75e0d0673d78a3bf203e7ca8d0f632789d0c32c7
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@redhat.com>
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
30 #include <asm/uaccess.h>
31 #include <asm/cacheflush.h>
32 #include <asm/tlb.h>
33 #include <asm/mmu_context.h>
34
35 #include "internal.h"
36
37 #ifndef arch_mmap_check
38 #define arch_mmap_check(addr, len, flags)       (0)
39 #endif
40
41 #ifndef arch_rebalance_pgtables
42 #define arch_rebalance_pgtables(addr, len)              (addr)
43 #endif
44
45 static void unmap_region(struct mm_struct *mm,
46                 struct vm_area_struct *vma, struct vm_area_struct *prev,
47                 unsigned long start, unsigned long end);
48
49 /*
50  * WARNING: the debugging will use recursive algorithms so never enable this
51  * unless you know what you are doing.
52  */
53 #undef DEBUG_MM_RB
54
55 /* description of effects of mapping type and prot in current implementation.
56  * this is due to the limited x86 page protection hardware.  The expected
57  * behavior is in parens:
58  *
59  * map_type     prot
60  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
61  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
62  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
63  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
64  *              
65  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *
69  */
70 pgprot_t protection_map[16] = {
71         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
72         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
73 };
74
75 pgprot_t vm_get_page_prot(unsigned long vm_flags)
76 {
77         return __pgprot(pgprot_val(protection_map[vm_flags &
78                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
79                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
80 }
81 EXPORT_SYMBOL(vm_get_page_prot);
82
83 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
84 int sysctl_overcommit_ratio = 50;       /* default is 50% */
85 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
86 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
87
88 /*
89  * Check that a process has enough memory to allocate a new virtual
90  * mapping. 0 means there is enough memory for the allocation to
91  * succeed and -ENOMEM implies there is not.
92  *
93  * We currently support three overcommit policies, which are set via the
94  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
95  *
96  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
97  * Additional code 2002 Jul 20 by Robert Love.
98  *
99  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
100  *
101  * Note this is a helper function intended to be used by LSMs which
102  * wish to use this logic.
103  */
104 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
105 {
106         unsigned long free, allowed;
107
108         vm_acct_memory(pages);
109
110         /*
111          * Sometimes we want to use more memory than we have
112          */
113         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
114                 return 0;
115
116         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
117                 unsigned long n;
118
119                 free = global_page_state(NR_FILE_PAGES);
120                 free += nr_swap_pages;
121
122                 /*
123                  * Any slabs which are created with the
124                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
125                  * which are reclaimable, under pressure.  The dentry
126                  * cache and most inode caches should fall into this
127                  */
128                 free += global_page_state(NR_SLAB_RECLAIMABLE);
129
130                 /*
131                  * Leave the last 3% for root
132                  */
133                 if (!cap_sys_admin)
134                         free -= free / 32;
135
136                 if (free > pages)
137                         return 0;
138
139                 /*
140                  * nr_free_pages() is very expensive on large systems,
141                  * only call if we're about to fail.
142                  */
143                 n = nr_free_pages();
144
145                 /*
146                  * Leave reserved pages. The pages are not for anonymous pages.
147                  */
148                 if (n <= totalreserve_pages)
149                         goto error;
150                 else
151                         n -= totalreserve_pages;
152
153                 /*
154                  * Leave the last 3% for root
155                  */
156                 if (!cap_sys_admin)
157                         n -= n / 32;
158                 free += n;
159
160                 if (free > pages)
161                         return 0;
162
163                 goto error;
164         }
165
166         allowed = (totalram_pages - hugetlb_total_pages())
167                 * sysctl_overcommit_ratio / 100;
168         /*
169          * Leave the last 3% for root
170          */
171         if (!cap_sys_admin)
172                 allowed -= allowed / 32;
173         allowed += total_swap_pages;
174
175         /* Don't let a single process grow too big:
176            leave 3% of the size of this process for other processes */
177         allowed -= mm->total_vm / 32;
178
179         /*
180          * cast `allowed' as a signed long because vm_committed_space
181          * sometimes has a negative value
182          */
183         if (atomic_long_read(&vm_committed_space) < (long)allowed)
184                 return 0;
185 error:
186         vm_unacct_memory(pages);
187
188         return -ENOMEM;
189 }
190
191 /*
192  * Requires inode->i_mapping->i_mmap_lock
193  */
194 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
195                 struct file *file, struct address_space *mapping)
196 {
197         if (vma->vm_flags & VM_DENYWRITE)
198                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
199         if (vma->vm_flags & VM_SHARED)
200                 mapping->i_mmap_writable--;
201
202         flush_dcache_mmap_lock(mapping);
203         if (unlikely(vma->vm_flags & VM_NONLINEAR))
204                 list_del_init(&vma->shared.vm_set.list);
205         else
206                 vma_prio_tree_remove(vma, &mapping->i_mmap);
207         flush_dcache_mmap_unlock(mapping);
208 }
209
210 /*
211  * Unlink a file-based vm structure from its prio_tree, to hide
212  * vma from rmap and vmtruncate before freeing its page tables.
213  */
214 void unlink_file_vma(struct vm_area_struct *vma)
215 {
216         struct file *file = vma->vm_file;
217
218         if (file) {
219                 struct address_space *mapping = file->f_mapping;
220                 spin_lock(&mapping->i_mmap_lock);
221                 __remove_shared_vm_struct(vma, file, mapping);
222                 spin_unlock(&mapping->i_mmap_lock);
223         }
224 }
225
226 /*
227  * Close a vm structure and free it, returning the next.
228  */
229 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
230 {
231         struct vm_area_struct *next = vma->vm_next;
232
233         might_sleep();
234         if (vma->vm_ops && vma->vm_ops->close)
235                 vma->vm_ops->close(vma);
236         if (vma->vm_file) {
237                 fput(vma->vm_file);
238                 if (vma->vm_flags & VM_EXECUTABLE)
239                         removed_exe_file_vma(vma->vm_mm);
240         }
241         mpol_put(vma_policy(vma));
242         kmem_cache_free(vm_area_cachep, vma);
243         return next;
244 }
245
246 asmlinkage unsigned long sys_brk(unsigned long brk)
247 {
248         unsigned long rlim, retval;
249         unsigned long newbrk, oldbrk;
250         struct mm_struct *mm = current->mm;
251         unsigned long min_brk;
252
253         down_write(&mm->mmap_sem);
254
255 #ifdef CONFIG_COMPAT_BRK
256         min_brk = mm->end_code;
257 #else
258         min_brk = mm->start_brk;
259 #endif
260         if (brk < min_brk)
261                 goto out;
262
263         /*
264          * Check against rlimit here. If this check is done later after the test
265          * of oldbrk with newbrk then it can escape the test and let the data
266          * segment grow beyond its set limit the in case where the limit is
267          * not page aligned -Ram Gupta
268          */
269         rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
270         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
271                         (mm->end_data - mm->start_data) > rlim)
272                 goto out;
273
274         newbrk = PAGE_ALIGN(brk);
275         oldbrk = PAGE_ALIGN(mm->brk);
276         if (oldbrk == newbrk)
277                 goto set_brk;
278
279         /* Always allow shrinking brk. */
280         if (brk <= mm->brk) {
281                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
282                         goto set_brk;
283                 goto out;
284         }
285
286         /* Check against existing mmap mappings. */
287         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
288                 goto out;
289
290         /* Ok, looks good - let it rip. */
291         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
292                 goto out;
293 set_brk:
294         mm->brk = brk;
295 out:
296         retval = mm->brk;
297         up_write(&mm->mmap_sem);
298         return retval;
299 }
300
301 #ifdef DEBUG_MM_RB
302 static int browse_rb(struct rb_root *root)
303 {
304         int i = 0, j;
305         struct rb_node *nd, *pn = NULL;
306         unsigned long prev = 0, pend = 0;
307
308         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
309                 struct vm_area_struct *vma;
310                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
311                 if (vma->vm_start < prev)
312                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
313                 if (vma->vm_start < pend)
314                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
315                 if (vma->vm_start > vma->vm_end)
316                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
317                 i++;
318                 pn = nd;
319                 prev = vma->vm_start;
320                 pend = vma->vm_end;
321         }
322         j = 0;
323         for (nd = pn; nd; nd = rb_prev(nd)) {
324                 j++;
325         }
326         if (i != j)
327                 printk("backwards %d, forwards %d\n", j, i), i = 0;
328         return i;
329 }
330
331 void validate_mm(struct mm_struct *mm)
332 {
333         int bug = 0;
334         int i = 0;
335         struct vm_area_struct *tmp = mm->mmap;
336         while (tmp) {
337                 tmp = tmp->vm_next;
338                 i++;
339         }
340         if (i != mm->map_count)
