mmu-notifiers: add mm_take_all_locks() operation
[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 (flags & MAP_NORESERVE)
1114                 vm_flags |= VM_NORESERVE;
1115
1116         if (accountable && (!(flags & MAP_NORESERVE) ||
1117                             sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1118                 if (vm_flags & VM_SHARED) {
1119                         /* Check memory availability in shmem_file_setup? */
1120                         vm_flags |= VM_ACCOUNT;
1121                 } else if (vm_flags & VM_WRITE) {
1122                         /*
1123                          * Private writable mapping: check memory availability
1124                          */
1125                         charged = len >> PAGE_SHIFT;
1126                         if (security_vm_enough_memory(charged))
1127                                 return -ENOMEM;
1128                         vm_flags |= VM_ACCOUNT;
1129                 }
1130         }
1131
1132         /*
1133          * Can we just expand an old private anonymous mapping?
1134          * The VM_SHARED test is necessary because shmem_zero_setup
1135          * will create the file object for a shared anonymous map below.
1136          */
1137         if (!file && !(vm_flags & VM_SHARED) &&
1138             vma_merge(mm, prev, addr, addr + len, vm_flags,
1139                                         NULL, NULL, pgoff, NULL))
1140                 goto out;
1141
1142         /*
1143          * Determine the object being mapped and call the appropriate
1144          * specific mapper. the address has already been validated, but
1145          * not unmapped, but the maps are removed from the list.
1146          */
1147         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1148         if (!vma) {
1149                 error = -ENOMEM;
1150                 goto unacct_error;
1151         }
1152
1153         vma->vm_mm = mm;
1154         vma->vm_start = addr;
1155         vma->vm_end = addr + len;
1156         vma->vm_flags = vm_flags;
1157         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1158         vma->vm_pgoff = pgoff;
1159
1160         if (file) {
1161                 error = -EINVAL;
1162                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1163                         goto free_vma;
1164                 if (vm_flags & VM_DENYWRITE) {
1165                         error = deny_write_access(file);
1166                         if (error)
1167                                 goto free_vma;
1168                         correct_wcount = 1;
1169                 }
1170                 vma->vm_file = file;
1171                 get_file(file);
1172                 error = file->f_op->mmap(file, vma);
1173                 if (error)
1174                         goto unmap_and_free_vma;
1175                 if (vm_flags & VM_EXECUTABLE)
1176                         added_exe_file_vma(mm);
1177         } else if (vm_flags & VM_SHARED) {
1178                 error = shmem_zero_setup(vma);
1179                 if (error)
1180                         goto free_vma;
1181         }
1182
1183         /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1184          * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1185          * that memory reservation must be checked; but that reservation
1186          * belongs to shared memory object, not to vma: so now clear it.
1187          */
1188         if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1189                 vma->vm_flags &= ~VM_ACCOUNT;
1190
1191         /* Can addr have changed??
1192          *
1193          * Answer: Yes, several device drivers can do it in their
1194          *         f_op->mmap method. -DaveM
1195          */
1196         addr = vma->vm_start;
1197         pgoff = vma->vm_pgoff;
1198         vm_flags = vma->vm_flags;
1199
1200         if (vma_wants_writenotify(vma))
1201                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1202
1203         if (file && vma_merge(mm, prev, addr, vma->vm_end,
1204                         vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1205                 mpol_put(vma_policy(vma));
1206                 kmem_cache_free(vm_area_cachep, vma);
1207                 fput(file);
1208                 if (vm_flags & VM_EXECUTABLE)
1209                         removed_exe_file_vma(mm);
1210         } else {
1211                 vma_link(mm, vma, prev, rb_link, rb_parent);
1212                 file = vma->vm_file;
1213         }
1214
1215         /* Once vma denies write, undo our temporary denial count */
1216         if (correct_wcount)
1217                 atomic_inc(&inode->i_writecount);
1218 out:
1219         mm->total_vm += len >> PAGE_SHIFT;
1220         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1221         if (vm_flags & VM_LOCKED) {
1222                 mm->locked_vm += len >> PAGE_SHIFT;
1223                 make_pages_present(addr, addr + len);
1224         }
1225         if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1226                 make_pages_present(addr, addr + len);
1227         return addr;
1228
1229 unmap_and_free_vma:
1230         if (correct_wcount)
1231                 atomic_inc(&inode->i_writecount);
1232         vma->vm_file = NULL;
1233         fput(file);
1234
1235         /* Undo any partial mapping done by a device driver. */
1236         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1237         charged = 0;
1238 free_vma:
1239         kmem_cache_free(vm_area_cachep, vma);
1240 unacct_error:
1241         if (charged)
1242                 vm_unacct_memory(charged);
1243         return error;
1244 }
1245
1246 /* Get an address range which is currently unmapped.
