fix a struct file leak in do_mmap_pgoff()
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31
32 #include <asm/uaccess.h>
33 #include <asm/cacheflush.h>
34 #include <asm/tlb.h>
35 #include <asm/mmu_context.h>
36
37 #include "internal.h"
38
39 #ifndef arch_mmap_check
40 #define arch_mmap_check(addr, len, flags)       (0)
41 #endif
42
43 #ifndef arch_rebalance_pgtables
44 #define arch_rebalance_pgtables(addr, len)              (addr)
45 #endif
46
47 static void unmap_region(struct mm_struct *mm,
48                 struct vm_area_struct *vma, struct vm_area_struct *prev,
49                 unsigned long start, unsigned long end);
50
51 /*
52  * WARNING: the debugging will use recursive algorithms so never enable this
53  * unless you know what you are doing.
54  */
55 #undef DEBUG_MM_RB
56
57 /* description of effects of mapping type and prot in current implementation.
58  * this is due to the limited x86 page protection hardware.  The expected
59  * behavior is in parens:
60  *
61  * map_type     prot
62  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
63  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
64  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
65  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
66  *              
67  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
68  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
69  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
70  *
71  */
72 pgprot_t protection_map[16] = {
73         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
74         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
75 };
76
77 pgprot_t vm_get_page_prot(unsigned long vm_flags)
78 {
79         return __pgprot(pgprot_val(protection_map[vm_flags &
80                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
81                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
82 }
83 EXPORT_SYMBOL(vm_get_page_prot);
84
85 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
86 int sysctl_overcommit_ratio = 50;       /* default is 50% */
87 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
88 struct percpu_counter vm_committed_as;
89
90 /*
91  * Check that a process has enough memory to allocate a new virtual
92  * mapping. 0 means there is enough memory for the allocation to
93  * succeed and -ENOMEM implies there is not.
94  *
95  * We currently support three overcommit policies, which are set via the
96  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
97  *
98  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
99  * Additional code 2002 Jul 20 by Robert Love.
100  *
101  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
102  *
103  * Note this is a helper function intended to be used by LSMs which
104  * wish to use this logic.
105  */
106 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
107 {
108         unsigned long free, allowed;
109
110         vm_acct_memory(pages);
111
112         /*
113          * Sometimes we want to use more memory than we have
114          */
115         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
116                 return 0;
117
118         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
119                 unsigned long n;
120
121                 free = global_page_state(NR_FILE_PAGES);
122                 free += nr_swap_pages;
123
124                 /*
125                  * Any slabs which are created with the
126                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
127                  * which are reclaimable, under pressure.  The dentry
128                  * cache and most inode caches should fall into this
129                  */
130                 free += global_page_state(NR_SLAB_RECLAIMABLE);
131
132                 /*
133                  * Leave the last 3% for root
134                  */
135                 if (!cap_sys_admin)
136                         free -= free / 32;
137
138                 if (free > pages)
139                         return 0;
140
141                 /*
142                  * nr_free_pages() is very expensive on large systems,
143                  * only call if we're about to fail.
144                  */
145                 n = nr_free_pages();
146
147                 /*
148                  * Leave reserved pages. The pages are not for anonymous pages.
149                  */
150                 if (n <= totalreserve_pages)
151                         goto error;
152                 else
153                         n -= totalreserve_pages;
154
155                 /*
156                  * Leave the last 3% for root
157                  */
158                 if (!cap_sys_admin)
159                         n -= n / 32;
160                 free += n;
161
162                 if (free > pages)
163                         return 0;
164
165                 goto error;
166         }
167
168         allowed = (totalram_pages - hugetlb_total_pages())
169                 * sysctl_overcommit_ratio / 100;
170         /*
171          * Leave the last 3% for root
172          */
173         if (!cap_sys_admin)
174                 allowed -= allowed / 32;
175         allowed += total_swap_pages;
176
177         /* Don't let a single process grow too big:
178            leave 3% of the size of this process for other processes */
179         if (mm)
180                 allowed -= mm->total_vm / 32;
181
182         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
183                 return 0;
184 error:
185         vm_unacct_memory(pages);
186
187         return -ENOMEM;
188 }
189
190 /*
191  * Requires inode->i_mapping->i_mmap_lock
192  */
193 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
194                 struct file *file, struct address_space *mapping)
195 {
196         if (vma->vm_flags & VM_DENYWRITE)
197                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
198         if (vma->vm_flags & VM_SHARED)
199                 mapping->i_mmap_writable--;
200
201         flush_dcache_mmap_lock(mapping);
202         if (unlikely(vma->vm_flags & VM_NONLINEAR))
203                 list_del_init(&vma->shared.vm_set.list);
204         else
205                 vma_prio_tree_remove(vma, &mapping->i_mmap);
206         flush_dcache_mmap_unlock(mapping);
207 }
208
209 /*
210  * Unlink a file-based vm structure from its prio_tree, to hide
211  * vma from rmap and vmtruncate before freeing its page tables.
212  */
213 void unlink_file_vma(struct vm_area_struct *vma)
214 {
215         struct file *file = vma->vm_file;
216
217         if (file) {
218                 struct address_space *mapping = file->f_mapping;
219                 spin_lock(&mapping->i_mmap_lock);
220                 __remove_shared_vm_struct(vma, file, mapping);
221                 spin_unlock(&mapping->i_mmap_lock);
222         }
223 }
224
225 /*
226  * Close a vm structure and free it, returning the next.
227  */
228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
229 {
230         struct vm_area_struct *next = vma->vm_next;
231
232         might_sleep();
233         if (vma->vm_ops && vma->vm_ops->close)
234                 vma->vm_ops->close(vma);
235         if (vma->vm_file) {
236                 fput(vma->vm_file);
237                 if (vma->vm_flags & VM_EXECUTABLE)
238                         removed_exe_file_vma(vma->vm_mm);
239         }
240         mpol_put(vma_policy(vma));
241         kmem_cache_free(vm_area_cachep, vma);
242         return next;
243 }
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         min_brk = mm->end_code;
256 #else
257         min_brk = mm->start_brk;
258 #endif
259         if (brk < min_brk)
260                 goto out;
261
262         /*
263          * Check against rlimit here. If this check is done later after the test
264          * of oldbrk with newbrk then it can escape the test and let the data
265          * segment grow beyond its set limit the in case where the limit is
266          * not page aligned -Ram Gupta
267          */
268         rlim = 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         error = arch_mmap_check(addr, len, flags);
935         if (error)
936                 return error;
937
938         /* Careful about overflows.. */
939         len = PAGE_ALIGN(len);
940         if (!len || len > TASK_SIZE)
941                 return -ENOMEM;
942
943         /* offset overflow? */
944         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
945                return -EOVERFLOW;
946
947         /* Too many mappings? */
948         if (mm->map_count > sysctl_max_map_count)
949                 return -ENOMEM;
950
951         /* Obtain the address to map to. we verify (or select) it and ensure
952          * that it represents a valid section of the address space.
