mm: use rlimit helpers
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31
32 #include <asm/uaccess.h>
33 #include <asm/cacheflush.h>
34 #include <asm/tlb.h>
35 #include <asm/mmu_context.h>
36
37 #include "internal.h"
38
39 #ifndef arch_mmap_check
40 #define arch_mmap_check(addr, len, flags)       (0)
41 #endif
42
43 #ifndef arch_rebalance_pgtables
44 #define arch_rebalance_pgtables(addr, len)              (addr)
45 #endif
46
47 static void unmap_region(struct mm_struct *mm,
48                 struct vm_area_struct *vma, struct vm_area_struct *prev,
49                 unsigned long start, unsigned long end);
50
51 /*
52  * WARNING: the debugging will use recursive algorithms so never enable this
53  * unless you know what you are doing.
54  */
55 #undef DEBUG_MM_RB
56
57 /* description of effects of mapping type and prot in current implementation.
58  * this is due to the limited x86 page protection hardware.  The expected
59  * behavior is in parens:
60  *
61  * map_type     prot
62  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
63  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
64  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
65  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
66  *              
67  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
68  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
69  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
70  *
71  */
72 pgprot_t protection_map[16] = {
73         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
74         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
75 };
76
77 pgprot_t vm_get_page_prot(unsigned long vm_flags)
78 {
79         return __pgprot(pgprot_val(protection_map[vm_flags &
80                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
81                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
82 }
83 EXPORT_SYMBOL(vm_get_page_prot);
84
85 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
86 int sysctl_overcommit_ratio = 50;       /* default is 50% */
87 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
88 struct percpu_counter vm_committed_as;
89
90 /*
91  * Check that a process has enough memory to allocate a new virtual
92  * mapping. 0 means there is enough memory for the allocation to
93  * succeed and -ENOMEM implies there is not.
94  *
95  * We currently support three overcommit policies, which are set via the
96  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
97  *
98  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
99  * Additional code 2002 Jul 20 by Robert Love.
100  *
101  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
102  *
103  * Note this is a helper function intended to be used by LSMs which
104  * wish to use this logic.
105  */
106 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
107 {
108         unsigned long free, allowed;
109
110         vm_acct_memory(pages);
111
112         /*
113          * Sometimes we want to use more memory than we have
114          */
115         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
116                 return 0;
117
118         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
119                 unsigned long n;
120
121                 free = global_page_state(NR_FILE_PAGES);
122                 free += nr_swap_pages;
123
124                 /*
125                  * Any slabs which are created with the
126                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
127                  * which are reclaimable, under pressure.  The dentry
128                  * cache and most inode caches should fall into this
129                  */
130                 free += global_page_state(NR_SLAB_RECLAIMABLE);
131
132                 /*
133                  * Leave the last 3% for root
134                  */
135                 if (!cap_sys_admin)
136                         free -= free / 32;
137
138                 if (free > pages)
139                         return 0;
140
141                 /*
142                  * nr_free_pages() is very expensive on large systems,
143                  * only call if we're about to fail.
144                  */
145                 n = nr_free_pages();
146
147                 /*
148                  * Leave reserved pages. The pages are not for anonymous pages.
149                  */
150                 if (n <= totalreserve_pages)
151                         goto error;
152                 else
153                         n -= totalreserve_pages;
154
155                 /*
156                  * Leave the last 3% for root
157                  */
158                 if (!cap_sys_admin)
159                         n -= n / 32;
160                 free += n;
161
162                 if (free > pages)
163                         return 0;
164
165                 goto error;
166         }
167
168         allowed = (totalram_pages - hugetlb_total_pages())
169                 * sysctl_overcommit_ratio / 100;
170         /*
171          * Leave the last 3% for root
172          */
173         if (!cap_sys_admin)
174                 allowed -= allowed / 32;
175         allowed += total_swap_pages;
176
177         /* Don't let a single process grow too big:
178            leave 3% of the size of this process for other processes */
179         if (mm)
180                 allowed -= mm->total_vm / 32;
181
182         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
183                 return 0;
184 error:
185         vm_unacct_memory(pages);
186
187         return -ENOMEM;
188 }
189
190 /*
191  * Requires inode->i_mapping->i_mmap_lock
192  */
193 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
194                 struct file *file, struct address_space *mapping)
195 {
196         if (vma->vm_flags & VM_DENYWRITE)
197                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
198         if (vma->vm_flags & VM_SHARED)
199                 mapping->i_mmap_writable--;
200
201         flush_dcache_mmap_lock(mapping);
202         if (unlikely(vma->vm_flags & VM_NONLINEAR))
203                 list_del_init(&vma->shared.vm_set.list);
204         else
205                 vma_prio_tree_remove(vma, &mapping->i_mmap);
206         flush_dcache_mmap_unlock(mapping);
207 }
208
209 /*
210  * Unlink a file-based vm structure from its prio_tree, to hide
211  * vma from rmap and vmtruncate before freeing its page tables.
212  */
213 void unlink_file_vma(struct vm_area_struct *vma)
214 {
215         struct file *file = vma->vm_file;
216
217         if (file) {
218                 struct address_space *mapping = file->f_mapping;
219                 spin_lock(&mapping->i_mmap_lock);
220                 __remove_shared_vm_struct(vma, file, mapping);
221                 spin_unlock(&mapping->i_mmap_lock);
222         }
223 }
224
225 /*
226  * Close a vm structure and free it, returning the next.
227  */
228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
229 {
230         struct vm_area_struct *next = vma->vm_next;
231
232         might_sleep();
233         if (vma->vm_ops && vma->vm_ops->close)
234                 vma->vm_ops->close(vma);
235         if (vma->vm_file) {
236                 fput(vma->vm_file);
237                 if (vma->vm_flags & VM_EXECUTABLE)
238                         removed_exe_file_vma(vma->vm_mm);
239         }
240         mpol_put(vma_policy(vma));
241         kmem_cache_free(vm_area_cachep, vma);
242         return next;
243 }
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         min_brk = mm->end_code;
256 #else
257         min_brk = mm->start_brk;
258 #endif
259         if (brk < min_brk)
260                 goto out;
261
262         /*
263          * Check against rlimit here. If this check is done later after the test
264          * of oldbrk with newbrk then it can escape the test and let the data
265          * segment grow beyond its set limit the in case where the limit is
266          * not page aligned -Ram Gupta
267          */
268         rlim = rlimit(RLIMIT_DATA);
269         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
270                         (mm->end_data - mm->start_data) > rlim)
271                 goto out;
272
273         newbrk = PAGE_ALIGN(brk);
274         oldbrk = PAGE_ALIGN(mm->brk);
275         if (oldbrk == newbrk)
276                 goto set_brk;
277
278         /* Always allow shrinking brk. */
279         if (brk <= mm->brk) {
280                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
281                         goto set_brk;
282                 goto out;
283         }
284
285         /* Check against existing mmap mappings. */
286         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
287                 goto out;
288
289         /* Ok, looks good - let it rip. */
290         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
291                 goto out;
292 set_brk:
293         mm->brk = brk;
294 out:
295         retval = mm->brk;
296         up_write(&mm->mmap_sem);
297         return retval;
298 }
299
300 #ifdef DEBUG_MM_RB
301 static int browse_rb(struct rb_root *root)
302 {
303         int i = 0, j;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev)
311                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
312                 if (vma->vm_start < pend)
313                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
314                 if (vma->vm_start > vma->vm_end)
315                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
316                 i++;
317                 pn = nd;
318                 prev = vma->vm_start;
319                 pend = vma->vm_end;
320         }
321         j = 0;
322         for (nd = pn; nd; nd = rb_prev(nd)) {
323                 j++;
324         }
325         if (i != j)
326                 printk("backwards %d, forwards %d\n", j, i), i = 0;
327         return i;
328 }
329
330 void validate_mm(struct mm_struct *mm)
331 {
332         int bug = 0;
333         int i = 0;
334         struct vm_area_struct *tmp = mm->mmap;
335         while (tmp) {
336                 tmp = tmp->vm_next;
337                 i++;
338         }
339         if (i != mm->map_count)
340                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
341         i = browse_rb(&mm->mm_rb);
342         if (i != mm->map_count)
343                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
344         BUG_ON(bug);
345 }
346 #else
347 #define validate_mm(mm) do { } while (0)
348 #endif
349
350 static struct vm_area_struct *
351 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
352                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
353                 struct rb_node ** rb_parent)
354 {
355         struct vm_area_struct * vma;
356         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
357
358         __rb_link = &mm->mm_rb.rb_node;
359         rb_prev = __rb_parent = NULL;
360         vma = NULL;
361
362         while (*__rb_link) {
363                 struct vm_area_struct *vma_tmp;
364
365                 __rb_parent = *__rb_link;
366                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
367
368                 if (vma_tmp->vm_end > addr) {
369                         vma = vma_tmp;
370                         if (vma_tmp->vm_start <= addr)
371                                 break;
372                         __rb_link = &__rb_parent->rb_left;
373                 } else {
374                         rb_prev = __rb_parent;
375                         __rb_link = &__rb_parent->rb_right;
376                 }
377         }
378
379         *pprev = NULL;
380         if (rb_prev)
381                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
382         *rb_link = __rb_link;
383         *rb_parent = __rb_parent;
384         return vma;
385 }
386
387 static inline void
388 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
389                 struct vm_area_struct *prev, struct rb_node *rb_parent)
390 {
391         if (prev) {
392                 vma->vm_next = prev->vm_next;
393                 prev->vm_next = vma;
394         } else {
395                 mm->mmap = vma;
396                 if (rb_parent)
397                         vma->vm_next = rb_entry(rb_parent,
398                                         struct vm_area_struct, vm_rb);
399                 else
400                         vma->vm_next = NULL;
401         }
402 }
403
404 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
405                 struct rb_node **rb_link, struct rb_node *rb_parent)
406 {
407         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
408         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
409 }
410
411 static void __vma_link_file(struct vm_area_struct *vma)
412 {
413         struct file *file;
414
415         file = vma->vm_file;
416         if (file) {
417                 struct address_space *mapping = file->f_mapping;
418
419                 if (vma->vm_flags & VM_DENYWRITE)
420                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
421                 if (vma->vm_flags & VM_SHARED)
422                         mapping->i_mmap_writable++;
423
424                 flush_dcache_mmap_lock(mapping);
425                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
426                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
427                 else
428                         vma_prio_tree_insert(vma, &mapping->i_mmap);
429                 flush_dcache_mmap_unlock(mapping);
430         }
431 }
432
433 static void
434 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435         struct vm_area_struct *prev, struct rb_node **rb_link,
436         struct rb_node *rb_parent)
437 {
438         __vma_link_list(mm, vma, prev, rb_parent);
439         __vma_link_rb(mm, vma, rb_link, rb_parent);
440         __anon_vma_link(vma);
441 }
442
443 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
444                         struct vm_area_struct *prev, struct rb_node **rb_link,
445                         struct rb_node *rb_parent)
446 {
447         struct address_space *mapping = NULL;
448
449         if (vma->vm_file)
450                 mapping = vma->vm_file->f_mapping;
451
452         if (mapping) {
453                 spin_lock(&mapping->i_mmap_lock);
454                 vma->vm_truncate_count = mapping->truncate_count;
455         }
456         anon_vma_lock(vma);
457
458         __vma_link(mm, vma, prev, rb_link, rb_parent);
459         __vma_link_file(vma);
460
461         anon_vma_unlock(vma);
462         if (mapping)
463                 spin_unlock(&mapping->i_mmap_lock);
464
465         mm->map_count++;
466         validate_mm(mm);
467 }
468
469 /*
470  * Helper for vma_adjust in the split_vma insert case:
471  * insert vm structure into list and rbtree and anon_vma,
472  * but it has already been inserted into prio_tree earlier.