1247  * For shmat() with addr=0.
1248  *
1249  * Ugly calling convention alert:
1250  * Return value with the low bits set means error value,
1251  * ie
1252  *      if (ret & ~PAGE_MASK)
1253  *              error = ret;
1254  *
1255  * This function "knows" that -ENOMEM has the bits set.
1256  */
1257 #ifndef HAVE_ARCH_UNMAPPED_AREA
1258 unsigned long
1259 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1260                 unsigned long len, unsigned long pgoff, unsigned long flags)
1261 {
1262         struct mm_struct *mm = current->mm;
1263         struct vm_area_struct *vma;
1264         unsigned long start_addr;
1265
1266         if (len > TASK_SIZE)
1267                 return -ENOMEM;
1268
1269         if (flags & MAP_FIXED)
1270                 return addr;
1271
1272         if (addr) {
1273                 addr = PAGE_ALIGN(addr);
1274                 vma = find_vma(mm, addr);
1275                 if (TASK_SIZE - len >= addr &&
1276                     (!vma || addr + len <= vma->vm_start))
1277                         return addr;
1278         }
1279         if (len > mm->cached_hole_size) {
1280                 start_addr = addr = mm->free_area_cache;
1281         } else {
1282                 start_addr = addr = TASK_UNMAPPED_BASE;
1283                 mm->cached_hole_size = 0;
1284         }
1285
1286 full_search:
1287         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1288                 /* At this point:  (!vma || addr < vma->vm_end). */
1289                 if (TASK_SIZE - len < addr) {
1290                         /*
1291                          * Start a new search - just in case we missed
1292                          * some holes.
1293                          */
1294                         if (start_addr != TASK_UNMAPPED_BASE) {
1295                                 addr = TASK_UNMAPPED_BASE;
1296                                 start_addr = addr;
1297                                 mm->cached_hole_size = 0;
1298                                 goto full_search;
1299                         }
1300                         return -ENOMEM;
1301                 }
1302                 if (!vma || addr + len <= vma->vm_start) {
1303                         /*
1304                          * Remember the place where we stopped the search:
1305                          */
1306                         mm->free_area_cache = addr + len;
1307                         return addr;
1308                 }
1309                 if (addr + mm->cached_hole_size < vma->vm_start)
1310                         mm->cached_hole_size = vma->vm_start - addr;
1311                 addr = vma->vm_end;
1312         }
1313 }
1314 #endif  
1315
1316 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1317 {
1318         /*
1319          * Is this a new hole at the lowest possible address?
