perf probe: Fix need_dwarf flag if lazy matching is used
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/module.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31
32 #include <asm/uaccess.h>
33 #include <asm/cacheflush.h>
34 #include <asm/tlb.h>
35 #include <asm/mmu_context.h>
36
37 #include "internal.h"
38
39 #ifndef arch_mmap_check
40 #define arch_mmap_check(addr, len, flags)       (0)
41 #endif
42
43 #ifndef arch_rebalance_pgtables
44 #define arch_rebalance_pgtables(addr, len)              (addr)
45 #endif
46
47 static void unmap_region(struct mm_struct *mm,
48                 struct vm_area_struct *vma, struct vm_area_struct *prev,
49                 unsigned long start, unsigned long end);
50
51 /*
52  * WARNING: the debugging will use recursive algorithms so never enable this
53  * unless you know what you are doing.
54  */
55 #undef DEBUG_MM_RB
56
57 /* description of effects of mapping type and prot in current implementation.
58  * this is due to the limited x86 page protection hardware.  The expected
59  * behavior is in parens:
60  *
61  * map_type     prot
62  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
63  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
64  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
65  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
66  *              
67  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
68  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
69  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
70  *
71  */
72 pgprot_t protection_map[16] = {
73         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
74         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
75 };
76
77 pgprot_t vm_get_page_prot(unsigned long vm_flags)
78 {
79         return __pgprot(pgprot_val(protection_map[vm_flags &
80                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
81                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
82 }
83 EXPORT_SYMBOL(vm_get_page_prot);
84
85 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
86 int sysctl_overcommit_ratio = 50;       /* default is 50% */
87 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
88 struct percpu_counter vm_committed_as;
89
90 /*
91  * Check that a process has enough memory to allocate a new virtual
92  * mapping. 0 means there is enough memory for the allocation to
93  * succeed and -ENOMEM implies there is not.
94  *
95  * We currently support three overcommit policies, which are set via the
96  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
97  *
98  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
99  * Additional code 2002 Jul 20 by Robert Love.
100  *
101  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
102  *
103  * Note this is a helper function intended to be used by LSMs which
104  * wish to use this logic.
105  */
106 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
107 {
108         unsigned long free, allowed;
109
110         vm_acct_memory(pages);
111
112         /*
113          * Sometimes we want to use more memory than we have
114          */
115         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
116                 return 0;
117
118         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
119                 unsigned long n;
120
121                 free = global_page_state(NR_FILE_PAGES);
122                 free += nr_swap_pages;
123
124                 /*
125                  * Any slabs which are created with the
126                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
127                  * which are reclaimable, under pressure.  The dentry
128                  * cache and most inode caches should fall into this
129                  */
130                 free += global_page_state(NR_SLAB_RECLAIMABLE);
131
132                 /*
133                  * Leave the last 3% for root
134                  */
135                 if (!cap_sys_admin)
136                         free -= free / 32;
137
138                 if (free > pages)
139                         return 0;
140
141                 /*
142                  * nr_free_pages() is very expensive on large systems,
143                  * only call if we're about to fail.
144                  */
145                 n = nr_free_pages();
146
147                 /*
148                  * Leave reserved pages. The pages are not for anonymous pages.
149                  */
150                 if (n <= totalreserve_pages)
151                         goto error;
152                 else
153                         n -= totalreserve_pages;
154
155                 /*
156                  * Leave the last 3% for root
157                  */
158                 if (!cap_sys_admin)
159                         n -= n / 32;
160                 free += n;
161
162                 if (free > pages)
163                         return 0;
164
165                 goto error;
166         }
167
168         allowed = (totalram_pages - hugetlb_total_pages())
169                 * sysctl_overcommit_ratio / 100;
170         /*
171          * Leave the last 3% for root
172          */
173         if (!cap_sys_admin)
174                 allowed -= allowed / 32;
175         allowed += total_swap_pages;
176
177         /* Don't let a single process grow too big:
178            leave 3% of the size of this process for other processes */
179         if (mm)
180                 allowed -= mm->total_vm / 32;
181
182         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
183                 return 0;
184 error:
185         vm_unacct_memory(pages);
186
187         return -ENOMEM;
188 }
189
190 /*
191  * Requires inode->i_mapping->i_mmap_lock
192  */
193 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
194                 struct file *file, struct address_space *mapping)
195 {
196         if (vma->vm_flags & VM_DENYWRITE)
197                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
198         if (vma->vm_flags & VM_SHARED)
199                 mapping->i_mmap_writable--;
200
201         flush_dcache_mmap_lock(mapping);
202         if (unlikely(vma->vm_flags & VM_NONLINEAR))
203                 list_del_init(&vma->shared.vm_set.list);
204         else
205                 vma_prio_tree_remove(vma, &mapping->i_mmap);
206         flush_dcache_mmap_unlock(mapping);
207 }
208
209 /*
210  * Unlink a file-based vm structure from its prio_tree, to hide
211  * vma from rmap and vmtruncate before freeing its page tables.
212  */
213 void unlink_file_vma(struct vm_area_struct *vma)
214 {
215         struct file *file = vma->vm_file;
216
217         if (file) {
218                 struct address_space *mapping = file->f_mapping;
219                 spin_lock(&mapping->i_mmap_lock);
220                 __remove_shared_vm_struct(vma, file, mapping);
221                 spin_unlock(&mapping->i_mmap_lock);
222         }
223 }
224
225 /*
226  * Close a vm structure and free it, returning the next.
227  */
228 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
229 {
230         struct vm_area_struct *next = vma->vm_next;
231
232         might_sleep();
233         if (vma->vm_ops && vma->vm_ops->close)
234                 vma->vm_ops->close(vma);
235         if (vma->vm_file) {
236                 fput(vma->vm_file);
237                 if (vma->vm_flags & VM_EXECUTABLE)
238                         removed_exe_file_vma(vma->vm_mm);
239         }
240         mpol_put(vma_policy(vma));
241         kmem_cache_free(vm_area_cachep, vma);
242         return next;
243 }
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         min_brk = mm->end_code;
256 #else
257         min_brk = mm->start_brk;
258 #endif
259         if (brk < min_brk)
260                 goto out;
261
262         /*
263          * Check against rlimit here. If this check is done later after the test
264          * of oldbrk with newbrk then it can escape the test and let the data
265          * segment grow beyond its set limit the in case where the limit is
266          * not page aligned -Ram Gupta
267          */
268         rlim = rlimit(RLIMIT_DATA);
269         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
270                         (mm->end_data - mm->start_data) > rlim)
271                 goto out;
272
273         newbrk = PAGE_ALIGN(brk);
274         oldbrk = PAGE_ALIGN(mm->brk);
275         if (oldbrk == newbrk)
276                 goto set_brk;
277
278         /* Always allow shrinking brk. */
279         if (brk <= mm->brk) {
280                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
281                         goto set_brk;
282                 goto out;
283         }
284
285         /* Check against existing mmap mappings. */
286         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
287                 goto out;
288
289         /* Ok, looks good - let it rip. */
290         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
291                 goto out;
292 set_brk:
293         mm->brk = brk;
294 out:
295         retval = mm->brk;
296         up_write(&mm->mmap_sem);
297         return retval;
298 }
299
300 #ifdef DEBUG_MM_RB
301 static int browse_rb(struct rb_root *root)
302 {
303         int i = 0, j;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev)
311                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
312                 if (vma->vm_start < pend)
313                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
314                 if (vma->vm_start > vma->vm_end)
315                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
316                 i++;
317                 pn = nd;
318                 prev = vma->vm_start;
319                 pend = vma->vm_end;
320         }
321         j = 0;
322         for (nd = pn; nd; nd = rb_prev(nd)) {
323                 j++;
324         }
325         if (i != j)
326                 printk("backwards %d, forwards %d\n", j, i), i = 0;
327         return i;
328 }
329
330 void validate_mm(struct mm_struct *mm)
331 {
332         int bug = 0;
333         int i = 0;
334         struct vm_area_struct *tmp = mm->mmap;
335         while (tmp) {
336                 tmp = tmp->vm_next;
337                 i++;
338         }
339         if (i != mm->map_count)
340                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
341         i = browse_rb(&mm->mm_rb);
342         if (i != mm->map_count)
343                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
344         BUG_ON(bug);
345 }
346 #else
347 #define validate_mm(mm) do { } while (0)
348 #endif
349
350 static struct vm_area_struct *
351 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
352                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
353                 struct rb_node ** rb_parent)
354 {
355         struct vm_area_struct * vma;
356         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
357
358         __rb_link = &mm->mm_rb.rb_node;
359         rb_prev = __rb_parent = NULL;
360         vma = NULL;
361
362         while (*__rb_link) {
363                 struct vm_area_struct *vma_tmp;
364
365                 __rb_parent = *__rb_link;
366                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
367
368                 if (vma_tmp->vm_end > addr) {
369                         vma = vma_tmp;
370                         if (vma_tmp->vm_start <= addr)
371                                 break;
372                         __rb_link = &__rb_parent->rb_left;
373                 } else {
374                         rb_prev = __rb_parent;
375                         __rb_link = &__rb_parent->rb_right;
376                 }
377         }
378
379         *pprev = NULL;
380         if (rb_prev)
381                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
382         *rb_link = __rb_link;
383         *rb_parent = __rb_parent;
384         return vma;
385 }
386
387 static inline void
388 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
389                 struct vm_area_struct *prev, struct rb_node *rb_parent)
390 {
391         if (prev) {
392                 vma->vm_next = prev->vm_next;
393                 prev->vm_next = vma;
394         } else {
395                 mm->mmap = vma;
396                 if (rb_parent)
397                         vma->vm_next = rb_entry(rb_parent,
398                                         struct vm_area_struct, vm_rb);
399                 else
400                         vma->vm_next = NULL;
401         }
402 }
403
404 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
405                 struct rb_node **rb_link, struct rb_node *rb_parent)
406 {
407         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
408         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
409 }
410
411 static void __vma_link_file(struct vm_area_struct *vma)
412 {
413         struct file *file;
414
415         file = vma->vm_file;
416         if (file) {
417                 struct address_space *mapping = file->f_mapping;
418
419                 if (vma->vm_flags & VM_DENYWRITE)
420                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
421                 if (vma->vm_flags & VM_SHARED)
422                         mapping->i_mmap_writable++;
423
424                 flush_dcache_mmap_lock(mapping);
425                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
426                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
427                 else
428                         vma_prio_tree_insert(vma, &mapping->i_mmap);
429                 flush_dcache_mmap_unlock(mapping);
430         }
431 }
432
433 static void
434 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
435         struct vm_area_struct *prev, struct rb_node **rb_link,
436         struct rb_node *rb_parent)
437 {
438         __vma_link_list(mm, vma, prev, rb_parent);
439         __vma_link_rb(mm, vma, rb_link, rb_parent);
440 }
441
442 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
443                         struct vm_area_struct *prev, struct rb_node **rb_link,
444                         struct rb_node *rb_parent)
445 {
446         struct address_space *mapping = NULL;
447
448         if (vma->vm_file)
449                 mapping = vma->vm_file->f_mapping;
450
451         if (mapping) {
452                 spin_lock(&mapping->i_mmap_lock);
453                 vma->vm_truncate_count = mapping->truncate_count;
454         }
455         anon_vma_lock(vma);
456
457         __vma_link(mm, vma, prev, rb_link, rb_parent);
458         __vma_link_file(vma);
459
460         anon_vma_unlock(vma);
461         if (mapping)
462                 spin_unlock(&mapping->i_mmap_lock);
463
464         mm->map_count++;
465         validate_mm(mm);
466 }
467
468 /*
469  * Helper for vma_adjust in the split_vma insert case:
470  * insert vm structure into list and rbtree and anon_vma,
471  * but it has already been inserted into prio_tree earlier.
