) is
746  * certain not to be mapped by the time vma_merge is called; but when
747  * called for mprotect, it is certain to be already mapped (either at
748  * an offset within prev, or at the start of next), and the flags of
749  * this area are about to be changed to vm_flags - and the no-change
750  * case has already been eliminated.
751  *
752  * The following mprotect cases have to be considered, where AAAA is
753  * the area passed down from mprotect_fixup, never extending beyond one
754  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
755  *
756  *     AAAA             AAAA                AAAA          AAAA
757  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
758  *    cannot merge    might become    might become    might become
759  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
760  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
761  *    mremap move:                                    PPPPNNNNNNNN 8
762  *        AAAA
763  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
764  *    might become    case 1 below    case 2 below    case 3 below
765  *
766  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
767  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
768  */
769 struct vm_area_struct *vma_merge(struct mm_struct *mm,
770                         struct vm_area_struct *prev, unsigned long addr,
771                         unsigned long end, unsigned long vm_flags,
772                         struct anon_vma *anon_vma, struct file *file,
773                         pgoff_t pgoff, struct mempolicy *policy)
774 {
775         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
776         struct vm_area_struct *area, *next;
777         int err;
778
779         /*
780          * We later require that vma->vm_flags == vm_flags,
781          * so this tests vma->vm_flags & VM_SPECIAL, too.
782          */
783         if (vm_flags & VM_SPECIAL)
784                 return NULL;
785
786         if (prev)
787                 next = prev->vm_next;
788         else
789                 next = mm->mmap;
790         area = next;
791         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
792                 next = next->vm_next;
793
794         /*
795          * Can it merge with the predecessor?
796          */
797         if (prev && prev->vm_end == addr &&
798                         mpol_equal(vma_policy(prev), policy) &&
799                         can_vma_merge_after(prev, vm_flags,
800                                                 anon_vma, file, pgoff)) {
801                 /*
802                  * OK, it can.  Can we now merge in the successor as well?
803                  */
804                 if (next && end == next->vm_start &&
805                                 mpol_equal(policy, vma_policy(next)) &&
806                                 can_vma_merge_before(next, vm_flags,
807                                         anon_vma, file, pgoff+pglen) &&
808                                 is_mergeable_anon_vma(prev->anon_vma,
809                                                       next->anon_vma)) {
810                                                         /* cases 1, 6 */
811                         err = vma_adjust(prev, prev->vm_start,
812                                 next->vm_end, prev->vm_pgoff, NULL);
813                 } else                                  /* cases 2, 5, 7 */
814                         err = vma_adjust(prev, prev->vm_start,
815                                 end, prev->vm_pgoff, NULL);
816                 if (err)
817                         return NULL;
818                 return prev;
819         }
820
821         /*
822          * Can this new request be merged in front of next?
823          */
824         if (next && end == next->vm_start &&
825                         mpol_equal(policy, vma_policy(next)) &&
826                         can_vma_merge_before(next, vm_flags,
827                                         anon_vma, file, pgoff+pglen)) {
828                 if (prev && addr < prev->vm_end)        /* case 4 */
829                         err = vma_adjust(prev, prev->vm_start,
830                                 addr, prev->vm_pgoff, NULL);
831                 else                                    /* cases 3, 8 */
832                         err = vma_adjust(area, addr, next->vm_end,
833                                 next->vm_pgoff - pglen, NULL);
834                 if (err)
835                         return NULL;
836                 return area;
837         }
838
839         return NULL;
840 }
841
842 /*
843  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
844  * neighbouring vmas for a suitable anon_vma, before it goes off
845  * to allocate a new anon_vma.  It checks because a repetitive
846  * sequence of mprotects and faults may otherwise lead to distinct
847  * anon_vmas being allocated, preventing vma merge in subsequent
848  * mprotect.
849  */
850 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
851 {
852         struct vm_area_struct *near;
853         unsigned long vm_flags;
854
855         near = vma->vm_next;
856         if (!near)
857                 goto try_prev;
858
859         /*
860          * Since only mprotect tries to remerge vmas, match flags
861          * which might be mprotected into each other later on.
862          * Neither mlock nor madvise tries to remerge at present,
863          * so leave their flags as obstructing a merge.
864          */
865         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
866         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
867
868         if (near->anon_vma && vma->vm_end == near->vm_start &&
869                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
870                         can_vma_merge_before(near, vm_flags,
871                                 NULL, vma->vm_file, vma->vm_pgoff +
872                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
873                 return near->anon_vma;
874 try_prev:
875         /*
876          * It is potentially slow to have to call find_vma_prev here.
877          * But it's only on the first write fault on the vma, not
878          * every time, and we could devise a way to avoid it later
879          * (e.g. stash info in next's anon_vma_node when assigning
880          * an anon_vma, or when trying vma_merge).  Another time.
881          */
882         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
883         if (!near)
884                 goto none;
885
886         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
887         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
888
889         if (near->anon_vma && near->vm_end == vma->vm_start &&
890                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
891                         can_vma_merge_after(near, vm_flags,
892                                 NULL, vma->vm_file, vma->vm_pgoff))
893                 return near->anon_vma;
894 none:
895         /*
896          * There's no absolute need to look only at touching neighbours:
897          * we could search further afield for "compatible" anon_vmas.
898          * But it would probably just be a waste of time searching,
899          * or lead to too many vmas hanging off the same anon_vma.
900          * We're trying to allow mprotect remerging later on,
901          * not trying to minimize memory used for anon_vmas.
902          */
903         return NULL;
904 }
905
906 #ifdef CONFIG_PROC_FS
907 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
908                                                 struct file *file, long pages)
909 {
910         const unsigned long stack_flags
911                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
912
913         if (file) {
914                 mm->shared_vm += pages;
915                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
916                         mm->exec_vm += pages;
917         } else if (flags & stack_flags)
918                 mm->stack_vm += pages;
919         if (flags & (VM_RESERVED|VM_IO))
920                 mm->reserved_vm += pages;
921 }
922 #endif /* CONFIG_PROC_FS */
923
924 /*
925  * The caller must hold down_write(&current->mm->mmap_sem).
926  */
927
928 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
929                         unsigned long len, unsigned long prot,
930                         unsigned long flags, unsigned long pgoff)
931 {
932         struct mm_struct * mm = current->mm;
933         struct inode *inode;
934         unsigned int vm_flags;
935         int error;
936         unsigned long reqprot = prot;
937
938         /*
939          * Does the application expect PROT_READ to imply PROT_EXEC?
940          *
941          * (the exception is when the underlying filesystem is noexec
942          *  mounted, in which case we dont add PROT_EXEC.)
943          */
944         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
945                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
946                         prot |= PROT_EXEC;
947
948         if (!len)
949                 return -EINVAL;
950
951         if (!(flags & MAP_FIXED))
952                 addr = round_hint_to_min(addr);
953
954         /* Careful about overflows.. */
955         len = PAGE_ALIGN(len);
956         if (!len)
957                 return -ENOMEM;
958
959         /* offset overflow? */
960         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
961                return -EOVERFLOW;
962
963         /* Too many mappings? */
964         if (mm->map_count > sysctl_max_map_count)
965                 return -ENOMEM;
966
967         /* Obtain the address to map to. we verify (or select) it and ensure
968          * that it represents a valid section of the address space.