953          */
954         addr = get_unmapped_area(file, addr, len, pgoff, flags);
955         if (addr & ~PAGE_MASK)
956                 return addr;
957
958         /* Do simple checking here so the lower-level routines won't have
959          * to. we assume access permissions have been handled by the open
960          * of the memory object, so we don't do any here.
961          */
962         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
963                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
964
965         if (flags & MAP_LOCKED)
966                 if (!can_do_mlock())
967                         return -EPERM;
968
969         /* mlock MCL_FUTURE? */
970         if (vm_flags & VM_LOCKED) {
971                 unsigned long locked, lock_limit;
972                 locked = len >> PAGE_SHIFT;
973                 locked += mm->locked_vm;
974                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
975                 lock_limit >>= PAGE_SHIFT;
976                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
977                         return -EAGAIN;
978         }
979
980         inode = file ? file->f_path.dentry->d_inode : NULL;
981
982         if (file) {
983                 switch (flags & MAP_TYPE) {
984                 case MAP_SHARED:
985                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
986                                 return -EACCES;
987
988                         /*
989                          * Make sure we don't allow writing to an append-only
990                          * file..
991                          */
992                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
993                                 return -EACCES;
994
995                         /*
996                          * Make sure there are no mandatory locks on the file.
997                          */
998                         if (locks_verify_locked(inode))
999                                 return -EAGAIN;
1000
1001                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1002                         if (!(file->f_mode & FMODE_WRITE))
1003                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1004
1005                         /* fall through */
1006                 case MAP_PRIVATE:
1007                         if (!(file->f_mode & FMODE_READ))
1008                                 return -EACCES;
1009                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1010                                 if (vm_flags & VM_EXEC)
1011                                         return -EPERM;
1012                                 vm_flags &= ~VM_MAYEXEC;
1013                         }
1014
1015                         if (!file->f_op || !file->f_op->mmap)
1016                                 return -ENODEV;
1017                         break;
1018
1019                 default:
1020                         return -EINVAL;
1021                 }
1022         } else {
1023                 switch (flags & MAP_TYPE) {
1024                 case MAP_SHARED:
1025                         /*
1026                          * Ignore pgoff.
1027                          */
1028                         pgoff = 0;
1029                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1030                         break;
1031                 case MAP_PRIVATE:
1032                         /*
1033                          * Set pgoff according to addr for anon_vma.
1034                          */
1035                         pgoff = addr >> PAGE_SHIFT;
1036                         break;
1037                 default:
1038                         return -EINVAL;
1039                 }
1040         }
1041
1042         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1043         if (error)
1044                 return error;
1045
1046         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1047 }
1048 EXPORT_SYMBOL(do_mmap_pgoff);
1049
1050 /*
1051  * Some shared mappigns will want the pages marked read-only
1052  * to track write events. If so, we'll downgrade vm_page_prot
1053  * to the private version (using protection_map[] without the
1054  * VM_SHARED bit).
1055  */
1056 int vma_wants_writenotify(struct vm_area_struct *vma)
1057 {
1058         unsigned int vm_flags = vma->vm_flags;
1059
1060         /* If it was private or non-writable, the write bit is already clear */
1061         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1062                 return 0;
1063
1064         /* The backer wishes to know when pages are first written to? */
1065         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1066                 return 1;
1067
1068         /* The open routine did something to the protections already? */
1069         if (pgprot_val(vma->vm_page_prot) !=
1070             pgprot_val(vm_get_page_prot(vm_flags)))
1071                 return 0;
1072
1073         /* Specialty mapping? */
1074         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1075                 return 0;
1076
1077         /* Can the mapping track the dirty pages? */
1078         return vma->vm_file && vma->vm_file->f_mapping &&
1079                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1080 }
1081
1082 /*
1083  * We account for memory if it's a private writeable mapping,
1084  * not hugepages and VM_NORESERVE wasn't set.
1085  */
1086 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1087 {
1088         /*
1089          * hugetlb has its own accounting separate from the core VM
1090          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1091          */
1092         if (file && is_file_hugepages(file))
1093                 return 0;
1094
1095         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1096 }
1097
1098 unsigned long mmap_region(struct file *file, unsigned long addr,
1099                           unsigned long len, unsigned long flags,
1100                           unsigned int vm_flags, unsigned long pgoff)
1101 {
1102         struct mm_struct *mm = current->mm;
1103         struct vm_area_struct *vma, *prev;
1104         int correct_wcount = 0;
1105         int error;
1106         struct rb_node **rb_link, *rb_parent;
1107         unsigned long charged = 0;
1108         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1109
1110         /* Clear old maps */
1111         error = -ENOMEM;
1112 munmap_back:
1113         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1114         if (vma && vma->vm_start < addr + len) {
1115                 if (do_munmap(mm, addr, len))
1116                         return -ENOMEM;
1117                 goto munmap_back;
1118         }
1119
1120         /* Check against address space limit. */
1121         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1122                 return -ENOMEM;
1123
1124         /*
1125          * Set 'VM_NORESERVE' if we should not account for the
1126          * memory use of this mapping.
1127          */
1128         if ((flags & MAP_NORESERVE)) {
1129                 /* We honor MAP_NORESERVE if allowed to overcommit */
1130                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1131                         vm_flags |= VM_NORESERVE;
1132
1133                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1134                 if (file && is_file_hugepages(file))
1135                         vm_flags |= VM_NORESERVE;
1136         }
1137
1138         /*
1139          * Private writable mapping: check memory availability
1140          */
1141         if (accountable_mapping(file, vm_flags)) {
1142                 charged = len >> PAGE_SHIFT;
1143                 if (security_vm_enough_memory(charged))
1144                         return -ENOMEM;
1145                 vm_flags |= VM_ACCOUNT;
1146         }
1147
1148         /*
1149          * Can we just expand an old mapping?
1150          */
1151         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1152         if (vma)
1153                 goto out;
1154
1155         /*
1156          * Determine the object being mapped and call the appropriate
1157          * specific mapper. the address has already been validated, but
1158          * not unmapped, but the maps are removed from the list.
1159          */
1160         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1161         if (!vma) {
1162                 error = -ENOMEM;
1163                 goto unacct_error;
1164         }
1165
1166         vma->vm_mm = mm;
1167         vma->vm_start = addr;
1168         vma->vm_end = addr + len;
1169         vma->vm_flags = vm_flags;
1170         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1171         vma->vm_pgoff = pgoff;
1172
1173         if (file) {
1174                 error = -EINVAL;
1175                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1176                         goto free_vma;
1177                 if (vm_flags & VM_DENYWRITE) {
1178                         error = deny_write_access(file);
1179                         if (error)
1180                                 goto free_vma;
1181                         correct_wcount = 1;
1182                 }
1183                 vma->vm_file = file;
1184                 get_file(file);
1185                 error = file->f_op->mmap(file, vma);
1186                 if (error)
1187                         goto unmap_and_free_vma;
1188                 if (vm_flags & VM_EXECUTABLE)
1189                         added_exe_file_vma(mm);
1190
1191                 /* Can addr have changed??