472  */
473 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
474 {
475         struct vm_area_struct *__vma, *prev;
476         struct rb_node **rb_link, *rb_parent;
477
478         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
479         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
480         __vma_link(mm, vma, prev, rb_link, rb_parent);
481         mm->map_count++;
482 }
483
484 static inline void
485 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
486                 struct vm_area_struct *prev)
487 {
488         prev->vm_next = vma->vm_next;
489         rb_erase(&vma->vm_rb, &mm->mm_rb);
490         if (mm->mmap_cache == vma)
491                 mm->mmap_cache = prev;
492 }
493
494 /*
495  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
496  * is already present in an i_mmap tree without adjusting the tree.
497  * The following helper function should be used when such adjustments
498  * are necessary.  The "insert" vma (if any) is to be inserted
499  * before we drop the necessary locks.
500  */
501 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
502         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
503 {
504         struct mm_struct *mm = vma->vm_mm;
505         struct vm_area_struct *next = vma->vm_next;
506         struct vm_area_struct *importer = NULL;
507         struct address_space *mapping = NULL;
508         struct prio_tree_root *root = NULL;
509         struct file *file = vma->vm_file;
510         struct anon_vma *anon_vma = NULL;
511         long adjust_next = 0;
512         int remove_next = 0;
513
514         if (next && !insert) {
515                 if (end >= next->vm_end) {
516                         /*
517                          * vma expands, overlapping all the next, and
518                          * perhaps the one after too (mprotect case 6).
519                          */
520 again:                  remove_next = 1 + (end > next->vm_end);
521                         end = next->vm_end;
522                         anon_vma = next->anon_vma;
523                         importer = vma;
524                 } else if (end > next->vm_start) {
525                         /*
526                          * vma expands, overlapping part of the next:
527                          * mprotect case 5 shifting the boundary up.
528                          */
529                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
530                         anon_vma = next->anon_vma;
531                         importer = vma;
532                 } else if (end < vma->vm_end) {
533                         /*
534                          * vma shrinks, and !insert tells it's not
535                          * split_vma inserting another: so it must be
536                          * mprotect case 4 shifting the boundary down.
537                          */
538                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
539                         anon_vma = next->anon_vma;
540                         importer = next;
541                 }
542         }
543
544         /*
545          * When changing only vma->vm_end, we don't really need anon_vma lock.
546          */
547         if (vma->anon_vma && (insert || importer || start != vma->vm_start))
548                 anon_vma = vma->anon_vma;
549         if (anon_vma) {
550                 /*
551                  * Easily overlooked: when mprotect shifts the boundary,
552                  * make sure the expanding vma has anon_vma set if the
553                  * shrinking vma had, to cover any anon pages imported.
554                  */
555                 if (importer && !importer->anon_vma) {
556                         /* Block reverse map lookups until things are set up. */
557                         if (anon_vma_clone(importer, vma)) {
558                                 return -ENOMEM;
559                         }
560                         importer->anon_vma = anon_vma;
561                 }
562         }
563
564         if (file) {
565                 mapping = file->f_mapping;
566                 if (!(vma->vm_flags & VM_NONLINEAR))
567                         root = &mapping->i_mmap;
568                 spin_lock(&mapping->i_mmap_lock);
569                 if (importer &&
570                     vma->vm_truncate_count != next->vm_truncate_count) {
571                         /*
572                          * unmap_mapping_range might be in progress:
573                          * ensure that the expanding vma is rescanned.
574                          */
575                         importer->vm_truncate_count = 0;
576                 }
577                 if (insert) {
578                         insert->vm_truncate_count = vma->vm_truncate_count;
579                         /*
580                          * Put into prio_tree now, so instantiated pages
581                          * are visible to arm/parisc __flush_dcache_page
582                          * throughout; but we cannot insert into address
583                          * space until vma start or end is updated.
584                          */
585                         __vma_link_file(insert);
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         } else if (insert) {
620                 /*
621                  * split_vma has split insert from vma, and needs
622                  * us to insert it before dropping the locks
623                  * (it may either follow vma or precede it).
624                  */
625                 __insert_vm_struct(mm, insert);
626         }
627
628         if (mapping)
629                 spin_unlock(&mapping->i_mmap_lock);
630
631         if (remove_next) {
632                 if (file) {
633                         fput(file);
634                         if (next->vm_flags & VM_EXECUTABLE)
635                                 removed_exe_file_vma(mm);
636                 }
637                 if (next->anon_vma)
638                         anon_vma_merge(vma, next);
639                 mm->map_count--;
640                 mpol_put(vma_policy(next));
641                 kmem_cache_free(vm_area_cachep, next);
642                 /*
643                  * In mprotect's case 6 (see comments on vma_merge),
644                  * we must remove another next too. It would clutter
645                  * up the code too much to do both in one go.
646                  */
647                 if (remove_next == 2) {
648                         next = vma->vm_next;
649                         goto again;
650                 }
651         }
652
653         validate_mm(mm);
654
655         return 0;
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         int err;
763
764         /*
765          * We later require that vma->vm_flags == vm_flags,
766          * so this tests vma->vm_flags & VM_SPECIAL, too.
767          */
768         if (vm_flags & VM_SPECIAL)
769                 return NULL;
770
771         if (prev)
772                 next = prev->vm_next;
773         else
774                 next = mm->mmap;
775         area = next;
776         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
777                 next = next->vm_next;
778
779         /*
780          * Can it merge with the predecessor?
781          */
782         if (prev && prev->vm_end == addr &&
783                         mpol_equal(vma_policy(prev), policy) &&
784                         can_vma_merge_after(prev, vm_flags,
785                                                 anon_vma, file, pgoff)) {
786                 /*
787                  * OK, it can.  Can we now merge in the successor as well?
788                  */
789                 if (next && end == next->vm_start &&
790                                 mpol_equal(policy, vma_policy(next)) &&
791                                 can_vma_merge_before(next, vm_flags,
792                                         anon_vma, file, pgoff+pglen) &&
793                                 is_mergeable_anon_vma(prev->anon_vma,
794                                                       next->anon_vma)) {
795                                                         /* cases 1, 6 */
796                         err = vma_adjust(prev, prev->vm_start,
797                                 next->vm_end, prev->vm_pgoff, NULL);
798                 } else                                  /* cases 2, 5, 7 */
799                         err = vma_adjust(prev, prev->vm_start,
800                                 end, prev->vm_pgoff, NULL);
801                 if (err)
802                         return NULL;
803                 return prev;
804         }
805
806         /*
807          * Can this new request be merged in front of next?