969          */
970         addr = get_unmapped_area(file, addr, len, pgoff, flags);
971         if (addr & ~PAGE_MASK)
972                 return addr;
973
974         /* Do simple checking here so the lower-level routines won't have
975          * to. we assume access permissions have been handled by the open
976          * of the memory object, so we don't do any here.
977          */
978         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
979                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
980
981         if (flags & MAP_LOCKED)
982                 if (!can_do_mlock())
983                         return -EPERM;
984
985         /* mlock MCL_FUTURE? */
986         if (vm_flags & VM_LOCKED) {
987                 unsigned long locked, lock_limit;
988                 locked = len >> PAGE_SHIFT;
989                 locked += mm->locked_vm;
990                 lock_limit = rlimit(RLIMIT_MEMLOCK);
991                 lock_limit >>= PAGE_SHIFT;
992                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
993                         return -EAGAIN;
994         }
995
996         inode = file ? file->f_path.dentry->d_inode : NULL;
997
998         if (file) {
999                 switch (flags & MAP_TYPE) {
1000                 case MAP_SHARED:
1001                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1002                                 return -EACCES;
1003
1004                         /*
1005                          * Make sure we don't allow writing to an append-only
1006                          * file..
1007                          */
1008                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1009                                 return -EACCES;
1010
1011                         /*
1012                          * Make sure there are no mandatory locks on the file.
1013                          */
1014                         if (locks_verify_locked(inode))
1015                                 return -EAGAIN;
1016
1017                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1018                         if (!(file->f_mode & FMODE_WRITE))
1019                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1020
1021                         /* fall through */
1022                 case MAP_PRIVATE:
1023                         if (!(file->f_mode & FMODE_READ))
1024                                 return -EACCES;
1025                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1026                                 if (vm_flags & VM_EXEC)
1027                                         return -EPERM;
1028                                 vm_flags &= ~VM_MAYEXEC;
1029                         }
1030
1031                         if (!file->f_op || !file->f_op->mmap)
1032                                 return -ENODEV;
1033                         break;
1034
1035                 default:
1036                         return -EINVAL;
1037                 }
1038         } else {
1039                 switch (flags & MAP_TYPE) {
1040                 case MAP_SHARED:
1041                         /*
1042                          * Ignore pgoff.
1043                          */
1044                         pgoff = 0;
1045                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1046                         break;
1047                 case MAP_PRIVATE:
1048                         /*
1049                          * Set pgoff according to addr for anon_vma.
1050                          */
1051                         pgoff = addr >> PAGE_SHIFT;
1052                         break;
1053                 default:
1054                         return -EINVAL;
1055                 }
1056         }
1057
1058         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1059         if (error)
1060                 return error;
1061
1062         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1063 }
1064 EXPORT_SYMBOL(do_mmap_pgoff);
1065
1066 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1067                 unsigned long, prot, unsigned long, flags,
1068                 unsigned long, fd, unsigned long, pgoff)
1069 {
1070         struct file *file = NULL;
1071         unsigned long retval = -EBADF;
1072
1073         if (!(flags & MAP_ANONYMOUS)) {
1074                 if (unlikely(flags & MAP_HUGETLB))
1075                         return -EINVAL;
1076                 file = fget(fd);
1077                 if (!file)
1078                         goto out;
1079         } else if (flags & MAP_HUGETLB) {
1080                 struct user_struct *user = NULL;
1081                 /*
1082                  * VM_NORESERVE is used because the reservations will be
1083                  * taken when vm_ops->mmap() is called
1084                  * A dummy user value is used because we are not locking
1085                  * memory so no accounting is necessary
1086                  */
1087                 len = ALIGN(len, huge_page_size(&default_hstate));
1088                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1089                                                 &user, HUGETLB_ANONHUGE_INODE);
1090                 if (IS_ERR(file))
1091                         return PTR_ERR(file);
1092         }
1093
1094         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1095
1096         down_write(&current->mm->mmap_sem);
1097         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1098         up_write(&current->mm->mmap_sem);
1099
1100         if (file)
1101                 fput(file);
1102 out:
1103         return retval;
1104 }
1105
1106 /*
1107  * Some shared mappigns will want the pages marked read-only
1108  * to track write events. If so, we'll downgrade vm_page_prot
1109  * to the private version (using protection_map[] without the
1110  * VM_SHARED bit).
1111  */
1112 int vma_wants_writenotify(struct vm_area_struct *vma)
1113 {
1114         unsigned int vm_flags = vma->vm_flags;
1115
1116         /* If it was private or non-writable, the write bit is already clear */
1117         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1118                 return 0;
1119
1120         /* The backer wishes to know when pages are first written to? */
1121         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1122                 return 1;
1123
1124         /* The open routine did something to the protections already? */
1125         if (pgprot_val(vma->vm_page_prot) !=
1126             pgprot_val(vm_get_page_prot(vm_flags)))
1127                 return 0;
1128
1129         /* Specialty mapping? */
1130         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1131                 return 0;
1132
1133         /* Can the mapping track the dirty pages? */
1134         return vma->vm_file && vma->vm_file->f_mapping &&
1135                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1136 }
1137
1138 /*
1139  * We account for memory if it's a private writeable mapping,
1140  * not hugepages and VM_NORESERVE wasn't set.
1141  */
1142 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1143 {
1144         /*
1145          * hugetlb has its own accounting separate from the core VM
1146          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1147          */
1148         if (file && is_file_hugepages(file))
1149                 return 0;
1150
1151         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1152 }
1153
1154 unsigned long mmap_region(struct file *file, unsigned long addr,
1155                           unsigned long len, unsigned long flags,
1156                           unsigned int vm_flags, unsigned long pgoff)
1157 {
1158         struct mm_struct *mm = current->mm;
1159         struct vm_area_struct *vma, *prev;
1160         int correct_wcount = 0;
1161         int error;
1162         struct rb_node **rb_link, *rb_parent;
1163         unsigned long charged = 0;
1164         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1165
1166         /* Clear old maps */
1167         error = -ENOMEM;
1168 munmap_back:
1169         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1170         if (vma && vma->vm_start < addr + len) {
1171                 if (do_munmap(mm, addr, len))
1172                         return -ENOMEM;
1173                 goto munmap_back;
1174         }
1175
1176         /* Check against address space limit. */
1177         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1178                 return -ENOMEM;
1179
1180         /*
1181          * Set 'VM_NORESERVE' if we should not account for the
1182          * memory use of this mapping.
1183          */
1184         if ((flags & MAP_NORESERVE)) {
1185                 /* We honor MAP_NORESERVE if allowed to overcommit */
1186                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1187                         vm_flags |= VM_NORESERVE;
1188
1189                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1190                 if (file && is_file_hugepages(file))
1191                         vm_flags |= VM_NORESERVE;
1192         }
1193
1194         /*
1195          * Private writable mapping: check memory availability
1196          */
1197         if (accountable_mapping(file, vm_flags)) {
1198                 charged = len >> PAGE_SHIFT;
1199                 if (security_vm_enough_memory(charged))
1200                         return -ENOMEM;
1201                 vm_flags |= VM_ACCOUNT;
1202         }
1203
1204         /*
1205          * Can we just expand an old mapping?