1192                  *
1193                  * Answer: Yes, several device drivers can do it in their
1194                  *         f_op->mmap method. -DaveM
1195                  */
1196                 addr = vma->vm_start;
1197                 pgoff = vma->vm_pgoff;
1198                 vm_flags = vma->vm_flags;
1199         } else if (vm_flags & VM_SHARED) {
1200                 error = shmem_zero_setup(vma);
1201                 if (error)
1202                         goto free_vma;
1203         }
1204
1205         if (vma_wants_writenotify(vma))
1206                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1207
1208         vma_link(mm, vma, prev, rb_link, rb_parent);
1209         file = vma->vm_file;
1210
1211         /* Once vma denies write, undo our temporary denial count */
1212         if (correct_wcount)
1213                 atomic_inc(&inode->i_writecount);
1214 out:
1215         perf_event_mmap(vma);
1216
1217         mm->total_vm += len >> PAGE_SHIFT;
1218         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1219         if (vm_flags & VM_LOCKED) {
1220                 /*
1221                  * makes pages present; downgrades, drops, reacquires mmap_sem
1222                  */
1223                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1224                 if (nr_pages < 0)
1225                         return nr_pages;        /* vma gone! */
1226                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1227         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1228                 make_pages_present(addr, addr + len);
1229         return addr;
1230
1231 unmap_and_free_vma:
1232         if (correct_wcount)
1233                 atomic_inc(&inode->i_writecount);
1234         vma->vm_file = NULL;
1235         fput(file);
1236
1237         /* Undo any partial mapping done by a device driver. */
1238         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1239         charged = 0;
1240 free_vma:
1241         kmem_cache_free(vm_area_cachep, vma);
1242 unacct_error:
1243         if (charged)
1244                 vm_unacct_memory(charged);
1245         return error;
1246 }
1247
1248 /* Get an address range which is currently unmapped.
1249  * For shmat() with addr=0.
1250  *
1251  * Ugly calling convention alert:
1252  * Return value with the low bits set means error value,
1253  * ie
1254  *      if (ret & ~PAGE_MASK)
1255  *              error = ret;
1256  *
1257  * This function "knows" that -ENOMEM has the bits set.
1258  */
1259 #ifndef HAVE_ARCH_UNMAPPED_AREA
1260 unsigned long
1261 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1262                 unsigned long len, unsigned long pgoff, unsigned long flags)
1263 {
1264         struct mm_struct *mm = current->mm;
1265         struct vm_area_struct *vma;
1266         unsigned long start_addr;
1267
1268         if (len > TASK_SIZE)
1269                 return -ENOMEM;
1270
1271         if (flags & MAP_FIXED)
1272                 return addr;
1273
1274         if (addr) {
1275                 addr = PAGE_ALIGN(addr);
1276                 vma = find_vma(mm, addr);
1277                 if (TASK_SIZE - len >= addr &&
1278                     (!vma || addr + len <= vma->vm_start))
1279                         return addr;
1280         }
1281         if (len > mm->cached_hole_size) {
1282                 start_addr = addr = mm->free_area_cache;
1283         } else {
1284                 start_addr = addr = TASK_UNMAPPED_BASE;
1285                 mm->cached_hole_size = 0;
1286         }
1287
1288 full_search:
1289         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1290                 /* At this point:  (!vma || addr < vma->vm_end). */
1291                 if (TASK_SIZE - len < addr) {
1292                         /*
1293                          * Start a new search - just in case we missed
1294                          * some holes.
1295                          */
1296                         if (start_addr != TASK_UNMAPPED_BASE) {
1297                                 addr = TASK_UNMAPPED_BASE;
1298                                 start_addr = addr;
1299                                 mm->cached_hole_size = 0;
1300                                 goto full_search;
1301                         }
1302                         return -ENOMEM;
1303                 }
1304                 if (!vma || addr + len <= vma->vm_start) {
1305                         /*
1306                          * Remember the place where we stopped the search:
1307                          */
1308                         mm->free_area_cache = addr + len;
1309                         return addr;
1310                 }
1311                 if (addr + mm->cached_hole_size < vma->vm_start)
1312                         mm->cached_hole_size = vma->vm_start - addr;
1313                 addr = vma->vm_end;
1314         }
1315 }
1316 #endif  
1317
1318 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1319 {
1320         /*
1321          * Is this a new hole at the lowest possible address?
1322          */
1323         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1324                 mm->free_area_cache = addr;
1325                 mm->cached_hole_size = ~0UL;
1326         }
1327 }
1328
1329 /*
1330  * This mmap-allocator allocates new areas top-down from below the
1331  * stack's low limit (the base):
1332  */
1333 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1334 unsigned long
1335 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1336                           const unsigned long len, const unsigned long pgoff,
1337                           const unsigned long flags)
1338 {
1339         struct vm_area_struct *vma;
1340         struct mm_struct *mm = current->mm;
1341         unsigned long addr = addr0;
1342
1343         /* requested length too big for entire address space */
1344         if (len > TASK_SIZE)
1345                 return -ENOMEM;
1346
1347         if (flags & MAP_FIXED)
1348                 return addr;
1349
1350         /* requesting a specific address */
1351         if (addr) {
1352                 addr = PAGE_ALIGN(addr);
1353                 vma = find_vma(mm, addr);
1354                 if (TASK_SIZE - len >= addr &&
1355                                 (!vma || addr + len <= vma->vm_start))
1356                         return addr;
1357         }
1358
1359         /* check if free_area_cache is useful for us */
1360         if (len <= mm->cached_hole_size) {
1361                 mm->cached_hole_size = 0;
1362                 mm->free_area_cache = mm->mmap_base;
1363         }
1364
1365         /* either no address requested or can't fit in requested address hole */
1366         addr = mm->free_area_cache;
1367
1368         /* make sure it can fit in the remaining address space */
1369         if (addr > len) {
1370                 vma = find_vma(mm, addr-len);
1371                 if (!vma || addr <= vma->vm_start)
1372                         /* remember the address as a hint for next time */
1373                         return (mm->free_area_cache = addr-len);
1374         }
1375
1376         if (mm->mmap_base < len)
1377                 goto bottomup;
1378
1379         addr = mm->mmap_base-len;
1380
1381         do {
1382                 /*
1383                  * Lookup failure means no vma is above this address,
1384                  * else if new region fits below vma->vm_start,
1385                  * return with success:
1386                  */
1387                 vma = find_vma(mm, addr);
1388                 if (!vma || addr+len <= vma->vm_start)
1389                         /* remember the address as a hint for next time */
1390                         return (mm->free_area_cache = addr);
1391
1392                 /* remember the largest hole we saw so far */
1393                 if (addr + mm->cached_hole_size < vma->vm_start)
1394                         mm->cached_hole_size = vma->vm_start - addr;
1395
1396                 /* try just below the current vma->vm_start */
1397                 addr = vma->vm_start-len;
1398         } while (len < vma->vm_start);
1399
1400 bottomup:
1401         /*
1402          * A failed mmap() very likely causes application failure,
1403          * so fall back to the bottom-up function here. This scenario
1404          * can happen with large stack limits and large mmap()
1405          * allocations.