808          */
809         if (next && end == next->vm_start &&
810                         mpol_equal(policy, vma_policy(next)) &&
811                         can_vma_merge_before(next, vm_flags,
812                                         anon_vma, file, pgoff+pglen)) {
813                 if (prev && addr < prev->vm_end)        /* case 4 */
814                         err = vma_adjust(prev, prev->vm_start,
815                                 addr, prev->vm_pgoff, NULL);
816                 else                                    /* cases 3, 8 */
817                         err = vma_adjust(area, addr, next->vm_end,
818                                 next->vm_pgoff - pglen, NULL);
819                 if (err)
820                         return NULL;
821                 return area;
822         }
823
824         return NULL;
825 }
826
827 /*
828  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
829  * neighbouring vmas for a suitable anon_vma, before it goes off
830  * to allocate a new anon_vma.  It checks because a repetitive
831  * sequence of mprotects and faults may otherwise lead to distinct
832  * anon_vmas being allocated, preventing vma merge in subsequent
833  * mprotect.
834  */
835 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
836 {
837         struct vm_area_struct *near;
838         unsigned long vm_flags;
839
840         near = vma->vm_next;
841         if (!near)
842                 goto try_prev;
843
844         /*
845          * Since only mprotect tries to remerge vmas, match flags
846          * which might be mprotected into each other later on.
847          * Neither mlock nor madvise tries to remerge at present,
848          * so leave their flags as obstructing a merge.
849          */
850         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
851         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
852
853         if (near->anon_vma && vma->vm_end == near->vm_start &&
854                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
855                         can_vma_merge_before(near, vm_flags,
856                                 NULL, vma->vm_file, vma->vm_pgoff +
857                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
858                 return near->anon_vma;
859 try_prev:
860         /*
861          * It is potentially slow to have to call find_vma_prev here.
862          * But it's only on the first write fault on the vma, not
863          * every time, and we could devise a way to avoid it later
864          * (e.g. stash info in next's anon_vma_node when assigning
865          * an anon_vma, or when trying vma_merge).  Another time.
866          */
867         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
868         if (!near)
869                 goto none;
870
871         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
872         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
873
874         if (near->anon_vma && near->vm_end == vma->vm_start &&
875                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
876                         can_vma_merge_after(near, vm_flags,
877                                 NULL, vma->vm_file, vma->vm_pgoff))
878                 return near->anon_vma;
879 none:
880         /*
881          * There's no absolute need to look only at touching neighbours:
882          * we could search further afield for "compatible" anon_vmas.
883          * But it would probably just be a waste of time searching,
884          * or lead to too many vmas hanging off the same anon_vma.
885          * We're trying to allow mprotect remerging later on,
886          * not trying to minimize memory used for anon_vmas.
887          */
888         return NULL;
889 }
890
891 #ifdef CONFIG_PROC_FS
892 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
893                                                 struct file *file, long pages)
894 {
895         const unsigned long stack_flags
896                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
897
898         if (file) {
899                 mm->shared_vm += pages;
900                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
901                         mm->exec_vm += pages;
902         } else if (flags & stack_flags)
903                 mm->stack_vm += pages;
904         if (flags & (VM_RESERVED|VM_IO))
905                 mm->reserved_vm += pages;
906 }
907 #endif /* CONFIG_PROC_FS */
908
909 /*
910  * The caller must hold down_write(&current->mm->mmap_sem).
911  */
912
913 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
914                         unsigned long len, unsigned long prot,
915                         unsigned long flags, unsigned long pgoff)
916 {
917         struct mm_struct * mm = current->mm;
918         struct inode *inode;
919         unsigned int vm_flags;
920         int error;
921         unsigned long reqprot = prot;
922
923         /*
924          * Does the application expect PROT_READ to imply PROT_EXEC?
925          *
926          * (the exception is when the underlying filesystem is noexec
927          *  mounted, in which case we dont add PROT_EXEC.)
928          */
929         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
930                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
931                         prot |= PROT_EXEC;
932
933         if (!len)
934                 return -EINVAL;
935
936         if (!(flags & MAP_FIXED))
937                 addr = round_hint_to_min(addr);
938
939         /* Careful about overflows.. */
940         len = PAGE_ALIGN(len);
941         if (!len)
942                 return -ENOMEM;
943
944         /* offset overflow? */
945         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
946                return -EOVERFLOW;
947
948         /* Too many mappings? */
949         if (mm->map_count > sysctl_max_map_count)
950                 return -ENOMEM;
951
952         /* Obtain the address to map to. we verify (or select) it and ensure
953          * that it represents a valid section of the address space.
954          */
955         addr = get_unmapped_area(file, addr, len, pgoff, flags);
956         if (addr & ~PAGE_MASK)
957                 return addr;
958
959         /* Do simple checking here so the lower-level routines won't have
960          * to. we assume access permissions have been handled by the open
961          * of the memory object, so we don't do any here.
962          */
963         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
964                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
965
966         if (flags & MAP_LOCKED)
967                 if (!can_do_mlock())
968                         return -EPERM;
969
970         /* mlock MCL_FUTURE? */
971         if (vm_flags & VM_LOCKED) {
972                 unsigned long locked, lock_limit;
973                 locked = len >> PAGE_SHIFT;
974                 locked += mm->locked_vm;
975                 lock_limit = rlimit(RLIMIT_MEMLOCK);
976                 lock_limit >>= PAGE_SHIFT;
977                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
978                         return -EAGAIN;
979         }
980
981         inode = file ? file->f_path.dentry->d_inode : NULL;
982
983         if (file) {
984                 switch (flags & MAP_TYPE) {
985                 case MAP_SHARED:
986                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
987                                 return -EACCES;
988
989                         /*
990                          * Make sure we don't allow writing to an append-only
991                          * file..
992                          */
993                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
994                                 return -EACCES;
995
996                         /*
997                          * Make sure there are no mandatory locks on the file.
998                          */
999                         if (locks_verify_locked(inode))
1000                                 return -EAGAIN;
1001
1002                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1003                         if (!(file->f_mode & FMODE_WRITE))
1004                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1005
1006                         /* fall through */
1007                 case MAP_PRIVATE:
1008                         if (!(file->f_mode & FMODE_READ))
1009                                 return -EACCES;
1010                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1011                                 if (vm_flags & VM_EXEC)
1012                                         return -EPERM;
1013                                 vm_flags &= ~VM_MAYEXEC;
1014                         }
1015
1016                         if (!file->f_op || !file->f_op->mmap)
1017                                 return -ENODEV;
1018                         break;
1019
1020                 default:
1021                         return -EINVAL;
1022                 }
1023         } else {
1024                 switch (flags & MAP_TYPE) {
1025                 case MAP_SHARED:
1026                         /*
1027                          * Ignore pgoff.
1028                          */
1029                         pgoff = 0;
1030                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1031                         break;
1032                 case MAP_PRIVATE:
1033                         /*
1034                          * Set pgoff according to addr for anon_vma.
1035                          */
1036                         pgoff = addr >> PAGE_SHIFT;
1037                         break;
1038                 default:
1039                         return -EINVAL;
1040                 }
1041         }
1042
1043         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1044         if (error)
1045                 return error;
1046
1047         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1048 }
1049 EXPORT_SYMBOL(do_mmap_pgoff);
1050
1051 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1052                 unsigned long, prot, unsigned long, flags,
1053                 unsigned long, fd, unsigned long, pgoff)
1054 {
1055         struct file *file = NULL;
1056         unsigned long retval = -EBADF;
1057
1058         if (!(flags & MAP_ANONYMOUS)) {
1059                 if (unlikely(flags & MAP_HUGETLB))
1060                         return -EINVAL;
1061                 file = fget(fd);
1062                 if (!file)
1063                         goto out;
1064         } else if (flags & MAP_HUGETLB) {
1065                 struct user_struct *user = NULL;
1066                 /*
1067                  * VM_NORESERVE is used because the reservations will be
1068                  * taken when vm_ops->mmap() is called
1069                  * A dummy user value is used because we are not locking
1070                  * memory so no accounting is necessary
1071                  */
1072                 len = ALIGN(len, huge_page_size(&default_hstate));
1073                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
1074                                                 &user, HUGETLB_ANONHUGE_INODE);
1075                 if (IS_ERR(file))
1076                         return PTR_ERR(file);
1077         }
1078
1079         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1080
1081         down_write(&current->mm->mmap_sem);
1082         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1083         up_write(&current->mm->mmap_sem);
1084
1085         if (file)
1086                 fput(file);
1087 out:
1088         return retval;
1089 }
1090
1091 /*
1092  * Some shared mappigns will want the pages marked read-only
1093  * to track write events. If so, we'll downgrade vm_page_prot
1094  * to the private version (using protection_map[] without the
1095  * VM_SHARED bit).