473  */
474 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
475 {
476         struct vm_area_struct *__vma, *prev;
477         struct rb_node **rb_link, *rb_parent;
478
479         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
480         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
481         __vma_link(mm, vma, prev, rb_link, rb_parent);
482         mm->map_count++;
483 }
484
485 static inline void
486 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
487                 struct vm_area_struct *prev)
488 {
489         prev->vm_next = vma->vm_next;
490         rb_erase(&vma->vm_rb, &mm->mm_rb);
491         if (mm->mmap_cache == vma)
492                 mm->mmap_cache = prev;
493 }
494
495 /*
496  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
497  * is already present in an i_mmap tree without adjusting the tree.
498  * The following helper function should be used when such adjustments
499  * are necessary.  The "insert" vma (if any) is to be inserted
500  * before we drop the necessary locks.
501  */
502 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
503         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
504 {
505         struct mm_struct *mm = vma->vm_mm;
506         struct vm_area_struct *next = vma->vm_next;
507         struct vm_area_struct *importer = NULL;
508         struct address_space *mapping = NULL;
509         struct prio_tree_root *root = NULL;
510         struct file *file = vma->vm_file;
511         struct anon_vma *anon_vma = NULL;
512         long adjust_next = 0;
513         int remove_next = 0;
514
515         if (next && !insert) {
516                 if (end >= next->vm_end) {
517                         /*
518                          * vma expands, overlapping all the next, and
519                          * perhaps the one after too (mprotect case 6).
520                          */
521 again:                  remove_next = 1 + (end > next->vm_end);
522                         end = next->vm_end;
523                         anon_vma = next->anon_vma;
524                         importer = vma;
525                 } else if (end > next->vm_start) {
526                         /*
527                          * vma expands, overlapping part of the next:
528                          * mprotect case 5 shifting the boundary up.
529                          */
530                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
531                         anon_vma = next->anon_vma;
532                         importer = vma;
533                 } else if (end < vma->vm_end) {
534                         /*
535                          * vma shrinks, and !insert tells it's not
536                          * split_vma inserting another: so it must be
537                          * mprotect case 4 shifting the boundary down.
538                          */
539                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
540                         anon_vma = next->anon_vma;
541                         importer = next;
542                 }
543         }
544
545         if (file) {
546                 mapping = file->f_mapping;
547                 if (!(vma->vm_flags & VM_NONLINEAR))
548                         root = &mapping->i_mmap;
549                 spin_lock(&mapping->i_mmap_lock);
550                 if (importer &&
551                     vma->vm_truncate_count != next->vm_truncate_count) {
552                         /*
553                          * unmap_mapping_range might be in progress:
554                          * ensure that the expanding vma is rescanned.
555                          */
556                         importer->vm_truncate_count = 0;
557                 }
558                 if (insert) {
559                         insert->vm_truncate_count = vma->vm_truncate_count;
560                         /*
561                          * Put into prio_tree now, so instantiated pages
562                          * are visible to arm/parisc __flush_dcache_page
563                          * throughout; but we cannot insert into address
564                          * space until vma start or end is updated.
565                          */
566                         __vma_link_file(insert);
567                 }
568         }
569
570         /*
571          * When changing only vma->vm_end, we don't really need
572          * anon_vma lock.
573          */
574         if (vma->anon_vma && (insert || importer || start != vma->vm_start))
575                 anon_vma = vma->anon_vma;
576         if (anon_vma) {
577                 spin_lock(&anon_vma->lock);
578                 /*
579                  * Easily overlooked: when mprotect shifts the boundary,
580                  * make sure the expanding vma has anon_vma set if the
581                  * shrinking vma had, to cover any anon pages imported.
582                  */
583                 if (importer && !importer->anon_vma) {
584                         importer->anon_vma = anon_vma;
585                         __anon_vma_link(importer);
586                 }
587         }
588
589         if (root) {
590                 flush_dcache_mmap_lock(mapping);
591                 vma_prio_tree_remove(vma, root);
592                 if (adjust_next)
593                         vma_prio_tree_remove(next, root);
594         }
595
596         vma->vm_start = start;
597         vma->vm_end = end;
598         vma->vm_pgoff = pgoff;
599         if (adjust_next) {
600                 next->vm_start += adjust_next << PAGE_SHIFT;
601                 next->vm_pgoff += adjust_next;
602         }
603
604         if (root) {
605                 if (adjust_next)
606                         vma_prio_tree_insert(next, root);
607                 vma_prio_tree_insert(vma, root);
608                 flush_dcache_mmap_unlock(mapping);
609         }
610
611         if (remove_next) {
612                 /*
613                  * vma_merge has merged next into vma, and needs
614                  * us to remove next before dropping the locks.
615                  */
616                 __vma_unlink(mm, next, vma);
617                 if (file)
618                         __remove_shared_vm_struct(next, file, mapping);
619                 if (next->anon_vma)
620                         __anon_vma_merge(vma, next);
621         } else if (insert) {
622                 /*
623                  * split_vma has split insert from vma, and needs
624                  * us to insert it before dropping the locks
625                  * (it may either follow vma or precede it).
626                  */
627                 __insert_vm_struct(mm, insert);
628         }
629
630         if (anon_vma)
631                 spin_unlock(&anon_vma->lock);
632         if (mapping)
633                 spin_unlock(&mapping->i_mmap_lock);
634
635         if (remove_next) {
636                 if (file) {
637                         fput(file);
638                         if (next->vm_flags & VM_EXECUTABLE)
639                                 removed_exe_file_vma(mm);
640                 }
641                 mm->map_count--;
642                 mpol_put(vma_policy(next));
643                 kmem_cache_free(vm_area_cachep, next);
644                 /*
645                  * In mprotect's case 6 (see comments on vma_merge),
646                  * we must remove another next too. It would clutter
647                  * up the code too much to do both in one go.
648                  */
649                 if (remove_next == 2) {
650                         next = vma->vm_next;
651                         goto again;
652                 }
653         }
654
655         validate_mm(mm);
656 }
657
658 /*
659  * If the vma has a ->close operation then the driver probably needs to release
660  * per-vma resources, so we don't attempt to merge those.
661  */
662 static inline int is_mergeable_vma(struct vm_area_struct *vma,
663                         struct file *file, unsigned long vm_flags)
664 {
665         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
666         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
667                 return 0;
668         if (vma->vm_file != file)
669                 return 0;
670         if (vma->vm_ops && vma->vm_ops->close)
671                 return 0;
672         return 1;
673 }
674
675 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
676                                         struct anon_vma *anon_vma2)
677 {
678         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
679 }
680
681 /*
682  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
683  * in front of (at a lower virtual address and file offset than) the vma.
684  *
685  * We cannot merge two vmas if they have differently assigned (non-NULL)
686  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
687  *
688  * We don't check here for the merged mmap wrapping around the end of pagecache
689  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
690  * wrap, nor mmaps which cover the final page at index -1UL.
691  */
692 static int
693 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
694         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
695 {
696         if (is_mergeable_vma(vma, file, vm_flags) &&
697             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
698                 if (vma->vm_pgoff == vm_pgoff)
699                         return 1;
700         }
701         return 0;
702 }
703
704 /*
705  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
706  * beyond (at a higher virtual address and file offset than) the vma.