1406          */
1407         mm->cached_hole_size = ~0UL;
1408         mm->free_area_cache = TASK_UNMAPPED_BASE;
1409         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1410         /*
1411          * Restore the topdown base:
1412          */
1413         mm->free_area_cache = mm->mmap_base;
1414         mm->cached_hole_size = ~0UL;
1415
1416         return addr;
1417 }
1418 #endif
1419
1420 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1421 {
1422         /*
1423          * Is this a new hole at the highest possible address?
1424          */
1425         if (addr > mm->free_area_cache)
1426                 mm->free_area_cache = addr;
1427
1428         /* dont allow allocations above current base */
1429         if (mm->free_area_cache > mm->mmap_base)
1430                 mm->free_area_cache = mm->mmap_base;
1431 }
1432
1433 unsigned long
1434 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1435                 unsigned long pgoff, unsigned long flags)
1436 {
1437         unsigned long (*get_area)(struct file *, unsigned long,
1438                                   unsigned long, unsigned long, unsigned long);
1439
1440         get_area = current->mm->get_unmapped_area;
1441         if (file && file->f_op && file->f_op->get_unmapped_area)
1442                 get_area = file->f_op->get_unmapped_area;
1443         addr = get_area(file, addr, len, pgoff, flags);
1444         if (IS_ERR_VALUE(addr))
1445                 return addr;
1446
1447         if (addr > TASK_SIZE - len)
1448                 return -ENOMEM;
1449         if (addr & ~PAGE_MASK)
1450                 return -EINVAL;
1451
1452         return arch_rebalance_pgtables(addr, len);
1453 }
1454
1455 EXPORT_SYMBOL(get_unmapped_area);
1456
1457 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1458 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1459 {
1460         struct vm_area_struct *vma = NULL;
1461
1462         if (mm) {
1463                 /* Check the cache first. */
1464                 /* (Cache hit rate is typically around 35%.) */
1465                 vma = mm->mmap_cache;
1466                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1467                         struct rb_node * rb_node;
1468
1469                         rb_node = mm->mm_rb.rb_node;
1470                         vma = NULL;
1471
1472                         while (rb_node) {
1473                                 struct vm_area_struct * vma_tmp;
1474
1475                                 vma_tmp = rb_entry(rb_node,
1476                                                 struct vm_area_struct, vm_rb);
1477
1478                                 if (vma_tmp->vm_end > addr) {
1479                                         vma = vma_tmp;
1480                                         if (vma_tmp->vm_start <= addr)
1481                                                 break;
1482                                         rb_node = rb_node->rb_left;
1483                                 } else
1484                                         rb_node = rb_node->rb_right;
1485                         }
1486                         if (vma)
1487                                 mm->mmap_cache = vma;
1488                 }
1489         }
1490         return vma;
1491 }
1492
1493 EXPORT_SYMBOL(find_vma);
1494
1495 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1496 struct vm_area_struct *
1497 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1498                         struct vm_area_struct **pprev)
1499 {
1500         struct vm_area_struct *vma = NULL, *prev = NULL;
1501         struct rb_node *rb_node;
1502         if (!mm)
1503                 goto out;
1504
1505         /* Guard against addr being lower than the first VMA */
1506         vma = mm->mmap;
1507
1508         /* Go through the RB tree quickly. */
1509         rb_node = mm->mm_rb.rb_node;
1510
1511         while (rb_node) {
1512                 struct vm_area_struct *vma_tmp;
1513                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1514
1515                 if (addr < vma_tmp->vm_end) {
1516                         rb_node = rb_node->rb_left;
1517                 } else {
1518                         prev = vma_tmp;
1519                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1520                                 break;
1521                         rb_node = rb_node->rb_right;
1522                 }
1523         }
1524
1525 out:
1526         *pprev = prev;
1527         return prev ? prev->vm_next : vma;
1528 }
1529
1530 /*
1531  * Verify that the stack growth is acceptable and
1532  * update accounting. This is shared with both the
1533  * grow-up and grow-down cases.
1534  */
1535 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1536 {
1537         struct mm_struct *mm = vma->vm_mm;
1538         struct rlimit *rlim = current->signal->rlim;
1539         unsigned long new_start;
1540
1541         /* address space limit tests */
1542         if (!may_expand_vm(mm, grow))
1543                 return -ENOMEM;
1544
1545         /* Stack limit test */
1546         if (size > rlim[RLIMIT_STACK].rlim_cur)
1547                 return -ENOMEM;
1548
1549         /* mlock limit tests */
1550         if (vma->vm_flags & VM_LOCKED) {
1551                 unsigned long locked;
1552                 unsigned long limit;
1553                 locked = mm->locked_vm + grow;
1554                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1555                 if (locked > limit && !capable(CAP_IPC_LOCK))
1556                         return -ENOMEM;
1557         }
1558
1559         /* Check to ensure the stack will not grow into a hugetlb-only region */
1560         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1561                         vma->vm_end - size;
1562         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1563                 return -EFAULT;
1564
1565         /*
1566          * Overcommit..  This must be the final test, as it will
1567          * update security statistics.
1568          */
1569         if (security_vm_enough_memory_mm(mm, grow))
1570                 return -ENOMEM;
1571
1572         /* Ok, everything looks good - let it rip */
1573         mm->total_vm += grow;
1574         if (vma->vm_flags & VM_LOCKED)
1575                 mm->locked_vm += grow;
1576         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1577         return 0;
1578 }
1579
1580 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1581 /*
1582  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1583  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1584  */
1585 #ifndef CONFIG_IA64
1586 static
1587 #endif
1588 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1589 {
1590         int error;
1591
1592         if (!(vma->vm_flags & VM_GROWSUP))
1593                 return -EFAULT;
1594
1595         /*
1596          * We must make sure the anon_vma is allocated
1597          * so that the anon_vma locking is not a noop.