1096  */
1097 int vma_wants_writenotify(struct vm_area_struct *vma)
1098 {
1099         unsigned int vm_flags = vma->vm_flags;
1100
1101         /* If it was private or non-writable, the write bit is already clear */
1102         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1103                 return 0;
1104
1105         /* The backer wishes to know when pages are first written to? */
1106         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1107                 return 1;
1108
1109         /* The open routine did something to the protections already? */
1110         if (pgprot_val(vma->vm_page_prot) !=
1111             pgprot_val(vm_get_page_prot(vm_flags)))
1112                 return 0;
1113
1114         /* Specialty mapping? */
1115         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1116                 return 0;
1117
1118         /* Can the mapping track the dirty pages? */
1119         return vma->vm_file && vma->vm_file->f_mapping &&
1120                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1121 }
1122
1123 /*
1124  * We account for memory if it's a private writeable mapping,
1125  * not hugepages and VM_NORESERVE wasn't set.
1126  */
1127 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1128 {
1129         /*
1130          * hugetlb has its own accounting separate from the core VM
1131          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1132          */
1133         if (file && is_file_hugepages(file))
1134                 return 0;
1135
1136         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1137 }
1138
1139 unsigned long mmap_region(struct file *file, unsigned long addr,
1140                           unsigned long len, unsigned long flags,
1141                           unsigned int vm_flags, unsigned long pgoff)
1142 {
1143         struct mm_struct *mm = current->mm;
1144         struct vm_area_struct *vma, *prev;
1145         int correct_wcount = 0;
1146         int error;
1147         struct rb_node **rb_link, *rb_parent;
1148         unsigned long charged = 0;
1149         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1150
1151         /* Clear old maps */
1152         error = -ENOMEM;
1153 munmap_back:
1154         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1155         if (vma && vma->vm_start < addr + len) {
1156                 if (do_munmap(mm, addr, len))
1157                         return -ENOMEM;
1158                 goto munmap_back;
1159         }
1160
1161         /* Check against address space limit. */
1162         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1163                 return -ENOMEM;
1164
1165         /*
1166          * Set 'VM_NORESERVE' if we should not account for the
1167          * memory use of this mapping.
1168          */
1169         if ((flags & MAP_NORESERVE)) {
1170                 /* We honor MAP_NORESERVE if allowed to overcommit */
1171                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1172                         vm_flags |= VM_NORESERVE;
1173
1174                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1175                 if (file && is_file_hugepages(file))
1176                         vm_flags |= VM_NORESERVE;
1177         }
1178
1179         /*
1180          * Private writable mapping: check memory availability
1181          */
1182         if (accountable_mapping(file, vm_flags)) {
1183                 charged = len >> PAGE_SHIFT;
1184                 if (security_vm_enough_memory(charged))
1185                         return -ENOMEM;
1186                 vm_flags |= VM_ACCOUNT;
1187         }
1188
1189         /*
1190          * Can we just expand an old mapping?
1191          */
1192         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1193         if (vma)
1194                 goto out;
1195
1196         /*
1197          * Determine the object being mapped and call the appropriate
1198          * specific mapper. the address has already been validated, but
1199          * not unmapped, but the maps are removed from the list.
1200          */
1201         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1202         if (!vma) {
1203                 error = -ENOMEM;
1204                 goto unacct_error;
1205         }
1206
1207         vma->vm_mm = mm;
1208         vma->vm_start = addr;
1209         vma->vm_end = addr + len;
1210         vma->vm_flags = vm_flags;
1211         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1212         vma->vm_pgoff = pgoff;
1213         INIT_LIST_HEAD(&vma->anon_vma_chain);
1214
1215         if (file) {
1216                 error = -EINVAL;
1217                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1218                         goto free_vma;
1219                 if (vm_flags & VM_DENYWRITE) {
1220                         error = deny_write_access(file);
1221                         if (error)
1222                                 goto free_vma;
1223                         correct_wcount = 1;
1224                 }
1225                 vma->vm_file = file;
1226                 get_file(file);
1227                 error = file->f_op->mmap(file, vma);
1228                 if (error)
1229                         goto unmap_and_free_vma;
1230                 if (vm_flags & VM_EXECUTABLE)
1231                         added_exe_file_vma(mm);
1232
1233                 /* Can addr have changed??
1234                  *
1235                  * Answer: Yes, several device drivers can do it in their
1236                  *         f_op->mmap method. -DaveM
1237                  */
1238                 addr = vma->vm_start;
1239                 pgoff = vma->vm_pgoff;
1240                 vm_flags = vma->vm_flags;
1241         } else if (vm_flags & VM_SHARED) {
1242                 error = shmem_zero_setup(vma);
1243                 if (error)
1244                         goto free_vma;
1245         }
1246
1247         if (vma_wants_writenotify(vma)) {
1248                 pgprot_t pprot = vma->vm_page_prot;
1249
1250                 /* Can vma->vm_page_prot have changed??
1251                  *
1252                  * Answer: Yes, drivers may have changed it in their
1253                  *         f_op->mmap method.
1254                  *
1255                  * Ensures that vmas marked as uncached stay that way.
1256                  */
1257                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1258                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1259                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1260         }
1261
1262         vma_link(mm, vma, prev, rb_link, rb_parent);
1263         file = vma->vm_file;
1264
1265         /* Once vma denies write, undo our temporary denial count */
1266         if (correct_wcount)
1267                 atomic_inc(&inode->i_writecount);
1268 out:
1269         perf_event_mmap(vma);
1270
1271         mm->total_vm += len >> PAGE_SHIFT;
1272         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1273         if (vm_flags & VM_LOCKED) {
1274                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1275                         mm->locked_vm += (len >> PAGE_SHIFT);
1276         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1277                 make_pages_present(addr, addr + len);
1278         return addr;
1279
1280 unmap_and_free_vma:
1281         if (correct_wcount)
1282                 atomic_inc(&inode->i_writecount);
1283         vma->vm_file = NULL;
1284         fput(file);
1285
1286         /* Undo any partial mapping done by a device driver. */
1287         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1288         charged = 0;
1289 free_vma:
1290         kmem_cache_free(vm_area_cachep, vma);
1291 unacct_error:
1292         if (charged)
1293                 vm_unacct_memory(charged);
1294         return error;
1295 }
1296
1297 /* Get an address range which is currently unmapped.
1298  * For shmat() with addr=0.
1299  *
1300  * Ugly calling convention alert:
1301  * Return value with the low bits set means error value,
1302  * ie
1303  *      if (ret & ~PAGE_MASK)
1304  *              error = ret;
1305  *
1306  * This function "knows" that -ENOMEM has the bits set.
1307  */
1308 #ifndef HAVE_ARCH_UNMAPPED_AREA
1309 unsigned long
1310 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1311                 unsigned long len, unsigned long pgoff, unsigned long flags)
1312 {
1313         struct mm_struct *mm = current->mm;
1314         struct vm_area_struct *vma;
1315         unsigned long start_addr;
1316
1317         if (len > TASK_SIZE)
1318                 return -ENOMEM;
1319
1320         if (flags & MAP_FIXED)
1321                 return addr;
1322
1323         if (addr) {
1324                 addr = PAGE_ALIGN(addr);
1325                 vma = find_vma(mm, addr);
1326                 if (TASK_SIZE - len >= addr &&
1327                     (!vma || addr + len <= vma->vm_start))
1328                         return addr;
1329         }
1330         if (len > mm->cached_hole_size) {
1331                 start_addr = addr = mm->free_area_cache;
1332         } else {
1333                 start_addr = addr = TASK_UNMAPPED_BASE;
1334                 mm->cached_hole_size = 0;
1335         }
1336
1337 full_search:
1338         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1339                 /* At this point:  (!vma || addr < vma->vm_end). */
1340                 if (TASK_SIZE - len < addr) {
1341                         /*
1342                          * Start a new search - just in case we missed
1343                          * some holes.
1344                          */
1345                         if (start_addr != TASK_UNMAPPED_BASE) {
1346                                 addr = TASK_UNMAPPED_BASE;
1347                                 start_addr = addr;
1348                                 mm->cached_hole_size = 0;
1349                                 goto full_search;
1350                         }
1351                         return -ENOMEM;
1352                 }
1353                 if (!vma || addr + len <= vma->vm_start) {
1354                         /*
1355                          * Remember the place where we stopped the search:
1356                          */
1357                         mm->free_area_cache = addr + len;
1358                         return addr;
1359                 }
1360                 if (addr + mm->cached_hole_size < vma->vm_start)
1361                         mm->cached_hole_size = vma->vm_start - addr;
1362                 addr = vma->vm_end;
1363         }
1364 }
1365 #endif  
1366
1367 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1368 {
1369         /*
1370          * Is this a new hole at the lowest possible address?