707  *
708  * We cannot merge two vmas if they have differently assigned (non-NULL)
709  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
710  */
711 static int
712 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
713         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
714 {
715         if (is_mergeable_vma(vma, file, vm_flags) &&
716             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
717                 pgoff_t vm_pglen;
718                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
719                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
720                         return 1;
721         }
722         return 0;
723 }
724
725 /*
726  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
727  * whether that can be merged with its predecessor or its successor.
728  * Or both (it neatly fills a hole).
729  *
730  * In most cases - when called for mmap, brk or mremap - [addr,end) is
731  * certain not to be mapped by the time vma_merge is called; but when
732  * called for mprotect, it is certain to be already mapped (either at
733  * an offset within prev, or at the start of next), and the flags of
734  * this area are about to be changed to vm_flags - and the no-change
735  * case has already been eliminated.
736  *
737  * The following mprotect cases have to be considered, where AAAA is
738  * the area passed down from mprotect_fixup, never extending beyond one
739  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
740  *
741  *     AAAA             AAAA                AAAA          AAAA
742  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
743  *    cannot merge    might become    might become    might become
744  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
745  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
746  *    mremap move:                                    PPPPNNNNNNNN 8
747  *        AAAA
748  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
749  *    might become    case 1 below    case 2 below    case 3 below
750  *
751  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
752  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
753  */
754 struct vm_area_struct *vma_merge(struct mm_struct *mm,
755                         struct vm_area_struct *prev, unsigned long addr,
756                         unsigned long end, unsigned long vm_flags,
757                         struct anon_vma *anon_vma, struct file *file,
758                         pgoff_t pgoff, struct mempolicy *policy)
759 {
760         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
761         struct vm_area_struct *area, *next;
762
763         /*
764          * We later require that vma->vm_flags == vm_flags,
765          * so this tests vma->vm_flags & VM_SPECIAL, too.
766          */
767         if (vm_flags & VM_SPECIAL)
768                 return NULL;
769
770         if (prev)
771                 next = prev->vm_next;
772         else
773                 next = mm->mmap;
774         area = next;
775         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
776                 next = next->vm_next;
777
778         /*
779          * Can it merge with the predecessor?
780          */
781         if (prev && prev->vm_end == addr &&
782                         mpol_equal(vma_policy(prev), policy) &&
783                         can_vma_merge_after(prev, vm_flags,
784                                                 anon_vma, file, pgoff)) {
785                 /*
786                  * OK, it can.  Can we now merge in the successor as well?
787                  */
788                 if (next && end == next->vm_start &&
789                                 mpol_equal(policy, vma_policy(next)) &&
790                                 can_vma_merge_before(next, vm_flags,
791                                         anon_vma, file, pgoff+pglen) &&
792                                 is_mergeable_anon_vma(prev->anon_vma,
793                                                       next->anon_vma)) {
794                                                         /* cases 1, 6 */
795                         vma_adjust(prev, prev->vm_start,
796                                 next->vm_end, prev->vm_pgoff, NULL);
797                 } else                                  /* cases 2, 5, 7 */
798                         vma_adjust(prev, prev->vm_start,
799                                 end, prev->vm_pgoff, NULL);
800                 return prev;
801         }
802
803         /*
804          * Can this new request be merged in front of next?
805          */
806         if (next && end == next->vm_start &&
807                         mpol_equal(policy, vma_policy(next)) &&
808                         can_vma_merge_before(next, vm_flags,
809                                         anon_vma, file, pgoff+pglen)) {
810                 if (prev && addr < prev->vm_end)        /* case 4 */
811                         vma_adjust(prev, prev->vm_start,
812                                 addr, prev->vm_pgoff, NULL);
813                 else                                    /* cases 3, 8 */
814                         vma_adjust(area, addr, next->vm_end,
815                                 next->vm_pgoff - pglen, NULL);
816                 return area;
817         }
818
819         return NULL;
820 }
821
822 /*
823  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
824  * neighbouring vmas for a suitable anon_vma, before it goes off
825  * to allocate a new anon_vma.  It checks because a repetitive
826  * sequence of mprotects and faults may otherwise lead to distinct
827  * anon_vmas being allocated, preventing vma merge in subsequent
828  * mprotect.
829  */
830 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
831 {
832         struct vm_area_struct *near;
833         unsigned long vm_flags;
834
835         near = vma->vm_next;
836         if (!near)
837                 goto try_prev;
838
839         /*
840          * Since only mprotect tries to remerge vmas, match flags
841          * which might be mprotected into each other later on.
842          * Neither mlock nor madvise tries to remerge at present,
843          * so leave their flags as obstructing a merge.
844          */
845         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
846         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
847
848         if (near->anon_vma && vma->vm_end == near->vm_start &&
849                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
850                         can_vma_merge_before(near, vm_flags,
851                                 NULL, vma->vm_file, vma->vm_pgoff +
852                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
853                 return near->anon_vma;
854 try_prev:
855         /*
856          * It is potentially slow to have to call find_vma_prev here.
857          * But it's only on the first write fault on the vma, not
858          * every time, and we could devise a way to avoid it later
859          * (e.g. stash info in next's anon_vma_node when assigning
860          * an anon_vma, or when trying vma_merge).  Another time.
861          */
862         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
863         if (!near)
864                 goto none;
865
866         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
867         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
868
869         if (near->anon_vma && near->vm_end == vma->vm_start &&
870                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
871                         can_vma_merge_after(near, vm_flags,
872                                 NULL, vma->vm_file, vma->vm_pgoff))
873                 return near->anon_vma;
874 none:
875         /*
876          * There's no absolute need to look only at touching neighbours:
877          * we could search further afield for "compatible" anon_vmas.
878          * But it would probably just be a waste of time searching,
879          * or lead to too many vmas hanging off the same anon_vma.
880          * We're trying to allow mprotect remerging later on,
881          * not trying to minimize memory used for anon_vmas.
882          */
883         return NULL;
884 }
885
886 #ifdef CONFIG_PROC_FS
887 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
888                                                 struct file *file, long pages)
889 {
890         const unsigned long stack_flags
891                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
892
893         if (file) {
894                 mm->shared_vm += pages;
895                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
896                         mm->exec_vm += pages;
897         } else if (flags & stack_flags)
898                 mm->stack_vm += pages;
899         if (flags & (VM_RESERVED|VM_IO))
900                 mm->reserved_vm += pages;
901 }
902 #endif /* CONFIG_PROC_FS */
903
904 /*
905  * The caller must hold down_write(&current->mm->mmap_sem).
906  */
907
908 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
909                         unsigned long len, unsigned long prot,
910                         unsigned long flags, unsigned long pgoff)
911 {
912         struct mm_struct * mm = current->mm;
913         struct inode *inode;
914         unsigned int vm_flags;
915         int error;
916         unsigned long reqprot = prot;
917
918         /*
919          * Does the application expect PROT_READ to imply PROT_EXEC?
920          *
921          * (the exception is when the underlying filesystem is noexec
922          *  mounted, in which case we dont add PROT_EXEC.)
923          */
924         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
925                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
926                         prot |= PROT_EXEC;
927
928         if (!len)
929                 return -EINVAL;
930
931         if (!(flags & MAP_FIXED))
932                 addr = round_hint_to_min(addr);
933
934         /* Careful about overflows.. */
935         len = PAGE_ALIGN(len);
936         if (!len)
937                 return -ENOMEM;
938
939         /* offset overflow? */
940         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
941                return -EOVERFLOW;
942
943         /* Too many mappings? */
944         if (mm->map_count > sysctl_max_map_count)
945                 return -ENOMEM;
946
947         /* Obtain the address to map to. we verify (or select) it and ensure
948          * that it represents a valid section of the address space.
949          */
950         addr = get_unmapped_area(file, addr, len, pgoff, flags);
951         if (addr & ~PAGE_MASK)
952                 return addr;
953
954         /* Do simple checking here so the lower-level routines won't have
955          * to. we assume access permissions have been handled by the open
956          * of the memory object, so we don't do any here.
957          */
958         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
959                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
960
961         if (flags & MAP_LOCKED)
962                 if (!can_do_mlock())
963                         return -EPERM;
964
965         /* mlock MCL_FUTURE? */
966         if (vm_flags & VM_LOCKED) {
967                 unsigned long locked, lock_limit;
968                 locked = len >> PAGE_SHIFT;
969                 locked += mm->locked_vm;
970                 lock_limit = rlimit(RLIMIT_MEMLOCK);
971                 lock_limit >>= PAGE_SHIFT;
972                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
973                         return -EAGAIN;
974         }
975
976         inode = file ? file->f_path.dentry->d_inode : NULL;
977
978         if (file) {
979                 switch (flags & MAP_TYPE) {
980                 case MAP_SHARED:
981                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
982                                 return -EACCES;
983
984                         /*
985                          * Make sure we don't allow writing to an append-only
986                          * file..
987                          */
988                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
989                                 return -EACCES;
990
991                         /*
992                          * Make sure there are no mandatory locks on the file.
