Merge branches 'slab/cleanups', 'slab/failslab', 'slab/fixes' and 'slub/percpu' into...
[linux-3.10.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 = current->signal->rlim[RLIMIT_DATA].rlim_cur;
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 = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
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                 /*
1269                  * makes pages present; downgrades, drops, reacquires mmap_sem
1270                  */
1271                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1272                 if (nr_pages < 0)
1273                         return nr_pages;        /* vma gone! */
1274                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1275         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1276                 make_pages_present(addr, addr + len);
1277         return addr;
1278
1279 unmap_and_free_vma:
1280         if (correct_wcount)
1281                 atomic_inc(&inode->i_writecount);
1282         vma->vm_file = NULL;
1283         fput(file);
1284
1285         /* Undo any partial mapping done by a device driver. */
1286         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1287         charged = 0;
1288 free_vma:
1289         kmem_cache_free(vm_area_cachep, vma);
1290 unacct_error:
1291         if (charged)
1292                 vm_unacct_memory(charged);
1293         return error;
1294 }
1295
1296 /* Get an address range which is currently unmapped.
1297  * For shmat() with addr=0.
1298  *
1299  * Ugly calling convention alert:
1300  * Return value with the low bits set means error value,
1301  * ie
1302  *      if (ret & ~PAGE_MASK)
1303  *              error = ret;
1304  *
1305  * This function "knows" that -ENOMEM has the bits set.
1306  */
1307 #ifndef HAVE_ARCH_UNMAPPED_AREA
1308 unsigned long
1309 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1310                 unsigned long len, unsigned long pgoff, unsigned long flags)
1311 {
1312         struct mm_struct *mm = current->mm;
1313         struct vm_area_struct *vma;
1314         unsigned long start_addr;
1315
1316         if (len > TASK_SIZE)
1317                 return -ENOMEM;
1318
1319         if (flags & MAP_FIXED)
1320                 return addr;
1321
1322         if (addr) {
1323                 addr = PAGE_ALIGN(addr);
1324                 vma = find_vma(mm, addr);
1325                 if (TASK_SIZE - len >= addr &&
1326                     (!vma || addr + len <= vma->vm_start))
1327                         return addr;
1328         }
1329         if (len > mm->cached_hole_size) {
1330                 start_addr = addr = mm->free_area_cache;
1331         } else {
1332                 start_addr = addr = TASK_UNMAPPED_BASE;
1333                 mm->cached_hole_size = 0;
1334         }
1335
1336 full_search:
1337         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1338                 /* At this point:  (!vma || addr < vma->vm_end). */
1339                 if (TASK_SIZE - len < addr) {
1340                         /*
1341                          * Start a new search - just in case we missed
1342                          * some holes.
1343                          */
1344                         if (start_addr != TASK_UNMAPPED_BASE) {
1345                                 addr = TASK_UNMAPPED_BASE;
1346                                 start_addr = addr;
1347                                 mm->cached_hole_size = 0;
1348                                 goto full_search;
1349                         }
1350                         return -ENOMEM;
1351                 }
1352                 if (!vma || addr + len <= vma->vm_start) {
1353                         /*
1354                          * Remember the place where we stopped the search:
1355                          */
1356                         mm->free_area_cache = addr + len;
1357                         return addr;
1358                 }
1359                 if (addr + mm->cached_hole_size < vma->vm_start)
1360                         mm->cached_hole_size = vma->vm_start - addr;
1361                 addr = vma->vm_end;
1362         }
1363 }
1364 #endif  
1365
1366 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1367 {
1368         /*
1369          * Is this a new hole at the lowest possible address?
1370          */
1371         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1372                 mm->free_area_cache = addr;
1373                 mm->cached_hole_size = ~0UL;
1374         }
1375 }
1376
1377 /*
1378  * This mmap-allocator allocates new areas top-down from below the
1379  * stack's low limit (the base):
1380  */
1381 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1382 unsigned long
1383 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1384                           const unsigned long len, const unsigned long pgoff,
1385                           const unsigned long flags)
1386 {
1387         struct vm_area_struct *vma;
1388         struct mm_struct *mm = current->mm;
1389         unsigned long addr = addr0;
1390
1391         /* requested length too big for entire address space */
1392         if (len > TASK_SIZE)
1393                 return -ENOMEM;
1394
1395         if (flags & MAP_FIXED)
1396                 return addr;
1397
1398         /* requesting a specific address */
1399         if (addr) {
1400                 addr = PAGE_ALIGN(addr);
1401                 vma = find_vma(mm, addr);
1402                 if (TASK_SIZE - len >= addr &&
1403                                 (!vma || addr + len <= vma->vm_start))
1404                         return addr;
1405         }
1406
1407         /* check if free_area_cache is useful for us */
1408         if (len <= mm->cached_hole_size) {
1409                 mm->cached_hole_size = 0;
1410                 mm->free_area_cache = mm->mmap_base;
1411         }
1412
1413         /* either no address requested or can't fit in requested address hole */
1414         addr = mm->free_area_cache;
1415
1416         /* make sure it can fit in the remaining address space */
1417         if (addr > len) {
1418                 vma = find_vma(mm, addr-len);
1419                 if (!vma || addr <= vma->vm_start)
1420                         /* remember the address as a hint for next time */
1421                         return (mm->free_area_cache = addr-len);
1422         }
1423
1424         if (mm->mmap_base < len)
1425                 goto bottomup;
1426
1427         addr = mm->mmap_base-len;
1428
1429         do {
1430                 /*
1431                  * Lookup failure means no vma is above this address,
1432                  * else if new region fits below vma->vm_start,
1433                  * return with success:
1434                  */
1435                 vma = find_vma(mm, addr);
1436                 if (!vma || addr+len <= vma->vm_start)
1437                         /* remember the address as a hint for next time */
1438                         return (mm->free_area_cache = addr);
1439
1440                 /* remember the largest hole we saw so far */
1441                 if (addr + mm->cached_hole_size < vma->vm_start)
1442                         mm->cached_hole_size = vma->vm_start - addr;
1443
1444                 /* try just below the current vma->vm_start */
1445                 addr = vma->vm_start-len;
1446         } while (len < vma->vm_start);
1447
1448 bottomup:
1449         /*
1450          * A failed mmap() very likely causes application failure,
1451          * so fall back to the bottom-up function here. This scenario
1452          * can happen with large stack limits and large mmap()
1453          * allocations.