1371          */
1372         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1373                 mm->free_area_cache = addr;
1374                 mm->cached_hole_size = ~0UL;
1375         }
1376 }
1377
1378 /*
1379  * This mmap-allocator allocates new areas top-down from below the
1380  * stack's low limit (the base):
1381  */
1382 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1383 unsigned long
1384 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1385                           const unsigned long len, const unsigned long pgoff,
1386                           const unsigned long flags)
1387 {
1388         struct vm_area_struct *vma;
1389         struct mm_struct *mm = current->mm;
1390         unsigned long addr = addr0;
1391
1392         /* requested length too big for entire address space */
1393         if (len > TASK_SIZE)
1394                 return -ENOMEM;
1395
1396         if (flags & MAP_FIXED)
1397                 return addr;
1398
1399         /* requesting a specific address */
1400         if (addr) {
1401                 addr = PAGE_ALIGN(addr);
1402                 vma = find_vma(mm, addr);
1403                 if (TASK_SIZE - len >= addr &&
1404                                 (!vma || addr + len <= vma->vm_start))
1405                         return addr;
1406         }
1407
1408         /* check if free_area_cache is useful for us */
1409         if (len <= mm->cached_hole_size) {
1410                 mm->cached_hole_size = 0;
1411                 mm->free_area_cache = mm->mmap_base;
1412         }
1413
1414         /* either no address requested or can't fit in requested address hole */
1415         addr = mm->free_area_cache;
1416
1417         /* make sure it can fit in the remaining address space */
1418         if (addr > len) {
1419                 vma = find_vma(mm, addr-len);
1420                 if (!vma || addr <= vma->vm_start)
1421                         /* remember the address as a hint for next time */
1422                         return (mm->free_area_cache = addr-len);
1423         }
1424
1425         if (mm->mmap_base < len)
1426                 goto bottomup;
1427
1428         addr = mm->mmap_base-len;
1429
1430         do {
1431                 /*
1432                  * Lookup failure means no vma is above this address,
1433                  * else if new region fits below vma->vm_start,
1434                  * return with success:
1435                  */
1436                 vma = find_vma(mm, addr);
1437                 if (!vma || addr+len <= vma->vm_start)
1438                         /* remember the address as a hint for next time */
1439                         return (mm->free_area_cache = addr);
1440
1441                 /* remember the largest hole we saw so far */
1442                 if (addr + mm->cached_hole_size < vma->vm_start)
1443                         mm->cached_hole_size = vma->vm_start - addr;
1444
1445                 /* try just below the current vma->vm_start */
1446                 addr = vma->vm_start-len;
1447         } while (len < vma->vm_start);
1448
1449 bottomup:
1450         /*
1451          * A failed mmap() very likely causes application failure,
1452          * so fall back to the bottom-up function here. This scenario
1453          * can happen with large stack limits and large mmap()
1454          * allocations.
1455          */
1456         mm->cached_hole_size = ~0UL;
1457         mm->free_area_cache = TASK_UNMAPPED_BASE;
1458         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1459         /*
1460          * Restore the topdown base:
1461          */
1462         mm->free_area_cache = mm->mmap_base;
1463         mm->cached_hole_size = ~0UL;
1464
1465         return addr;
1466 }
1467 #endif
1468
1469 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1470 {
1471         /*
1472          * Is this a new hole at the highest possible address?
1473          */
1474         if (addr > mm->free_area_cache)
1475                 mm->free_area_cache = addr;
1476
1477         /* dont allow allocations above current base */
1478         if (mm->free_area_cache > mm->mmap_base)
1479                 mm->free_area_cache = mm->mmap_base;
1480 }
1481
1482 unsigned long
1483 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1484                 unsigned long pgoff, unsigned long flags)
1485 {
1486         unsigned long (*get_area)(struct file *, unsigned long,
1487                                   unsigned long, unsigned long, unsigned long);
1488
1489         unsigned long error = arch_mmap_check(addr, len, flags);
1490         if (error)
1491                 return error;
1492
1493         /* Careful about overflows.. */
1494         if (len > TASK_SIZE)
1495                 return -ENOMEM;
1496
1497         get_area = current->mm->get_unmapped_area;
1498         if (file && file->f_op && file->f_op->get_unmapped_area)
1499                 get_area = file->f_op->get_unmapped_area;
1500         addr = get_area(file, addr, len, pgoff, flags);
1501         if (IS_ERR_VALUE(addr))
1502                 return addr;
1503
1504         if (addr > TASK_SIZE - len)
1505                 return -ENOMEM;
1506         if (addr & ~PAGE_MASK)
1507                 return -EINVAL;
1508
1509         return arch_rebalance_pgtables(addr, len);
1510 }
1511
1512 EXPORT_SYMBOL(get_unmapped_area);
1513
1514 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1515 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1516 {
1517         struct vm_area_struct *vma = NULL;
1518
1519         if (mm) {
1520                 /* Check the cache first. */
1521                 /* (Cache hit rate is typically around 35%.) */
1522                 vma = mm->mmap_cache;
1523                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1524                         struct rb_node * rb_node;
1525
1526                         rb_node = mm->mm_rb.rb_node;
1527                         vma = NULL;
1528
1529                         while (rb_node) {
1530                                 struct vm_area_struct * vma_tmp;
1531
1532                                 vma_tmp = rb_entry(rb_node,
1533                                                 struct vm_area_struct, vm_rb);
1534
1535                                 if (vma_tmp->vm_end > addr) {
1536                                         vma = vma_tmp;
1537                                         if (vma_tmp->vm_start <= addr)
1538                                                 break;
1539                                         rb_node = rb_node->rb_left;
1540                                 } else
1541                                         rb_node = rb_node->rb_right;
1542                         }
1543                         if (vma)
1544                                 mm->mmap_cache = vma;
1545                 }
1546         }
1547         return vma;
1548 }
1549
1550 EXPORT_SYMBOL(find_vma);
1551
1552 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1553 struct vm_area_struct *
1554 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1555                         struct vm_area_struct **pprev)
1556 {
1557         struct vm_area_struct *vma = NULL, *prev = NULL;
1558         struct rb_node *rb_node;
1559         if (!mm)
1560                 goto out;
1561
1562         /* Guard against addr being lower than the first VMA */
1563         vma = mm->mmap;
1564
1565         /* Go through the RB tree quickly. */
1566         rb_node = mm->mm_rb.rb_node;
1567
1568         while (rb_node) {
1569                 struct vm_area_struct *vma_tmp;
1570                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1571
1572                 if (addr < vma_tmp->vm_end) {
1573                         rb_node = rb_node->rb_left;
1574                 } else {
1575                         prev = vma_tmp;
1576                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1577                                 break;
1578                         rb_node = rb_node->rb_right;
1579                 }
1580         }
1581
1582 out:
1583         *pprev = prev;
1584         return prev ? prev->vm_next : vma;
1585 }
1586
1587 /*
1588  * Verify that the stack growth is acceptable and
1589  * update accounting. This is shared with both the
1590  * grow-up and grow-down cases.
1591  */
1592 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1593 {
1594         struct mm_struct *mm = vma->vm_mm;
1595         struct rlimit *rlim = current->signal->rlim;
1596         unsigned long new_start;
1597
1598         /* address space limit tests */
1599         if (!may_expand_vm(mm, grow))
1600                 return -ENOMEM;
1601
1602         /* Stack limit test */
1603         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1604                 return -ENOMEM;
1605
1606         /* mlock limit tests */
1607         if (vma->vm_flags & VM_LOCKED) {
1608                 unsigned long locked;
1609                 unsigned long limit;
1610                 locked = mm->locked_vm + grow;
1611                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1612                 limit >>= PAGE_SHIFT;
1613                 if (locked > limit && !capable(CAP_IPC_LOCK))
1614                         return -ENOMEM;
1615         }
1616
1617         /* Check to ensure the stack will not grow into a hugetlb-only region */
1618         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1619                         vma->vm_end - size;
1620         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1621                 return -EFAULT;
1622
1623         /*
1624          * Overcommit..  This must be the final test, as it will
1625          * update security statistics.
1626          */
1627         if (security_vm_enough_memory_mm(mm, grow))
1628                 return -ENOMEM;
1629
1630         /* Ok, everything looks good - let it rip */
1631         mm->total_vm += grow;
1632         if (vma->vm_flags & VM_LOCKED)
1633                 mm->locked_vm += grow;
1634         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1635         return 0;
1636 }
1637
1638 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1639 /*
1640  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1641  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1642  */
1643 #ifndef CONFIG_IA64
1644 static
1645 #endif
1646 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1647 {
1648         int error;
1649
1650         if (!(vma->vm_flags & VM_GROWSUP))
1651                 return -EFAULT;
1652
1653         /*
1654          * We must make sure the anon_vma is allocated
1655          * so that the anon_vma locking is not a noop.
1656          */
1657         if (unlikely(anon_vma_prepare(vma)))
1658                 return -ENOMEM;
1659         anon_vma_lock(vma);
1660
1661         /*
1662          * vma->vm_start/vm_end cannot change under us because the caller
1663          * is required to hold the mmap_sem in read mode.  We need the
1664          * anon_vma lock to serialize against concurrent expand_stacks.
1665          * Also guard against wrapping around to address 0.
1666          */
1667         if (address < PAGE_ALIGN(address+4))
1668                 address = PAGE_ALIGN(address+4);
1669         else {
1670                 anon_vma_unlock(vma);
1671                 return -ENOMEM;
1672         }
1673         error = 0;
1674
1675         /* Somebody else might have raced and expanded it already */
1676         if (address > vma->vm_end) {
1677                 unsigned long size, grow;
1678
1679                 size = address - vma->vm_start;
1680                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1681
1682                 error = acct_stack_growth(vma, size, grow);
1683                 if (!error)
1684                         vma->vm_end = address;
1685         }
1686         anon_vma_unlock(vma);
1687         return error;
1688 }
1689 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1690
1691 /*
1692  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1693  */
1694 static int expand_downwards(struct vm_area_struct *vma,
1695                                    unsigned long address)
1696 {
1697         int error;
1698
1699         /*
1700          * We must make sure the anon_vma is allocated
1701          * so that the anon_vma locking is not a noop.