341                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
342         i = browse_rb(&mm->mm_rb);
343         if (i != mm->map_count)
344                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
345         BUG_ON(bug);
346 }
347 #else
348 #define validate_mm(mm) do { } while (0)
349 #endif
350
351 static struct vm_area_struct *
352 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
353                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
354                 struct rb_node ** rb_parent)
355 {
356         struct vm_area_struct * vma;
357         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
358
359         __rb_link = &mm->mm_rb.rb_node;
360         rb_prev = __rb_parent = NULL;
361         vma = NULL;
362
363         while (*__rb_link) {
364                 struct vm_area_struct *vma_tmp;
365
366                 __rb_parent = *__rb_link;
367                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
368
369                 if (vma_tmp->vm_end > addr) {
370                         vma = vma_tmp;
371                         if (vma_tmp->vm_start <= addr)
372                                 return vma;
373                         __rb_link = &__rb_parent->rb_left;
374                 } else {
375                         rb_prev = __rb_parent;
376                         __rb_link = &__rb_parent->rb_right;
377                 }
378         }
379
380         *pprev = NULL;
381         if (rb_prev)
382                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
383         *rb_link = __rb_link;
384         *rb_parent = __rb_parent;
385         return vma;
386 }
387
388 static inline void
389 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
390                 struct vm_area_struct *prev, struct rb_node *rb_parent)
391 {
392         if (prev) {
393                 vma->vm_next = prev->vm_next;
394                 prev->vm_next = vma;
395         } else {
396                 mm->mmap = vma;
397                 if (rb_parent)
398                         vma->vm_next = rb_entry(rb_parent,
399                                         struct vm_area_struct, vm_rb);
400                 else
401                         vma->vm_next = NULL;
402         }
403 }
404
405 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
406                 struct rb_node **rb_link, struct rb_node *rb_parent)
407 {
408         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
409         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
410 }
411
412 static inline void __vma_link_file(struct vm_area_struct *vma)
413 {
414         struct file * file;
415
416         file = vma->vm_file;
417         if (file) {
418                 struct address_space *mapping = file->f_mapping;
419
420                 if (vma->vm_flags & VM_DENYWRITE)
421                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
422                 if (vma->vm_flags & VM_SHARED)
423                         mapping->i_mmap_writable++;
424
425                 flush_dcache_mmap_lock(mapping);
426                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
427                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
428                 else
429                         vma_prio_tree_insert(vma, &mapping->i_mmap);
430                 flush_dcache_mmap_unlock(mapping);
431         }
432 }
433
434 static void
435 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
436         struct vm_area_struct *prev, struct rb_node **rb_link,
437         struct rb_node *rb_parent)
438 {
439         __vma_link_list(mm, vma, prev, rb_parent);
440         __vma_link_rb(mm, vma, rb_link, rb_parent);
441         __anon_vma_link(vma);
442 }
443
444 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
445                         struct vm_area_struct *prev, struct rb_node **rb_link,
446                         struct rb_node *rb_parent)
447 {
448         struct address_space *mapping = NULL;
449
450         if (vma->vm_file)
451                 mapping = vma->vm_file->f_mapping;
452
453         if (mapping) {
454                 spin_lock(&mapping->i_mmap_lock);
455                 vma->vm_truncate_count = mapping->truncate_count;
456         }
457         anon_vma_lock(vma);
458
459         __vma_link(mm, vma, prev, rb_link, rb_parent);
460         __vma_link_file(vma);
461
462         anon_vma_unlock(vma);
463         if (mapping)
464                 spin_unlock(&mapping->i_mmap_lock);
465
466         mm->map_count++;
467         validate_mm(mm);
468 }
469
470 /*
471  * Helper for vma_adjust in the split_vma insert case:
472  * insert vm structure into list and rbtree and anon_vma,
473  * but it has already been inserted into prio_tree earlier.
474  */
475 static void
476 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
477 {
478         struct vm_area_struct * __vma, * prev;
479         struct rb_node ** rb_link, * rb_parent;
480
481         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
482         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
483         __vma_link(mm, vma, prev, rb_link, rb_parent);
484         mm->map_count++;
485 }
486
487 static inline void
488 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
489                 struct vm_area_struct *prev)
490 {
491         prev->vm_next = vma->vm_next;
492         rb_erase(&vma->vm_rb, &mm->mm_rb);
493         if (mm->mmap_cache == vma)
494                 mm->mmap_cache = prev;
495 }
496
497 /*
498  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
499  * is already present in an i_mmap tree without adjusting the tree.
500  * The following helper function should be used when such adjustments
501  * are necessary.  The "insert" vma (if any) is to be inserted
502  * before we drop the necessary locks.
503  */
504 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
505         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
506 {
507         struct mm_struct *mm = vma->vm_mm;
508         struct vm_area_struct *next = vma->vm_next;
509         struct vm_area_struct *importer = NULL;
510         struct address_space *mapping = NULL;
511         struct prio_tree_root *root = NULL;
512         struct file *file = vma->vm_file;
513         struct anon_vma *anon_vma = NULL;
514         long adjust_next = 0;
515         int remove_next = 0;
516
517         if (next && !insert) {
518                 if (end >= next->vm_end) {
519                         /*
520                          * vma expands, overlapping all the next, and
521                          * perhaps the one after too (mprotect case 6).
522                          */
523 again:                  remove_next = 1 + (end > next->vm_end);
524                         end = next->vm_end;
525                         anon_vma = next->anon_vma;
526                         importer = vma;
527                 } else if (end > next->vm_start) {
528                         /*
529                          * vma expands, overlapping part of the next:
530                          * mprotect case 5 shifting the boundary up.
531                          */
532                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
533                         anon_vma = next->anon_vma;
534                         importer = vma;
535                 } else if (end < vma->vm_end) {
536                         /*
537                          * vma shrinks, and !insert tells it's not
538                          * split_vma inserting another: so it must be
539                          * mprotect case 4 shifting the boundary down.
540                          */
541                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
542                         anon_vma = next->anon_vma;
543                         importer = next;
544                 }
545         }
546
547         if (file) {
548                 mapping = file->f_mapping;
549                 if (!(vma->vm_flags & VM_NONLINEAR))
550                         root = &mapping->i_mmap;
551                 spin_lock(&mapping->i_mmap_lock);
552                 if (importer &&
553                     vma->vm_truncate_count != next->vm_truncate_count) {
554                         /*
555                          * unmap_mapping_range might be in progress:
556                          * ensure that the expanding vma is rescanned.
557                          */
558                         importer->vm_truncate_count = 0;
559                 }
560                 if (insert) {
561                         insert->vm_truncate_count = vma->vm_truncate_count;
562                         /*
563                          * Put into prio_tree now, so instantiated pages
564                          * are visible to arm/parisc __flush_dcache_page
565                          * throughout; but we cannot insert into address
566                          * space until vma start or end is updated.
567                          */
568                         __vma_link_file(insert);
569                 }
570         }
571
572         /*
573          * When changing only vma->vm_end, we don't really need
574          * anon_vma lock: but is that case worth optimizing out?
575          */
576         if (vma->anon_vma)
577                 anon_vma = vma->anon_vma;
578         if (anon_vma) {
579                 spin_lock(&anon_vma->lock);
580                 /*
581                  * Easily overlooked: when mprotect shifts the boundary,
582                  * make sure the expanding vma has anon_vma set if the
583                  * shrinking vma had, to cover any anon pages imported.
584                  */
585                 if (importer && !importer->anon_vma) {
586                         importer->anon_vma = anon_vma;
587                         __anon_vma_link(importer);
588                 }
589         }
590
591         if (root) {
592                 flush_dcache_mmap_lock(mapping);
593                 vma_prio_tree_remove(vma, root);
594                 if (adjust_next)
595                         vma_prio_tree_remove(next, root);
596         }
597
598         vma->vm_start = start;
599         vma->vm_end = end;
600         vma->vm_pgoff = pgoff;
601         if (adjust_next) {
602                 next->vm_start += adjust_next << PAGE_SHIFT;
603                 next->vm_pgoff += adjust_next;
604         }
605
606         if (root) {
607                 if (adjust_next)
608                         vma_prio_tree_insert(next, root);
609                 vma_prio_tree_insert(vma, root);
610                 flush_dcache_mmap_unlock(mapping);
611         }
612
613         if (remove_next) {
614                 /*
615                  * vma_merge has merged next into vma, and needs
616                  * us to remove next before dropping the locks.