1454          */
1455         mm->cached_hole_size = ~0UL;
1456         mm->free_area_cache = TASK_UNMAPPED_BASE;
1457         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1458         /*
1459          * Restore the topdown base:
1460          */
1461         mm->free_area_cache = mm->mmap_base;
1462         mm->cached_hole_size = ~0UL;
1463
1464         return addr;
1465 }
1466 #endif
1467
1468 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1469 {
1470         /*
1471          * Is this a new hole at the highest possible address?
1472          */
1473         if (addr > mm->free_area_cache)
1474                 mm->free_area_cache = addr;
1475
1476         /* dont allow allocations above current base */
1477         if (mm->free_area_cache > mm->mmap_base)
1478                 mm->free_area_cache = mm->mmap_base;
1479 }
1480
1481 unsigned long
1482 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1483                 unsigned long pgoff, unsigned long flags)
1484 {
1485         unsigned long (*get_area)(struct file *, unsigned long,
1486                                   unsigned long, unsigned long, unsigned long);
1487
1488         unsigned long error = arch_mmap_check(addr, len, flags);
1489         if (error)
1490                 return error;
1491
1492         /* Careful about overflows.. */
1493         if (len > TASK_SIZE)
1494                 return -ENOMEM;
1495
1496         get_area = current->mm->get_unmapped_area;
1497         if (file && file->f_op && file->f_op->get_unmapped_area)
1498                 get_area = file->f_op->get_unmapped_area;
1499         addr = get_area(file, addr, len, pgoff, flags);
1500         if (IS_ERR_VALUE(addr))
1501                 return addr;
1502
1503         if (addr > TASK_SIZE - len)
1504                 return -ENOMEM;
1505         if (addr & ~PAGE_MASK)
1506                 return -EINVAL;
1507
1508         return arch_rebalance_pgtables(addr, len);
1509 }
1510
1511 EXPORT_SYMBOL(get_unmapped_area);
1512
1513 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1514 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1515 {
1516         struct vm_area_struct *vma = NULL;
1517
1518         if (mm) {
1519                 /* Check the cache first. */
1520                 /* (Cache hit rate is typically around 35%.) */
1521                 vma = mm->mmap_cache;
1522                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1523                         struct rb_node * rb_node;
1524
1525                         rb_node = mm->mm_rb.rb_node;
1526                         vma = NULL;
1527
1528                         while (rb_node) {
1529                                 struct vm_area_struct * vma_tmp;
1530
1531                                 vma_tmp = rb_entry(rb_node,
1532                                                 struct vm_area_struct, vm_rb);
1533
1534                                 if (vma_tmp->vm_end > addr) {
1535                                         vma = vma_tmp;
1536                                         if (vma_tmp->vm_start <= addr)
1537                                                 break;
1538                                         rb_node = rb_node->rb_left;
1539                                 } else
1540                                         rb_node = rb_node->rb_right;
1541                         }
1542                         if (vma)
1543                                 mm->mmap_cache = vma;
1544                 }
1545         }
1546         return vma;
1547 }
1548
1549 EXPORT_SYMBOL(find_vma);
1550
1551 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1552 struct vm_area_struct *
1553 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1554                         struct vm_area_struct **pprev)
1555 {
1556         struct vm_area_struct *vma = NULL, *prev = NULL;
1557         struct rb_node *rb_node;
1558         if (!mm)
1559                 goto out;
1560
1561         /* Guard against addr being lower than the first VMA */
1562         vma = mm->mmap;
1563
1564         /* Go through the RB tree quickly. */
1565         rb_node = mm->mm_rb.rb_node;
1566
1567         while (rb_node) {
1568                 struct vm_area_struct *vma_tmp;
1569                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1570
1571                 if (addr < vma_tmp->vm_end) {
1572                         rb_node = rb_node->rb_left;
1573                 } else {
1574                         prev = vma_tmp;
1575                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1576                                 break;
1577                         rb_node = rb_node->rb_right;
1578                 }
1579         }
1580
1581 out:
1582         *pprev = prev;
1583         return prev ? prev->vm_next : vma;
1584 }
1585
1586 /*
1587  * Verify that the stack growth is acceptable and
1588  * update accounting. This is shared with both the
1589  * grow-up and grow-down cases.