1206          */
1207         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1208         if (vma)
1209                 goto out;
1210
1211         /*
1212          * Determine the object being mapped and call the appropriate
1213          * specific mapper. the address has already been validated, but
1214          * not unmapped, but the maps are removed from the list.
1215          */
1216         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1217         if (!vma) {
1218                 error = -ENOMEM;
1219                 goto unacct_error;
1220         }
1221
1222         vma->vm_mm = mm;
1223         vma->vm_start = addr;
1224         vma->vm_end = addr + len;
1225         vma->vm_flags = vm_flags;
1226         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1227         vma->vm_pgoff = pgoff;
1228         INIT_LIST_HEAD(&vma->anon_vma_chain);
1229
1230         if (file) {
1231                 error = -EINVAL;
1232                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1233                         goto free_vma;
1234                 if (vm_flags & VM_DENYWRITE) {
1235                         error = deny_write_access(file);
1236                         if (error)
1237                                 goto free_vma;
1238                         correct_wcount = 1;
1239                 }
1240                 vma->vm_file = file;
1241                 get_file(file);
1242                 error = file->f_op->mmap(file, vma);
1243                 if (error)
1244                         goto unmap_and_free_vma;
1245                 if (vm_flags & VM_EXECUTABLE)
1246                         added_exe_file_vma(mm);
1247
1248                 /* Can addr have changed??
1249                  *
1250                  * Answer: Yes, several device drivers can do it in their
1251                  *         f_op->mmap method. -DaveM
1252                  */
1253                 addr = vma->vm_start;
1254                 pgoff = vma->vm_pgoff;
1255                 vm_flags = vma->vm_flags;
1256         } else if (vm_flags & VM_SHARED) {
1257                 error = shmem_zero_setup(vma);
1258                 if (error)
1259                         goto free_vma;
1260         }
1261
1262         if (vma_wants_writenotify(vma)) {
1263                 pgprot_t pprot = vma->vm_page_prot;
1264
1265                 /* Can vma->vm_page_prot have changed??
1266                  *
1267                  * Answer: Yes, drivers may have changed it in their
1268                  *         f_op->mmap method.
1269                  *
1270                  * Ensures that vmas marked as uncached stay that way.
1271                  */
1272                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1273                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1274                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1275         }
1276
1277         vma_link(mm, vma, prev, rb_link, rb_parent);
1278         file = vma->vm_file;
1279
1280         /* Once vma denies write, undo our temporary denial count */
1281         if (correct_wcount)
1282                 atomic_inc(&inode->i_writecount);
1283 out:
1284         perf_event_mmap(vma);
1285
1286         mm->total_vm += len >> PAGE_SHIFT;
1287         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1288         if (vm_flags & VM_LOCKED) {
1289                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1290                         mm->locked_vm += (len >> PAGE_SHIFT);
1291         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1292                 make_pages_present(addr, addr + len);
1293         return addr;
1294
1295 unmap_and_free_vma:
1296         if (correct_wcount)
1297                 atomic_inc(&inode->i_writecount);
1298         vma->vm_file = NULL;
1299         fput(file);
1300
1301         /* Undo any partial mapping done by a device driver. */
1302         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1303         charged = 0;
1304 free_vma:
1305         kmem_cache_free(vm_area_cachep, vma);
1306 unacct_error:
1307         if (charged)
1308                 vm_unacct_memory(charged);
1309         return error;
1310 }
1311
1312 /* Get an address range which is currently unmapped.
1313  * For shmat() with addr=0.
1314  *
1315  * Ugly calling convention alert:
1316  * Return value with the low bits set means error value,
1317  * ie
1318  *      if (ret & ~PAGE_MASK)
1319  *              error = ret;
1320  *
1321  * This function "knows" that -ENOMEM has the bits set.
1322  */
1323 #ifndef HAVE_ARCH_UNMAPPED_AREA
1324 unsigned long
1325 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1326                 unsigned long len, unsigned long pgoff, unsigned long flags)
1327 {
1328         struct mm_struct *mm = current->mm;
1329         struct vm_area_struct *vma;
1330         unsigned long start_addr;
1331
1332         if (len > TASK_SIZE)
1333                 return -ENOMEM;
1334
1335         if (flags & MAP_FIXED)
1336                 return addr;
1337
1338         if (addr) {
1339                 addr = PAGE_ALIGN(addr);
1340                 vma = find_vma(mm, addr);
1341                 if (TASK_SIZE - len >= addr &&
1342                     (!vma || addr + len <= vma->vm_start))
1343                         return addr;
1344         }
1345         if (len > mm->cached_hole_size) {
1346                 start_addr = addr = mm->free_area_cache;
1347         } else {
1348                 start_addr = addr = TASK_UNMAPPED_BASE;
1349                 mm->cached_hole_size = 0;
1350         }
1351
1352 full_search:
1353         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1354                 /* At this point:  (!vma || addr < vma->vm_end). */
1355                 if (TASK_SIZE - len < addr) {
1356                         /*
1357                          * Start a new search - just in case we missed
1358                          * some holes.
1359                          */
1360                         if (start_addr != TASK_UNMAPPED_BASE) {
1361                                 addr = TASK_UNMAPPED_BASE;
1362                                 start_addr = addr;
1363                                 mm->cached_hole_size = 0;
1364                                 goto full_search;
1365                         }
1366                         return -ENOMEM;
1367                 }
1368                 if (!vma || addr + len <= vma->vm_start) {
1369                         /*
1370                          * Remember the place where we stopped the search:
1371                          */
1372                         mm->free_area_cache = addr + len;
1373                         return addr;
1374                 }
1375                 if (addr + mm->cached_hole_size < vma->vm_start)
1376                         mm->cached_hole_size = vma->vm_start - addr;
1377                 addr = vma->vm_end;
1378         }
1379 }
1380 #endif  
1381
1382 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1383 {
1384         /*
1385          * Is this a new hole at the lowest possible address?
