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