617                  */
618                 __vma_unlink(mm, next, vma);
619                 if (file)
620                         __remove_shared_vm_struct(next, file, mapping);
621                 if (next->anon_vma)
622                         __anon_vma_merge(vma, next);
623         } else if (insert) {
624                 /*
625                  * split_vma has split insert from vma, and needs
626                  * us to insert it before dropping the locks
627                  * (it may either follow vma or precede it).
628                  */
629                 __insert_vm_struct(mm, insert);
630         }
631
632         if (anon_vma)
633                 spin_unlock(&anon_vma->lock);
634         if (mapping)
635                 spin_unlock(&mapping->i_mmap_lock);
636
637         if (remove_next) {
638                 if (file) {
639                         fput(file);
640                         if (next->vm_flags & VM_EXECUTABLE)
641                                 removed_exe_file_vma(mm);
642                 }
643                 mm->map_count--;
644                 mpol_put(vma_policy(next));
645                 kmem_cache_free(vm_area_cachep, next);
646                 /*
647                  * In mprotect's case 6 (see comments on vma_merge),
648                  * we must remove another next too. It would clutter
649                  * up the code too much to do both in one go.
650                  */
651                 if (remove_next == 2) {
652                         next = vma->vm_next;
653                         goto again;
654                 }
655         }
656
657         validate_mm(mm);
658 }
659
660 /*
661  * If the vma has a ->close operation then the driver probably needs to release
662  * per-vma resources, so we don't attempt to merge those.
663  */
664 #define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
665
666 static inline int is_mergeable_vma(struct vm_area_struct *vma,
667                         struct file *file, unsigned long vm_flags)
668 {
669         if (vma->vm_flags != vm_flags)
670                 return 0;
671         if (vma->vm_file != file)
672                 return 0;
673         if (vma->vm_ops && vma->vm_ops->close)
674                 return 0;
675         return 1;
676 }
677
678 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
679                                         struct anon_vma *anon_vma2)
680 {
681         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
682 }
683
684 /*
685  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
686  * in front of (at a lower virtual address and file offset than) the vma.
687  *
688  * We cannot merge two vmas if they have differently assigned (non-NULL)
689  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
690  *
691  * We don't check here for the merged mmap wrapping around the end of pagecache
692  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
693  * wrap, nor mmaps which cover the final page at index -1UL.
694  */
695 static int
696 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
697         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
698 {
699         if (is_mergeable_vma(vma, file, vm_flags) &&
700             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
701                 if (vma->vm_pgoff == vm_pgoff)
702                         return 1;
703         }
704         return 0;
705 }
706
707 /*
708  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
709  * beyond (at a higher virtual address and file offset than) the vma.
710  *
711  * We cannot merge two vmas if they have differently assigned (non-NULL)
712  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
713  */
714 static int
715 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
716         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
717 {
718         if (is_mergeable_vma(vma, file, vm_flags) &&
719             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
720                 pgoff_t vm_pglen;
721                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
722                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
723                         return 1;
724         }
725         return 0;
726 }
727
728 /*
729  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
730  * whether that can be merged with its predecessor or its successor.
731  * Or both (it neatly fills a hole).
732  *
733  * In most cases - when called for mmap, brk or mremap - [addr,end) is
734  * certain not to be mapped by the time vma_merge is called; but when
735  * called for mprotect, it is certain to be already mapped (either at
736  * an offset within prev, or at the start of next), and the flags of
737  * this area are about to be changed to vm_flags - and the no-change
738  * case has already been eliminated.
739  *
740  * The following mprotect cases have to be considered, where AAAA is
741  * the area passed down from mprotect_fixup, never extending beyond one
742  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
743  *
744  *     AAAA             AAAA                AAAA          AAAA
745  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
746  *    cannot merge    might become    might become    might become
747  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
748  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
749  *    mremap move:                                    PPPPNNNNNNNN 8
750  *        AAAA
751  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
752  *    might become    case 1 below    case 2 below    case 3 below
753  *
754  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
755  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
756  */
757 struct vm_area_struct *vma_merge(struct mm_struct *mm,
758                         struct vm_area_struct *prev, unsigned long addr,
759                         unsigned long end, unsigned long vm_flags,
760                         struct anon_vma *anon_vma, struct file *file,
761                         pgoff_t pgoff, struct mempolicy *policy)
762 {
763         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
764         struct vm_area_struct *area, *next;
765
766         /*
767          * We later require that vma->vm_flags == vm_flags,
768          * so this tests vma->vm_flags & VM_SPECIAL, too.
769          */
770         if (vm_flags & VM_SPECIAL)
771                 return NULL;
772
773         if (prev)
774                 next = prev->vm_next;
775         else
776                 next = mm->mmap;
777         area = next;
778         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
779                 next = next->vm_next;
780
781         /*
782          * Can it merge with the predecessor?
783          */
784         if (prev && prev->vm_end == addr &&
785                         mpol_equal(vma_policy(prev), policy) &&
786                         can_vma_merge_after(prev, vm_flags,
787                                                 anon_vma, file, pgoff)) {
788                 /*
789                  * OK, it can.  Can we now merge in the successor as well?
790                  */
791                 if (next && end == next->vm_start &&
792                                 mpol_equal(policy, vma_policy(next)) &&
793                                 can_vma_merge_before(next, vm_flags,
794                                         anon_vma, file, pgoff+pglen) &&
795                                 is_mergeable_anon_vma(prev->anon_vma,
796                                                       next->anon_vma)) {
797                                                         /* cases 1, 6 */
798                         vma_adjust(prev, prev->vm_start,
799                                 next->vm_end, prev->vm_pgoff, NULL);
800                 } else                                  /* cases 2, 5, 7 */
801                         vma_adjust(prev, prev->vm_start,
802                                 end, prev->vm_pgoff, 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                         vma_adjust(prev, prev->vm_start,
815                                 addr, prev->vm_pgoff, NULL);
816                 else                                    /* cases 3, 8 */
817                         vma_adjust(area, addr, next->vm_end,
818                                 next->vm_pgoff - pglen, NULL);
819                 return area;
820         }
821
822         return NULL;
823 }
824
825 /*
826  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
827  * neighbouring vmas for a suitable anon_vma, before it goes off
828  * to allocate a new anon_vma.  It checks because a repetitive
829  * sequence of mprotects and faults may otherwise lead to distinct
830  * anon_vmas being allocated, preventing vma merge in subsequent
831  * mprotect.
832  */
833 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
834 {
835         struct vm_area_struct *near;
836         unsigned long vm_flags;
837
838         near = vma->vm_next;
839         if (!near)
840                 goto try_prev;
841
842         /*
843          * Since only mprotect tries to remerge vmas, match flags
844          * which might be mprotected into each other later on.
845          * Neither mlock nor madvise tries to remerge at present,
846          * so leave their flags as obstructing a merge.
847          */
848         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
849         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
850
851         if (near->anon_vma && vma->vm_end == near->vm_start &&
852                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
853                         can_vma_merge_before(near, vm_flags,
854                                 NULL, vma->vm_file, vma->vm_pgoff +
855                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
856                 return near->anon_vma;
857 try_prev:
858         /*
859          * It is potentially slow to have to call find_vma_prev here.
860          * But it's only on the first write fault on the vma, not
861          * every time, and we could devise a way to avoid it later
862          * (e.g. stash info in next's anon_vma_node when assigning
863          * an anon_vma, or when trying vma_merge).  Another time.
864          */
865         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
866         if (!near)
867                 goto none;
868
869         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
870         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
871
872         if (near->anon_vma && near->vm_end == vma->vm_start &&
873                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
874                         can_vma_merge_after(near, vm_flags,
875                                 NULL, vma->vm_file, vma->vm_pgoff))
876                 return near->anon_vma;
877 none:
878         /*
879          * There's no absolute need to look only at touching neighbours:
880          * we could search further afield for "compatible" anon_vmas.
881          * But it would probably just be a waste of time searching,
882          * or lead to too many vmas hanging off the same anon_vma.
883          * We're trying to allow mprotect remerging later on,
884          * not trying to minimize memory used for anon_vmas.