1590  */
1591 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1592 {
1593         struct mm_struct *mm = vma->vm_mm;
1594         struct rlimit *rlim = current->signal->rlim;
1595         unsigned long new_start;
1596
1597         /* address space limit tests */
1598         if (!may_expand_vm(mm, grow))
1599                 return -ENOMEM;
1600
1601         /* Stack limit test */
1602         if (size > rlim[RLIMIT_STACK].rlim_cur)
1603                 return -ENOMEM;
1604
1605         /* mlock limit tests */
1606         if (vma->vm_flags & VM_LOCKED) {
1607                 unsigned long locked;
1608                 unsigned long limit;
1609                 locked = mm->locked_vm + grow;
1610                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1611                 if (locked > limit && !capable(CAP_IPC_LOCK))
1612                         return -ENOMEM;
1613         }
1614
1615         /* Check to ensure the stack will not grow into a hugetlb-only region */
1616         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1617                         vma->vm_end - size;
1618         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1619                 return -EFAULT;
1620
1621         /*
1622          * Overcommit..  This must be the final test, as it will
1623          * update security statistics.
1624          */
1625         if (security_vm_enough_memory_mm(mm, grow))
1626                 return -ENOMEM;
1627
1628         /* Ok, everything looks good - let it rip */
1629         mm->total_vm += grow;
1630         if (vma->vm_flags & VM_LOCKED)
1631                 mm->locked_vm += grow;
1632         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1633         return 0;
1634 }
1635
1636 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1637 /*
1638  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1639  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1640  */
1641 #ifndef CONFIG_IA64
1642 static
1643 #endif
1644 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1645 {
1646         int error;
1647
1648         if (!(vma->vm_flags & VM_GROWSUP))
1649                 return -EFAULT;
1650
1651         /*
1652          * We must make sure the anon_vma is allocated
1653          * so that the anon_vma locking is not a noop.
1654          */
1655         if (unlikely(anon_vma_prepare(vma)))
1656                 return -ENOMEM;
1657         anon_vma_lock(vma);
1658
1659         /*
1660          * vma->vm_start/vm_end cannot change under us because the caller
1661          * is required to hold the mmap_sem in read mode.  We need the
1662          * anon_vma lock to serialize against concurrent expand_stacks.
1663          * Also guard against wrapping around to address 0.
1664          */
1665         if (address < PAGE_ALIGN(address+4))
1666                 address = PAGE_ALIGN(address+4);
1667         else {
1668                 anon_vma_unlock(vma);
1669                 return -ENOMEM;
1670         }
1671         error = 0;
1672
1673         /* Somebody else might have raced and expanded it already */
1674         if (address > vma->vm_end) {
1675                 unsigned long size, grow;
1676
1677                 size = address - vma->vm_start;
1678                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1679
1680                 error = acct_stack_growth(vma, size, grow);
1681                 if (!error)
1682                         vma->vm_end = address;
1683         }
1684         anon_vma_unlock(vma);
1685         return error;
1686 }
1687 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1688
1689 /*
1690  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1691  */
1692 static int expand_downwards(struct vm_area_struct *vma,
1693                                    unsigned long address)
1694 {
1695         int error;
1696
1697         /*
1698          * We must make sure the anon_vma is allocated
1699          * so that the anon_vma locking is not a noop.
1700          */
1701         if (unlikely(anon_vma_prepare(vma)))
1702                 return -ENOMEM;
1703
1704         address &= PAGE_MASK;
1705         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1706         if (error)
1707                 return error;
1708
1709         anon_vma_lock(vma);
1710
1711         /*
1712          * vma->vm_start/vm_end cannot change under us because the caller
1713          * is required to hold the mmap_sem in read mode.  We need the
1714          * anon_vma lock to serialize against concurrent expand_stacks.