1320          */
1321         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1322                 mm->free_area_cache = addr;
1323                 mm->cached_hole_size = ~0UL;
1324         }
1325 }
1326
1327 /*
1328  * This mmap-allocator allocates new areas top-down from below the
1329  * stack's low limit (the base):
1330  */
1331 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1332 unsigned long
1333 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1334                           const unsigned long len, const unsigned long pgoff,
1335                           const unsigned long flags)
1336 {
1337         struct vm_area_struct *vma;
1338         struct mm_struct *mm = current->mm;
1339         unsigned long addr = addr0;
1340
1341         /* requested length too big for entire address space */
1342         if (len > TASK_SIZE)
1343                 return -ENOMEM;
1344
1345         if (flags & MAP_FIXED)
1346                 return addr;
1347
1348         /* requesting a specific address */
1349         if (addr) {
1350                 addr = PAGE_ALIGN(addr);
1351                 vma = find_vma(mm, addr);
1352                 if (TASK_SIZE - len >= addr &&
1353                                 (!vma || addr + len <= vma->vm_start))
1354                         return addr;
1355         }
1356
1357         /* check if free_area_cache is useful for us */
1358         if (len <= mm->cached_hole_size) {
1359                 mm->cached_hole_size = 0;
1360                 mm->free_area_cache = mm->mmap_base;
1361         }
1362
1363         /* either no address requested or can't fit in requested address hole */
1364         addr = mm->free_area_cache;
1365
1366         /* make sure it can fit in the remaining address space */
1367         if (addr > len) {
1368                 vma = find_vma(mm, addr-len);
1369                 if (!vma || addr <= vma->vm_start)
1370                         /* remember the address as a hint for next time */
1371                         return (mm->free_area_cache = addr-len);
1372         }
1373
1374         if (mm->mmap_base < len)
1375                 goto bottomup;
1376
1377         addr = mm->mmap_base-len;
1378
1379         do {
1380                 /*
1381                  * Lookup failure means no vma is above this address,
1382                  * else if new region fits below vma->vm_start,
1383                  * return with success:
1384                  */
1385                 vma = find_vma(mm, addr);
1386                 if (!vma || addr+len <= vma->vm_start)
1387                         /* remember the address as a hint for next time */
1388                         return (mm->free_area_cache = addr);
1389
1390                 /* remember the largest hole we saw so far */
1391                 if (addr + mm->cached_hole_size < vma->vm_start)
1392                         mm->cached_hole_size = vma->vm_start - addr;
1393
1394                 /* try just below the current vma->vm_start */
1395                 addr = vma->vm_start-len;
1396         } while (len < vma->vm_start);
1397
1398 bottomup:
1399         /*
1400          * A failed mmap() very likely causes application failure,
1401          * so fall back to the bottom-up function here. This scenario
1402          * can happen with large stack limits and large mmap()
1403          * allocations.
1404          */
1405         mm->cached_hole_size = ~0UL;
1406         mm->free_area_cache = TASK_UNMAPPED_BASE;
1407         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1408         /*
1409          * Restore the topdown base:
1410          */
1411         mm->free_area_cache = mm->mmap_base;
1412         mm->cached_hole_size = ~0UL;
1413
1414         return addr;
1415 }
1416 #endif
1417
1418 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1419 {
1420         /*
1421          * Is this a new hole at the highest possible address?
1422          */
1423         if (addr > mm->free_area_cache)
1424                 mm->free_area_cache = addr;
1425
1426         /* dont allow allocations above current base */
1427         if (mm->free_area_cache > mm->mmap_base)
1428                 mm->free_area_cache = mm->mmap_base;
1429 }
1430
1431 unsigned long
1432 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1433                 unsigned long pgoff, unsigned long flags)
1434 {
1435         unsigned long (*get_area)(struct file *, unsigned long,
1436                                   unsigned long, unsigned long, unsigned long);
1437
1438         get_area = current->mm->get_unmapped_area;
1439         if (file && file->f_op && file->f_op->get_unmapped_area)
1440                 get_area = file->f_op->get_unmapped_area;
1441         addr = get_area(file, addr, len, pgoff, flags);
1442         if (IS_ERR_VALUE(addr))
1443                 return addr;
1444
1445         if (addr > TASK_SIZE - len)
1446                 return -ENOMEM;
1447         if (addr & ~PAGE_MASK)
1448                 return -EINVAL;
1449
1450         return arch_rebalance_pgtables(addr, len);
1451 }
1452
1453 EXPORT_SYMBOL(get_unmapped_area);
1454
1455 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1456 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1457 {
1458         struct vm_area_struct *vma = NULL;
1459
1460         if (mm) {
1461                 /* Check the cache first. */
1462                 /* (Cache hit rate is typically around 35%.) */