1598          */
1599         if (unlikely(anon_vma_prepare(vma)))
1600                 return -ENOMEM;
1601         anon_vma_lock(vma);
1602
1603         /*
1604          * vma->vm_start/vm_end cannot change under us because the caller
1605          * is required to hold the mmap_sem in read mode.  We need the
1606          * anon_vma lock to serialize against concurrent expand_stacks.
1607          * Also guard against wrapping around to address 0.
1608          */
1609         if (address < PAGE_ALIGN(address+4))
1610                 address = PAGE_ALIGN(address+4);
1611         else {
1612                 anon_vma_unlock(vma);
1613                 return -ENOMEM;
1614         }
1615         error = 0;
1616
1617         /* Somebody else might have raced and expanded it already */
1618         if (address > vma->vm_end) {
1619                 unsigned long size, grow;
1620
1621                 size = address - vma->vm_start;
1622                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1623
1624                 error = acct_stack_growth(vma, size, grow);
1625                 if (!error)
1626                         vma->vm_end = address;
1627         }
1628         anon_vma_unlock(vma);
1629         return error;
1630 }
1631 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1632
1633 /*
1634  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1635  */
1636 static int expand_downwards(struct vm_area_struct *vma,
1637                                    unsigned long address)
1638 {
1639         int error;
1640
1641         /*
1642          * We must make sure the anon_vma is allocated
1643          * so that the anon_vma locking is not a noop.
1644          */
1645         if (unlikely(anon_vma_prepare(vma)))
1646                 return -ENOMEM;
1647
1648         address &= PAGE_MASK;
1649         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1650         if (error)
1651                 return error;
1652
1653         anon_vma_lock(vma);
1654
1655         /*
1656          * vma->vm_start/vm_end cannot change under us because the caller
1657          * is required to hold the mmap_sem in read mode.  We need the
1658          * anon_vma lock to serialize against concurrent expand_stacks.
1659          */
1660
1661         /* Somebody else might have raced and expanded it already */
1662         if (address < vma->vm_start) {
1663                 unsigned long size, grow;
1664
1665                 size = vma->vm_end - address;
1666                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1667
1668                 error = acct_stack_growth(vma, size, grow);
1669                 if (!error) {
1670                         vma->vm_start = address;
1671                         vma->vm_pgoff -= grow;
1672                 }
1673         }
1674         anon_vma_unlock(vma);
1675         return error;
1676 }
1677
1678 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1679 {
1680         return expand_downwards(vma, address);
1681 }
1682
1683 #ifdef CONFIG_STACK_GROWSUP
1684 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1685 {
1686         return expand_upwards(vma, address);
1687 }
1688
1689 struct vm_area_struct *
1690 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1691 {
1692         struct vm_area_struct *vma, *prev;
1693
1694         addr &= PAGE_MASK;
1695         vma = find_vma_prev(mm, addr, &prev);
1696         if (vma && (vma->vm_start <= addr))
1697                 return vma;
1698         if (!prev || expand_stack(prev, addr))
1699                 return NULL;
1700         if (prev->vm_flags & VM_LOCKED) {
1701                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1702                         return NULL;    /* vma gone! */
1703         }
1704         return prev;
1705 }
1706 #else
1707 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1708 {
1709         return expand_downwards(vma, address);
1710 }
1711
1712 struct vm_area_struct *
1713 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1714 {
1715         struct vm_area_struct * vma;
1716         unsigned long start;
1717
1718         addr &= PAGE_MASK;
1719         vma = find_vma(mm,addr);
1720         if (!vma)
1721                 return NULL;
1722         if (vma->vm_start <= addr)
1723                 return vma;
1724         if (!(vma->vm_flags & VM_GROWSDOWN))
1725                 return NULL;
1726         start = vma->vm_start;
1727         if (expand_stack(vma, addr))
1728                 return NULL;
1729         if (vma->vm_flags & VM_LOCKED) {
1730                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1731                         return NULL;    /* vma gone! */
1732         }
1733         return vma;
1734 }
1735 #endif
1736
1737 /*
1738  * Ok - we have the memory areas we should free on the vma list,
1739  * so release them, and do the vma updates.
1740  *
1741  * Called with the mm semaphore held.
1742  */
1743 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1744 {
1745         /* Update high watermark before we lower total_vm */
1746         update_hiwater_vm(mm);
1747         do {
1748                 long nrpages = vma_pages(vma);
1749
1750                 mm->total_vm -= nrpages;
1751                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1752                 vma = remove_vma(vma);
1753         } while (vma);
1754         validate_mm(mm);
1755 }
1756
1757 /*
1758  * Get rid of page table information in the indicated region.
1759  *
1760  * Called with the mm semaphore held.
1761  */
1762 static void unmap_region(struct mm_struct *mm,
1763                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1764                 unsigned long start, unsigned long end)
1765 {
1766         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1767         struct mmu_gather *tlb;
1768         unsigned long nr_accounted = 0;
1769
1770         lru_add_drain();
1771         tlb = tlb_gather_mmu(mm, 0);
1772         update_hiwater_rss(mm);
1773         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1774         vm_unacct_memory(nr_accounted);
1775         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1776                                  next? next->vm_start: 0);
1777         tlb_finish_mmu(tlb, start, end);
1778 }
1779
1780 /*
1781  * Create a list of vma's touched by the unmap, removing them from the mm's
1782  * vma list as we go..
1783  */
1784 static void
1785 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1786         struct vm_area_struct *prev, unsigned long end)
1787 {
1788         struct vm_area_struct **insertion_point;
1789         struct vm_area_struct *tail_vma = NULL;
1790         unsigned long addr;
1791
1792         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1793         do {
1794                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1795                 mm->map_count--;
1796                 tail_vma = vma;
1797                 vma = vma->vm_next;
1798         } while (vma && vma->vm_start < end);
1799         *insertion_point = vma;
1800         tail_vma->vm_next = NULL;
1801         if (mm->unmap_area == arch_unmap_area)
1802                 addr = prev ? prev->vm_end : mm->mmap_base;
1803         else
1804                 addr = vma ?  vma->vm_start : mm->mmap_base;
1805         mm->unmap_area(mm, addr);
1806         mm->mmap_cache = NULL;          /* Kill the cache. */
1807 }
1808
1809 /*
1810  * Split a vma into two pieces at address 'addr', a new vma is allocated
1811  * either for the first part or the tail.