1715          */
1716
1717         /* Somebody else might have raced and expanded it already */
1718         if (address < vma->vm_start) {
1719                 unsigned long size, grow;
1720
1721                 size = vma->vm_end - address;
1722                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1723
1724                 error = acct_stack_growth(vma, size, grow);
1725                 if (!error) {
1726                         vma->vm_start = address;
1727                         vma->vm_pgoff -= grow;
1728                 }
1729         }
1730         anon_vma_unlock(vma);
1731         return error;
1732 }
1733
1734 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1735 {
1736         return expand_downwards(vma, address);
1737 }
1738
1739 #ifdef CONFIG_STACK_GROWSUP
1740 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1741 {
1742         return expand_upwards(vma, address);
1743 }
1744
1745 struct vm_area_struct *
1746 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1747 {
1748         struct vm_area_struct *vma, *prev;
1749
1750         addr &= PAGE_MASK;
1751         vma = find_vma_prev(mm, addr, &prev);
1752         if (vma && (vma->vm_start <= addr))
1753                 return vma;
1754         if (!prev || expand_stack(prev, addr))
1755                 return NULL;
1756         if (prev->vm_flags & VM_LOCKED) {
1757                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1758                         return NULL;    /* vma gone! */
1759         }
1760         return prev;
1761 }
1762 #else
1763 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1764 {
1765         return expand_downwards(vma, address);
1766 }
1767
1768 struct vm_area_struct *
1769 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1770 {
1771         struct vm_area_struct * vma;
1772         unsigned long start;
1773
1774         addr &= PAGE_MASK;
1775         vma = find_vma(mm,addr);
1776         if (!vma)
1777                 return NULL;
1778         if (vma->vm_start <= addr)
1779                 return vma;
1780         if (!(vma->vm_flags & VM_GROWSDOWN))
1781                 return NULL;
1782         start = vma->vm_start;
1783         if (expand_stack(vma, addr))
1784                 return NULL;
1785         if (vma->vm_flags & VM_LOCKED) {
1786                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1787                         return NULL;    /* vma gone! */
1788         }
1789         return vma;
1790 }
1791 #endif
1792
1793 /*
1794  * Ok - we have the memory areas we should free on the vma list,
1795  * so release them, and do the vma updates.
1796  *
1797  * Called with the mm semaphore held.
1798  */
1799 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1800 {
1801         /* Update high watermark before we lower total_vm */
1802         update_hiwater_vm(mm);
1803         do {
1804                 long nrpages = vma_pages(vma);
1805
1806                 mm->total_vm -= nrpages;
1807                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1808                 vma = remove_vma(vma);
1809         } while (vma);
1810         validate_mm(mm);
1811 }
1812
1813 /*
1814  * Get rid of page table information in the indicated region.
1815  *
1816  * Called with the mm semaphore held.
1817  */
1818 static void unmap_region(struct mm_struct *mm,
1819                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1820                 unsigned long start, unsigned long end)
1821 {
1822         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1823         struct mmu_gather *tlb;
1824         unsigned long nr_accounted = 0;
1825
1826         lru_add_drain();
1827         tlb = tlb_gather_mmu(mm, 0);
1828         update_hiwater_rss(mm);
1829         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1830         vm_unacct_memory(nr_accounted);
1831         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1832                                  next? next->vm_start: 0);
1833         tlb_finish_mmu(tlb, start, end);
1834 }
1835
1836 /*
1837  * Create a list of vma's touched by the unmap, removing them from the mm's
1838  * vma list as we go..
1839  */
1840 static void
1841 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1842         struct vm_area_struct *prev, unsigned long end)
1843 {
1844         struct vm_area_struct **insertion_point;
1845         struct vm_area_struct *tail_vma = NULL;
1846         unsigned long addr;
1847
1848         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1849         do {
1850                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1851                 mm->map_count--;
1852                 tail_vma = vma;
1853                 vma = vma->vm_next;
1854         } while (vma && vma->vm_start < end);
1855         *insertion_point = vma;
1856         tail_vma->vm_next = NULL;
1857         if (mm->unmap_area == arch_unmap_area)
1858                 addr = prev ? prev->vm_end : mm->mmap_base;
1859         else
1860                 addr = vma ?  vma->vm_start : mm->mmap_base;
1861         mm->unmap_area(mm, addr);
1862         mm->mmap_cache = NULL;          /* Kill the cache. */
1863 }
1864
1865 /*
1866  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1867  * munmap path where it doesn't make sense to fail.
1868  */
1869 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1870               unsigned long addr, int new_below)
1871 {
1872         struct mempolicy *pol;
1873         struct vm_area_struct *new;
1874
1875         if (is_vm_hugetlb_page(vma) && (addr &
1876                                         ~(huge_page_mask(hstate_vma(vma)))))
1877                 return -EINVAL;
1878
1879         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1880         if (!new)
1881                 return -ENOMEM;
1882
1883         /* most fields are the same, copy all, and then fixup */
1884         *new = *vma;
1885
1886         if (new_below)
1887                 new->vm_end = addr;
1888         else {
1889                 new->vm_start = addr;
1890                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1891         }
1892
1893         pol = mpol_dup(vma_policy(vma));
1894         if (IS_ERR(pol)) {
1895                 kmem_cache_free(vm_area_cachep, new);
1896                 return PTR_ERR(pol);
1897         }
1898         vma_set_policy(new, pol);
1899
1900         if (new->vm_file) {
1901                 get_file(new->vm_file);
1902                 if (vma->vm_flags & VM_EXECUTABLE)
1903                         added_exe_file_vma(mm);
1904         }
1905
1906         if (new->vm_ops && new->vm_ops->open)
1907                 new->vm_ops->open(new);
1908
1909         if (new_below)
1910                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1911                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1912         else
1913                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1914
1915         return 0;
1916 }
1917
1918 /*
1919  * Split a vma into two pieces at address 'addr', a new vma is allocated
1920  * either for the first part or the tail.