1702          */
1703         if (unlikely(anon_vma_prepare(vma)))
1704                 return -ENOMEM;
1705
1706         address &= PAGE_MASK;
1707         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1708         if (error)
1709                 return error;
1710
1711         anon_vma_lock(vma);
1712
1713         /*
1714          * vma->vm_start/vm_end cannot change under us because the caller
1715          * is required to hold the mmap_sem in read mode.  We need the
1716          * anon_vma lock to serialize against concurrent expand_stacks.
1717          */
1718
1719         /* Somebody else might have raced and expanded it already */
1720         if (address < vma->vm_start) {
1721                 unsigned long size, grow;
1722
1723                 size = vma->vm_end - address;
1724                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1725
1726                 error = acct_stack_growth(vma, size, grow);
1727                 if (!error) {
1728                         vma->vm_start = address;
1729                         vma->vm_pgoff -= grow;
1730                 }
1731         }
1732         anon_vma_unlock(vma);
1733         return error;
1734 }
1735
1736 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1737 {
1738         return expand_downwards(vma, address);
1739 }
1740
1741 #ifdef CONFIG_STACK_GROWSUP
1742 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1743 {
1744         return expand_upwards(vma, address);
1745 }
1746
1747 struct vm_area_struct *
1748 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1749 {
1750         struct vm_area_struct *vma, *prev;
1751
1752         addr &= PAGE_MASK;
1753         vma = find_vma_prev(mm, addr, &prev);
1754         if (vma && (vma->vm_start <= addr))
1755                 return vma;
1756         if (!prev || expand_stack(prev, addr))
1757                 return NULL;
1758         if (prev->vm_flags & VM_LOCKED) {
1759                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1760         }
1761         return prev;
1762 }
1763 #else
1764 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1765 {
1766         return expand_downwards(vma, address);
1767 }
1768
1769 struct vm_area_struct *
1770 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1771 {
1772         struct vm_area_struct * vma;
1773         unsigned long start;
1774
1775         addr &= PAGE_MASK;
1776         vma = find_vma(mm,addr);
1777         if (!vma)
1778                 return NULL;
1779         if (vma->vm_start <= addr)
1780                 return vma;
1781         if (!(vma->vm_flags & VM_GROWSDOWN))
1782                 return NULL;
1783         start = vma->vm_start;
1784         if (expand_stack(vma, addr))
1785                 return NULL;
1786         if (vma->vm_flags & VM_LOCKED) {
1787                 mlock_vma_pages_range(vma, addr, start);
1788         }
1789         return vma;
1790 }
1791 #endif
1792
1793 /*
1794  * Ok - we have the memory areas we should free on the vma list,
1795  * so release them, and do the vma updates.
1796  *
1797  * Called with the mm semaphore held.
1798  */
1799 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1800 {
1801         /* Update high watermark before we lower total_vm */
1802         update_hiwater_vm(mm);
1803         do {
1804                 long nrpages = vma_pages(vma);
1805
1806                 mm->total_vm -= nrpages;
1807                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1808                 vma = remove_vma(vma);
1809         } while (vma);
1810         validate_mm(mm);
1811 }
1812
1813 /*
1814  * Get rid of page table information in the indicated region.
1815  *
1816  * Called with the mm semaphore held.
1817  */
1818 static void unmap_region(struct mm_struct *mm,
1819                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1820                 unsigned long start, unsigned long end)
1821 {
1822         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1823         struct mmu_gather *tlb;
1824         unsigned long nr_accounted = 0;
1825
1826         lru_add_drain();
1827         tlb = tlb_gather_mmu(mm, 0);
1828         update_hiwater_rss(mm);
1829         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1830         vm_unacct_memory(nr_accounted);
1831         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1832                                  next? next->vm_start: 0);
1833         tlb_finish_mmu(tlb, start, end);
1834 }
1835
1836 /*
1837  * Create a list of vma's touched by the unmap, removing them from the mm's
1838  * vma list as we go..
1839  */
1840 static void
1841 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1842         struct vm_area_struct *prev, unsigned long end)
1843 {
1844         struct vm_area_struct **insertion_point;
1845         struct vm_area_struct *tail_vma = NULL;
1846         unsigned long addr;
1847
1848         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1849         do {
1850                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1851                 mm->map_count--;
1852                 tail_vma = vma;
1853                 vma = vma->vm_next;
1854         } while (vma && vma->vm_start < end);
1855         *insertion_point = vma;
1856         tail_vma->vm_next = NULL;
1857         if (mm->unmap_area == arch_unmap_area)
1858                 addr = prev ? prev->vm_end : mm->mmap_base;
1859         else
1860                 addr = vma ?  vma->vm_start : mm->mmap_base;
1861         mm->unmap_area(mm, addr);
1862         mm->mmap_cache = NULL;          /* Kill the cache. */
1863 }
1864
1865 /*
1866  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1867  * munmap path where it doesn't make sense to fail.
1868  */
1869 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1870               unsigned long addr, int new_below)
1871 {
1872         struct mempolicy *pol;
1873         struct vm_area_struct *new;
1874         int err = -ENOMEM;
1875
1876         if (is_vm_hugetlb_page(vma) && (addr &
1877                                         ~(huge_page_mask(hstate_vma(vma)))))
1878                 return -EINVAL;
1879
1880         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1881         if (!new)
1882                 goto out_err;
1883
1884         /* most fields are the same, copy all, and then fixup */
1885         *new = *vma;
1886
1887         INIT_LIST_HEAD(&new->anon_vma_chain);
1888
1889         if (new_below)
1890                 new->vm_end = addr;
1891         else {
1892                 new->vm_start = addr;
1893                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1894         }
1895
1896         pol = mpol_dup(vma_policy(vma));
1897         if (IS_ERR(pol)) {
1898                 err = PTR_ERR(pol);
1899                 goto out_free_vma;
1900         }
1901         vma_set_policy(new, pol);
1902
1903         if (anon_vma_clone(new, vma))
1904                 goto out_free_mpol;
1905
1906         if (new->vm_file) {
1907                 get_file(new->vm_file);
1908                 if (vma->vm_flags & VM_EXECUTABLE)
1909                         added_exe_file_vma(mm);
1910         }
1911
1912         if (new->vm_ops && new->vm_ops->open)
1913                 new->vm_ops->open(new);
1914
1915         if (new_below)
1916                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1917                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1918         else
1919                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1920
1921         /* Success. */
1922         if (!err)
1923                 return 0;
1924
1925         /* Clean everything up if vma_adjust failed. */
1926         new->vm_ops->close(new);
1927         if (new->vm_file) {
1928                 if (vma->vm_flags & VM_EXECUTABLE)
1929                         removed_exe_file_vma(mm);
1930                 fput(new->vm_file);
1931         }
1932  out_free_mpol:
1933         mpol_put(pol);
1934  out_free_vma:
1935         kmem_cache_free(vm_area_cachep, new);
1936  out_err:
1937         return err;
1938 }
1939
1940 /*
1941  * Split a vma into two pieces at address 'addr', a new vma is allocated
1942  * either for the first part or the tail.
1943  */
1944 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1945               unsigned long addr, int new_below)
1946 {
1947         if (mm->map_count >= sysctl_max_map_count)
1948                 return -ENOMEM;
1949
1950         return __split_vma(mm, vma, addr, new_below);
1951 }
1952
1953 /* Munmap is split into 2 main parts -- this part which finds
1954  * what needs doing, and the areas themselves, which do the
1955  * work.  This now handles partial unmappings.
1956  * Jeremy Fitzhardinge <jeremy@goop.org>
1957  */
1958 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1959 {
1960         unsigned long end;
1961         struct vm_area_struct *vma, *prev, *last;
1962
1963         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1964                 return -EINVAL;
1965
1966         if ((len = PAGE_ALIGN(len)) == 0)
1967                 return -EINVAL;
1968
1969         /* Find the first overlapping VMA */
1970         vma = find_vma_prev(mm, start, &prev);
1971         if (!vma)
1972                 return 0;
1973         /* we have  start < vma->vm_end  */
1974
1975         /* if it doesn't overlap, we have nothing.. */
1976         end = start + len;
1977         if (vma->vm_start >= end)
1978                 return 0;
1979
1980         /*
1981          * If we need to split any vma, do it now to save pain later.
1982          *
1983          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1984          * unmapped vm_area_struct will remain in use: so lower split_vma
1985          * places tmp vma above, and higher split_vma places tmp vma below.
1986          */
1987         if (start > vma->vm_start) {
1988                 int error;
1989
1990                 /*
1991                  * Make sure that map_count on return from munmap() will
1992                  * not exceed its limit; but let map_count go just above
1993                  * its limit temporarily, to help free resources as expected.