993                          */
994                         if (locks_verify_locked(inode))
995                                 return -EAGAIN;
996
997                         vm_flags |= VM_SHARED | VM_MAYSHARE;
998                         if (!(file->f_mode & FMODE_WRITE))
999                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1000
1001                         /* fall through */
1002                 case MAP_PRIVATE:
1003                         if (!(file->f_mode & FMODE_READ))
1004                                 return -EACCES;
1005                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1006                                 if (vm_flags & VM_EXEC)
1007                                         return -EPERM;
1008                                 vm_flags &= ~VM_MAYEXEC;
1009                         }
1010
1011                         if (!file->f_op || !file->f_op->mmap)
1012                                 return -ENODEV;
1013                         break;
1014
1015                 default:
1016                         return -EINVAL;
1017                 }
1018         } else {
1019                 switch (flags & MAP_TYPE) {
1020                 case MAP_SHARED:
1021                         /*
1022                          * Ignore pgoff.
1023                          */
1024                         pgoff = 0;
1025                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1026                         break;
1027                 case MAP_PRIVATE:
1028                         /*
1029                          * Set pgoff according to addr for anon_vma.
1030                          */
1031                         pgoff = addr >> PAGE_SHIFT;
1032                         break;
1033                 default:
1034                         return -EINVAL;
1035                 }
1036         }
1037
1038         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1039         if (error)
1040                 return error;
1041
1042         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1043 }
1044 EXPORT_SYMBOL(do_mmap_pgoff);
1045
1046 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1047                 unsigned long, prot, unsigned long, flags,
1048                 unsigned long, fd, unsigned long, pgoff)
1049 {
1050         struct file *file = NULL;
1051         unsigned long retval = -EBADF;
1052
1053         if (!(flags & MAP_ANONYMOUS)) {
1054                 if (unlikely(flags & MAP_HUGETLB))
1055                         return -EINVAL;
1056                 file = fget(fd);
1057                 if (!file)
1058                         goto out;
1059         } else if (flags & MAP_HUGETLB) {
1060                 struct user_struct *user = NULL;
1061                 /*
1062                  * VM_NORESERVE is used because the reservations will be
1063                  * taken when vm_ops->mmap() is called
1064                  * A dummy user value is used because we are not locking
1065                  * memory so no accounting is necessary
1066                  */
1067                 len = ALIGN(len, huge_page_size(&default_hstate));
1068                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1069                                                 &user, HUGETLB_ANONHUGE_INODE);
1070                 if (IS_ERR(file))
1071                         return PTR_ERR(file);
1072         }
1073
1074         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1075
1076         down_write(&current->mm->mmap_sem);
1077         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1078         up_write(&current->mm->mmap_sem);
1079
1080         if (file)
1081                 fput(file);
1082 out:
1083         return retval;
1084 }
1085
1086 /*
1087  * Some shared mappigns will want the pages marked read-only
1088  * to track write events. If so, we'll downgrade vm_page_prot
1089  * to the private version (using protection_map[] without the
1090  * VM_SHARED bit).
1091  */
1092 int vma_wants_writenotify(struct vm_area_struct *vma)
1093 {
1094         unsigned int vm_flags = vma->vm_flags;
1095
1096         /* If it was private or non-writable, the write bit is already clear */
1097         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1098                 return 0;
1099
1100         /* The backer wishes to know when pages are first written to? */
1101         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1102                 return 1;
1103
1104         /* The open routine did something to the protections already? */
1105         if (pgprot_val(vma->vm_page_prot) !=
1106             pgprot_val(vm_get_page_prot(vm_flags)))
1107                 return 0;
1108
1109         /* Specialty mapping? */
1110         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1111                 return 0;
1112
1113         /* Can the mapping track the dirty pages? */
1114         return vma->vm_file && vma->vm_file->f_mapping &&
1115                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1116 }
1117
1118 /*
1119  * We account for memory if it's a private writeable mapping,
1120  * not hugepages and VM_NORESERVE wasn't set.
1121  */
1122 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1123 {
1124         /*
1125          * hugetlb has its own accounting separate from the core VM
1126          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1127          */
1128         if (file && is_file_hugepages(file))
1129                 return 0;
1130
1131         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1132 }
1133
1134 unsigned long mmap_region(struct file *file, unsigned long addr,
1135                           unsigned long len, unsigned long flags,
1136                           unsigned int vm_flags, unsigned long pgoff)
1137 {
1138         struct mm_struct *mm = current->mm;
1139         struct vm_area_struct *vma, *prev;
1140         int correct_wcount = 0;
1141         int error;
1142         struct rb_node **rb_link, *rb_parent;
1143         unsigned long charged = 0;
1144         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1145
1146         /* Clear old maps */
1147         error = -ENOMEM;
1148 munmap_back:
1149         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1150         if (vma && vma->vm_start < addr + len) {
1151                 if (do_munmap(mm, addr, len))
1152                         return -ENOMEM;
1153                 goto munmap_back;
1154         }
1155
1156         /* Check against address space limit. */
1157         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1158                 return -ENOMEM;
1159
1160         /*
1161          * Set 'VM_NORESERVE' if we should not account for the
1162          * memory use of this mapping.
1163          */
1164         if ((flags & MAP_NORESERVE)) {
1165                 /* We honor MAP_NORESERVE if allowed to overcommit */
1166                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1167                         vm_flags |= VM_NORESERVE;
1168
1169                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1170                 if (file && is_file_hugepages(file))
1171                         vm_flags |= VM_NORESERVE;
1172         }
1173
1174         /*
1175          * Private writable mapping: check memory availability
1176          */
1177         if (accountable_mapping(file, vm_flags)) {
1178                 charged = len >> PAGE_SHIFT;
1179                 if (security_vm_enough_memory(charged))
1180                         return -ENOMEM;
1181                 vm_flags |= VM_ACCOUNT;
1182         }
1183
1184         /*
1185          * Can we just expand an old mapping?
1186          */
1187         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1188         if (vma)
1189                 goto out;
1190
1191         /*
1192          * Determine the object being mapped and call the appropriate
1193          * specific mapper. the address has already been validated, but
1194          * not unmapped, but the maps are removed from the list.
1195          */
1196         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1197         if (!vma) {
1198                 error = -ENOMEM;
1199                 goto unacct_error;
1200         }
1201
1202         vma->vm_mm = mm;
1203         vma->vm_start = addr;
1204         vma->vm_end = addr + len;
1205         vma->vm_flags = vm_flags;
1206         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1207         vma->vm_pgoff = pgoff;
1208
1209         if (file) {
1210                 error = -EINVAL;
1211                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1212                         goto free_vma;
1213                 if (vm_flags & VM_DENYWRITE) {
1214                         error = deny_write_access(file);
1215                         if (error)
1216                                 goto free_vma;
1217                         correct_wcount = 1;
1218                 }
1219                 vma->vm_file = file;
1220                 get_file(file);
1221                 error = file->f_op->mmap(file, vma);
1222                 if (error)
1223                         goto unmap_and_free_vma;
1224                 if (vm_flags & VM_EXECUTABLE)
1225                         added_exe_file_vma(mm);
1226
1227                 /* Can addr have changed??
1228                  *
1229                  * Answer: Yes, several device drivers can do it in their
1230                  *         f_op->mmap method. -DaveM
1231                  */
1232                 addr = vma->vm_start;
1233                 pgoff = vma->vm_pgoff;
1234                 vm_flags = vma->vm_flags;
1235         } else if (vm_flags & VM_SHARED) {
1236                 error = shmem_zero_setup(vma);
1237                 if (error)
1238                         goto free_vma;
1239         }
1240
1241         if (vma_wants_writenotify(vma)) {
1242                 pgprot_t pprot = vma->vm_page_prot;
1243
1244                 /* Can vma->vm_page_prot have changed??
1245                  *
1246                  * Answer: Yes, drivers may have changed it in their
1247                  *         f_op->mmap method.
1248                  *
1249                  * Ensures that vmas marked as uncached stay that way.
1250                  */
1251                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1252                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1253                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1254         }
1255
1256         vma_link(mm, vma, prev, rb_link, rb_parent);
1257         file = vma->vm_file;
1258
1259         /* Once vma denies write, undo our temporary denial count */
1260         if (correct_wcount)
1261                 atomic_inc(&inode->i_writecount);
1262 out:
1263         perf_event_mmap(vma);
1264
1265         mm->total_vm += len >> PAGE_SHIFT;
1266         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1267         if (vm_flags & VM_LOCKED) {
1268                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1269                         mm->locked_vm += (len >> PAGE_SHIFT);
1270         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1271                 make_pages_present(addr, addr + len);
1272         return addr;
1273
1274 unmap_and_free_vma:
1275         if (correct_wcount)
1276                 atomic_inc(&inode->i_writecount);
1277         vma->vm_file = NULL;
1278         fput(file);
1279
1280         /* Undo any partial mapping done by a device driver. */
1281         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1282         charged = 0;
1283 free_vma:
1284         kmem_cache_free(vm_area_cachep, vma);
1285 unacct_error:
1286         if (charged)
1287                 vm_unacct_memory(charged);
1288         return error;
1289 }
1290
1291 /* Get an address range which is currently unmapped.
1292  * For shmat() with addr=0.
1293  *
1294  * Ugly calling convention alert:
1295  * Return value with the low bits set means error value,
1296  * ie
1297  *      if (ret & ~PAGE_MASK)
1298  *              error = ret;
1299  *
1300  * This function "knows" that -ENOMEM has the bits set.