1921  */
1922 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1923               unsigned long addr, int new_below)
1924 {
1925         if (mm->map_count >= sysctl_max_map_count)
1926                 return -ENOMEM;
1927
1928         return __split_vma(mm, vma, addr, new_below);
1929 }
1930
1931 /* Munmap is split into 2 main parts -- this part which finds
1932  * what needs doing, and the areas themselves, which do the
1933  * work.  This now handles partial unmappings.
1934  * Jeremy Fitzhardinge <jeremy@goop.org>
1935  */
1936 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1937 {
1938         unsigned long end;
1939         struct vm_area_struct *vma, *prev, *last;
1940
1941         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1942                 return -EINVAL;
1943
1944         if ((len = PAGE_ALIGN(len)) == 0)
1945                 return -EINVAL;
1946
1947         /* Find the first overlapping VMA */
1948         vma = find_vma_prev(mm, start, &prev);
1949         if (!vma)
1950                 return 0;
1951         /* we have  start < vma->vm_end  */
1952
1953         /* if it doesn't overlap, we have nothing.. */
1954         end = start + len;
1955         if (vma->vm_start >= end)
1956                 return 0;
1957
1958         /*
1959          * If we need to split any vma, do it now to save pain later.
1960          *
1961          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1962          * unmapped vm_area_struct will remain in use: so lower split_vma
1963          * places tmp vma above, and higher split_vma places tmp vma below.
1964          */
1965         if (start > vma->vm_start) {
1966                 int error;
1967
1968                 /*
1969                  * Make sure that map_count on return from munmap() will
1970                  * not exceed its limit; but let map_count go just above
1971                  * its limit temporarily, to help free resources as expected.
1972                  */
1973                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
1974                         return -ENOMEM;
1975
1976                 error = __split_vma(mm, vma, start, 0);
1977                 if (error)
1978                         return error;
1979                 prev = vma;
1980         }
1981
1982         /* Does it split the last one? */
1983         last = find_vma(mm, end);
1984         if (last && end > last->vm_start) {
1985                 int error = __split_vma(mm, last, end, 1);
1986                 if (error)
1987                         return error;
1988         }
1989         vma = prev? prev->vm_next: mm->mmap;
1990
1991         /*
1992          * unlock any mlock()ed ranges before detaching vmas
1993          */
1994         if (mm->locked_vm) {
1995                 struct vm_area_struct *tmp = vma;
1996                 while (tmp && tmp->vm_start < end) {
1997                         if (tmp->vm_flags & VM_LOCKED) {
1998                                 mm->locked_vm -= vma_pages(tmp);
1999                                 munlock_vma_pages_all(tmp);
2000                         }
2001                         tmp = tmp->vm_next;
2002                 }
2003         }
2004
2005         /*
2006          * Remove the vma's, and unmap the actual pages
2007          */
2008         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2009         unmap_region(mm, vma, prev, start, end);
2010
2011         /* Fix up all other VM information */
2012         remove_vma_list(mm, vma);
2013
2014         return 0;
2015 }
2016
2017 EXPORT_SYMBOL(do_munmap);
2018
2019 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2020 {
2021         int ret;
2022         struct mm_struct *mm = current->mm;
2023
2024         profile_munmap(addr);
2025
2026         down_write(&mm->mmap_sem);
2027         ret = do_munmap(mm, addr, len);
2028         up_write(&mm->mmap_sem);
2029         return ret;
2030 }
2031
2032 static inline void verify_mm_writelocked(struct mm_struct *mm)
2033 {
2034 #ifdef CONFIG_DEBUG_VM
2035         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2036                 WARN_ON(1);
2037                 up_read(&mm->mmap_sem);
2038         }
2039 #endif
2040 }
2041
2042 /*
2043  *  this is really a simplified "do_mmap".  it only handles
2044  *  anonymous maps.  eventually we may be able to do some
2045  *  brk-specific accounting here.
2046  */
2047 unsigned long do_brk(unsigned long addr, unsigned long len)
2048 {
2049         struct mm_struct * mm = current->mm;
2050         struct vm_area_struct * vma, * prev;
2051         unsigned long flags;
2052         struct rb_node ** rb_link, * rb_parent;
2053         pgoff_t pgoff = addr >> PAGE_SHIFT;
2054         int error;
2055
2056         len = PAGE_ALIGN(len);
2057         if (!len)
2058                 return addr;
2059
2060         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2061         if (error)
2062                 return error;
2063
2064         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2065
2066         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2067         if (error & ~PAGE_MASK)
2068                 return error;
2069
2070         /*
2071          * mlock MCL_FUTURE?
2072          */
2073         if (mm->def_flags & VM_LOCKED) {
2074                 unsigned long locked, lock_limit;
2075                 locked = len >> PAGE_SHIFT;
2076                 locked += mm->locked_vm;
2077                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2078                 lock_limit >>= PAGE_SHIFT;
2079                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2080                         return -EAGAIN;
2081         }
2082
2083         /*
2084          * mm->mmap_sem is required to protect against another thread
2085          * changing the mappings in case we sleep.