885          */
886         return NULL;
887 }
888
889 #ifdef CONFIG_PROC_FS
890 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
891                                                 struct file *file, long pages)
892 {
893         const unsigned long stack_flags
894                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
895
896         if (file) {
897                 mm->shared_vm += pages;
898                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
899                         mm->exec_vm += pages;
900         } else if (flags & stack_flags)
901                 mm->stack_vm += pages;
902         if (flags & (VM_RESERVED|VM_IO))
903                 mm->reserved_vm += pages;
904 }
905 #endif /* CONFIG_PROC_FS */
906
907 /*
908  * The caller must hold down_write(current->mm->mmap_sem).
909  */
910
911 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
912                         unsigned long len, unsigned long prot,
913                         unsigned long flags, unsigned long pgoff)
914 {
915         struct mm_struct * mm = current->mm;
916         struct inode *inode;
917         unsigned int vm_flags;
918         int error;
919         int accountable = 1;
920         unsigned long reqprot = prot;
921
922         /*
923          * Does the application expect PROT_READ to imply PROT_EXEC?
924          *
925          * (the exception is when the underlying filesystem is noexec
926          *  mounted, in which case we dont add PROT_EXEC.)
927          */
928         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
929                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
930                         prot |= PROT_EXEC;
931
932         if (!len)
933                 return -EINVAL;
934
935         if (!(flags & MAP_FIXED))
936                 addr = round_hint_to_min(addr);
937
938         error = arch_mmap_check(addr, len, flags);
939         if (error)
940                 return error;
941
942         /* Careful about overflows.. */
943         len = PAGE_ALIGN(len);
944         if (!len || len > TASK_SIZE)
945                 return -ENOMEM;
946
947         /* offset overflow? */
948         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
949                return -EOVERFLOW;
950
951         /* Too many mappings? */
952         if (mm->map_count > sysctl_max_map_count)
953                 return -ENOMEM;
954
955         /* Obtain the address to map to. we verify (or select) it and ensure
956          * that it represents a valid section of the address space.
957          */
958         addr = get_unmapped_area(file, addr, len, pgoff, flags);
959         if (addr & ~PAGE_MASK)
960                 return addr;
961
962         /* Do simple checking here so the lower-level routines won't have
963          * to. we assume access permissions have been handled by the open
964          * of the memory object, so we don't do any here.
965          */
966         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
967                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
968
969         if (flags & MAP_LOCKED) {
970                 if (!can_do_mlock())
971                         return -EPERM;
972                 vm_flags |= VM_LOCKED;
973         }
974         /* mlock MCL_FUTURE? */
975         if (vm_flags & VM_LOCKED) {
976                 unsigned long locked, lock_limit;
977                 locked = len >> PAGE_SHIFT;
978                 locked += mm->locked_vm;
979                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
980                 lock_limit >>= PAGE_SHIFT;
981                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
982                         return -EAGAIN;
983         }
984
985         inode = file ? file->f_path.dentry->d_inode : NULL;
986
987         if (file) {
988                 switch (flags & MAP_TYPE) {
989                 case MAP_SHARED:
990                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
991                                 return -EACCES;
992
993                         /*
994                          * Make sure we don't allow writing to an append-only
995                          * file..
996                          */
997                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
998                                 return -EACCES;
999
1000                         /*
1001                          * Make sure there are no mandatory locks on the file.
1002                          */
1003                         if (locks_verify_locked(inode))
1004                                 return -EAGAIN;
1005
1006                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1007                         if (!(file->f_mode & FMODE_WRITE))
1008                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1009
1010                         /* fall through */
1011                 case MAP_PRIVATE:
1012                         if (!(file->f_mode & FMODE_READ))
1013                                 return -EACCES;
1014                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1015                                 if (vm_flags & VM_EXEC)
1016                                         return -EPERM;
1017                                 vm_flags &= ~VM_MAYEXEC;
1018                         }
1019                         if (is_file_hugepages(file))
1020                                 accountable = 0;
1021
1022                         if (!file->f_op || !file->f_op->mmap)
1023                                 return -ENODEV;
1024                         break;
1025
1026                 default:
1027                         return -EINVAL;
1028                 }
1029         } else {
1030                 switch (flags & MAP_TYPE) {
1031                 case MAP_SHARED:
1032                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1033                         break;
1034                 case MAP_PRIVATE:
1035                         /*
1036                          * Set pgoff according to addr for anon_vma.
1037                          */
1038                         pgoff = addr >> PAGE_SHIFT;
1039                         break;
1040                 default:
1041                         return -EINVAL;
1042                 }
1043         }
1044
1045         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1046         if (error)
1047                 return error;
1048
1049         return mmap_region(file, addr, len, flags, vm_flags, pgoff,
1050                            accountable);
1051 }
1052 EXPORT_SYMBOL(do_mmap_pgoff);
1053
1054 /*
1055  * Some shared mappigns will want the pages marked read-only
1056  * to track write events. If so, we'll downgrade vm_page_prot
1057  * to the private version (using protection_map[] without the
1058  * VM_SHARED bit).
1059  */
1060 int vma_wants_writenotify(struct vm_area_struct *vma)
1061 {
1062         unsigned int vm_flags = vma->vm_flags;
1063
1064         /* If it was private or non-writable, the write bit is already clear */
1065         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1066                 return 0;
1067
1068         /* The backer wishes to know when pages are first written to? */
1069         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1070                 return 1;
1071
1072         /* The open routine did something to the protections already? */
1073         if (pgprot_val(vma->vm_page_prot) !=
1074             pgprot_val(vm_get_page_prot(vm_flags)))
1075                 return 0;
1076
1077         /* Specialty mapping? */
1078         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1079                 return 0;
1080
1081         /* Can the mapping track the dirty pages? */
1082         return vma->vm_file && vma->vm_file->f_mapping &&
1083                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1084 }
1085
1086 unsigned long mmap_region(struct file *file, unsigned long addr,
1087                           unsigned long len, unsigned long flags,
1088                           unsigned int vm_flags, unsigned long pgoff,
1089                           int accountable)
1090 {
1091         struct mm_struct *mm = current->mm;
1092         struct vm_area_struct *vma, *prev;
1093         int correct_wcount = 0;
1094         int error;
1095         struct rb_node **rb_link, *rb_parent;
1096         unsigned long charged = 0;
1097         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1098
1099         /* Clear old maps */
1100         error = -ENOMEM;
1101 munmap_back:
1102         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1103         if (vma && vma->vm_start < addr + len) {
1104                 if (do_munmap(mm, addr, len))
1105                         return -ENOMEM;
1106                 goto munmap_back;
1107         }
1108
1109         /* Check against address space limit. */
1110         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1111                 return -ENOMEM;
1112
1113         if (accountable && (!(flags & MAP_NORESERVE) ||
1114                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1115                 if (vm_flags & VM_SHARED) {
1116                         /* Check memory availability in shmem_file_setup? */
1117                         vm_flags |= VM_ACCOUNT;
1118                 } else if (vm_flags & VM_WRITE) {
1119                         /*
1120                          * Private writable mapping: check memory availability
1121                          */
1122                         charged = len >> PAGE_SHIFT;
1123                         if (security_vm_enough_memory(charged))
1124                                 return -ENOMEM;
1125                         vm_flags |= VM_ACCOUNT;
1126                 }
1127         }
1128
1129         /*
1130          * Can we just expand an old private anonymous mapping?
1131          * The VM_SHARED test is necessary because shmem_zero_setup
1132          * will create the file object for a shared anonymous map below.
1133          */
1134         if (!file && !(vm_flags & VM_SHARED) &&
1135             vma_merge(mm, prev, addr, addr + len, vm_flags,
1136                                         NULL, NULL, pgoff, NULL))
1137                 goto out;
1138
1139         /*
1140          * Determine the object being mapped and call the appropriate
1141          * specific mapper. the address has already been validated, but
1142          * not unmapped, but the maps are removed from the list.
1143          */
1144         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1145         if (!vma) {
1146                 error = -ENOMEM;
1147                 goto unacct_error;
1148         }
1149
1150         vma->vm_mm = mm;
1151         vma->vm_start = addr;
1152         vma->vm_end = addr + len;
1153         vma->vm_flags = vm_flags;
1154         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1155         vma->vm_pgoff = pgoff;
1156
1157         if (file) {
1158                 error = -EINVAL;
1159                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1160                         goto free_vma;
1161                 if (vm_flags & VM_DENYWRITE) {
1162                         error = deny_write_access(file);
1163                         if (error)
1164                                 goto free_vma;
1165                         correct_wcount = 1;
1166                 }
1167                 vma->vm_file = file;
1168                 get_file(file);
1169                 error = file->f_op->mmap(file, vma);
1170                 if (error)
1171                         goto unmap_and_free_vma;
1172                 if (vm_flags & VM_EXECUTABLE)
1173                         added_exe_file_vma(mm);
1174         } else if (vm_flags & VM_SHARED) {
1175                 error = shmem_zero_setup(vma);
1176                 if (error)
1177                         goto free_vma;
1178         }
1179
1180         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1181          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1182          * that memory reservation must be checked; but that reservation
1183          * belongs to shared memory object, not to vma: so now clear it.