1301  */
1302 #ifndef HAVE_ARCH_UNMAPPED_AREA
1303 unsigned long
1304 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1305                 unsigned long len, unsigned long pgoff, unsigned long flags)
1306 {
1307         struct mm_struct *mm = current->mm;
1308         struct vm_area_struct *vma;
1309         unsigned long start_addr;
1310
1311         if (len > TASK_SIZE)
1312                 return -ENOMEM;
1313
1314         if (flags & MAP_FIXED)
1315                 return addr;
1316
1317         if (addr) {
1318                 addr = PAGE_ALIGN(addr);
1319                 vma = find_vma(mm, addr);
1320                 if (TASK_SIZE - len >= addr &&
1321                     (!vma || addr + len <= vma->vm_start))
1322                         return addr;
1323         }
1324         if (len > mm->cached_hole_size) {
1325                 start_addr = addr = mm->free_area_cache;
1326         } else {
1327                 start_addr = addr = TASK_UNMAPPED_BASE;
1328                 mm->cached_hole_size = 0;
1329         }
1330
1331 full_search:
1332         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1333                 /* At this point:  (!vma || addr < vma->vm_end). */
1334                 if (TASK_SIZE - len < addr) {
1335                         /*
1336                          * Start a new search - just in case we missed
1337                          * some holes.
1338                          */
1339                         if (start_addr != TASK_UNMAPPED_BASE) {
1340                                 addr = TASK_UNMAPPED_BASE;
1341                                 start_addr = addr;
1342                                 mm->cached_hole_size = 0;
1343                                 goto full_search;
1344                         }
1345                         return -ENOMEM;
1346                 }
1347                 if (!vma || addr + len <= vma->vm_start) {
1348                         /*
1349                          * Remember the place where we stopped the search:
1350                          */
1351                         mm->free_area_cache = addr + len;
1352                         return addr;
1353                 }
1354                 if (addr + mm->cached_hole_size < vma->vm_start)
1355                         mm->cached_hole_size = vma->vm_start - addr;
1356                 addr = vma->vm_end;
1357         }
1358 }
1359 #endif  
1360
1361 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1362 {
1363         /*
1364          * Is this a new hole at the lowest possible address?
1365          */
1366         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1367                 mm->free_area_cache = addr;
1368                 mm->cached_hole_size = ~0UL;
1369         }
1370 }
1371
1372 /*
1373  * This mmap-allocator allocates new areas top-down from below the
1374  * stack's low limit (the base):
1375  */
1376 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1377 unsigned long
1378 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1379                           const unsigned long len, const unsigned long pgoff,
1380                           const unsigned long flags)
1381 {
1382         struct vm_area_struct *vma;
1383         struct mm_struct *mm = current->mm;
1384         unsigned long addr = addr0;
1385
1386         /* requested length too big for entire address space */
1387         if (len > TASK_SIZE)
1388                 return -ENOMEM;
1389
1390         if (flags & MAP_FIXED)
1391                 return addr;
1392
1393         /* requesting a specific address */
1394         if (addr) {
1395                 addr = PAGE_ALIGN(addr);
1396                 vma = find_vma(mm, addr);
1397                 if (TASK_SIZE - len >= addr &&
1398                                 (!vma || addr + len <= vma->vm_start))
1399                         return addr;
1400         }
1401
1402         /* check if free_area_cache is useful for us */
1403         if (len <= mm->cached_hole_size) {
1404                 mm->cached_hole_size = 0;
1405                 mm->free_area_cache = mm->mmap_base;
1406         }
1407
1408         /* either no address requested or can't fit in requested address hole */
1409         addr = mm->free_area_cache;
1410
1411         /* make sure it can fit in the remaining address space */
1412         if (addr > len) {
1413                 vma = find_vma(mm, addr-len);
1414                 if (!vma || addr <= vma->vm_start)
1415                         /* remember the address as a hint for next time */
1416                         return (mm->free_area_cache = addr-len);
1417         }
1418
1419         if (mm->mmap_base < len)
1420                 goto bottomup;
1421
1422         addr = mm->mmap_base-len;
1423
1424         do {
1425                 /*
1426                  * Lookup failure means no vma is above this address,
1427                  * else if new region fits below vma->vm_start,
1428                  * return with success:
1429                  */
1430                 vma = find_vma(mm, addr);
1431                 if (!vma || addr+len <= vma->vm_start)
1432                         /* remember the address as a hint for next time */
1433                         return (mm->free_area_cache = addr);
1434
1435                 /* remember the largest hole we saw so far */
1436                 if (addr + mm->cached_hole_size < vma->vm_start)
1437                         mm->cached_hole_size = vma->vm_start - addr;
1438
1439                 /* try just below the current vma->vm_start */
1440                 addr = vma->vm_start-len;
1441         } while (len < vma->vm_start);
1442
1443 bottomup:
1444         /*
1445          * A failed mmap() very likely causes application failure,
1446          * so fall back to the bottom-up function here. This scenario
1447          * can happen with large stack limits and large mmap()
1448          * allocations.
1449          */
1450         mm->cached_hole_size = ~0UL;
1451         mm->free_area_cache = TASK_UNMAPPED_BASE;
1452         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1453         /*
1454          * Restore the topdown base:
1455          */
1456         mm->free_area_cache = mm->mmap_base;
1457         mm->cached_hole_size = ~0UL;
1458
1459         return addr;
1460 }
1461 #endif
1462
1463 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1464 {
1465         /*
1466          * Is this a new hole at the highest possible address?
1467          */
1468         if (addr > mm->free_area_cache)
1469                 mm->free_area_cache = addr;
1470
1471         /* dont allow allocations above current base */
1472         if (mm->free_area_cache > mm->mmap_base)
1473                 mm->free_area_cache = mm->mmap_base;
1474 }
1475
1476 unsigned long
1477 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1478                 unsigned long pgoff, unsigned long flags)
1479 {
1480         unsigned long (*get_area)(struct file *, unsigned long,
1481                                   unsigned long, unsigned long, unsigned long);
1482
1483         unsigned long error = arch_mmap_check(addr, len, flags);
1484         if (error)
1485                 return error;
1486
1487         /* Careful about overflows.. */
1488         if (len > TASK_SIZE)
1489                 return -ENOMEM;
1490
1491         get_area = current->mm->get_unmapped_area;
1492         if (file && file->f_op && file->f_op->get_unmapped_area)
1493                 get_area = file->f_op->get_unmapped_area;
1494         addr = get_area(file, addr, len, pgoff, flags);
1495         if (IS_ERR_VALUE(addr))
1496                 return addr;
1497
1498         if (addr > TASK_SIZE - len)
1499                 return -ENOMEM;
1500         if (addr & ~PAGE_MASK)
1501                 return -EINVAL;
1502
1503         return arch_rebalance_pgtables(addr, len);
1504 }
1505
1506 EXPORT_SYMBOL(get_unmapped_area);
1507
1508 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1509 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1510 {
1511         struct vm_area_struct *vma = NULL;
1512
1513         if (mm) {
1514                 /* Check the cache first. */
1515                 /* (Cache hit rate is typically around 35%.) */
1516                 vma = mm->mmap_cache;
1517                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1518                         struct rb_node * rb_node;
1519
1520                         rb_node = mm->mm_rb.rb_node;
1521                         vma = NULL;
1522
1523                         while (rb_node) {
1524                                 struct vm_area_struct * vma_tmp;
1525
1526                                 vma_tmp = rb_entry(rb_node,
1527                                                 struct vm_area_struct, vm_rb);
1528
1529                                 if (vma_tmp->vm_end > addr) {
1530                                         vma = vma_tmp;
1531                                         if (vma_tmp->vm_start <= addr)
1532                                                 break;
1533                                         rb_node = rb_node->rb_left;
1534                                 } else
1535                                         rb_node = rb_node->rb_right;
1536                         }
1537                         if (vma)
1538                                 mm->mmap_cache = vma;
1539                 }
1540         }
1541         return vma;
1542 }
1543
1544 EXPORT_SYMBOL(find_vma);
1545
1546 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1547 struct vm_area_struct *
1548 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1549                         struct vm_area_struct **pprev)
1550 {
1551         struct vm_area_struct *vma = NULL, *prev = NULL;
1552         struct rb_node *rb_node;
1553         if (!mm)
1554                 goto out;
1555
1556         /* Guard against addr being lower than the first VMA */
1557         vma = mm->mmap;
1558
1559         /* Go through the RB tree quickly. */
1560         rb_node = mm->mm_rb.rb_node;
1561
1562         while (rb_node) {
1563                 struct vm_area_struct *vma_tmp;
1564                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1565
1566                 if (addr < vma_tmp->vm_end) {
1567                         rb_node = rb_node->rb_left;
1568                 } else {
1569                         prev = vma_tmp;
1570                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1571                                 break;
1572                         rb_node = rb_node->rb_right;
1573                 }
1574         }
1575
1576 out:
1577         *pprev = prev;
1578         return prev ? prev->vm_next : vma;
1579 }
1580
1581 /*
1582  * Verify that the stack growth is acceptable and
1583  * update accounting. This is shared with both the
1584  * grow-up and grow-down cases.
1585  */
1586 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1587 {
1588         struct mm_struct *mm = vma->vm_mm;
1589         struct rlimit *rlim = current->signal->rlim;
1590         unsigned long new_start;
1591
1592         /* address space limit tests */
1593         if (!may_expand_vm(mm, grow))
1594                 return -ENOMEM;
1595
1596         /* Stack limit test */
1597         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1598                 return -ENOMEM;
1599
1600         /* mlock limit tests */
1601         if (vma->vm_flags & VM_LOCKED) {
1602                 unsigned long locked;
1603                 unsigned long limit;
1604                 locked = mm->locked_vm + grow;
1605                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1606                 limit >>= PAGE_SHIFT;
1607                 if (locked > limit && !capable(CAP_IPC_LOCK))
1608                         return -ENOMEM;
1609         }
1610
1611         /* Check to ensure the stack will not grow into a hugetlb-only region */
1612         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1613                         vma->vm_end - size;
1614         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1615                 return -EFAULT;
1616
1617         /*
1618          * Overcommit..  This must be the final test, as it will
1619          * update security statistics.