1812  */
1813 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1814               unsigned long addr, int new_below)
1815 {
1816         struct mempolicy *pol;
1817         struct vm_area_struct *new;
1818
1819         if (is_vm_hugetlb_page(vma) && (addr &
1820                                         ~(huge_page_mask(hstate_vma(vma)))))
1821                 return -EINVAL;
1822
1823         if (mm->map_count >= sysctl_max_map_count)
1824                 return -ENOMEM;
1825
1826         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1827         if (!new)
1828                 return -ENOMEM;
1829
1830         /* most fields are the same, copy all, and then fixup */
1831         *new = *vma;
1832
1833         if (new_below)
1834                 new->vm_end = addr;
1835         else {
1836                 new->vm_start = addr;
1837                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1838         }
1839
1840         pol = mpol_dup(vma_policy(vma));
1841         if (IS_ERR(pol)) {
1842                 kmem_cache_free(vm_area_cachep, new);
1843                 return PTR_ERR(pol);
1844         }
1845         vma_set_policy(new, pol);
1846
1847         if (new->vm_file) {
1848                 get_file(new->vm_file);
1849                 if (vma->vm_flags & VM_EXECUTABLE)
1850                         added_exe_file_vma(mm);
1851         }
1852
1853         if (new->vm_ops && new->vm_ops->open)
1854                 new->vm_ops->open(new);
1855
1856         if (new_below)
1857                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1858                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1859         else
1860                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1861
1862         return 0;
1863 }
1864
1865 /* Munmap is split into 2 main parts -- this part which finds
1866  * what needs doing, and the areas themselves, which do the
1867  * work.  This now handles partial unmappings.
1868  * Jeremy Fitzhardinge <jeremy@goop.org>
1869  */
1870 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1871 {
1872         unsigned long end;
1873         struct vm_area_struct *vma, *prev, *last;
1874
1875         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1876                 return -EINVAL;
1877
1878         if ((len = PAGE_ALIGN(len)) == 0)
1879                 return -EINVAL;
1880
1881         /* Find the first overlapping VMA */
1882         vma = find_vma_prev(mm, start, &prev);
1883         if (!vma)
1884                 return 0;
1885         /* we have  start < vma->vm_end  */
1886
1887         /* if it doesn't overlap, we have nothing.. */
1888         end = start + len;
1889         if (vma->vm_start >= end)
1890                 return 0;
1891
1892         /*
1893          * If we need to split any vma, do it now to save pain later.
1894          *
1895          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1896          * unmapped vm_area_struct will remain in use: so lower split_vma
1897          * places tmp vma above, and higher split_vma places tmp vma below.
1898          */
1899         if (start > vma->vm_start) {
1900                 int error = split_vma(mm, vma, start, 0);
1901                 if (error)
1902                         return error;
1903                 prev = vma;
1904         }
1905
1906         /* Does it split the last one? */
1907         last = find_vma(mm, end);
1908         if (last && end > last->vm_start) {
1909                 int error = split_vma(mm, last, end, 1);
1910                 if (error)
1911                         return error;
1912         }
1913         vma = prev? prev->vm_next: mm->mmap;
1914
1915         /*
1916          * unlock any mlock()ed ranges before detaching vmas
1917          */
1918         if (mm->locked_vm) {
1919                 struct vm_area_struct *tmp = vma;
1920                 while (tmp && tmp->vm_start < end) {
1921                         if (tmp->vm_flags & VM_LOCKED) {
1922                                 mm->locked_vm -= vma_pages(tmp);
1923                                 munlock_vma_pages_all(tmp);
1924                         }
1925                         tmp = tmp->vm_next;
1926                 }
1927         }
1928
1929         /*
1930          * Remove the vma's, and unmap the actual pages
1931          */
1932         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1933         unmap_region(mm, vma, prev, start, end);
1934
1935         /* Fix up all other VM information */
1936         remove_vma_list(mm, vma);
1937
1938         return 0;
1939 }
1940
1941 EXPORT_SYMBOL(do_munmap);
1942
1943 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1944 {
1945         int ret;
1946         struct mm_struct *mm = current->mm;
1947
1948         profile_munmap(addr);
1949
1950         down_write(&mm->mmap_sem);
1951         ret = do_munmap(mm, addr, len);
1952         up_write(&mm->mmap_sem);
1953         return ret;
1954 }
1955
1956 static inline void verify_mm_writelocked(struct mm_struct *mm)
1957 {
1958 #ifdef CONFIG_DEBUG_VM
1959         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1960                 WARN_ON(1);
1961                 up_read(&mm->mmap_sem);
1962         }
1963 #endif
1964 }
1965
1966 /*
1967  *  this is really a simplified "do_mmap".  it only handles
1968  *  anonymous maps.  eventually we may be able to do some
1969  *  brk-specific accounting here.
1970  */
1971 unsigned long do_brk(unsigned long addr, unsigned long len)
1972 {
1973         struct mm_struct * mm = current->mm;
1974         struct vm_area_struct * vma, * prev;
1975         unsigned long flags;
1976         struct rb_node ** rb_link, * rb_parent;
1977         pgoff_t pgoff = addr >> PAGE_SHIFT;
1978         int error;
1979
1980         len = PAGE_ALIGN(len);
1981         if (!len)
1982                 return addr;
1983
1984         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1985                 return -EINVAL;
1986
1987         if (is_hugepage_only_range(mm, addr, len))
1988                 return -EINVAL;
1989
1990         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1991         if (error)
1992                 return error;
1993
1994         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1995
1996         error = arch_mmap_check(addr, len, flags);
1997         if (error)
1998                 return error;
1999
2000         /*
2001          * mlock MCL_FUTURE?
2002          */
2003         if (mm->def_flags & VM_LOCKED) {
2004                 unsigned long locked, lock_limit;
2005                 locked = len >> PAGE_SHIFT;
2006                 locked += mm->locked_vm;
2007                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2008                 lock_limit >>= PAGE_SHIFT;
2009                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2010                         return -EAGAIN;
2011         }
2012
2013         /*
2014          * mm->mmap_sem is required to protect against another thread
2015          * changing the mappings in case we sleep.
