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