2086          */
2087         verify_mm_writelocked(mm);
2088
2089         /*
2090          * Clear old maps.  this also does some error checking for us
2091          */
2092  munmap_back:
2093         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2094         if (vma && vma->vm_start < addr + len) {
2095                 if (do_munmap(mm, addr, len))
2096                         return -ENOMEM;
2097                 goto munmap_back;
2098         }
2099
2100         /* Check against address space limits *after* clearing old maps... */
2101         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2102                 return -ENOMEM;
2103
2104         if (mm->map_count > sysctl_max_map_count)
2105                 return -ENOMEM;
2106
2107         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2108                 return -ENOMEM;
2109
2110         /* Can we just expand an old private anonymous mapping? */
2111         vma = vma_merge(mm, prev, addr, addr + len, flags,
2112                                         NULL, NULL, pgoff, NULL);
2113         if (vma)
2114                 goto out;
2115
2116         /*
2117          * create a vma struct for an anonymous mapping
2118          */
2119         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2120         if (!vma) {
2121                 vm_unacct_memory(len >> PAGE_SHIFT);
2122                 return -ENOMEM;
2123         }
2124
2125         vma->vm_mm = mm;
2126         vma->vm_start = addr;
2127         vma->vm_end = addr + len;
2128         vma->vm_pgoff = pgoff;
2129         vma->vm_flags = flags;
2130         vma->vm_page_prot = vm_get_page_prot(flags);
2131         vma_link(mm, vma, prev, rb_link, rb_parent);
2132 out:
2133         mm->total_vm += len >> PAGE_SHIFT;
2134         if (flags & VM_LOCKED) {
2135                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2136                         mm->locked_vm += (len >> PAGE_SHIFT);
2137         }
2138         return addr;
2139 }
2140
2141 EXPORT_SYMBOL(do_brk);
2142
2143 /* Release all mmaps. */
2144 void exit_mmap(struct mm_struct *mm)
2145 {
2146         struct mmu_gather *tlb;
2147         struct vm_area_struct *vma;
2148         unsigned long nr_accounted = 0;
2149         unsigned long end;
2150
2151         /* mm's last user has gone, and its about to be pulled down */
2152         mmu_notifier_release(mm);
2153
2154         if (mm->locked_vm) {
2155                 vma = mm->mmap;
2156                 while (vma) {
2157                         if (vma->vm_flags & VM_LOCKED)
2158                                 munlock_vma_pages_all(vma);
2159                         vma = vma->vm_next;
2160                 }
2161         }
2162
2163         arch_exit_mmap(mm);
2164
2165         vma = mm->mmap;
2166         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2167                 return;
2168
2169         lru_add_drain();
2170         flush_cache_mm(mm);
2171         tlb = tlb_gather_mmu(mm, 1);
2172         /* update_hiwater_rss(mm) here? but nobody should be looking */
2173         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2174         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2175         vm_unacct_memory(nr_accounted);
2176
2177         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2178         tlb_finish_mmu(tlb, 0, end);
2179
2180         /*
2181          * Walk the list again, actually closing and freeing it,
2182          * with preemption enabled, without holding any MM locks.
2183          */
2184         while (vma)
2185                 vma = remove_vma(vma);
2186
2187         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2188 }
2189
2190 /* Insert vm structure into process list sorted by address
2191  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2192  * then i_mmap_lock is taken here.
2193  */
2194 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2195 {
2196         struct vm_area_struct * __vma, * prev;
2197         struct rb_node ** rb_link, * rb_parent;
2198
2199         /*
2200          * The vm_pgoff of a purely anonymous vma should be irrelevant
2201          * until its first write fault, when page's anon_vma and index
2202          * are set.  But now set the vm_pgoff it will almost certainly
2203          * end up with (unless mremap moves it elsewhere before that
2204          * first wfault), so /proc/pid/maps tells a consistent story.
2205          *
2206          * By setting it to reflect the virtual start address of the
2207          * vma, merges and splits can happen in a seamless way, just
2208          * using the existing file pgoff checks and manipulations.
2209          * Similarly in do_mmap_pgoff and in do_brk.
2210          */
2211         if (!vma->vm_file) {
2212                 BUG_ON(vma->anon_vma);
2213                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2214         }
2215         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2216         if (__vma && __vma->vm_start < vma->vm_end)
2217                 return -ENOMEM;
2218         if ((vma->vm_flags & VM_ACCOUNT) &&
2219              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2220                 return -ENOMEM;
2221         vma_link(mm, vma, prev, rb_link, rb_parent);
2222         return 0;
2223 }
2224
2225 /*
2226  * Copy the vma structure to a new location in the same mm,
2227  * prior to moving page table entries, to effect an mremap move.