1994                  */
1995                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
1996                         return -ENOMEM;
1997
1998                 error = __split_vma(mm, vma, start, 0);
1999                 if (error)
2000                         return error;
2001                 prev = vma;
2002         }
2003
2004         /* Does it split the last one? */
2005         last = find_vma(mm, end);
2006         if (last && end > last->vm_start) {
2007                 int error = __split_vma(mm, last, end, 1);
2008                 if (error)
2009                         return error;
2010         }
2011         vma = prev? prev->vm_next: mm->mmap;
2012
2013         /*
2014          * unlock any mlock()ed ranges before detaching vmas
2015          */
2016         if (mm->locked_vm) {
2017                 struct vm_area_struct *tmp = vma;
2018                 while (tmp && tmp->vm_start < end) {
2019                         if (tmp->vm_flags & VM_LOCKED) {
2020                                 mm->locked_vm -= vma_pages(tmp);
2021                                 munlock_vma_pages_all(tmp);
2022                         }
2023                         tmp = tmp->vm_next;
2024                 }
2025         }
2026
2027         /*
2028          * Remove the vma's, and unmap the actual pages
2029          */
2030         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2031         unmap_region(mm, vma, prev, start, end);
2032
2033         /* Fix up all other VM information */
2034         remove_vma_list(mm, vma);
2035
2036         return 0;
2037 }
2038
2039 EXPORT_SYMBOL(do_munmap);
2040
2041 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2042 {
2043         int ret;
2044         struct mm_struct *mm = current->mm;
2045
2046         profile_munmap(addr);
2047
2048         down_write(&mm->mmap_sem);
2049         ret = do_munmap(mm, addr, len);
2050         up_write(&mm->mmap_sem);
2051         return ret;
2052 }
2053
2054 static inline void verify_mm_writelocked(struct mm_struct *mm)
2055 {
2056 #ifdef CONFIG_DEBUG_VM
2057         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2058                 WARN_ON(1);
2059                 up_read(&mm->mmap_sem);
2060         }
2061 #endif
2062 }
2063
2064 /*
2065  *  this is really a simplified "do_mmap".  it only handles
2066  *  anonymous maps.  eventually we may be able to do some
2067  *  brk-specific accounting here.
2068  */
2069 unsigned long do_brk(unsigned long addr, unsigned long len)
2070 {
2071         struct mm_struct * mm = current->mm;
2072         struct vm_area_struct * vma, * prev;
2073         unsigned long flags;
2074         struct rb_node ** rb_link, * rb_parent;
2075         pgoff_t pgoff = addr >> PAGE_SHIFT;
2076         int error;
2077
2078         len = PAGE_ALIGN(len);
2079         if (!len)
2080                 return addr;
2081
2082         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2083         if (error)
2084                 return error;
2085
2086         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2087
2088         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2089         if (error & ~PAGE_MASK)
2090                 return error;
2091
2092         /*
2093          * mlock MCL_FUTURE?
2094          */
2095         if (mm->def_flags & VM_LOCKED) {
2096                 unsigned long locked, lock_limit;
2097                 locked = len >> PAGE_SHIFT;
2098                 locked += mm->locked_vm;
2099                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2100                 lock_limit >>= PAGE_SHIFT;
2101                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2102                         return -EAGAIN;
2103         }
2104
2105         /*
2106          * mm->mmap_sem is required to protect against another thread
2107          * changing the mappings in case we sleep.
2108          */
2109         verify_mm_writelocked(mm);
2110
2111         /*
2112          * Clear old maps.  this also does some error checking for us
2113          */
2114  munmap_back:
2115         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2116         if (vma && vma->vm_start < addr + len) {
2117                 if (do_munmap(mm, addr, len))
2118                         return -ENOMEM;
2119                 goto munmap_back;
2120         }
2121
2122         /* Check against address space limits *after* clearing old maps... */
2123         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2124                 return -ENOMEM;
2125
2126         if (mm->map_count > sysctl_max_map_count)
2127                 return -ENOMEM;
2128
2129         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2130                 return -ENOMEM;
2131
2132         /* Can we just expand an old private anonymous mapping? */
2133         vma = vma_merge(mm, prev, addr, addr + len, flags,
2134                                         NULL, NULL, pgoff, NULL);
2135         if (vma)
2136                 goto out;
2137
2138         /*
2139          * create a vma struct for an anonymous mapping
2140          */
2141         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2142         if (!vma) {
2143                 vm_unacct_memory(len >> PAGE_SHIFT);
2144                 return -ENOMEM;
2145         }
2146
2147         INIT_LIST_HEAD(&vma->anon_vma_chain);
2148         vma->vm_mm = mm;
2149         vma->vm_start = addr;
2150         vma->vm_end = addr + len;
2151         vma->vm_pgoff = pgoff;
2152         vma->vm_flags = flags;
2153         vma->vm_page_prot = vm_get_page_prot(flags);
2154         vma_link(mm, vma, prev, rb_link, rb_parent);
2155 out:
2156         mm->total_vm += len >> PAGE_SHIFT;
2157         if (flags & VM_LOCKED) {
2158                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2159                         mm->locked_vm += (len >> PAGE_SHIFT);
2160         }
2161         return addr;
2162 }
2163
2164 EXPORT_SYMBOL(do_brk);
2165
2166 /* Release all mmaps. */
2167 void exit_mmap(struct mm_struct *mm)
2168 {
2169         struct mmu_gather *tlb;
2170         struct vm_area_struct *vma;
2171         unsigned long nr_accounted = 0;
2172         unsigned long end;
2173
2174         /* mm's last user has gone, and its about to be pulled down */
2175         mmu_notifier_release(mm);
2176
2177         if (mm->locked_vm) {
2178                 vma = mm->mmap;
2179                 while (vma) {
2180                         if (vma->vm_flags & VM_LOCKED)
2181                                 munlock_vma_pages_all(vma);
2182                         vma = vma->vm_next;
2183                 }
2184         }
2185
2186         arch_exit_mmap(mm);
2187
2188         vma = mm->mmap;
2189         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2190                 return;
2191
2192         lru_add_drain();
2193         flush_cache_mm(mm);
2194         tlb = tlb_gather_mmu(mm, 1);
2195         /* update_hiwater_rss(mm) here? but nobody should be looking */
2196         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2197         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2198         vm_unacct_memory(nr_accounted);
2199
2200         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2201         tlb_finish_mmu(tlb, 0, end);
2202
2203         /*
2204          * Walk the list again, actually closing and freeing it,
2205          * with preemption enabled, without holding any MM locks.
2206          */
2207         while (vma)
2208                 vma = remove_vma(vma);
2209
2210         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2211 }
2212
2213 /* Insert vm structure into process list sorted by address
2214  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2215  * then i_mmap_lock is taken here.
2216  */
2217 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2218 {
2219         struct vm_area_struct * __vma, * prev;
2220         struct rb_node ** rb_link, * rb_parent;
2221
2222         /*
2223          * The vm_pgoff of a purely anonymous vma should be irrelevant
2224          * until its first write fault, when page's anon_vma and index
2225          * are set.  But now set the vm_pgoff it will almost certainly
2226          * end up with (unless mremap moves it elsewhere before that
2227          * first wfault), so /proc/pid/maps tells a consistent story.
2228          *
2229          * By setting it to reflect the virtual start address of the
2230          * vma, merges and splits can happen in a seamless way, just
2231          * using the existing file pgoff checks and manipulations.
2232          * Similarly in do_mmap_pgoff and in do_brk.
2233          */
2234         if (!vma->vm_file) {
2235                 BUG_ON(vma->anon_vma);
2236                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2237         }
2238         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2239         if (__vma && __vma->vm_start < vma->vm_end)
2240                 return -ENOMEM;
2241         if ((vma->vm_flags & VM_ACCOUNT) &&
2242              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2243                 return -ENOMEM;
2244         vma_link(mm, vma, prev, rb_link, rb_parent);
2245         return 0;
2246 }
2247
2248 /*
2249  * Copy the vma structure to a new location in the same mm,
2250  * prior to moving page table entries, to effect an mremap move.
2251  */
2252 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2253         unsigned long addr, unsigned long len, pgoff_t pgoff)
2254 {
2255         struct vm_area_struct *vma = *vmap;
2256         unsigned long vma_start = vma->vm_start;
2257         struct mm_struct *mm = vma->vm_mm;
2258         struct vm_area_struct *new_vma, *prev;
2259         struct rb_node **rb_link, *rb_parent;
2260         struct mempolicy *pol;
2261
2262         /*
2263          * If anonymous vma has not yet been faulted, update new pgoff
2264          * to match new location, to increase its chance of merging.