1386          */
1387         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1388                 mm->free_area_cache = addr;
1389                 mm->cached_hole_size = ~0UL;
1390         }
1391 }
1392
1393 /*
1394  * This mmap-allocator allocates new areas top-down from below the
1395  * stack's low limit (the base):
1396  */
1397 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1398 unsigned long
1399 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1400                           const unsigned long len, const unsigned long pgoff,
1401                           const unsigned long flags)
1402 {
1403         struct vm_area_struct *vma;
1404         struct mm_struct *mm = current->mm;
1405         unsigned long addr = addr0;
1406
1407         /* requested length too big for entire address space */
1408         if (len > TASK_SIZE)
1409                 return -ENOMEM;
1410
1411         if (flags & MAP_FIXED)
1412                 return addr;
1413
1414         /* requesting a specific address */
1415         if (addr) {
1416                 addr = PAGE_ALIGN(addr);
1417                 vma = find_vma(mm, addr);
1418                 if (TASK_SIZE - len >= addr &&
1419                                 (!vma || addr + len <= vma->vm_start))
1420                         return addr;
1421         }
1422
1423         /* check if free_area_cache is useful for us */
1424         if (len <= mm->cached_hole_size) {
1425                 mm->cached_hole_size = 0;
1426                 mm->free_area_cache = mm->mmap_base;
1427         }
1428
1429         /* either no address requested or can't fit in requested address hole */
1430         addr = mm->free_area_cache;
1431
1432         /* make sure it can fit in the remaining address space */
1433         if (addr > len) {
1434                 vma = find_vma(mm, addr-len);
1435                 if (!vma || addr <= vma->vm_start)
1436                         /* remember the address as a hint for next time */
1437                         return (mm->free_area_cache = addr-len);
1438         }
1439
1440         if (mm->mmap_base < len)
1441                 goto bottomup;
1442
1443         addr = mm->mmap_base-len;
1444
1445         do {
1446                 /*
1447                  * Lookup failure means no vma is above this address,
1448                  * else if new region fits below vma->vm_start,
1449                  * return with success:
1450                  */
1451                 vma = find_vma(mm, addr);
1452                 if (!vma || addr+len <= vma->vm_start)
1453                         /* remember the address as a hint for next time */
1454                         return (mm->free_area_cache = addr);
1455
1456                 /* remember the largest hole we saw so far */
1457                 if (addr + mm->cached_hole_size < vma->vm_start)
1458                         mm->cached_hole_size = vma->vm_start - addr;
1459
1460                 /* try just below the current vma->vm_start */
1461                 addr = vma->vm_start-len;
1462         } while (len < vma->vm_start);
1463
1464 bottomup:
1465         /*
1466          * A failed mmap() very likely causes application failure,
1467          * so fall back to the bottom-up function here. This scenario
1468          * can happen with large stack limits and large mmap()
1469          * allocations.
1470          */
1471         mm->cached_hole_size = ~0UL;
1472         mm->free_area_cache = TASK_UNMAPPED_BASE;
1473         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1474         /*
1475          * Restore the topdown base:
1476          */
1477         mm->free_area_cache = mm->mmap_base;
1478         mm->cached_hole_size = ~0UL;
1479
1480         return addr;
1481 }
1482 #endif
1483
1484 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1485 {
1486         /*
1487          * Is this a new hole at the highest possible address?
1488          */
1489         if (addr > mm->free_area_cache)
1490                 mm->free_area_cache = addr;
1491
1492         /* dont allow allocations above current base */
1493         if (mm->free_area_cache > mm->mmap_base)
1494                 mm->free_area_cache = mm->mmap_base;
1495 }
1496
1497 unsigned long
1498 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1499                 unsigned long pgoff, unsigned long flags)
1500 {
1501         unsigned long (*get_area)(struct file *, unsigned long,
1502                                   unsigned long, unsigned long, unsigned long);
1503
1504         unsigned long error = arch_mmap_check(addr, len, flags);
1505         if (error)
1506                 return error;
1507
1508         /* Careful about overflows.. */
1509         if (len > TASK_SIZE)
1510                 return -ENOMEM;
1511
1512         get_area = current->mm->get_unmapped_area;
1513         if (file && file->f_op && file->f_op->get_unmapped_area)
1514                 get_area = file->f_op->get_unmapped_area;
1515         addr = get_area(file, addr, len, pgoff, flags);
1516         if (IS_ERR_VALUE(addr))
1517                 return addr;
1518
1519         if (addr > TASK_SIZE - len)
1520                 return -ENOMEM;
1521         if (addr & ~PAGE_MASK)
1522                 return -EINVAL;
1523
1524         return arch_rebalance_pgtables(addr, len);
1525 }
1526
1527 EXPORT_SYMBOL(get_unmapped_area);
1528
1529 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1530 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1531 {
1532         struct vm_area_struct *vma = NULL;
1533
1534         if (mm) {
1535                 /* Check the cache first. */
1536                 /* (Cache hit rate is typically around 35%.) */
1537                 vma = mm->mmap_cache;
1538                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1539                         struct rb_node * rb_node;
1540
1541                         rb_node = mm->mm_rb.rb_node;
1542                         vma = NULL;
1543
1544                         while (rb_node) {
1545                                 struct vm_area_struct * vma_tmp;
1546
1547                                 vma_tmp = rb_entry(rb_node,
1548                                                 struct vm_area_struct, vm_rb);
1549
1550                                 if (vma_tmp->vm_end > addr) {
1551                                         vma = vma_tmp;
1552                                         if (vma_tmp->vm_start <= addr)
1553                                                 break;
1554                                         rb_node = rb_node->rb_left;
1555                                 } else
1556                                         rb_node = rb_node->rb_right;
1557                         }
1558                         if (vma)
1559                                 mm->mmap_cache = vma;
1560                 }
1561         }
1562         return vma;
1563 }
1564
1565 EXPORT_SYMBOL(find_vma);
1566
1567 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1568 struct vm_area_struct *
1569 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1570                         struct vm_area_struct **pprev)
1571 {
1572         struct vm_area_struct *vma = NULL, *prev = NULL;
1573         struct rb_node *rb_node;
1574         if (!mm)
1575                 goto out;
1576
1577         /* Guard against addr being lower than the first VMA */
1578         vma = mm->mmap;
1579
1580         /* Go through the RB tree quickly. */
1581         rb_node = mm->mm_rb.rb_node;
1582
1583         while (rb_node) {
1584                 struct vm_area_struct *vma_tmp;
1585                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1586
1587                 if (addr < vma_tmp->vm_end) {
1588                         rb_node = rb_node->rb_left;
1589                 } else {
1590                         prev = vma_tmp;
1591                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1592                                 break;
1593                         rb_node = rb_node->rb_right;
1594                 }
1595         }
1596
1597 out:
1598         *pprev = prev;
1599         return prev ? prev->vm_next : vma;
1600 }
1601
1602 /*
1603  * Verify that the stack growth is acceptable and
1604  * update accounting. This is shared with both the
1605  * grow-up and grow-down cases.
1606  */
1607 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1608 {
1609         struct mm_struct *mm = vma->vm_mm;
1610         struct rlimit *rlim = current->signal->rlim;
1611         unsigned long new_start;
1612
1613         /* address space limit tests */
1614         if (!may_expand_vm(mm, grow))
1615                 return -ENOMEM;
1616
1617         /* Stack limit test */
1618         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1619                 return -ENOMEM;
1620
1621         /* mlock limit tests */
1622         if (vma->vm_flags & VM_LOCKED) {
1623                 unsigned long locked;
1624                 unsigned long limit;
1625                 locked = mm->locked_vm + grow;
1626                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1627                 limit >>= PAGE_SHIFT;
1628                 if (locked > limit && !capable(CAP_IPC_LOCK))
1629                         return -ENOMEM;
1630         }
1631
1632         /* Check to ensure the stack will not grow into a hugetlb-only region */
1633         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1634                         vma->vm_end - size;
1635         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1636                 return -EFAULT;
1637
1638         /*
1639          * Overcommit..  This must be the final test, as it will
1640          * update security statistics.
1641          */
1642         if (security_vm_enough_memory_mm(mm, grow))
1643                 return -ENOMEM;
1644
1645         /* Ok, everything looks good - let it rip */
1646         mm->total_vm += grow;
1647         if (vma->vm_flags & VM_LOCKED)
1648                 mm->locked_vm += grow;
1649         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1650         return 0;
1651 }
1652
1653 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1654 /*
1655  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1656  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1657  */
1658 #ifndef CONFIG_IA64
1659 static
1660 #endif
1661 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1662 {
1663         int error;
1664
1665         if (!(vma->vm_flags & VM_GROWSUP))
1666                 return -EFAULT;
1667
1668         /*
1669          * We must make sure the anon_vma is allocated
1670          * so that the anon_vma locking is not a noop.