1184          */
1185         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1186                 vma->vm_flags &= ~VM_ACCOUNT;
1187
1188         /* Can addr have changed??
1189          *
1190          * Answer: Yes, several device drivers can do it in their
1191          *         f_op->mmap method. -DaveM
1192          */
1193         addr = vma->vm_start;
1194         pgoff = vma->vm_pgoff;
1195         vm_flags = vma->vm_flags;
1196
1197         if (vma_wants_writenotify(vma))
1198                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1199
1200         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1201                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1202                 mpol_put(vma_policy(vma));
1203                 kmem_cache_free(vm_area_cachep, vma);
1204                 fput(file);
1205                 if (vm_flags & VM_EXECUTABLE)
1206                         removed_exe_file_vma(mm);
1207         } else {
1208                 vma_link(mm, vma, prev, rb_link, rb_parent);
1209                 file = vma->vm_file;
1210         }
1211
1212         /* Once vma denies write, undo our temporary denial count */
1213         if (correct_wcount)
1214                 atomic_inc(&inode->i_writecount);
1215 out:
1216         mm->total_vm += len >> PAGE_SHIFT;
1217         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1218         if (vm_flags & VM_LOCKED) {
1219                 mm->locked_vm += len >> PAGE_SHIFT;
1220                 make_pages_present(addr, addr + len);
1221         }
1222         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1223                 make_pages_present(addr, addr + len);
1224         return addr;
1225
1226 unmap_and_free_vma:
1227         if (correct_wcount)
1228                 atomic_inc(&inode->i_writecount);
1229         vma->vm_file = NULL;
1230         fput(file);
1231
1232         /* Undo any partial mapping done by a device driver. */
1233         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1234         charged = 0;
1235 free_vma:
1236         kmem_cache_free(vm_area_cachep, vma);
1237 unacct_error:
1238         if (charged)
1239                 vm_unacct_memory(charged);
1240         return error;
1241 }
1242
1243 /* Get an address range which is currently unmapped.
1244  * For shmat() with addr=0.
1245  *
1246  * Ugly calling convention alert:
1247  * Return value with the low bits set means error value,
1248  * ie
1249  *      if (ret & ~PAGE_MASK)
1250  *              error = ret;
1251  *
1252  * This function "knows" that -ENOMEM has the bits set.
1253  */
1254 #ifndef HAVE_ARCH_UNMAPPED_AREA
1255 unsigned long
1256 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1257                 unsigned long len, unsigned long pgoff, unsigned long flags)
1258 {
1259         struct mm_struct *mm = current->mm;
1260         struct vm_area_struct *vma;
1261         unsigned long start_addr;
1262
1263         if (len > TASK_SIZE)
1264                 return -ENOMEM;
1265
1266         if (flags & MAP_FIXED)
1267                 return addr;
1268
1269         if (addr) {
1270                 addr = PAGE_ALIGN(addr);
1271                 vma = find_vma(mm, addr);
1272                 if (TASK_SIZE - len >= addr &&
1273                     (!vma || addr + len <= vma->vm_start))
1274                         return addr;
1275         }
1276         if (len > mm->cached_hole_size) {
1277                 start_addr = addr = mm->free_area_cache;
1278         } else {
1279                 start_addr = addr = TASK_UNMAPPED_BASE;
1280                 mm->cached_hole_size = 0;
1281         }
1282
1283 full_search:
1284         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1285                 /* At this point:  (!vma || addr < vma->vm_end). */
1286                 if (TASK_SIZE - len < addr) {
1287                         /*
1288                          * Start a new search - just in case we missed
1289                          * some holes.
1290                          */
1291                         if (start_addr != TASK_UNMAPPED_BASE) {
1292                                 addr = TASK_UNMAPPED_BASE;
1293                                 start_addr = addr;
1294                                 mm->cached_hole_size = 0;
1295                                 goto full_search;
1296                         }
1297                         return -ENOMEM;
1298                 }
1299                 if (!vma || addr + len <= vma->vm_start) {
1300                         /*
1301                          * Remember the place where we stopped the search:
1302                          */
1303                         mm->free_area_cache = addr + len;
1304                         return addr;
1305                 }
1306                 if (addr + mm->cached_hole_size < vma->vm_start)
1307                         mm->cached_hole_size = vma->vm_start - addr;
1308                 addr = vma->vm_end;
1309         }
1310 }
1311 #endif  
1312
1313 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1314 {
1315         /*
1316          * Is this a new hole at the lowest possible address?
1317          */
1318         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1319                 mm->free_area_cache = addr;
1320                 mm->cached_hole_size = ~0UL;
1321         }
1322 }
1323
1324 /*
1325  * This mmap-allocator allocates new areas top-down from below the
1326  * stack's low limit (the base):
1327  */
1328 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1329 unsigned long
1330 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1331                           const unsigned long len, const unsigned long pgoff,
1332                           const unsigned long flags)
1333 {
1334         struct vm_area_struct *vma;
1335         struct mm_struct *mm = current->mm;
1336         unsigned long addr = addr0;
1337
1338         /* requested length too big for entire address space */
1339         if (len > TASK_SIZE)
1340                 return -ENOMEM;
1341
1342         if (flags & MAP_FIXED)
1343                 return addr;
1344
1345         /* requesting a specific address */
1346         if (addr) {
1347                 addr = PAGE_ALIGN(addr);
1348                 vma = find_vma(mm, addr);
1349                 if (TASK_SIZE - len >= addr &&
1350                                 (!vma || addr + len <= vma->vm_start))
1351                         return addr;
1352         }
1353
1354         /* check if free_area_cache is useful for us */
1355         if (len <= mm->cached_hole_size) {
1356                 mm->cached_hole_size = 0;
1357                 mm->free_area_cache = mm->mmap_base;
1358         }
1359
1360         /* either no address requested or can't fit in requested address hole */
1361         addr = mm->free_area_cache;
1362
1363         /* make sure it can fit in the remaining address space */
1364         if (addr > len) {
1365                 vma = find_vma(mm, addr-len);
1366                 if (!vma || addr <= vma->vm_start)
1367                         /* remember the address as a hint for next time */
1368                         return (mm->free_area_cache = addr-len);
1369         }
1370
1371         if (mm->mmap_base < len)
1372                 goto bottomup;
1373
1374         addr = mm->mmap_base-len;
1375
1376         do {
1377                 /*
1378                  * Lookup failure means no vma is above this address,
1379                  * else if new region fits below vma->vm_start,
1380                  * return with success:
1381                  */
1382                 vma = find_vma(mm, addr);
1383                 if (!vma || addr+len <= vma->vm_start)
1384                         /* remember the address as a hint for next time */
1385                         return (mm->free_area_cache = addr);
1386
1387                 /* remember the largest hole we saw so far */
1388                 if (addr + mm->cached_hole_size < vma->vm_start)
1389                         mm->cached_hole_size = vma->vm_start - addr;
1390
1391                 /* try just below the current vma->vm_start */
1392                 addr = vma->vm_start-len;
1393         } while (len < vma->vm_start);
1394
1395 bottomup:
1396         /*
1397          * A failed mmap() very likely causes application failure,
1398          * so fall back to the bottom-up function here. This scenario
1399          * can happen with large stack limits and large mmap()
1400          * allocations.
1401          */
1402         mm->cached_hole_size = ~0UL;
1403         mm->free_area_cache = TASK_UNMAPPED_BASE;
1404         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1405         /*
1406          * Restore the topdown base:
1407          */
1408         mm->free_area_cache = mm->mmap_base;
1409         mm->cached_hole_size = ~0UL;
1410
1411         return addr;
1412 }
1413 #endif
1414
1415 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1416 {
1417         /*
1418          * Is this a new hole at the highest possible address?