1620          */
1621         if (security_vm_enough_memory_mm(mm, grow))
1622                 return -ENOMEM;
1623
1624         /* Ok, everything looks good - let it rip */
1625         mm->total_vm += grow;
1626         if (vma->vm_flags & VM_LOCKED)
1627                 mm->locked_vm += grow;
1628         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1629         return 0;
1630 }
1631
1632 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1633 /*
1634  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1635  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1636  */
1637 #ifndef CONFIG_IA64
1638 static
1639 #endif
1640 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1641 {
1642         int error;
1643
1644         if (!(vma->vm_flags & VM_GROWSUP))
1645                 return -EFAULT;
1646
1647         /*
1648          * We must make sure the anon_vma is allocated
1649          * so that the anon_vma locking is not a noop.
1650          */
1651         if (unlikely(anon_vma_prepare(vma)))
1652                 return -ENOMEM;
1653         anon_vma_lock(vma);
1654
1655         /*
1656          * vma->vm_start/vm_end cannot change under us because the caller
1657          * is required to hold the mmap_sem in read mode.  We need the
1658          * anon_vma lock to serialize against concurrent expand_stacks.
1659          * Also guard against wrapping around to address 0.
1660          */
1661         if (address < PAGE_ALIGN(address+4))
1662                 address = PAGE_ALIGN(address+4);
1663         else {
1664                 anon_vma_unlock(vma);
1665                 return -ENOMEM;
1666         }
1667         error = 0;
1668
1669         /* Somebody else might have raced and expanded it already */
1670         if (address > vma->vm_end) {
1671                 unsigned long size, grow;
1672
1673                 size = address - vma->vm_start;
1674                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1675
1676                 error = acct_stack_growth(vma, size, grow);
1677                 if (!error)
1678                         vma->vm_end = address;
1679         }
1680         anon_vma_unlock(vma);
1681         return error;
1682 }
1683 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1684
1685 /*
1686  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1687  */
1688 static int expand_downwards(struct vm_area_struct *vma,
1689                                    unsigned long address)
1690 {
1691         int error;
1692
1693         /*
1694          * We must make sure the anon_vma is allocated
1695          * so that the anon_vma locking is not a noop.
1696          */
1697         if (unlikely(anon_vma_prepare(vma)))
1698                 return -ENOMEM;
1699
1700         address &= PAGE_MASK;
1701         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1702         if (error)
1703                 return error;
1704
1705         anon_vma_lock(vma);
1706
1707         /*
1708          * vma->vm_start/vm_end cannot change under us because the caller
1709          * is required to hold the mmap_sem in read mode.  We need the
1710          * anon_vma lock to serialize against concurrent expand_stacks.
1711          */
1712
1713         /* Somebody else might have raced and expanded it already */
1714         if (address < vma->vm_start) {
1715                 unsigned long size, grow;
1716
1717                 size = vma->vm_end - address;
1718                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1719
1720                 error = acct_stack_growth(vma, size, grow);
1721                 if (!error) {
1722                         vma->vm_start = address;
1723                         vma->vm_pgoff -= grow;
1724                 }
1725         }
1726         anon_vma_unlock(vma);
1727         return error;
1728 }
1729
1730 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1731 {
1732         return expand_downwards(vma, address);
1733 }
1734
1735 #ifdef CONFIG_STACK_GROWSUP
1736 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1737 {
1738         return expand_upwards(vma, address);
1739 }
1740
1741 struct vm_area_struct *
1742 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1743 {
1744         struct vm_area_struct *vma, *prev;
1745
1746         addr &= PAGE_MASK;
1747         vma = find_vma_prev(mm, addr, &prev);
1748         if (vma && (vma->vm_start <= addr))
1749                 return vma;
1750         if (!prev || expand_stack(prev, addr))
1751                 return NULL;
1752         if (prev->vm_flags & VM_LOCKED) {
1753                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1754         }
1755         return prev;
1756 }
1757 #else
1758 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1759 {
1760         return expand_downwards(vma, address);
1761 }
1762
1763 struct vm_area_struct *
1764 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1765 {
1766         struct vm_area_struct * vma;
1767         unsigned long start;
1768
1769         addr &= PAGE_MASK;
1770         vma = find_vma(mm,addr);
1771         if (!vma)
1772                 return NULL;
1773         if (vma->vm_start <= addr)
1774                 return vma;
1775         if (!(vma->vm_flags & VM_GROWSDOWN))
1776                 return NULL;
1777         start = vma->vm_start;
1778         if (expand_stack(vma, addr))
1779                 return NULL;
1780         if (vma->vm_flags & VM_LOCKED) {
1781                 mlock_vma_pages_range(vma, addr, start);
1782         }
1783         return vma;
1784 }
1785 #endif
1786
1787 /*
1788  * Ok - we have the memory areas we should free on the vma list,
1789  * so release them, and do the vma updates.
1790  *
1791  * Called with the mm semaphore held.
1792  */
1793 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1794 {
1795         /* Update high watermark before we lower total_vm */
1796         update_hiwater_vm(mm);
1797         do {
1798                 long nrpages = vma_pages(vma);
1799
1800                 mm->total_vm -= nrpages;
1801                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1802                 vma = remove_vma(vma);
1803         } while (vma);
1804         validate_mm(mm);
1805 }
1806
1807 /*
1808  * Get rid of page table information in the indicated region.
1809  *
1810  * Called with the mm semaphore held.
1811  */
1812 static void unmap_region(struct mm_struct *mm,
1813                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1814                 unsigned long start, unsigned long end)
1815 {
1816         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1817         struct mmu_gather *tlb;
1818         unsigned long nr_accounted = 0;
1819
1820         lru_add_drain();
1821         tlb = tlb_gather_mmu(mm, 0);
1822         update_hiwater_rss(mm);
1823         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1824         vm_unacct_memory(nr_accounted);
1825         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1826                                  next? next->vm_start: 0);
1827         tlb_finish_mmu(tlb, start, end);
1828 }
1829
1830 /*
1831  * Create a list of vma's touched by the unmap, removing them from the mm's
1832  * vma list as we go..
1833  */
1834 static void
1835 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1836         struct vm_area_struct *prev, unsigned long end)
1837 {
1838         struct vm_area_struct **insertion_point;
1839         struct vm_area_struct *tail_vma = NULL;
1840         unsigned long addr;
1841
1842         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1843         do {
1844                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1845                 mm->map_count--;
1846                 tail_vma = vma;
1847                 vma = vma->vm_next;
1848         } while (vma && vma->vm_start < end);
1849         *insertion_point = vma;
1850         tail_vma->vm_next = NULL;
1851         if (mm->unmap_area == arch_unmap_area)
1852                 addr = prev ? prev->vm_end : mm->mmap_base;
1853         else
1854                 addr = vma ?  vma->vm_start : mm->mmap_base;
1855         mm->unmap_area(mm, addr);
1856         mm->mmap_cache = NULL;          /* Kill the cache. */
1857 }
1858
1859 /*
1860  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1861  * munmap path where it doesn't make sense to fail.
1862  */
1863 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1864               unsigned long addr, int new_below)
1865 {
1866         struct mempolicy *pol;
1867         struct vm_area_struct *new;
1868
1869         if (is_vm_hugetlb_page(vma) && (addr &
1870                                         ~(huge_page_mask(hstate_vma(vma)))))
1871                 return -EINVAL;
1872
1873         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1874         if (!new)
1875                 return -ENOMEM;
1876
1877         /* most fields are the same, copy all, and then fixup */
1878         *new = *vma;
1879
1880         if (new_below)
1881                 new->vm_end = addr;
1882         else {
1883                 new->vm_start = addr;
1884                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1885         }
1886
1887         pol = mpol_dup(vma_policy(vma));
1888         if (IS_ERR(pol)) {
1889                 kmem_cache_free(vm_area_cachep, new);
1890                 return PTR_ERR(pol);
1891         }
1892         vma_set_policy(new, pol);
1893
1894         if (new->vm_file) {
1895                 get_file(new->vm_file);
1896                 if (vma->vm_flags & VM_EXECUTABLE)
1897                         added_exe_file_vma(mm);
1898         }
1899
1900         if (new->vm_ops && new->vm_ops->open)
1901                 new->vm_ops->open(new);
1902
1903         if (new_below)
1904                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1905                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1906         else
1907                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1908
1909         return 0;
1910 }
1911
1912 /*
1913  * Split a vma into two pieces at address 'addr', a new vma is allocated
1914  * either for the first part or the tail.
1915  */
1916 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1917               unsigned long addr, int new_below)
1918 {
1919         if (mm->map_count >= sysctl_max_map_count)
1920                 return -ENOMEM;
1921
1922         return __split_vma(mm, vma, addr, new_below);
1923 }
1924
1925 /* Munmap is split into 2 main parts -- this part which finds
1926  * what needs doing, and the areas themselves, which do the
1927  * work.  This now handles partial unmappings.