1671          */
1672         if (unlikely(anon_vma_prepare(vma)))
1673                 return -ENOMEM;
1674         anon_vma_lock(vma);
1675
1676         /*
1677          * vma->vm_start/vm_end cannot change under us because the caller
1678          * is required to hold the mmap_sem in read mode.  We need the
1679          * anon_vma lock to serialize against concurrent expand_stacks.
1680          * Also guard against wrapping around to address 0.
1681          */
1682         if (address < PAGE_ALIGN(address+4))
1683                 address = PAGE_ALIGN(address+4);
1684         else {
1685                 anon_vma_unlock(vma);
1686                 return -ENOMEM;
1687         }
1688         error = 0;
1689
1690         /* Somebody else might have raced and expanded it already */
1691         if (address > vma->vm_end) {
1692                 unsigned long size, grow;
1693
1694                 size = address - vma->vm_start;
1695                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1696
1697                 error = acct_stack_growth(vma, size, grow);
1698                 if (!error)
1699                         vma->vm_end = address;
1700         }
1701         anon_vma_unlock(vma);
1702         return error;
1703 }
1704 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1705
1706 /*
1707  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1708  */
1709 static int expand_downwards(struct vm_area_struct *vma,
1710                                    unsigned long address)
1711 {
1712         int error;
1713
1714         /*
1715          * We must make sure the anon_vma is allocated
1716          * so that the anon_vma locking is not a noop.
1717          */
1718         if (unlikely(anon_vma_prepare(vma)))
1719                 return -ENOMEM;
1720
1721         address &= PAGE_MASK;
1722         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1723         if (error)
1724                 return error;
1725
1726         anon_vma_lock(vma);
1727
1728         /*
1729          * vma->vm_start/vm_end cannot change under us because the caller
1730          * is required to hold the mmap_sem in read mode.  We need the
1731          * anon_vma lock to serialize against concurrent expand_stacks.
1732          */
1733
1734         /* Somebody else might have raced and expanded it already */
1735         if (address < vma->vm_start) {
1736                 unsigned long size, grow;
1737
1738                 size = vma->vm_end - address;
1739                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1740
1741                 error = acct_stack_growth(vma, size, grow);
1742                 if (!error) {
1743                         vma->vm_start = address;
1744                         vma->vm_pgoff -= grow;
1745                 }
1746         }
1747         anon_vma_unlock(vma);
1748         return error;
1749 }
1750
1751 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1752 {
1753         return expand_downwards(vma, address);
1754 }
1755
1756 #ifdef CONFIG_STACK_GROWSUP
1757 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1758 {
1759         return expand_upwards(vma, address);
1760 }
1761
1762 struct vm_area_struct *
1763 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1764 {
1765         struct vm_area_struct *vma, *prev;
1766
1767         addr &= PAGE_MASK;
1768         vma = find_vma_prev(mm, addr, &prev);
1769         if (vma && (vma->vm_start <= addr))
1770                 return vma;
1771         if (!prev || expand_stack(prev, addr))
1772                 return NULL;
1773         if (prev->vm_flags & VM_LOCKED) {
1774                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1775         }
1776         return prev;
1777 }
1778 #else
1779 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1780 {
1781         return expand_downwards(vma, address);
1782 }
1783
1784 struct vm_area_struct *
1785 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1786 {
1787         struct vm_area_struct * vma;
1788         unsigned long start;
1789
1790         addr &= PAGE_MASK;
1791         vma = find_vma(mm,addr);
1792         if (!vma)
1793                 return NULL;
1794         if (vma->vm_start <= addr)
1795                 return vma;
1796         if (!(vma->vm_flags & VM_GROWSDOWN))
1797                 return NULL;
1798         start = vma->vm_start;
1799         if (expand_stack(vma, addr))
1800                 return NULL;
1801         if (vma->vm_flags & VM_LOCKED) {
1802                 mlock_vma_pages_range(vma, addr, start);
1803         }
1804         return vma;
1805 }
1806 #endif
1807
1808 /*
1809  * Ok - we have the memory areas we should free on the vma list,
1810  * so release them, and do the vma updates.
1811  *
1812  * Called with the mm semaphore held.
1813  */
1814 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1815 {
1816         /* Update high watermark before we lower total_vm */
1817         update_hiwater_vm(mm);
1818         do {
1819                 long nrpages = vma_pages(vma);
1820
1821                 mm->total_vm -= nrpages;
1822                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1823                 vma = remove_vma(vma);
1824         } while (vma);
1825         validate_mm(mm);
1826 }
1827
1828 /*
1829  * Get rid of page table information in the indicated region.
1830  *
1831  * Called with the mm semaphore held.
1832  */
1833 static void unmap_region(struct mm_struct *mm,
1834                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1835                 unsigned long start, unsigned long end)
1836 {
1837         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1838         struct mmu_gather *tlb;
1839         unsigned long nr_accounted = 0;
1840
1841         lru_add_drain();
1842         tlb = tlb_gather_mmu(mm, 0);
1843         update_hiwater_rss(mm);
1844         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1845         vm_unacct_memory(nr_accounted);
1846         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1847                                  next? next->vm_start: 0);
1848         tlb_finish_mmu(tlb, start, end);
1849 }
1850
1851 /*
1852  * Create a list of vma's touched by the unmap, removing them from the mm's
1853  * vma list as we go..
1854  */
1855 static void
1856 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1857         struct vm_area_struct *prev, unsigned long end)
1858 {
1859         struct vm_area_struct **insertion_point;
1860         struct vm_area_struct *tail_vma = NULL;
1861         unsigned long addr;
1862
1863         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1864         do {
1865                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1866                 mm->map_count--;
1867                 tail_vma = vma;
1868                 vma = vma->vm_next;
1869         } while (vma && vma->vm_start < end);
1870         *insertion_point = vma;
1871         tail_vma->vm_next = NULL;
1872         if (mm->unmap_area == arch_unmap_area)
1873                 addr = prev ? prev->vm_end : mm->mmap_base;
1874         else
1875                 addr = vma ?  vma->vm_start : mm->mmap_base;
1876         mm->unmap_area(mm, addr);
1877         mm->mmap_cache = NULL;          /* Kill the cache. */
1878 }
1879
1880 /*
1881  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1882  * munmap path where it doesn't make sense to fail.