1419          */
1420         if (addr > mm->free_area_cache)
1421                 mm->free_area_cache = addr;
1422
1423         /* dont allow allocations above current base */
1424         if (mm->free_area_cache > mm->mmap_base)
1425                 mm->free_area_cache = mm->mmap_base;
1426 }
1427
1428 unsigned long
1429 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1430                 unsigned long pgoff, unsigned long flags)
1431 {
1432         unsigned long (*get_area)(struct file *, unsigned long,
1433                                   unsigned long, unsigned long, unsigned long);
1434
1435         get_area = current->mm->get_unmapped_area;
1436         if (file && file->f_op && file->f_op->get_unmapped_area)
1437                 get_area = file->f_op->get_unmapped_area;
1438         addr = get_area(file, addr, len, pgoff, flags);
1439         if (IS_ERR_VALUE(addr))
1440                 return addr;
1441
1442         if (addr > TASK_SIZE - len)
1443                 return -ENOMEM;
1444         if (addr & ~PAGE_MASK)
1445                 return -EINVAL;
1446
1447         return arch_rebalance_pgtables(addr, len);
1448 }
1449
1450 EXPORT_SYMBOL(get_unmapped_area);
1451
1452 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1453 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1454 {
1455         struct vm_area_struct *vma = NULL;
1456
1457         if (mm) {
1458                 /* Check the cache first. */
1459                 /* (Cache hit rate is typically around 35%.) */
1460                 vma = mm->mmap_cache;
1461                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1462                         struct rb_node * rb_node;
1463
1464                         rb_node = mm->mm_rb.rb_node;
1465                         vma = NULL;
1466
1467                         while (rb_node) {
1468                                 struct vm_area_struct * vma_tmp;
1469
1470                                 vma_tmp = rb_entry(rb_node,
1471                                                 struct vm_area_struct, vm_rb);
1472
1473                                 if (vma_tmp->vm_end > addr) {
1474                                         vma = vma_tmp;
1475                                         if (vma_tmp->vm_start <= addr)
1476                                                 break;
1477                                         rb_node = rb_node->rb_left;
1478                                 } else
1479                                         rb_node = rb_node->rb_right;
1480                         }
1481                         if (vma)
1482                                 mm->mmap_cache = vma;
1483                 }
1484         }
1485         return vma;
1486 }
1487
1488 EXPORT_SYMBOL(find_vma);
1489
1490 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1491 struct vm_area_struct *
1492 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1493                         struct vm_area_struct **pprev)
1494 {
1495         struct vm_area_struct *vma = NULL, *prev = NULL;
1496         struct rb_node * rb_node;
1497         if (!mm)
1498                 goto out;
1499
1500         /* Guard against addr being lower than the first VMA */
1501         vma = mm->mmap;
1502
1503         /* Go through the RB tree quickly. */
1504         rb_node = mm->mm_rb.rb_node;
1505
1506         while (rb_node) {
1507                 struct vm_area_struct *vma_tmp;
1508                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1509
1510                 if (addr < vma_tmp->vm_end) {
1511                         rb_node = rb_node->rb_left;
1512                 } else {
1513                         prev = vma_tmp;
1514                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1515                                 break;
1516                         rb_node = rb_node->rb_right;
1517                 }
1518         }
1519
1520 out:
1521         *pprev = prev;
1522         return prev ? prev->vm_next : vma;
1523 }
1524
1525 /*
1526  * Verify that the stack growth is acceptable and
1527  * update accounting. This is shared with both the
1528  * grow-up and grow-down cases.
1529  */
1530 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1531 {
1532         struct mm_struct *mm = vma->vm_mm;
1533         struct rlimit *rlim = current->signal->rlim;
1534         unsigned long new_start;
1535
1536         /* address space limit tests */
1537         if (!may_expand_vm(mm, grow))
1538                 return -ENOMEM;
1539
1540         /* Stack limit test */
1541         if (size > rlim[RLIMIT_STACK].rlim_cur)
1542                 return -ENOMEM;
1543
1544         /* mlock limit tests */
1545         if (vma->vm_flags & VM_LOCKED) {
1546                 unsigned long locked;
1547                 unsigned long limit;
1548                 locked = mm->locked_vm + grow;
1549                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1550                 if (locked > limit && !capable(CAP_IPC_LOCK))
1551                         return -ENOMEM;
1552         }
1553
1554         /* Check to ensure the stack will not grow into a hugetlb-only region */
1555         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1556                         vma->vm_end - size;
1557         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1558                 return -EFAULT;
1559
1560         /*
1561          * Overcommit..  This must be the final test, as it will
1562          * update security statistics.
1563          */
1564         if (security_vm_enough_memory(grow))
1565                 return -ENOMEM;
1566
1567         /* Ok, everything looks good - let it rip */
1568         mm->total_vm += grow;
1569         if (vma->vm_flags & VM_LOCKED)
1570                 mm->locked_vm += grow;
1571         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1572         return 0;
1573 }
1574
1575 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1576 /*
1577  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1578  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1579  */
1580 #ifndef CONFIG_IA64
1581 static inline
1582 #endif
1583 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1584 {
1585         int error;
1586
1587         if (!(vma->vm_flags & VM_GROWSUP))
1588                 return -EFAULT;
1589
1590         /*
1591          * We must make sure the anon_vma is allocated
1592          * so that the anon_vma locking is not a noop.
1593          */
1594         if (unlikely(anon_vma_prepare(vma)))
1595                 return -ENOMEM;
1596         anon_vma_lock(vma);
1597
1598         /*
1599          * vma->vm_start/vm_end cannot change under us because the caller
1600          * is required to hold the mmap_sem in read mode.  We need the
1601          * anon_vma lock to serialize against concurrent expand_stacks.
1602          * Also guard against wrapping around to address 0.
1603          */
1604         if (address < PAGE_ALIGN(address+4))
1605                 address = PAGE_ALIGN(address+4);
1606         else {
1607                 anon_vma_unlock(vma);
1608                 return -ENOMEM;
1609         }
1610         error = 0;
1611
1612         /* Somebody else might have raced and expanded it already */
1613         if (address > vma->vm_end) {
1614                 unsigned long size, grow;
1615
1616                 size = address - vma->vm_start;
1617                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1618
1619                 error = acct_stack_growth(vma, size, grow);
1620                 if (!error)
1621                         vma->vm_end = address;
1622         }
1623         anon_vma_unlock(vma);
1624         return error;
1625 }
1626 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1627
1628 /*
1629  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1630  */
1631 static inline int expand_downwards(struct vm_area_struct *vma,
1632                                    unsigned long address)
1633 {
1634         int error;
1635
1636         /*
1637          * We must make sure the anon_vma is allocated
1638          * so that the anon_vma locking is not a noop.
1639          */
1640         if (unlikely(anon_vma_prepare(vma)))
1641                 return -ENOMEM;
1642
1643         address &= PAGE_MASK;
1644         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1645         if (error)
1646                 return error;
1647
1648         anon_vma_lock(vma);
1649
1650         /*
1651          * vma->vm_start/vm_end cannot change under us because the caller
1652          * is required to hold the mmap_sem in read mode.  We need the
1653          * anon_vma lock to serialize against concurrent expand_stacks.
1654          */
1655
1656         /* Somebody else might have raced and expanded it already */
1657         if (address < vma->vm_start) {
1658                 unsigned long size, grow;
1659
1660                 size = vma->vm_end - address;
1661                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1662
1663                 error = acct_stack_growth(vma, size, grow);
1664                 if (!error) {
1665                         vma->vm_start = address;
1666                         vma->vm_pgoff -= grow;
1667                 }
1668         }
1669         anon_vma_unlock(vma);
1670         return error;
1671 }
1672
1673 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1674 {
1675         return expand_downwards(vma, address);
1676 }
1677
1678 #ifdef CONFIG_STACK_GROWSUP
1679 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1680 {
1681         return expand_upwards(vma, address);
1682 }
1683
1684 struct vm_area_struct *
1685 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1686 {
1687         struct vm_area_struct *vma, *prev;
1688
1689         addr &= PAGE_MASK;
1690         vma = find_vma_prev(mm, addr, &prev);
1691         if (vma && (vma->vm_start <= addr))
1692                 return vma;
1693         if (!prev || expand_stack(prev, addr))
1694                 return NULL;
1695         if (prev->vm_flags & VM_LOCKED)
1696                 make_pages_present(addr, prev->vm_end);
1697         return prev;
1698 }
1699 #else
1700 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1701 {
1702         return expand_downwards(vma, address);
1703 }
1704
1705 struct vm_area_struct *
1706 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1707 {
1708         struct vm_area_struct * vma;
1709         unsigned long start;
1710
1711         addr &= PAGE_MASK;
1712         vma = find_vma(mm,addr);
1713         if (!vma)
1714                 return NULL;
1715         if (vma->vm_start <= addr)
1716                 return vma;
1717         if (!(vma->vm_flags & VM_GROWSDOWN))
1718                 return NULL;
1719         start = vma->vm_start;
1720         if (expand_stack(vma, addr))
1721                 return NULL;
1722         if (vma->vm_flags & VM_LOCKED)
1723                 make_pages_present(addr, start);
1724         return vma;
1725 }
1726 #endif
1727
1728 /*
1729  * Ok - we have the memory areas we should free on the vma list,
1730  * so release them, and do the vma updates.