2016          */
2017         verify_mm_writelocked(mm);
2018
2019         /*
2020          * Clear old maps.  this also does some error checking for us
2021          */
2022  munmap_back:
2023         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2024         if (vma && vma->vm_start < addr + len) {
2025                 if (do_munmap(mm, addr, len))
2026                         return -ENOMEM;
2027                 goto munmap_back;
2028         }
2029
2030         /* Check against address space limits *after* clearing old maps... */
2031         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2032                 return -ENOMEM;
2033
2034         if (mm->map_count > sysctl_max_map_count)
2035                 return -ENOMEM;
2036
2037         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2038                 return -ENOMEM;
2039
2040         /* Can we just expand an old private anonymous mapping? */
2041         vma = vma_merge(mm, prev, addr, addr + len, flags,
2042                                         NULL, NULL, pgoff, NULL);
2043         if (vma)
2044                 goto out;
2045
2046         /*
2047          * create a vma struct for an anonymous mapping
2048          */
2049         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2050         if (!vma) {
2051                 vm_unacct_memory(len >> PAGE_SHIFT);
2052                 return -ENOMEM;
2053         }
2054
2055         vma->vm_mm = mm;
2056         vma->vm_start = addr;
2057         vma->vm_end = addr + len;
2058         vma->vm_pgoff = pgoff;
2059         vma->vm_flags = flags;
2060         vma->vm_page_prot = vm_get_page_prot(flags);
2061         vma_link(mm, vma, prev, rb_link, rb_parent);
2062 out:
2063         mm->total_vm += len >> PAGE_SHIFT;
2064         if (flags & VM_LOCKED) {
2065                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2066                         mm->locked_vm += (len >> PAGE_SHIFT);
2067         }
2068         return addr;
2069 }
2070
2071 EXPORT_SYMBOL(do_brk);
2072
2073 /* Release all mmaps. */
2074 void exit_mmap(struct mm_struct *mm)
2075 {
2076         struct mmu_gather *tlb;
2077         struct vm_area_struct *vma;
2078         unsigned long nr_accounted = 0;
2079         unsigned long end;
2080
2081         /* mm's last user has gone, and its about to be pulled down */
2082         mmu_notifier_release(mm);
2083
2084         if (mm->locked_vm) {
2085                 vma = mm->mmap;
2086                 while (vma) {
2087                         if (vma->vm_flags & VM_LOCKED)
2088                                 munlock_vma_pages_all(vma);
2089                         vma = vma->vm_next;
2090                 }
2091         }
2092
2093         arch_exit_mmap(mm);
2094
2095         vma = mm->mmap;
2096         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2097                 return;
2098
2099         lru_add_drain();
2100         flush_cache_mm(mm);
2101         tlb = tlb_gather_mmu(mm, 1);
2102         /* update_hiwater_rss(mm) here? but nobody should be looking */
2103         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2104         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2105         vm_unacct_memory(nr_accounted);
2106
2107         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2108         tlb_finish_mmu(tlb, 0, end);
2109
2110         /*
2111          * Walk the list again, actually closing and freeing it,
2112          * with preemption enabled, without holding any MM locks.
2113          */
2114         while (vma)
2115                 vma = remove_vma(vma);
2116
2117         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2118 }
2119
2120 /* Insert vm structure into process list sorted by address
2121  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2122  * then i_mmap_lock is taken here.
2123  */
2124 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2125 {
2126         struct vm_area_struct * __vma, * prev;
2127         struct rb_node ** rb_link, * rb_parent;
2128
2129         /*
2130          * The vm_pgoff of a purely anonymous vma should be irrelevant
2131          * until its first write fault, when page's anon_vma and index
2132          * are set.  But now set the vm_pgoff it will almost certainly
2133          * end up with (unless mremap moves it elsewhere before that
2134          * first wfault), so /proc/pid/maps tells a consistent story.
2135          *
2136          * By setting it to reflect the virtual start address of the
2137          * vma, merges and splits can happen in a seamless way, just
2138          * using the existing file pgoff checks and manipulations.
2139          * Similarly in do_mmap_pgoff and in do_brk.
2140          */
2141         if (!vma->vm_file) {
2142                 BUG_ON(vma->anon_vma);
2143                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2144         }
2145         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2146         if (__vma && __vma->vm_start < vma->vm_end)
2147                 return -ENOMEM;
2148         if ((vma->vm_flags & VM_ACCOUNT) &&
2149              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2150                 return -ENOMEM;
2151         vma_link(mm, vma, prev, rb_link, rb_parent);
2152         return 0;
2153 }
2154
2155 /*
2156  * Copy the vma structure to a new location in the same mm,
2157  * prior to moving page table entries, to effect an mremap move.
2158  */
2159 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2160         unsigned long addr, unsigned long len, pgoff_t pgoff)
2161 {
2162         struct vm_area_struct *vma = *vmap;
2163         unsigned long vma_start = vma->vm_start;
2164         struct mm_struct *mm = vma->vm_mm;
2165         struct vm_area_struct *new_vma, *prev;
2166         struct rb_node **rb_link, *rb_parent;
2167         struct mempolicy *pol;
2168
2169         /*
2170          * If anonymous vma has not yet been faulted, update new pgoff
2171          * to match new location, to increase its chance of merging.
2172          */
2173         if (!vma->vm_file && !vma->anon_vma)
2174                 pgoff = addr >> PAGE_SHIFT;
2175
2176         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2177         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2178                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2179         if (new_vma) {
2180                 /*
2181                  * Source vma may have been merged into new_vma
2182                  */
2183                 if (vma_start >= new_vma->vm_start &&
2184                     vma_start < new_vma->vm_end)
2185                         *vmap = new_vma;
2186         } else {
2187                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2188                 if (new_vma) {
2189                         *new_vma = *vma;
2190                         pol = mpol_dup(vma_policy(vma));
2191                         if (IS_ERR(pol)) {
2192                                 kmem_cache_free(vm_area_cachep, new_vma);
2193                                 return NULL;
2194                         }
2195                         vma_set_policy(new_vma, pol);
2196                         new_vma->vm_start = addr;
2197                         new_vma->vm_end = addr + len;
2198                         new_vma->vm_pgoff = pgoff;
2199                         if (new_vma->vm_file) {
2200                                 get_file(new_vma->vm_file);
2201                                 if (vma->vm_flags & VM_EXECUTABLE)
2202                                         added_exe_file_vma(mm);
2203                         }
2204                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2205                                 new_vma->vm_ops->open(new_vma);
2206                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2207                 }
2208         }
2209         return new_vma;
2210 }
2211
2212 /*
2213  * Return true if the calling process may expand its vm space by the passed
2214  * number of pages
2215  */
2216 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2217 {
2218         unsigned long cur = mm->total_vm;       /* pages */
2219         unsigned long lim;
2220
2221         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2222
2223         if (cur + npages > lim)
2224                 return 0;
2225         return 1;
2226 }
2227
2228
2229 static int special_mapping_fault(struct vm_area_struct *vma,
2230                                 struct vm_fault *vmf)
2231 {
2232         pgoff_t pgoff;
2233         struct page **pages;
2234
2235         /*
2236          * special mappings have no vm_file, and in that case, the mm
2237          * uses vm_pgoff internally. So we have to subtract it from here.
2238          * We are allowed to do this because we are the mm; do not copy
2239          * this code into drivers!
2240          */
2241         pgoff = vmf->pgoff - vma->vm_pgoff;
2242
2243         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2244                 pgoff--;
2245
2246         if (*pages) {
2247                 struct page *page = *pages;
2248                 get_page(page);
2249                 vmf->page = page;
2250                 return 0;
2251         }
2252
2253         return VM_FAULT_SIGBUS;
2254 }
2255
2256 /*
2257  * Having a close hook prevents vma merging regardless of flags.
2258  */
2259 static void special_mapping_close(struct vm_area_struct *vma)
2260 {
2261 }
2262
2263 static const struct vm_operations_struct special_mapping_vmops = {
2264         .close = special_mapping_close,
2265         .fault = special_mapping_fault,
2266 };
2267
2268 /*
2269  * Called with mm->mmap_sem held for writing.