1883  */
1884 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1885               unsigned long addr, int new_below)
1886 {
1887         struct mempolicy *pol;
1888         struct vm_area_struct *new;
1889         int err = -ENOMEM;
1890
1891         if (is_vm_hugetlb_page(vma) && (addr &
1892                                         ~(huge_page_mask(hstate_vma(vma)))))
1893                 return -EINVAL;
1894
1895         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1896         if (!new)
1897                 goto out_err;
1898
1899         /* most fields are the same, copy all, and then fixup */
1900         *new = *vma;
1901
1902         INIT_LIST_HEAD(&new->anon_vma_chain);
1903
1904         if (new_below)
1905                 new->vm_end = addr;
1906         else {
1907                 new->vm_start = addr;
1908                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1909         }
1910
1911         pol = mpol_dup(vma_policy(vma));
1912         if (IS_ERR(pol)) {
1913                 err = PTR_ERR(pol);
1914                 goto out_free_vma;
1915         }
1916         vma_set_policy(new, pol);
1917
1918         if (anon_vma_clone(new, vma))
1919                 goto out_free_mpol;
1920
1921         if (new->vm_file) {
1922                 get_file(new->vm_file);
1923                 if (vma->vm_flags & VM_EXECUTABLE)
1924                         added_exe_file_vma(mm);
1925         }
1926
1927         if (new->vm_ops && new->vm_ops->open)
1928                 new->vm_ops->open(new);
1929
1930         if (new_below)
1931                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1932                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1933         else
1934                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1935
1936         /* Success. */
1937         if (!err)
1938                 return 0;
1939
1940         /* Clean everything up if vma_adjust failed. */
1941         new->vm_ops->close(new);
1942         if (new->vm_file) {
1943                 if (vma->vm_flags & VM_EXECUTABLE)
1944                         removed_exe_file_vma(mm);
1945                 fput(new->vm_file);
1946         }
1947  out_free_mpol:
1948         mpol_put(pol);
1949  out_free_vma:
1950         kmem_cache_free(vm_area_cachep, new);
1951  out_err:
1952         return err;
1953 }
1954
1955 /*
1956  * Split a vma into two pieces at address 'addr', a new vma is allocated
1957  * either for the first part or the tail.
1958  */
1959 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1960               unsigned long addr, int new_below)
1961 {
1962         if (mm->map_count >= sysctl_max_map_count)
1963                 return -ENOMEM;
1964
1965         return __split_vma(mm, vma, addr, new_below);
1966 }
1967
1968 /* Munmap is split into 2 main parts -- this part which finds
1969  * what needs doing, and the areas themselves, which do the
1970  * work.  This now handles partial unmappings.
1971  * Jeremy Fitzhardinge <jeremy@goop.org>
1972  */
1973 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1974 {
1975         unsigned long end;
1976         struct vm_area_struct *vma, *prev, *last;
1977
1978         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1979                 return -EINVAL;
1980
1981         if ((len = PAGE_ALIGN(len)) == 0)
1982                 return -EINVAL;
1983
1984         /* Find the first overlapping VMA */
1985         vma = find_vma_prev(mm, start, &prev);
1986         if (!vma)
1987                 return 0;
1988         /* we have  start < vma->vm_end  */
1989
1990         /* if it doesn't overlap, we have nothing.. */
1991         end = start + len;
1992         if (vma->vm_start >= end)
1993                 return 0;
1994
1995         /*
1996          * If we need to split any vma, do it now to save pain later.
1997          *
1998          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1999          * unmapped vm_area_struct will remain in use: so lower split_vma
2000          * places tmp vma above, and higher split_vma places tmp vma below.
2001          */
2002         if (start > vma->vm_start) {
2003                 int error;
2004
2005                 /*
2006                  * Make sure that map_count on return from munmap() will
2007                  * not exceed its limit; but let map_count go just above
2008                  * its limit temporarily, to help free resources as expected.
2009                  */
2010                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2011                         return -ENOMEM;
2012
2013                 error = __split_vma(mm, vma, start, 0);
2014                 if (error)
2015                         return error;
2016                 prev = vma;
2017         }
2018
2019         /* Does it split the last one? */
2020         last = find_vma(mm, end);
2021         if (last && end > last->vm_start) {
2022                 int error = __split_vma(mm, last, end, 1);
2023                 if (error)
2024                         return error;
2025         }
2026         vma = prev? prev->vm_next: mm->mmap;
2027
2028         /*
2029          * unlock any mlock()ed ranges before detaching vmas
2030          */
2031         if (mm->locked_vm) {
2032                 struct vm_area_struct *tmp = vma;
2033                 while (tmp && tmp->vm_start < end) {
2034                         if (tmp->vm_flags & VM_LOCKED) {
2035                                 mm->locked_vm -= vma_pages(tmp);
2036                                 munlock_vma_pages_all(tmp);
2037                         }
2038                         tmp = tmp->vm_next;
2039                 }
2040         }
2041
2042         /*
2043          * Remove the vma's, and unmap the actual pages
2044          */
2045         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2046         unmap_region(mm, vma, prev, start, end);
2047
2048         /* Fix up all other VM information */
2049         remove_vma_list(mm, vma);
2050
2051         return 0;
2052 }
2053
2054 EXPORT_SYMBOL(do_munmap);
2055
2056 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2057 {
2058         int ret;
2059         struct mm_struct *mm = current->mm;
2060
2061         profile_munmap(addr);
2062
2063         down_write(&mm->mmap_sem);
2064         ret = do_munmap(mm, addr, len);
2065         up_write(&mm->mmap_sem);
2066         return ret;
2067 }
2068
2069 static inline void verify_mm_writelocked(struct mm_struct *mm)
2070 {
2071 #ifdef CONFIG_DEBUG_VM
2072         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2073                 WARN_ON(1);
2074                 up_read(&mm->mmap_sem);
2075         }
2076 #endif
2077 }
2078
2079 /*
2080  *  this is really a simplified "do_mmap".  it only handles
2081  *  anonymous maps.  eventually we may be able to do some
2082  *  brk-specific accounting here.
2083  */
2084 unsigned long do_brk(unsigned long addr, unsigned long len)
2085 {
2086         struct mm_struct * mm = current->mm;
2087         struct vm_area_struct * vma, * prev;
2088         unsigned long flags;
2089         struct rb_node ** rb_link, * rb_parent;
2090         pgoff_t pgoff = addr >> PAGE_SHIFT;
2091         int error;
2092
2093         len = PAGE_ALIGN(len);
2094         if (!len)
2095                 return addr;
2096
2097         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2098         if (error)
2099                 return error;
2100
2101         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2102
2103         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2104         if (error & ~PAGE_MASK)
2105                 return error;
2106
2107         /*
2108          * mlock MCL_FUTURE?
2109          */
2110         if (mm->def_flags & VM_LOCKED) {
2111                 unsigned long locked, lock_limit;
2112                 locked = len >> PAGE_SHIFT;
2113                 locked += mm->locked_vm;
2114                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2115                 lock_limit >>= PAGE_SHIFT;
2116                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2117                         return -EAGAIN;
2118         }
2119
2120         /*
2121          * mm->mmap_sem is required to protect against another thread
2122          * changing the mappings in case we sleep.