1731  *
1732  * Called with the mm semaphore held.
1733  */
1734 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1735 {
1736         /* Update high watermark before we lower total_vm */
1737         update_hiwater_vm(mm);
1738         do {
1739                 long nrpages = vma_pages(vma);
1740
1741                 mm->total_vm -= nrpages;
1742                 if (vma->vm_flags & VM_LOCKED)
1743                         mm->locked_vm -= nrpages;
1744                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1745                 vma = remove_vma(vma);
1746         } while (vma);
1747         validate_mm(mm);
1748 }
1749
1750 /*
1751  * Get rid of page table information in the indicated region.
1752  *
1753  * Called with the mm semaphore held.
1754  */
1755 static void unmap_region(struct mm_struct *mm,
1756                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1757                 unsigned long start, unsigned long end)
1758 {
1759         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1760         struct mmu_gather *tlb;
1761         unsigned long nr_accounted = 0;
1762
1763         lru_add_drain();
1764         tlb = tlb_gather_mmu(mm, 0);
1765         update_hiwater_rss(mm);
1766         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1767         vm_unacct_memory(nr_accounted);
1768         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1769                                  next? next->vm_start: 0);
1770         tlb_finish_mmu(tlb, start, end);
1771 }
1772
1773 /*
1774  * Create a list of vma's touched by the unmap, removing them from the mm's
1775  * vma list as we go..
1776  */
1777 static void
1778 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1779         struct vm_area_struct *prev, unsigned long end)
1780 {
1781         struct vm_area_struct **insertion_point;
1782         struct vm_area_struct *tail_vma = NULL;
1783         unsigned long addr;
1784
1785         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1786         do {
1787                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1788                 mm->map_count--;
1789                 tail_vma = vma;
1790                 vma = vma->vm_next;
1791         } while (vma && vma->vm_start < end);
1792         *insertion_point = vma;
1793         tail_vma->vm_next = NULL;
1794         if (mm->unmap_area == arch_unmap_area)
1795                 addr = prev ? prev->vm_end : mm->mmap_base;
1796         else
1797                 addr = vma ?  vma->vm_start : mm->mmap_base;
1798         mm->unmap_area(mm, addr);
1799         mm->mmap_cache = NULL;          /* Kill the cache. */
1800 }
1801
1802 /*
1803  * Split a vma into two pieces at address 'addr', a new vma is allocated
1804  * either for the first part or the tail.
1805  */
1806 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1807               unsigned long addr, int new_below)
1808 {
1809         struct mempolicy *pol;
1810         struct vm_area_struct *new;
1811
1812         if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1813                 return -EINVAL;
1814
1815         if (mm->map_count >= sysctl_max_map_count)
1816                 return -ENOMEM;
1817
1818         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1819         if (!new)
1820                 return -ENOMEM;
1821
1822         /* most fields are the same, copy all, and then fixup */
1823         *new = *vma;
1824
1825         if (new_below)
1826                 new->vm_end = addr;
1827         else {
1828                 new->vm_start = addr;
1829                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1830         }
1831
1832         pol = mpol_dup(vma_policy(vma));
1833         if (IS_ERR(pol)) {
1834                 kmem_cache_free(vm_area_cachep, new);
1835                 return PTR_ERR(pol);
1836         }
1837         vma_set_policy(new, pol);
1838
1839         if (new->vm_file) {
1840                 get_file(new->vm_file);
1841                 if (vma->vm_flags & VM_EXECUTABLE)
1842                         added_exe_file_vma(mm);
1843         }
1844
1845         if (new->vm_ops && new->vm_ops->open)
1846                 new->vm_ops->open(new);
1847
1848         if (new_below)
1849                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1850                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1851         else
1852                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1853
1854         return 0;
1855 }
1856
1857 /* Munmap is split into 2 main parts -- this part which finds
1858  * what needs doing, and the areas themselves, which do the
1859  * work.  This now handles partial unmappings.
1860  * Jeremy Fitzhardinge <jeremy@goop.org>
1861  */
1862 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1863 {
1864         unsigned long end;
1865         struct vm_area_struct *vma, *prev, *last;
1866
1867         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1868                 return -EINVAL;
1869
1870         if ((len = PAGE_ALIGN(len)) == 0)
1871                 return -EINVAL;
1872
1873         /* Find the first overlapping VMA */
1874         vma = find_vma_prev(mm, start, &prev);
1875         if (!vma)
1876                 return 0;
1877         /* we have  start < vma->vm_end  */
1878
1879         /* if it doesn't overlap, we have nothing.. */
1880         end = start + len;
1881         if (vma->vm_start >= end)
1882                 return 0;
1883
1884         /*
1885          * If we need to split any vma, do it now to save pain later.
1886          *
1887          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1888          * unmapped vm_area_struct will remain in use: so lower split_vma
1889          * places tmp vma above, and higher split_vma places tmp vma below.
1890          */
1891         if (start > vma->vm_start) {
1892                 int error = split_vma(mm, vma, start, 0);
1893                 if (error)
1894                         return error;
1895                 prev = vma;
1896         }
1897
1898         /* Does it split the last one? */
1899         last = find_vma(mm, end);
1900         if (last && end > last->vm_start) {
1901                 int error = split_vma(mm, last, end, 1);
1902                 if (error)
1903                         return error;
1904         }
1905         vma = prev? prev->vm_next: mm->mmap;
1906
1907         /*
1908          * Remove the vma's, and unmap the actual pages
1909          */
1910         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1911         unmap_region(mm, vma, prev, start, end);
1912
1913         /* Fix up all other VM information */
1914         remove_vma_list(mm, vma);
1915
1916         return 0;
1917 }
1918
1919 EXPORT_SYMBOL(do_munmap);
1920
1921 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1922 {
1923         int ret;
1924         struct mm_struct *mm = current->mm;
1925
1926         profile_munmap(addr);
1927
1928         down_write(&mm->mmap_sem);
1929         ret = do_munmap(mm, addr, len);
1930         up_write(&mm->mmap_sem);
1931         return ret;
1932 }
1933
1934 static inline void verify_mm_writelocked(struct mm_struct *mm)
1935 {
1936 #ifdef CONFIG_DEBUG_VM
1937         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1938                 WARN_ON(1);
1939                 up_read(&mm->mmap_sem);
1940         }
1941 #endif
1942 }
1943
1944 /*
1945  *  this is really a simplified "do_mmap".  it only handles
1946  *  anonymous maps.  eventually we may be able to do some
1947  *  brk-specific accounting here.
1948  */
1949 unsigned long do_brk(unsigned long addr, unsigned long len)
1950 {
1951         struct mm_struct * mm = current->mm;
1952         struct vm_area_struct * vma, * prev;
1953         unsigned long flags;
1954         struct rb_node ** rb_link, * rb_parent;
1955         pgoff_t pgoff = addr >> PAGE_SHIFT;
1956         int error;
1957
1958         len = PAGE_ALIGN(len);
1959         if (!len)
1960                 return addr;
1961
1962         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1963                 return -EINVAL;
1964
1965         if (is_hugepage_only_range(mm, addr, len))
1966                 return -EINVAL;
1967
1968         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1969         if (error)
1970                 return error;
1971
1972         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1973
1974         error = arch_mmap_check(addr, len, flags);
1975         if (error)
1976                 return error;
1977
1978         /*
1979          * mlock MCL_FUTURE?
1980          */
1981         if (mm->def_flags & VM_LOCKED) {
1982                 unsigned long locked, lock_limit;
1983                 locked = len >> PAGE_SHIFT;
1984                 locked += mm->locked_vm;
1985                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1986                 lock_limit >>= PAGE_SHIFT;
1987                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1988                         return -EAGAIN;
1989         }
1990
1991         /*
1992          * mm->mmap_sem is required to protect against another thread
1993          * changing the mappings in case we sleep.