1928  * Jeremy Fitzhardinge <jeremy@goop.org>
1929  */
1930 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1931 {
1932         unsigned long end;
1933         struct vm_area_struct *vma, *prev, *last;
1934
1935         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1936                 return -EINVAL;
1937
1938         if ((len = PAGE_ALIGN(len)) == 0)
1939                 return -EINVAL;
1940
1941         /* Find the first overlapping VMA */
1942         vma = find_vma_prev(mm, start, &prev);
1943         if (!vma)
1944                 return 0;
1945         /* we have  start < vma->vm_end  */
1946
1947         /* if it doesn't overlap, we have nothing.. */
1948         end = start + len;
1949         if (vma->vm_start >= end)
1950                 return 0;
1951
1952         /*
1953          * If we need to split any vma, do it now to save pain later.
1954          *
1955          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1956          * unmapped vm_area_struct will remain in use: so lower split_vma
1957          * places tmp vma above, and higher split_vma places tmp vma below.
1958          */
1959         if (start > vma->vm_start) {
1960                 int error;
1961
1962                 /*
1963                  * Make sure that map_count on return from munmap() will
1964                  * not exceed its limit; but let map_count go just above
1965                  * its limit temporarily, to help free resources as expected.
1966                  */
1967                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
1968                         return -ENOMEM;
1969
1970                 error = __split_vma(mm, vma, start, 0);
1971                 if (error)
1972                         return error;
1973                 prev = vma;
1974         }
1975
1976         /* Does it split the last one? */
1977         last = find_vma(mm, end);
1978         if (last && end > last->vm_start) {
1979                 int error = __split_vma(mm, last, end, 1);
1980                 if (error)
1981                         return error;
1982         }
1983         vma = prev? prev->vm_next: mm->mmap;
1984
1985         /*
1986          * unlock any mlock()ed ranges before detaching vmas
1987          */
1988         if (mm->locked_vm) {
1989                 struct vm_area_struct *tmp = vma;
1990                 while (tmp && tmp->vm_start < end) {
1991                         if (tmp->vm_flags & VM_LOCKED) {
1992                                 mm->locked_vm -= vma_pages(tmp);
1993                                 munlock_vma_pages_all(tmp);
1994                         }
1995                         tmp = tmp->vm_next;
1996                 }
1997         }
1998
1999         /*
2000          * Remove the vma's, and unmap the actual pages
2001          */
2002         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2003         unmap_region(mm, vma, prev, start, end);
2004
2005         /* Fix up all other VM information */
2006         remove_vma_list(mm, vma);
2007
2008         return 0;
2009 }
2010
2011 EXPORT_SYMBOL(do_munmap);
2012
2013 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2014 {
2015         int ret;
2016         struct mm_struct *mm = current->mm;
2017
2018         profile_munmap(addr);
2019
2020         down_write(&mm->mmap_sem);
2021         ret = do_munmap(mm, addr, len);
2022         up_write(&mm->mmap_sem);
2023         return ret;
2024 }
2025
2026 static inline void verify_mm_writelocked(struct mm_struct *mm)
2027 {
2028 #ifdef CONFIG_DEBUG_VM
2029         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2030                 WARN_ON(1);
2031                 up_read(&mm->mmap_sem);
2032         }
2033 #endif
2034 }
2035
2036 /*
2037  *  this is really a simplified "do_mmap".  it only handles
2038  *  anonymous maps.  eventually we may be able to do some
2039  *  brk-specific accounting here.
2040  */
2041 unsigned long do_brk(unsigned long addr, unsigned long len)
2042 {
2043         struct mm_struct * mm = current->mm;
2044         struct vm_area_struct * vma, * prev;
2045         unsigned long flags;
2046         struct rb_node ** rb_link, * rb_parent;
2047         pgoff_t pgoff = addr >> PAGE_SHIFT;
2048         int error;
2049
2050         len = PAGE_ALIGN(len);
2051         if (!len)
2052                 return addr;
2053
2054         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2055         if (error)
2056                 return error;
2057
2058         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2059
2060         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2061         if (error & ~PAGE_MASK)
2062                 return error;
2063
2064         /*
2065          * mlock MCL_FUTURE?
2066          */
2067         if (mm->def_flags & VM_LOCKED) {
2068                 unsigned long locked, lock_limit;
2069                 locked = len >> PAGE_SHIFT;
2070                 locked += mm->locked_vm;
2071                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2072                 lock_limit >>= PAGE_SHIFT;
2073                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2074                         return -EAGAIN;
2075         }
2076
2077         /*
2078          * mm->mmap_sem is required to protect against another thread
2079          * changing the mappings in case we sleep.
2080          */
2081         verify_mm_writelocked(mm);
2082
2083         /*
2084          * Clear old maps.  this also does some error checking for us
2085          */
2086  munmap_back:
2087         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2088         if (vma && vma->vm_start < addr + len) {
2089                 if (do_munmap(mm, addr, len))
2090                         return -ENOMEM;
2091                 goto munmap_back;
2092         }
2093
2094         /* Check against address space limits *after* clearing old maps... */
2095         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2096                 return -ENOMEM;
2097
2098         if (mm->map_count > sysctl_max_map_count)
2099                 return -ENOMEM;
2100
2101         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2102                 return -ENOMEM;
2103
2104         /* Can we just expand an old private anonymous mapping? */
2105         vma = vma_merge(mm, prev, addr, addr + len, flags,
2106                                         NULL, NULL, pgoff, NULL);
2107         if (vma)
2108                 goto out;
2109
2110         /*
2111          * create a vma struct for an anonymous mapping
2112          */
2113         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2114         if (!vma) {
2115                 vm_unacct_memory(len >> PAGE_SHIFT);
2116                 return -ENOMEM;
2117         }
2118
2119         vma->vm_mm = mm;
2120         vma->vm_start = addr;
2121         vma->vm_end = addr + len;
2122         vma->vm_pgoff = pgoff;
2123         vma->vm_flags = flags;
2124         vma->vm_page_prot = vm_get_page_prot(flags);
2125         vma_link(mm, vma, prev, rb_link, rb_parent);
2126 out:
2127         mm->total_vm += len >> PAGE_SHIFT;
2128         if (flags & VM_LOCKED) {
2129                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2130                         mm->locked_vm += (len >> PAGE_SHIFT);
2131         }
2132         return addr;
2133 }
2134
2135 EXPORT_SYMBOL(do_brk);
2136
2137 /* Release all mmaps. */
2138 void exit_mmap(struct mm_struct *mm)
2139 {
2140         struct mmu_gather *tlb;
2141         struct vm_area_struct *vma;
2142         unsigned long nr_accounted = 0;
2143         unsigned long end;
2144
2145         /* mm's last user has gone, and its about to be pulled down */
2146         mmu_notifier_release(mm);
2147
2148         if (mm->locked_vm) {
2149                 vma = mm->mmap;
2150                 while (vma) {
2151                         if (vma->vm_flags & VM_LOCKED)
2152                                 munlock_vma_pages_all(vma);
2153                         vma = vma->vm_next;
2154                 }
2155         }
2156
2157         arch_exit_mmap(mm);
2158
2159         vma = mm->mmap;
2160         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2161                 return;
2162
2163         lru_add_drain();
2164         flush_cache_mm(mm);
2165         tlb = tlb_gather_mmu(mm, 1);
2166         /* update_hiwater_rss(mm) here? but nobody should be looking */
2167         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2168         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2169         vm_unacct_memory(nr_accounted);
2170
2171         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2172         tlb_finish_mmu(tlb, 0, end);
2173
2174         /*
2175          * Walk the list again, actually closing and freeing it,
2176          * with preemption enabled, without holding any MM locks.
2177          */
2178         while (vma)
2179                 vma = remove_vma(vma);
2180
2181         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2182 }
2183
2184 /* Insert vm structure into process list sorted by address
2185  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2186  * then i_mmap_lock is taken here.
2187  */
2188 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2189 {
2190         struct vm_area_struct * __vma, * prev;
2191         struct rb_node ** rb_link, * rb_parent;
2192
2193         /*
2194          * The vm_pgoff of a purely anonymous vma should be irrelevant
2195          * until its first write fault, when page's anon_vma and index
2196          * are set.  But now set the vm_pgoff it will almost certainly
2197          * end up with (unless mremap moves it elsewhere before that
2198          * first wfault), so /proc/pid/maps tells a consistent story.
2199          *
2200          * By setting it to reflect the virtual start address of the
2201          * vma, merges and splits can happen in a seamless way, just
2202          * using the existing file pgoff checks and manipulations.
2203          * Similarly in do_mmap_pgoff and in do_brk.
2204          */
2205         if (!vma->vm_file) {
2206                 BUG_ON(vma->anon_vma);
2207                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2208         }
2209         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2210         if (__vma && __vma->vm_start < vma->vm_end)
2211                 return -ENOMEM;
2212         if ((vma->vm_flags & VM_ACCOUNT) &&
2213              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2214                 return -ENOMEM;
2215         vma_link(mm, vma, prev, rb_link, rb_parent);
2216         return 0;
2217 }
2218
2219 /*
2220  * Copy the vma structure to a new location in the same mm,
2221  * prior to moving page table entries, to effect an mremap move.
2222  */
2223 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2224         unsigned long addr, unsigned long len, pgoff_t pgoff)
2225 {
2226         struct vm_area_struct *vma = *vmap;
2227         unsigned long vma_start = vma->vm_start;
2228         struct mm_struct *mm = vma->vm_mm;
2229         struct vm_area_struct *new_vma, *prev;
2230         struct rb_node **rb_link, *rb_parent;
2231         struct mempolicy *pol;
2232
2233         /*
2234          * If anonymous vma has not yet been faulted, update new pgoff
2235          * to match new location, to increase its chance of merging.