2228  */
2229 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2230         unsigned long addr, unsigned long len, pgoff_t pgoff)
2231 {
2232         struct vm_area_struct *vma = *vmap;
2233         unsigned long vma_start = vma->vm_start;
2234         struct mm_struct *mm = vma->vm_mm;
2235         struct vm_area_struct *new_vma, *prev;
2236         struct rb_node **rb_link, *rb_parent;
2237         struct mempolicy *pol;
2238
2239         /*
2240          * If anonymous vma has not yet been faulted, update new pgoff
2241          * to match new location, to increase its chance of merging.
2242          */
2243         if (!vma->vm_file && !vma->anon_vma)
2244                 pgoff = addr >> PAGE_SHIFT;
2245
2246         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2247         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2248                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2249         if (new_vma) {
2250                 /*
2251                  * Source vma may have been merged into new_vma
2252                  */
2253                 if (vma_start >= new_vma->vm_start &&
2254                     vma_start < new_vma->vm_end)
2255                         *vmap = new_vma;
2256         } else {
2257                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2258                 if (new_vma) {
2259                         *new_vma = *vma;
2260                         pol = mpol_dup(vma_policy(vma));
2261                         if (IS_ERR(pol)) {
2262                                 kmem_cache_free(vm_area_cachep, new_vma);
2263                                 return NULL;
2264                         }
2265                         vma_set_policy(new_vma, pol);
2266                         new_vma->vm_start = addr;
2267                         new_vma->vm_end = addr + len;
2268                         new_vma->vm_pgoff = pgoff;
2269                         if (new_vma->vm_file) {
2270                                 get_file(new_vma->vm_file);
2271                                 if (vma->vm_flags & VM_EXECUTABLE)
2272                                         added_exe_file_vma(mm);
2273                         }
2274                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2275                                 new_vma->vm_ops->open(new_vma);
2276                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2277                 }
2278         }
2279         return new_vma;
2280 }
2281
2282 /*
2283  * Return true if the calling process may expand its vm space by the passed
2284  * number of pages
2285  */
2286 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2287 {
2288         unsigned long cur = mm->total_vm;       /* pages */
2289         unsigned long lim;
2290
2291         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2292
2293         if (cur + npages > lim)
2294                 return 0;
2295         return 1;
2296 }
2297
2298
2299 static int special_mapping_fault(struct vm_area_struct *vma,
2300                                 struct vm_fault *vmf)
2301 {
2302         pgoff_t pgoff;
2303         struct page **pages;
2304
2305         /*
2306          * special mappings have no vm_file, and in that case, the mm
2307          * uses vm_pgoff internally. So we have to subtract it from here.
2308          * We are allowed to do this because we are the mm; do not copy
2309          * this code into drivers!
2310          */
2311         pgoff = vmf->pgoff - vma->vm_pgoff;
2312
2313         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2314                 pgoff--;
2315
2316         if (*pages) {
2317                 struct page *page = *pages;
2318                 get_page(page);
2319                 vmf->page = page;
2320                 return 0;
2321         }
2322
2323         return VM_FAULT_SIGBUS;
2324 }
2325
2326 /*
2327  * Having a close hook prevents vma merging regardless of flags.
2328  */
2329 static void special_mapping_close(struct vm_area_struct *vma)
2330 {
2331 }
2332
2333 static const struct vm_operations_struct special_mapping_vmops = {
2334         .close = special_mapping_close,
2335         .fault = special_mapping_fault,
2336 };
2337
2338 /*
2339  * Called with mm->mmap_sem held for writing.
2340  * Insert a new vma covering the given region, with the given flags.
2341  * Its pages are supplied by the given array of struct page *.
2342  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2343  * The region past the last page supplied will always produce SIGBUS.
2344  * The array pointer and the pages it points to are assumed to stay alive
2345  * for as long as this mapping might exist.
2346  */
2347 int install_special_mapping(struct mm_struct *mm,
2348                             unsigned long addr, unsigned long len,
2349                             unsigned long vm_flags, struct page **pages)
2350 {
2351         struct vm_area_struct *vma;
2352
2353         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2354         if (unlikely(vma == NULL))
2355                 return -ENOMEM;
2356
2357         vma->vm_mm = mm;
2358         vma->vm_start = addr;
2359         vma->vm_end = addr + len;
2360
2361         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2362         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2363
2364         vma->vm_ops = &special_mapping_vmops;
2365         vma->vm_private_data = pages;
2366
2367         if (unlikely(insert_vm_struct(mm, vma))) {
2368                 kmem_cache_free(vm_area_cachep, vma);
2369                 return -ENOMEM;
2370         }
2371
2372         mm->total_vm += len >> PAGE_SHIFT;
2373
2374         perf_event_mmap(vma);
2375
2376         return 0;
2377 }
2378
2379 static DEFINE_MUTEX(mm_all_locks_mutex);
2380
2381 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2382 {
2383         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2384                 /*
2385                  * The LSB of head.next can't change from under us
2386                  * because we hold the mm_all_locks_mutex.
2387                  */
2388                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2389                 /*
2390                  * We can safely modify head.next after taking the
2391                  * anon_vma->lock. If some other vma in this mm shares
2392                  * the same anon_vma we won't take it again.