1463                 vma = mm->mmap_cache;
1464                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1465                         struct rb_node * rb_node;
1466
1467                         rb_node = mm->mm_rb.rb_node;
1468                         vma = NULL;
1469
1470                         while (rb_node) {
1471                                 struct vm_area_struct * vma_tmp;
1472
1473                                 vma_tmp = rb_entry(rb_node,
1474                                                 struct vm_area_struct, vm_rb);
1475
1476                                 if (vma_tmp->vm_end > addr) {
1477                                         vma = vma_tmp;
1478                                         if (vma_tmp->vm_start <= addr)
1479                                                 break;
1480                                         rb_node = rb_node->rb_left;
1481                                 } else
1482                                         rb_node = rb_node->rb_right;
1483                         }
1484                         if (vma)
1485                                 mm->mmap_cache = vma;
1486                 }
1487         }
1488         return vma;
1489 }
1490
1491 EXPORT_SYMBOL(find_vma);
1492
1493 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1494 struct vm_area_struct *
1495 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1496                         struct vm_area_struct **pprev)
1497 {
1498         struct vm_area_struct *vma = NULL, *prev = NULL;
1499         struct rb_node * rb_node;
1500         if (!mm)
1501                 goto out;
1502
1503         /* Guard against addr being lower than the first VMA */
1504         vma = mm->mmap;
1505
1506         /* Go through the RB tree quickly. */
1507         rb_node = mm->mm_rb.rb_node;
1508
1509         while (rb_node) {
1510                 struct vm_area_struct *vma_tmp;
1511                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1512
1513                 if (addr < vma_tmp->vm_end) {
1514                         rb_node = rb_node->rb_left;
1515                 } else {
1516                         prev = vma_tmp;
1517                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1518                                 break;
1519                         rb_node = rb_node->rb_right;
1520                 }
1521         }
1522
1523 out:
1524         *pprev = prev;
1525         return prev ? prev->vm_next : vma;
1526 }
1527
1528 /*
1529  * Verify that the stack growth is acceptable and
1530  * update accounting. This is shared with both the
1531  * grow-up and grow-down cases.
1532  */
1533 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1534 {
1535         struct mm_struct *mm = vma->vm_mm;
1536         struct rlimit *rlim = current->signal->rlim;
1537         unsigned long new_start;
1538
1539         /* address space limit tests */
1540         if (!may_expand_vm(mm, grow))
1541                 return -ENOMEM;
1542
1543         /* Stack limit test */
1544         if (size > rlim[RLIMIT_STACK].rlim_cur)
1545                 return -ENOMEM;
1546
1547         /* mlock limit tests */
1548         if (vma->vm_flags & VM_LOCKED) {
1549                 unsigned long locked;
1550                 unsigned long limit;
1551                 locked = mm->locked_vm + grow;
1552                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1553                 if (locked > limit && !capable(CAP_IPC_LOCK))
1554                         return -ENOMEM;
1555         }
1556
1557         /* Check to ensure the stack will not grow into a hugetlb-only region */
1558         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1559                         vma->vm_end - size;
1560         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1561                 return -EFAULT;
1562
1563         /*
1564          * Overcommit..  This must be the final test, as it will
1565          * update security statistics.
1566          */
1567         if (security_vm_enough_memory(grow))
1568                 return -ENOMEM;
1569
1570         /* Ok, everything looks good - let it rip */
1571         mm->total_vm += grow;
1572         if (vma->vm_flags & VM_LOCKED)
1573                 mm->locked_vm += grow;
1574         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1575         return 0;
1576 }
1577
1578 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1579 /*
1580  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1581  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1582  */
1583 #ifndef CONFIG_IA64
1584 static inline
1585 #endif
1586 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1587 {
1588         int error;
1589
1590         if (!(vma->vm_flags & VM_GROWSUP))
1591                 return -EFAULT;
1592
1593         /*
1594          * We must make sure the anon_vma is allocated
1595          * so that the anon_vma locking is not a noop.
1596          */
1597         if (unlikely(anon_vma_prepare(vma)))
1598                 return -ENOMEM;
1599         anon_vma_lock(vma);
1600
1601         /*
1602          * vma->vm_start/vm_end cannot change under us because the caller
1603          * is required to hold the mmap_sem in read mode.  We need the
1604          * anon_vma lock to serialize against concurrent expand_stacks.
1605          * Also guard against wrapping around to address 0.