1994          */
1995         verify_mm_writelocked(mm);
1996
1997         /*
1998          * Clear old maps.  this also does some error checking for us
1999          */
2000  munmap_back:
2001         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2002         if (vma && vma->vm_start < addr + len) {
2003                 if (do_munmap(mm, addr, len))
2004                         return -ENOMEM;
2005                 goto munmap_back;
2006         }
2007
2008         /* Check against address space limits *after* clearing old maps... */
2009         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2010                 return -ENOMEM;
2011
2012         if (mm->map_count > sysctl_max_map_count)
2013                 return -ENOMEM;
2014
2015         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2016                 return -ENOMEM;
2017
2018         /* Can we just expand an old private anonymous mapping? */
2019         if (vma_merge(mm, prev, addr, addr + len, flags,
2020                                         NULL, NULL, pgoff, NULL))
2021                 goto out;
2022
2023         /*
2024          * create a vma struct for an anonymous mapping
2025          */
2026         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2027         if (!vma) {
2028                 vm_unacct_memory(len >> PAGE_SHIFT);
2029                 return -ENOMEM;
2030         }
2031
2032         vma->vm_mm = mm;
2033         vma->vm_start = addr;
2034         vma->vm_end = addr + len;
2035         vma->vm_pgoff = pgoff;
2036         vma->vm_flags = flags;
2037         vma->vm_page_prot = vm_get_page_prot(flags);
2038         vma_link(mm, vma, prev, rb_link, rb_parent);
2039 out:
2040         mm->total_vm += len >> PAGE_SHIFT;
2041         if (flags & VM_LOCKED) {
2042                 mm->locked_vm += len >> PAGE_SHIFT;
2043                 make_pages_present(addr, addr + len);
2044         }
2045         return addr;
2046 }
2047
2048 EXPORT_SYMBOL(do_brk);
2049
2050 /* Release all mmaps. */
2051 void exit_mmap(struct mm_struct *mm)
2052 {
2053         struct mmu_gather *tlb;
2054         struct vm_area_struct *vma = mm->mmap;
2055         unsigned long nr_accounted = 0;
2056         unsigned long end;
2057
2058         /* mm's last user has gone, and its about to be pulled down */
2059         arch_exit_mmap(mm);
2060
2061         lru_add_drain();
2062         flush_cache_mm(mm);
2063         tlb = tlb_gather_mmu(mm, 1);
2064         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2065         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2066         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2067         vm_unacct_memory(nr_accounted);
2068         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2069         tlb_finish_mmu(tlb, 0, end);
2070
2071         /*
2072          * Walk the list again, actually closing and freeing it,
2073          * with preemption enabled, without holding any MM locks.
2074          */
2075         while (vma)
2076                 vma = remove_vma(vma);
2077
2078         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2079 }
2080
2081 /* Insert vm structure into process list sorted by address
2082  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2083  * then i_mmap_lock is taken here.
2084  */
2085 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2086 {
2087         struct vm_area_struct * __vma, * prev;
2088         struct rb_node ** rb_link, * rb_parent;
2089
2090         /*
2091          * The vm_pgoff of a purely anonymous vma should be irrelevant
2092          * until its first write fault, when page's anon_vma and index
2093          * are set.  But now set the vm_pgoff it will almost certainly
2094          * end up with (unless mremap moves it elsewhere before that
2095          * first wfault), so /proc/pid/maps tells a consistent story.
2096          *
2097          * By setting it to reflect the virtual start address of the
2098          * vma, merges and splits can happen in a seamless way, just
2099          * using the existing file pgoff checks and manipulations.
2100          * Similarly in do_mmap_pgoff and in do_brk.
2101          */
2102         if (!vma->vm_file) {
2103                 BUG_ON(vma->anon_vma);
2104                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2105         }
2106         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2107         if (__vma && __vma->vm_start < vma->vm_end)
2108                 return -ENOMEM;
2109         if ((vma->vm_flags & VM_ACCOUNT) &&
2110              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2111                 return -ENOMEM;
2112         vma_link(mm, vma, prev, rb_link, rb_parent);
2113         return 0;
2114 }
2115
2116 /*
2117  * Copy the vma structure to a new location in the same mm,
2118  * prior to moving page table entries, to effect an mremap move.
2119  */
2120 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2121         unsigned long addr, unsigned long len, pgoff_t pgoff)
2122 {
2123         struct vm_area_struct *vma = *vmap;
2124         unsigned long vma_start = vma->vm_start;
2125         struct mm_struct *mm = vma->vm_mm;
2126         struct vm_area_struct *new_vma, *prev;
2127         struct rb_node **rb_link, *rb_parent;
2128         struct mempolicy *pol;
2129
2130         /*
2131          * If anonymous vma has not yet been faulted, update new pgoff
2132          * to match new location, to increase its chance of merging.
2133          */
2134         if (!vma->vm_file && !vma->anon_vma)
2135                 pgoff = addr >> PAGE_SHIFT;
2136
2137         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2138         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2139                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2140         if (new_vma) {
2141                 /*
2142                  * Source vma may have been merged into new_vma
2143                  */
2144                 if (vma_start >= new_vma->vm_start &&
2145                     vma_start < new_vma->vm_end)
2146                         *vmap = new_vma;
2147         } else {
2148                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2149                 if (new_vma) {
2150                         *new_vma = *vma;
2151                         pol = mpol_dup(vma_policy(vma));
2152                         if (IS_ERR(pol)) {
2153                                 kmem_cache_free(vm_area_cachep, new_vma);
2154                                 return NULL;
2155                         }
2156                         vma_set_policy(new_vma, pol);
2157                         new_vma->vm_start = addr;
2158                         new_vma->vm_end = addr + len;
2159                         new_vma->vm_pgoff = pgoff;
2160                         if (new_vma->vm_file) {
2161                                 get_file(new_vma->vm_file);
2162                                 if (vma->vm_flags & VM_EXECUTABLE)
2163                                         added_exe_file_vma(mm);
2164                         }
2165                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2166                                 new_vma->vm_ops->open(new_vma);
2167                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2168                 }
2169         }
2170         return new_vma;
2171 }
2172
2173 /*
2174  * Return true if the calling process may expand its vm space by the passed
2175  * number of pages
2176  */
2177 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2178 {
2179         unsigned long cur = mm->total_vm;       /* pages */
2180         unsigned long lim;
2181
2182         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2183
2184         if (cur + npages > lim)
2185                 return 0;
2186         return 1;
2187 }
2188
2189
2190 static int special_mapping_fault(struct vm_area_struct *vma,
2191                                 struct vm_fault *vmf)
2192 {
2193         pgoff_t pgoff;
2194         struct page **pages;
2195
2196         /*
2197          * special mappings have no vm_file, and in that case, the mm
2198          * uses vm_pgoff internally. So we have to subtract it from here.
2199          * We are allowed to do this because we are the mm; do not copy
2200          * this code into drivers!
2201          */
2202         pgoff = vmf->pgoff - vma->vm_pgoff;
2203
2204         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2205                 pgoff--;
2206
2207         if (*pages) {
2208                 struct page *page = *pages;
2209                 get_page(page);
2210                 vmf->page = page;
2211                 return 0;
2212         }
2213
2214         return VM_FAULT_SIGBUS;
2215 }
2216
2217 /*
2218  * Having a close hook prevents vma merging regardless of flags.
2219  */
2220 static void special_mapping_close(struct vm_area_struct *vma)
2221 {
2222 }
2223
2224 static struct vm_operations_struct special_mapping_vmops = {
2225         .close = special_mapping_close,
2226         .fault = special_mapping_fault,
2227 };
2228
2229 /*
2230  * Called with mm->mmap_sem held for writing.
2231  * Insert a new vma covering the given region, with the given flags.
2232  * Its pages are supplied by the given array of struct page *.
2233  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2234  * The region past the last page supplied will always produce SIGBUS.
2235  * The array pointer and the pages it points to are assumed to stay alive
2236  * for as long as this mapping might exist.
2237  */
2238 int install_special_mapping(struct mm_struct *mm,
2239                             unsigned long addr, unsigned long len,
2240                             unsigned long vm_flags, struct page **pages)
2241 {
2242         struct vm_area_struct *vma;
2243
2244         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2245         if (unlikely(vma == NULL))
2246                 return -ENOMEM;
2247
2248         vma->vm_mm = mm;
2249         vma->vm_start = addr;
2250         vma->vm_end = addr + len;
2251
2252         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2253         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2254
2255         vma->vm_ops = &special_mapping_vmops;
2256         vma->vm_private_data = pages;
2257
2258         if (unlikely(insert_vm_struct(mm, vma))) {
2259                 kmem_cache_free(vm_area_cachep, vma);
2260                 return -ENOMEM;
2261         }
2262
2263         mm->total_vm += len >> PAGE_SHIFT;
2264
2265         return 0;
2266 }