2123          */
2124         verify_mm_writelocked(mm);
2125
2126         /*
2127          * Clear old maps.  this also does some error checking for us
2128          */
2129  munmap_back:
2130         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2131         if (vma && vma->vm_start < addr + len) {
2132                 if (do_munmap(mm, addr, len))
2133                         return -ENOMEM;
2134                 goto munmap_back;
2135         }
2136
2137         /* Check against address space limits *after* clearing old maps... */
2138         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2139                 return -ENOMEM;
2140
2141         if (mm->map_count > sysctl_max_map_count)
2142                 return -ENOMEM;
2143
2144         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2145                 return -ENOMEM;
2146
2147         /* Can we just expand an old private anonymous mapping? */
2148         vma = vma_merge(mm, prev, addr, addr + len, flags,
2149                                         NULL, NULL, pgoff, NULL);
2150         if (vma)
2151                 goto out;
2152
2153         /*
2154          * create a vma struct for an anonymous mapping
2155          */
2156         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2157         if (!vma) {
2158                 vm_unacct_memory(len >> PAGE_SHIFT);
2159                 return -ENOMEM;
2160         }
2161
2162         INIT_LIST_HEAD(&vma->anon_vma_chain);
2163         vma->vm_mm = mm;
2164         vma->vm_start = addr;
2165         vma->vm_end = addr + len;
2166         vma->vm_pgoff = pgoff;
2167         vma->vm_flags = flags;
2168         vma->vm_page_prot = vm_get_page_prot(flags);
2169         vma_link(mm, vma, prev, rb_link, rb_parent);
2170 out:
2171         mm->total_vm += len >> PAGE_SHIFT;
2172         if (flags & VM_LOCKED) {
2173                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2174                         mm->locked_vm += (len >> PAGE_SHIFT);
2175         }
2176         return addr;
2177 }
2178
2179 EXPORT_SYMBOL(do_brk);
2180
2181 /* Release all mmaps. */
2182 void exit_mmap(struct mm_struct *mm)
2183 {
2184         struct mmu_gather *tlb;
2185         struct vm_area_struct *vma;
2186         unsigned long nr_accounted = 0;
2187         unsigned long end;
2188
2189         /* mm's last user has gone, and its about to be pulled down */
2190         mmu_notifier_release(mm);
2191
2192         if (mm->locked_vm) {
2193                 vma = mm->mmap;
2194                 while (vma) {
2195                         if (vma->vm_flags & VM_LOCKED)
2196                                 munlock_vma_pages_all(vma);
2197                         vma = vma->vm_next;
2198                 }
2199         }
2200
2201         arch_exit_mmap(mm);
2202
2203         vma = mm->mmap;
2204         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2205                 return;
2206
2207         lru_add_drain();
2208         flush_cache_mm(mm);
2209         tlb = tlb_gather_mmu(mm, 1);
2210         /* update_hiwater_rss(mm) here? but nobody should be looking */
2211         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2212         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2213         vm_unacct_memory(nr_accounted);
2214
2215         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2216         tlb_finish_mmu(tlb, 0, end);
2217
2218         /*
2219          * Walk the list again, actually closing and freeing it,
2220          * with preemption enabled, without holding any MM locks.
2221          */
2222         while (vma)
2223                 vma = remove_vma(vma);
2224
2225         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2226 }
2227
2228 /* Insert vm structure into process list sorted by address
2229  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2230  * then i_mmap_lock is taken here.
2231  */
2232 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2233 {
2234         struct vm_area_struct * __vma, * prev;
2235         struct rb_node ** rb_link, * rb_parent;
2236
2237         /*
2238          * The vm_pgoff of a purely anonymous vma should be irrelevant
2239          * until its first write fault, when page's anon_vma and index
2240          * are set.  But now set the vm_pgoff it will almost certainly
2241          * end up with (unless mremap moves it elsewhere before that
2242          * first wfault), so /proc/pid/maps tells a consistent story.
2243          *
2244          * By setting it to reflect the virtual start address of the
2245          * vma, merges and splits can happen in a seamless way, just
2246          * using the existing file pgoff checks and manipulations.
2247          * Similarly in do_mmap_pgoff and in do_brk.
2248          */
2249         if (!vma->vm_file) {
2250                 BUG_ON(vma->anon_vma);
2251                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2252         }
2253         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2254         if (__vma && __vma->vm_start < vma->vm_end)
2255                 return -ENOMEM;
2256         if ((vma->vm_flags & VM_ACCOUNT) &&
2257              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2258                 return -ENOMEM;
2259         vma_link(mm, vma, prev, rb_link, rb_parent);
2260         return 0;
2261 }
2262
2263 /*
2264  * Copy the vma structure to a new location in the same mm,
2265  * prior to moving page table entries, to effect an mremap move.
2266  */
2267 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2268         unsigned long addr, unsigned long len, pgoff_t pgoff)
2269 {
2270         struct vm_area_struct *vma = *vmap;
2271         unsigned long vma_start = vma->vm_start;
2272         struct mm_struct *mm = vma->vm_mm;
2273         struct vm_area_struct *new_vma, *prev;
2274         struct rb_node **rb_link, *rb_parent;
2275         struct mempolicy *pol;
2276
2277         /*
2278          * If anonymous vma has not yet been faulted, update new pgoff
2279          * to match new location, to increase its chance of merging.
2280          */
2281         if (!vma->vm_file && !vma->anon_vma)
2282                 pgoff = addr >> PAGE_SHIFT;
2283
2284         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2285         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2286                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2287         if (new_vma) {
2288                 /*
2289                  * Source vma may have been merged into new_vma
2290                  */
2291                 if (vma_start >= new_vma->vm_start &&
2292                     vma_start < new_vma->vm_end)
2293                         *vmap = new_vma;
2294         } else {
2295                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2296                 if (new_vma) {
2297                         *new_vma = *vma;
2298                         pol = mpol_dup(vma_policy(vma));
2299                         if (IS_ERR(pol))
2300                                 goto out_free_vma;
2301                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2302                         if (anon_vma_clone(new_vma, vma))
2303                                 goto out_free_mempol;
2304                         vma_set_policy(new_vma, pol);
2305                         new_vma->vm_start = addr;
2306                         new_vma->vm_end = addr + len;
2307                         new_vma->vm_pgoff = pgoff;
2308                         if (new_vma->vm_file) {
2309                                 get_file(new_vma->vm_file);
2310                                 if (vma->vm_flags & VM_EXECUTABLE)
2311                                         added_exe_file_vma(mm);
2312                         }
2313                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2314                                 new_vma->vm_ops->open(new_vma);
2315                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2316                 }
2317         }
2318         return new_vma;
2319
2320  out_free_mempol:
2321         mpol_put(pol);
2322  out_free_vma:
2323         kmem_cache_free(vm_area_cachep, new_vma);
2324         return NULL;
2325 }
2326
2327 /*
2328  * Return true if the calling process may expand its vm space by the passed
2329  * number of pages
2330  */
2331 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2332 {
2333         unsigned long cur = mm->total_vm;       /* pages */
2334         unsigned long lim;
2335
2336         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2337
2338         if (cur + npages > lim)
2339                 return 0;
2340         return 1;
2341 }
2342
2343
2344 static int special_mapping_fault(struct vm_area_struct *vma,
2345                                 struct vm_fault *vmf)
2346 {
2347         pgoff_t pgoff;
2348         struct page **pages;
2349
2350         /*
2351          * special mappings have no vm_file, and in that case, the mm
2352          * uses vm_pgoff internally. So we have to subtract it from here.
2353          * We are allowed to do this because we are the mm; do not copy
2354          * this code into drivers!
2355          */
2356         pgoff = vmf->pgoff - vma->vm_pgoff;
2357
2358         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2359                 pgoff--;
2360
2361         if (*pages) {
2362                 struct page *page = *pages;
2363                 get_page(page);
2364                 vmf->page = page;
2365                 return 0;
2366         }
2367
2368         return VM_FAULT_SIGBUS;
2369 }
2370
2371 /*
2372  * Having a close hook prevents vma merging regardless of flags.