2265          */
2266         if (!vma->vm_file && !vma->anon_vma)
2267                 pgoff = addr >> PAGE_SHIFT;
2268
2269         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2270         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2271                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2272         if (new_vma) {
2273                 /*
2274                  * Source vma may have been merged into new_vma
2275                  */
2276                 if (vma_start >= new_vma->vm_start &&
2277                     vma_start < new_vma->vm_end)
2278                         *vmap = new_vma;
2279         } else {
2280                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2281                 if (new_vma) {
2282                         *new_vma = *vma;
2283                         pol = mpol_dup(vma_policy(vma));
2284                         if (IS_ERR(pol))
2285                                 goto out_free_vma;
2286                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2287                         if (anon_vma_clone(new_vma, vma))
2288                                 goto out_free_mempol;
2289                         vma_set_policy(new_vma, pol);
2290                         new_vma->vm_start = addr;
2291                         new_vma->vm_end = addr + len;
2292                         new_vma->vm_pgoff = pgoff;
2293                         if (new_vma->vm_file) {
2294                                 get_file(new_vma->vm_file);
2295                                 if (vma->vm_flags & VM_EXECUTABLE)
2296                                         added_exe_file_vma(mm);
2297                         }
2298                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2299                                 new_vma->vm_ops->open(new_vma);
2300                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2301                 }
2302         }
2303         return new_vma;
2304
2305  out_free_mempol:
2306         mpol_put(pol);
2307  out_free_vma:
2308         kmem_cache_free(vm_area_cachep, new_vma);
2309         return NULL;
2310 }
2311
2312 /*
2313  * Return true if the calling process may expand its vm space by the passed
2314  * number of pages
2315  */
2316 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2317 {
2318         unsigned long cur = mm->total_vm;       /* pages */
2319         unsigned long lim;
2320
2321         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2322
2323         if (cur + npages > lim)
2324                 return 0;
2325         return 1;
2326 }
2327
2328
2329 static int special_mapping_fault(struct vm_area_struct *vma,
2330                                 struct vm_fault *vmf)
2331 {
2332         pgoff_t pgoff;
2333         struct page **pages;
2334
2335         /*
2336          * special mappings have no vm_file, and in that case, the mm
2337          * uses vm_pgoff internally. So we have to subtract it from here.
2338          * We are allowed to do this because we are the mm; do not copy
2339          * this code into drivers!
2340          */
2341         pgoff = vmf->pgoff - vma->vm_pgoff;
2342
2343         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2344                 pgoff--;
2345
2346         if (*pages) {
2347                 struct page *page = *pages;
2348                 get_page(page);
2349                 vmf->page = page;
2350                 return 0;
2351         }
2352
2353         return VM_FAULT_SIGBUS;
2354 }
2355
2356 /*
2357  * Having a close hook prevents vma merging regardless of flags.
2358  */
2359 static void special_mapping_close(struct vm_area_struct *vma)
2360 {
2361 }
2362
2363 static const struct vm_operations_struct special_mapping_vmops = {
2364         .close = special_mapping_close,
2365         .fault = special_mapping_fault,
2366 };
2367
2368 /*
2369  * Called with mm->mmap_sem held for writing.
2370  * Insert a new vma covering the given region, with the given flags.
2371  * Its pages are supplied by the given array of struct page *.
2372  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2373  * The region past the last page supplied will always produce SIGBUS.
2374  * The array pointer and the pages it points to are assumed to stay alive
2375  * for as long as this mapping might exist.
2376  */
2377 int install_special_mapping(struct mm_struct *mm,
2378                             unsigned long addr, unsigned long len,
2379                             unsigned long vm_flags, struct page **pages)
2380 {
2381         struct vm_area_struct *vma;
2382
2383         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2384         if (unlikely(vma == NULL))
2385                 return -ENOMEM;
2386
2387         INIT_LIST_HEAD(&vma->anon_vma_chain);
2388         vma->vm_mm = mm;
2389         vma->vm_start = addr;
2390         vma->vm_end = addr + len;
2391
2392         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2393         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2394
2395         vma->vm_ops = &special_mapping_vmops;
2396         vma->vm_private_data = pages;
2397
2398         if (unlikely(insert_vm_struct(mm, vma))) {
2399                 kmem_cache_free(vm_area_cachep, vma);
2400                 return -ENOMEM;
2401         }
2402
2403         mm->total_vm += len >> PAGE_SHIFT;
2404
2405         perf_event_mmap(vma);
2406
2407         return 0;
2408 }
2409
2410 static DEFINE_MUTEX(mm_all_locks_mutex);
2411
2412 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2413 {
2414         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2415                 /*
2416                  * The LSB of head.next can't change from under us
2417                  * because we hold the mm_all_locks_mutex.
2418                  */
2419                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2420                 /*
2421                  * We can safely modify head.next after taking the
2422                  * anon_vma->lock. If some other vma in this mm shares
2423                  * the same anon_vma we won't take it again.
2424                  *
2425                  * No need of atomic instructions here, head.next
2426                  * can't change from under us thanks to the
2427                  * anon_vma->lock.
2428                  */
2429                 if (__test_and_set_bit(0, (unsigned long *)
2430                                        &anon_vma->head.next))
2431                         BUG();
2432         }
2433 }
2434
2435 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2436 {
2437         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2438                 /*
2439                  * AS_MM_ALL_LOCKS can't change from under us because
2440                  * we hold the mm_all_locks_mutex.
2441                  *
2442                  * Operations on ->flags have to be atomic because
2443                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2444                  * mm_all_locks_mutex, there may be other cpus
2445                  * changing other bitflags in parallel to us.
2446                  */
2447                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2448                         BUG();
2449                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2450         }
2451 }
2452
2453 /*
2454  * This operation locks against the VM for all pte/vma/mm related
2455  * operations that could ever happen on a certain mm. This includes
2456  * vmtruncate, try_to_unmap, and all page faults.
2457  *
2458  * The caller must take the mmap_sem in write mode before calling
2459  * mm_take_all_locks(). The caller isn't allowed to release the
2460  * mmap_sem until mm_drop_all_locks() returns.
2461  *
2462  * mmap_sem in write mode is required in order to block all operations
2463  * that could modify pagetables and free pages without need of
2464  * altering the vma layout (for example populate_range() with
2465  * nonlinear vmas). It's also needed in write mode to avoid new
2466  * anon_vmas to be associated with existing vmas.
2467  *
2468  * A single task can't take more than one mm_take_all_locks() in a row
2469  * or it would deadlock.
2470  *
2471  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2472  * mapping->flags avoid to take the same lock twice, if more than one
2473  * vma in this mm is backed by the same anon_vma or address_space.
2474  *
2475  * We can take all the locks in random order because the VM code
2476  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2477  * takes more than one of them in a row. Secondly we're protected
2478  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2479  *
2480  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2481  * that may have to take thousand of locks.
2482  *
2483  * mm_take_all_locks() can fail if it's interrupted by signals.
2484  */
2485 int mm_take_all_locks(struct mm_struct *mm)
2486 {
2487         struct vm_area_struct *vma;
2488         struct anon_vma_chain *avc;
2489         int ret = -EINTR;
2490
2491         BUG_ON(down_read_trylock(&mm->mmap_sem));
2492
2493         mutex_lock(&mm_all_locks_mutex);
2494
2495         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2496                 if (signal_pending(current))
2497                         goto out_unlock;
2498                 if (vma->vm_file && vma->vm_file->f_mapping)
2499                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2500         }
2501
2502         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2503                 if (signal_pending(current))
2504                         goto out_unlock;
2505                 if (vma->anon_vma)
2506                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2507                                 vm_lock_anon_vma(mm, avc->anon_vma);
2508         }
2509
2510         ret = 0;
2511
2512 out_unlock:
2513         if (ret)
2514                 mm_drop_all_locks(mm);
2515
2516         return ret;
2517 }
2518
2519 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2520 {
2521         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2522                 /*
2523                  * The LSB of head.next can't change to 0 from under
2524                  * us because we hold the mm_all_locks_mutex.
2525                  *
2526                  * We must however clear the bitflag before unlocking
2527                  * the vma so the users using the anon_vma->head will
2528                  * never see our bitflag.
2529                  *
2530                  * No need of atomic instructions here, head.next
2531                  * can't change from under us until we release the
2532                  * anon_vma->lock.
2533                  */
2534                 if (!__test_and_clear_bit(0, (unsigned long *)
2535                                           &anon_vma->head.next))
2536                         BUG();
2537                 spin_unlock(&anon_vma->lock);
2538         }
2539 }
2540
2541 static void vm_unlock_mapping(struct address_space *mapping)
2542 {
2543         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2544                 /*
2545                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2546                  * because we hold the mm_all_locks_mutex.
2547                  */
2548                 spin_unlock(&mapping->i_mmap_lock);
2549                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2550                                         &mapping->flags))
2551                         BUG();
2552         }
2553 }
2554
2555 /*
2556  * The mmap_sem cannot be released by the caller until
2557  * mm_drop_all_locks() returns.
2558  */
2559 void mm_drop_all_locks(struct mm_struct *mm)
2560 {
2561         struct vm_area_struct *vma;
2562         struct anon_vma_chain *avc;
2563
2564         BUG_ON(down_read_trylock(&mm->mmap_sem));
2565         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2566
2567         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2568                 if (vma->anon_vma)
2569                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2570                                 vm_unlock_anon_vma(avc->anon_vma);
2571                 if (vma->vm_file && vma->vm_file->f_mapping)
2572                         vm_unlock_mapping(vma->vm_file->f_mapping);
2573         }
2574
2575         mutex_unlock(&mm_all_locks_mutex);
2576 }
2577
2578 /*
2579  * initialise the VMA slab
2580  */
2581 void __init mmap_init(void)
2582 {
2583         int ret;
2584
2585         ret = percpu_counter_init(&vm_committed_as, 0);
2586         VM_BUG_ON(ret);
2587 }