2393                  *
2394                  * No need of atomic instructions here, head.next
2395                  * can't change from under us thanks to the
2396                  * anon_vma->lock.
2397                  */
2398                 if (__test_and_set_bit(0, (unsigned long *)
2399                                        &anon_vma->head.next))
2400                         BUG();
2401         }
2402 }
2403
2404 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2405 {
2406         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2407                 /*
2408                  * AS_MM_ALL_LOCKS can't change from under us because
2409                  * we hold the mm_all_locks_mutex.
2410                  *
2411                  * Operations on ->flags have to be atomic because
2412                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2413                  * mm_all_locks_mutex, there may be other cpus
2414                  * changing other bitflags in parallel to us.
2415                  */
2416                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2417                         BUG();
2418                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2419         }
2420 }
2421
2422 /*
2423  * This operation locks against the VM for all pte/vma/mm related
2424  * operations that could ever happen on a certain mm. This includes
2425  * vmtruncate, try_to_unmap, and all page faults.
2426  *
2427  * The caller must take the mmap_sem in write mode before calling
2428  * mm_take_all_locks(). The caller isn't allowed to release the
2429  * mmap_sem until mm_drop_all_locks() returns.
2430  *
2431  * mmap_sem in write mode is required in order to block all operations
2432  * that could modify pagetables and free pages without need of
2433  * altering the vma layout (for example populate_range() with
2434  * nonlinear vmas). It's also needed in write mode to avoid new
2435  * anon_vmas to be associated with existing vmas.
2436  *
2437  * A single task can't take more than one mm_take_all_locks() in a row
2438  * or it would deadlock.
2439  *
2440  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2441  * mapping->flags avoid to take the same lock twice, if more than one
2442  * vma in this mm is backed by the same anon_vma or address_space.
2443  *
2444  * We can take all the locks in random order because the VM code
2445  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2446  * takes more than one of them in a row. Secondly we're protected
2447  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2448  *
2449  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2450  * that may have to take thousand of locks.
2451  *
2452  * mm_take_all_locks() can fail if it's interrupted by signals.
2453  */
2454 int mm_take_all_locks(struct mm_struct *mm)
2455 {
2456         struct vm_area_struct *vma;
2457         int ret = -EINTR;
2458
2459         BUG_ON(down_read_trylock(&mm->mmap_sem));
2460
2461         mutex_lock(&mm_all_locks_mutex);
2462
2463         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2464                 if (signal_pending(current))
2465                         goto out_unlock;
2466                 if (vma->vm_file && vma->vm_file->f_mapping)
2467                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2468         }
2469
2470         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2471                 if (signal_pending(current))
2472                         goto out_unlock;
2473                 if (vma->anon_vma)
2474                         vm_lock_anon_vma(mm, vma->anon_vma);
2475         }
2476
2477         ret = 0;
2478
2479 out_unlock:
2480         if (ret)
2481                 mm_drop_all_locks(mm);
2482
2483         return ret;
2484 }
2485
2486 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2487 {
2488         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2489                 /*
2490                  * The LSB of head.next can't change to 0 from under
2491                  * us because we hold the mm_all_locks_mutex.
2492                  *
2493                  * We must however clear the bitflag before unlocking
2494                  * the vma so the users using the anon_vma->head will
2495                  * never see our bitflag.
2496                  *
2497                  * No need of atomic instructions here, head.next
2498                  * can't change from under us until we release the
2499                  * anon_vma->lock.
2500                  */
2501                 if (!__test_and_clear_bit(0, (unsigned long *)
2502                                           &anon_vma->head.next))
2503                         BUG();
2504                 spin_unlock(&anon_vma->lock);
2505         }
2506 }
2507
2508 static void vm_unlock_mapping(struct address_space *mapping)
2509 {
2510         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2511                 /*
2512                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2513                  * because we hold the mm_all_locks_mutex.
2514                  */
2515                 spin_unlock(&mapping->i_mmap_lock);
2516                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2517                                         &mapping->flags))
2518                         BUG();
2519         }
2520 }
2521
2522 /*
2523  * The mmap_sem cannot be released by the caller until
2524  * mm_drop_all_locks() returns.
2525  */
2526 void mm_drop_all_locks(struct mm_struct *mm)
2527 {
2528         struct vm_area_struct *vma;
2529
2530         BUG_ON(down_read_trylock(&mm->mmap_sem));
2531         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2532
2533         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2534                 if (vma->anon_vma)
2535                         vm_unlock_anon_vma(vma->anon_vma);
2536                 if (vma->vm_file && vma->vm_file->f_mapping)
2537                         vm_unlock_mapping(vma->vm_file->f_mapping);
2538         }
2539
2540         mutex_unlock(&mm_all_locks_mutex);
2541 }
2542
2543 /*
2544  * initialise the VMA slab
2545  */
2546 void __init mmap_init(void)
2547 {
2548         int ret;
2549
2550         ret = percpu_counter_init(&vm_committed_as, 0);
2551         VM_BUG_ON(ret);
2552 }