2270  * Insert a new vma covering the given region, with the given flags.
2271  * Its pages are supplied by the given array of struct page *.
2272  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2273  * The region past the last page supplied will always produce SIGBUS.
2274  * The array pointer and the pages it points to are assumed to stay alive
2275  * for as long as this mapping might exist.
2276  */
2277 int install_special_mapping(struct mm_struct *mm,
2278                             unsigned long addr, unsigned long len,
2279                             unsigned long vm_flags, struct page **pages)
2280 {
2281         struct vm_area_struct *vma;
2282
2283         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2284         if (unlikely(vma == NULL))
2285                 return -ENOMEM;
2286
2287         vma->vm_mm = mm;
2288         vma->vm_start = addr;
2289         vma->vm_end = addr + len;
2290
2291         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2292         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2293
2294         vma->vm_ops = &special_mapping_vmops;
2295         vma->vm_private_data = pages;
2296
2297         if (unlikely(insert_vm_struct(mm, vma))) {
2298                 kmem_cache_free(vm_area_cachep, vma);
2299                 return -ENOMEM;
2300         }
2301
2302         mm->total_vm += len >> PAGE_SHIFT;
2303
2304         perf_event_mmap(vma);
2305
2306         return 0;
2307 }
2308
2309 static DEFINE_MUTEX(mm_all_locks_mutex);
2310
2311 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2312 {
2313         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2314                 /*
2315                  * The LSB of head.next can't change from under us
2316                  * because we hold the mm_all_locks_mutex.
2317                  */
2318                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2319                 /*
2320                  * We can safely modify head.next after taking the
2321                  * anon_vma->lock. If some other vma in this mm shares
2322                  * the same anon_vma we won't take it again.
2323                  *
2324                  * No need of atomic instructions here, head.next
2325                  * can't change from under us thanks to the
2326                  * anon_vma->lock.
2327                  */
2328                 if (__test_and_set_bit(0, (unsigned long *)
2329                                        &anon_vma->head.next))
2330                         BUG();
2331         }
2332 }
2333
2334 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2335 {
2336         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2337                 /*
2338                  * AS_MM_ALL_LOCKS can't change from under us because
2339                  * we hold the mm_all_locks_mutex.
2340                  *
2341                  * Operations on ->flags have to be atomic because
2342                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2343                  * mm_all_locks_mutex, there may be other cpus
2344                  * changing other bitflags in parallel to us.
2345                  */
2346                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2347                         BUG();
2348                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2349         }
2350 }
2351
2352 /*
2353  * This operation locks against the VM for all pte/vma/mm related
2354  * operations that could ever happen on a certain mm. This includes
2355  * vmtruncate, try_to_unmap, and all page faults.
2356  *
2357  * The caller must take the mmap_sem in write mode before calling
2358  * mm_take_all_locks(). The caller isn't allowed to release the
2359  * mmap_sem until mm_drop_all_locks() returns.
2360  *
2361  * mmap_sem in write mode is required in order to block all operations
2362  * that could modify pagetables and free pages without need of
2363  * altering the vma layout (for example populate_range() with
2364  * nonlinear vmas). It's also needed in write mode to avoid new
2365  * anon_vmas to be associated with existing vmas.
2366  *
2367  * A single task can't take more than one mm_take_all_locks() in a row
2368  * or it would deadlock.
2369  *
2370  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2371  * mapping->flags avoid to take the same lock twice, if more than one
2372  * vma in this mm is backed by the same anon_vma or address_space.
2373  *
2374  * We can take all the locks in random order because the VM code
2375  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2376  * takes more than one of them in a row. Secondly we're protected
2377  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2378  *
2379  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2380  * that may have to take thousand of locks.
2381  *
2382  * mm_take_all_locks() can fail if it's interrupted by signals.
2383  */
2384 int mm_take_all_locks(struct mm_struct *mm)
2385 {
2386         struct vm_area_struct *vma;
2387         int ret = -EINTR;
2388
2389         BUG_ON(down_read_trylock(&mm->mmap_sem));
2390
2391         mutex_lock(&mm_all_locks_mutex);
2392
2393         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2394                 if (signal_pending(current))
2395                         goto out_unlock;
2396                 if (vma->vm_file && vma->vm_file->f_mapping)
2397                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2398         }
2399
2400         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2401                 if (signal_pending(current))
2402                         goto out_unlock;
2403                 if (vma->anon_vma)
2404                         vm_lock_anon_vma(mm, vma->anon_vma);
2405         }
2406
2407         ret = 0;
2408
2409 out_unlock:
2410         if (ret)
2411                 mm_drop_all_locks(mm);
2412
2413         return ret;
2414 }
2415
2416 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2417 {
2418         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2419                 /*
2420                  * The LSB of head.next can't change to 0 from under
2421                  * us because we hold the mm_all_locks_mutex.
2422                  *
2423                  * We must however clear the bitflag before unlocking
2424                  * the vma so the users using the anon_vma->head will
2425                  * never see our bitflag.
2426                  *
2427                  * No need of atomic instructions here, head.next
2428                  * can't change from under us until we release the
2429                  * anon_vma->lock.
2430                  */
2431                 if (!__test_and_clear_bit(0, (unsigned long *)
2432                                           &anon_vma->head.next))
2433                         BUG();
2434                 spin_unlock(&anon_vma->lock);
2435         }
2436 }
2437
2438 static void vm_unlock_mapping(struct address_space *mapping)
2439 {
2440         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2441                 /*
2442                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2443                  * because we hold the mm_all_locks_mutex.
2444                  */
2445                 spin_unlock(&mapping->i_mmap_lock);
2446                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2447                                         &mapping->flags))
2448                         BUG();
2449         }
2450 }
2451
2452 /*
2453  * The mmap_sem cannot be released by the caller until
2454  * mm_drop_all_locks() returns.
2455  */
2456 void mm_drop_all_locks(struct mm_struct *mm)
2457 {
2458         struct vm_area_struct *vma;
2459
2460         BUG_ON(down_read_trylock(&mm->mmap_sem));
2461         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2462
2463         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2464                 if (vma->anon_vma)
2465                         vm_unlock_anon_vma(vma->anon_vma);
2466                 if (vma->vm_file && vma->vm_file->f_mapping)
2467                         vm_unlock_mapping(vma->vm_file->f_mapping);
2468         }
2469
2470         mutex_unlock(&mm_all_locks_mutex);
2471 }
2472
2473 /*
2474  * initialise the VMA slab
2475  */
2476 void __init mmap_init(void)
2477 {
2478         int ret;
2479
2480         ret = percpu_counter_init(&vm_committed_as, 0);
2481         VM_BUG_ON(ret);
2482 }