2236          */
2237         if (!vma->vm_file && !vma->anon_vma)
2238                 pgoff = addr >> PAGE_SHIFT;
2239
2240         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2241         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2242                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2243         if (new_vma) {
2244                 /*
2245                  * Source vma may have been merged into new_vma
2246                  */
2247                 if (vma_start >= new_vma->vm_start &&
2248                     vma_start < new_vma->vm_end)
2249                         *vmap = new_vma;
2250         } else {
2251                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2252                 if (new_vma) {
2253                         *new_vma = *vma;
2254                         pol = mpol_dup(vma_policy(vma));
2255                         if (IS_ERR(pol)) {
2256                                 kmem_cache_free(vm_area_cachep, new_vma);
2257                                 return NULL;
2258                         }
2259                         vma_set_policy(new_vma, pol);
2260                         new_vma->vm_start = addr;
2261                         new_vma->vm_end = addr + len;
2262                         new_vma->vm_pgoff = pgoff;
2263                         if (new_vma->vm_file) {
2264                                 get_file(new_vma->vm_file);
2265                                 if (vma->vm_flags & VM_EXECUTABLE)
2266                                         added_exe_file_vma(mm);
2267                         }
2268                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2269                                 new_vma->vm_ops->open(new_vma);
2270                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2271                 }
2272         }
2273         return new_vma;
2274 }
2275
2276 /*
2277  * Return true if the calling process may expand its vm space by the passed
2278  * number of pages
2279  */
2280 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2281 {
2282         unsigned long cur = mm->total_vm;       /* pages */
2283         unsigned long lim;
2284
2285         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2286
2287         if (cur + npages > lim)
2288                 return 0;
2289         return 1;
2290 }
2291
2292
2293 static int special_mapping_fault(struct vm_area_struct *vma,
2294                                 struct vm_fault *vmf)
2295 {
2296         pgoff_t pgoff;
2297         struct page **pages;
2298
2299         /*
2300          * special mappings have no vm_file, and in that case, the mm
2301          * uses vm_pgoff internally. So we have to subtract it from here.
2302          * We are allowed to do this because we are the mm; do not copy
2303          * this code into drivers!
2304          */
2305         pgoff = vmf->pgoff - vma->vm_pgoff;
2306
2307         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2308                 pgoff--;
2309
2310         if (*pages) {
2311                 struct page *page = *pages;
2312                 get_page(page);
2313                 vmf->page = page;
2314                 return 0;
2315         }
2316
2317         return VM_FAULT_SIGBUS;
2318 }
2319
2320 /*
2321  * Having a close hook prevents vma merging regardless of flags.
2322  */
2323 static void special_mapping_close(struct vm_area_struct *vma)
2324 {
2325 }
2326
2327 static const struct vm_operations_struct special_mapping_vmops = {
2328         .close = special_mapping_close,
2329         .fault = special_mapping_fault,
2330 };
2331
2332 /*
2333  * Called with mm->mmap_sem held for writing.
2334  * Insert a new vma covering the given region, with the given flags.
2335  * Its pages are supplied by the given array of struct page *.
2336  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2337  * The region past the last page supplied will always produce SIGBUS.
2338  * The array pointer and the pages it points to are assumed to stay alive
2339  * for as long as this mapping might exist.
2340  */
2341 int install_special_mapping(struct mm_struct *mm,
2342                             unsigned long addr, unsigned long len,
2343                             unsigned long vm_flags, struct page **pages)
2344 {
2345         struct vm_area_struct *vma;
2346
2347         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2348         if (unlikely(vma == NULL))
2349                 return -ENOMEM;
2350
2351         vma->vm_mm = mm;
2352         vma->vm_start = addr;
2353         vma->vm_end = addr + len;
2354
2355         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2356         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2357
2358         vma->vm_ops = &special_mapping_vmops;
2359         vma->vm_private_data = pages;
2360
2361         if (unlikely(insert_vm_struct(mm, vma))) {
2362                 kmem_cache_free(vm_area_cachep, vma);
2363                 return -ENOMEM;
2364         }
2365
2366         mm->total_vm += len >> PAGE_SHIFT;
2367
2368         perf_event_mmap(vma);
2369
2370         return 0;
2371 }
2372
2373 static DEFINE_MUTEX(mm_all_locks_mutex);
2374
2375 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2376 {
2377         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2378                 /*
2379                  * The LSB of head.next can't change from under us
2380                  * because we hold the mm_all_locks_mutex.
2381                  */
2382                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2383                 /*
2384                  * We can safely modify head.next after taking the
2385                  * anon_vma->lock. If some other vma in this mm shares
2386                  * the same anon_vma we won't take it again.
2387                  *
2388                  * No need of atomic instructions here, head.next
2389                  * can't change from under us thanks to the
2390                  * anon_vma->lock.
2391                  */
2392                 if (__test_and_set_bit(0, (unsigned long *)
2393                                        &anon_vma->head.next))
2394                         BUG();
2395         }
2396 }
2397
2398 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2399 {
2400         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2401                 /*
2402                  * AS_MM_ALL_LOCKS can't change from under us because
2403                  * we hold the mm_all_locks_mutex.
2404                  *
2405                  * Operations on ->flags have to be atomic because
2406                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2407                  * mm_all_locks_mutex, there may be other cpus
2408                  * changing other bitflags in parallel to us.
2409                  */
2410                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2411                         BUG();
2412                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2413         }
2414 }
2415
2416 /*
2417  * This operation locks against the VM for all pte/vma/mm related
2418  * operations that could ever happen on a certain mm. This includes
2419  * vmtruncate, try_to_unmap, and all page faults.
2420  *
2421  * The caller must take the mmap_sem in write mode before calling
2422  * mm_take_all_locks(). The caller isn't allowed to release the
2423  * mmap_sem until mm_drop_all_locks() returns.
2424  *
2425  * mmap_sem in write mode is required in order to block all operations
2426  * that could modify pagetables and free pages without need of
2427  * altering the vma layout (for example populate_range() with
2428  * nonlinear vmas). It's also needed in write mode to avoid new
2429  * anon_vmas to be associated with existing vmas.
2430  *
2431  * A single task can't take more than one mm_take_all_locks() in a row
2432  * or it would deadlock.
2433  *
2434  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2435  * mapping->flags avoid to take the same lock twice, if more than one
2436  * vma in this mm is backed by the same anon_vma or address_space.
2437  *
2438  * We can take all the locks in random order because the VM code
2439  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2440  * takes more than one of them in a row. Secondly we're protected
2441  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2442  *
2443  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2444  * that may have to take thousand of locks.
2445  *
2446  * mm_take_all_locks() can fail if it's interrupted by signals.
2447  */
2448 int mm_take_all_locks(struct mm_struct *mm)
2449 {
2450         struct vm_area_struct *vma;
2451         int ret = -EINTR;
2452
2453         BUG_ON(down_read_trylock(&mm->mmap_sem));
2454
2455         mutex_lock(&mm_all_locks_mutex);
2456
2457         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2458                 if (signal_pending(current))
2459                         goto out_unlock;
2460                 if (vma->vm_file && vma->vm_file->f_mapping)
2461                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2462         }
2463
2464         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2465                 if (signal_pending(current))
2466                         goto out_unlock;
2467                 if (vma->anon_vma)
2468                         vm_lock_anon_vma(mm, vma->anon_vma);
2469         }
2470
2471         ret = 0;
2472
2473 out_unlock:
2474         if (ret)
2475                 mm_drop_all_locks(mm);
2476
2477         return ret;
2478 }
2479
2480 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2481 {
2482         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2483                 /*
2484                  * The LSB of head.next can't change to 0 from under
2485                  * us because we hold the mm_all_locks_mutex.
2486                  *
2487                  * We must however clear the bitflag before unlocking
2488                  * the vma so the users using the anon_vma->head will
2489                  * never see our bitflag.
2490                  *
2491                  * No need of atomic instructions here, head.next
2492                  * can't change from under us until we release the
2493                  * anon_vma->lock.
2494                  */
2495                 if (!__test_and_clear_bit(0, (unsigned long *)
2496                                           &anon_vma->head.next))
2497                         BUG();
2498                 spin_unlock(&anon_vma->lock);
2499         }
2500 }
2501
2502 static void vm_unlock_mapping(struct address_space *mapping)
2503 {
2504         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2505                 /*
2506                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2507                  * because we hold the mm_all_locks_mutex.
2508                  */
2509                 spin_unlock(&mapping->i_mmap_lock);
2510                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2511                                         &mapping->flags))
2512                         BUG();
2513         }
2514 }
2515
2516 /*
2517  * The mmap_sem cannot be released by the caller until
2518  * mm_drop_all_locks() returns.
2519  */
2520 void mm_drop_all_locks(struct mm_struct *mm)
2521 {
2522         struct vm_area_struct *vma;
2523
2524         BUG_ON(down_read_trylock(&mm->mmap_sem));
2525         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2526
2527         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2528                 if (vma->anon_vma)
2529                         vm_unlock_anon_vma(vma->anon_vma);
2530                 if (vma->vm_file && vma->vm_file->f_mapping)
2531                         vm_unlock_mapping(vma->vm_file->f_mapping);
2532         }
2533
2534         mutex_unlock(&mm_all_locks_mutex);
2535 }
2536
2537 /*
2538  * initialise the VMA slab
2539  */
2540 void __init mmap_init(void)
2541 {
2542         int ret;
2543
2544         ret = percpu_counter_init(&vm_committed_as, 0);
2545         VM_BUG_ON(ret);
2546 }