2373  */
2374 static void special_mapping_close(struct vm_area_struct *vma)
2375 {
2376 }
2377
2378 static const struct vm_operations_struct special_mapping_vmops = {
2379         .close = special_mapping_close,
2380         .fault = special_mapping_fault,
2381 };
2382
2383 /*
2384  * Called with mm->mmap_sem held for writing.
2385  * Insert a new vma covering the given region, with the given flags.
2386  * Its pages are supplied by the given array of struct page *.
2387  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2388  * The region past the last page supplied will always produce SIGBUS.
2389  * The array pointer and the pages it points to are assumed to stay alive
2390  * for as long as this mapping might exist.
2391  */
2392 int install_special_mapping(struct mm_struct *mm,
2393                             unsigned long addr, unsigned long len,
2394                             unsigned long vm_flags, struct page **pages)
2395 {
2396         struct vm_area_struct *vma;
2397
2398         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2399         if (unlikely(vma == NULL))
2400                 return -ENOMEM;
2401
2402         INIT_LIST_HEAD(&vma->anon_vma_chain);
2403         vma->vm_mm = mm;
2404         vma->vm_start = addr;
2405         vma->vm_end = addr + len;
2406
2407         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2408         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2409
2410         vma->vm_ops = &special_mapping_vmops;
2411         vma->vm_private_data = pages;
2412
2413         if (unlikely(insert_vm_struct(mm, vma))) {
2414                 kmem_cache_free(vm_area_cachep, vma);
2415                 return -ENOMEM;
2416         }
2417
2418         mm->total_vm += len >> PAGE_SHIFT;
2419
2420         perf_event_mmap(vma);
2421
2422         return 0;
2423 }
2424
2425 static DEFINE_MUTEX(mm_all_locks_mutex);
2426
2427 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2428 {
2429         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2430                 /*
2431                  * The LSB of head.next can't change from under us
2432                  * because we hold the mm_all_locks_mutex.
2433                  */
2434                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2435                 /*
2436                  * We can safely modify head.next after taking the
2437                  * anon_vma->lock. If some other vma in this mm shares
2438                  * the same anon_vma we won't take it again.
2439                  *
2440                  * No need of atomic instructions here, head.next
2441                  * can't change from under us thanks to the
2442                  * anon_vma->lock.
2443                  */
2444                 if (__test_and_set_bit(0, (unsigned long *)
2445                                        &anon_vma->head.next))
2446                         BUG();
2447         }
2448 }
2449
2450 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2451 {
2452         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2453                 /*
2454                  * AS_MM_ALL_LOCKS can't change from under us because
2455                  * we hold the mm_all_locks_mutex.
2456                  *
2457                  * Operations on ->flags have to be atomic because
2458                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2459                  * mm_all_locks_mutex, there may be other cpus
2460                  * changing other bitflags in parallel to us.
2461                  */
2462                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2463                         BUG();
2464                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2465         }
2466 }
2467
2468 /*
2469  * This operation locks against the VM for all pte/vma/mm related
2470  * operations that could ever happen on a certain mm. This includes
2471  * vmtruncate, try_to_unmap, and all page faults.
2472  *
2473  * The caller must take the mmap_sem in write mode before calling
2474  * mm_take_all_locks(). The caller isn't allowed to release the
2475  * mmap_sem until mm_drop_all_locks() returns.
2476  *
2477  * mmap_sem in write mode is required in order to block all operations
2478  * that could modify pagetables and free pages without need of
2479  * altering the vma layout (for example populate_range() with
2480  * nonlinear vmas). It's also needed in write mode to avoid new
2481  * anon_vmas to be associated with existing vmas.
2482  *
2483  * A single task can't take more than one mm_take_all_locks() in a row
2484  * or it would deadlock.
2485  *
2486  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2487  * mapping->flags avoid to take the same lock twice, if more than one
2488  * vma in this mm is backed by the same anon_vma or address_space.
2489  *
2490  * We can take all the locks in random order because the VM code
2491  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2492  * takes more than one of them in a row. Secondly we're protected
2493  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2494  *
2495  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2496  * that may have to take thousand of locks.
2497  *
2498  * mm_take_all_locks() can fail if it's interrupted by signals.
2499  */
2500 int mm_take_all_locks(struct mm_struct *mm)
2501 {
2502         struct vm_area_struct *vma;
2503         struct anon_vma_chain *avc;
2504         int ret = -EINTR;
2505
2506         BUG_ON(down_read_trylock(&mm->mmap_sem));
2507
2508         mutex_lock(&mm_all_locks_mutex);
2509
2510         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2511                 if (signal_pending(current))
2512                         goto out_unlock;
2513                 if (vma->vm_file && vma->vm_file->f_mapping)
2514                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2515         }
2516
2517         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2518                 if (signal_pending(current))
2519                         goto out_unlock;
2520                 if (vma->anon_vma)
2521                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2522                                 vm_lock_anon_vma(mm, avc->anon_vma);
2523         }
2524
2525         ret = 0;
2526
2527 out_unlock:
2528         if (ret)
2529                 mm_drop_all_locks(mm);
2530
2531         return ret;
2532 }
2533
2534 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2535 {
2536         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2537                 /*
2538                  * The LSB of head.next can't change to 0 from under
2539                  * us because we hold the mm_all_locks_mutex.
2540                  *
2541                  * We must however clear the bitflag before unlocking
2542                  * the vma so the users using the anon_vma->head will
2543                  * never see our bitflag.
2544                  *
2545                  * No need of atomic instructions here, head.next
2546                  * can't change from under us until we release the
2547                  * anon_vma->lock.
2548                  */
2549                 if (!__test_and_clear_bit(0, (unsigned long *)
2550                                           &anon_vma->head.next))
2551                         BUG();
2552                 spin_unlock(&anon_vma->lock);
2553         }
2554 }
2555
2556 static void vm_unlock_mapping(struct address_space *mapping)
2557 {
2558         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2559                 /*
2560                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2561                  * because we hold the mm_all_locks_mutex.
2562                  */
2563                 spin_unlock(&mapping->i_mmap_lock);
2564                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2565                                         &mapping->flags))
2566                         BUG();
2567         }
2568 }
2569
2570 /*
2571  * The mmap_sem cannot be released by the caller until
2572  * mm_drop_all_locks() returns.
2573  */
2574 void mm_drop_all_locks(struct mm_struct *mm)
2575 {
2576         struct vm_area_struct *vma;
2577         struct anon_vma_chain *avc;
2578
2579         BUG_ON(down_read_trylock(&mm->mmap_sem));
2580         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2581
2582         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2583                 if (vma->anon_vma)
2584                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2585                                 vm_unlock_anon_vma(avc->anon_vma);
2586                 if (vma->vm_file && vma->vm_file->f_mapping)
2587                         vm_unlock_mapping(vma->vm_file->f_mapping);
2588         }
2589
2590         mutex_unlock(&mm_all_locks_mutex);
2591 }
2592
2593 /*
2594  * initialise the VMA slab
2595  */
2596 void __init mmap_init(void)
2597 {
2598         int ret;
2599
2600         ret = percpu_counter_init(&vm_committed_as, 0);
2601         VM_BUG_ON(ret);
2602 }