1606          */
1607         if (address < PAGE_ALIGN(address+4))
1608                 address = PAGE_ALIGN(address+4);
1609         else {
1610                 anon_vma_unlock(vma);
1611                 return -ENOMEM;
1612         }
1613         error = 0;
1614
1615         /* Somebody else might have raced and expanded it already */
1616         if (address > vma->vm_end) {
1617                 unsigned long size, grow;
1618
1619                 size = address - vma->vm_start;
1620                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1621
1622                 error = acct_stack_growth(vma, size, grow);
1623                 if (!error)
1624                         vma->vm_end = address;
1625         }
1626         anon_vma_unlock(vma);
1627         return error;
1628 }
1629 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1630
1631 /*
1632  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1633  */
1634 static inline int expand_downwards(struct vm_area_struct *vma,
1635                                    unsigned long address)
1636 {
1637         int error;
1638
1639         /*
1640          * We must make sure the anon_vma is allocated
1641          * so that the anon_vma locking is not a noop.
1642          */
1643         if (unlikely(anon_vma_prepare(vma)))
1644                 return -ENOMEM;
1645
1646         address &= PAGE_MASK;
1647         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1648         if (error)
1649                 return error;
1650
1651         anon_vma_lock(vma);
1652
1653         /*
1654          * vma->vm_start/vm_end cannot change under us because the caller
1655          * is required to hold the mmap_sem in read mode.  We need the
1656          * anon_vma lock to serialize against concurrent expand_stacks.
1657          */
1658
1659         /* Somebody else might have raced and expanded it already */
1660         if (address < vma->vm_start) {
1661                 unsigned long size, grow;
1662
1663                 size = vma->vm_end - address;
1664                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1665
1666                 error = acct_stack_growth(vma, size, grow);
1667                 if (!error) {
1668                         vma->vm_start = address;
1669                         vma->vm_pgoff -= grow;
1670                 }
1671         }
1672         anon_vma_unlock(vma);
1673         return error;
1674 }
1675
1676 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1677 {
1678         return expand_downwards(vma, address);
1679 }
1680
1681 #ifdef CONFIG_STACK_GROWSUP
1682 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1683 {
1684         return expand_upwards(vma, address);
1685 }
1686
1687 struct vm_area_struct *
1688 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1689 {
1690         struct vm_area_struct *vma, *prev;
1691
1692         addr &= PAGE_MASK;
1693         vma = find_vma_prev(mm, addr, &prev);
1694         if (vma && (vma->vm_start <= addr))
1695                 return vma;
1696         if (!prev || expand_stack(prev, addr))
1697                 return NULL;
1698         if (prev->vm_flags & VM_LOCKED)
1699                 make_pages_present(addr, prev->vm_end);
1700         return prev;
1701 }
1702 #else
1703 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1704 {
1705         return expand_downwards(vma, address);
1706 }
1707
1708 struct vm_area_struct *
1709 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1710 {
1711         struct vm_area_struct * vma;
1712         unsigned long start;
1713
1714         addr &= PAGE_MASK;
1715         vma = find_vma(mm,addr);
1716         if (!vma)
1717                 return NULL;
1718         if (vma->vm_start <= addr)
1719                 return vma;
1720         if (!(vma->vm_flags & VM_GROWSDOWN))
1721                 return NULL;
1722         start = vma->vm_start;
1723         if (expand_stack(vma, addr))
1724                 return NULL;
1725         if (vma->vm_flags & VM_LOCKED)
1726                 make_pages_present(addr, start);
1727         return vma;
1728 }
1729 #endif
1730
1731 /*
1732  * Ok - we have the memory areas we should free on the vma list,
1733  * so release them, and do the vma updates.
1734  *
1735  * Called with the mm semaphore held.
1736  */
1737 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1738 {
1739         /* Update high watermark before we lower total_vm */
1740         update_hiwater_vm(mm);
1741         do {
1742                 long nrpages = vma_pages(vma);
1743
1744                 mm->total_vm -= nrpages;
1745                 if (vma->vm_flags & VM_LOCKED)
1746                         mm->locked_vm -= nrpages;
1747                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1748                 vma = remove_vma(vma);
1749         } while (vma);
1750         validate_mm(mm);
1751 }
1752
1753 /*
1754  * Get rid of page table information in the indicated region.
1755  *
1756  * Called with the mm semaphore held.
1757  */
1758 static void unmap_region(struct mm_struct *mm,
1759                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1760                 unsigned long start, unsigned long end)
1761 {
1762         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1763         struct mmu_gather *tlb;
1764         unsigned long nr_accounted = 0;
1765
1766         lru_add_drain();
1767         tlb = tlb_gather_mmu(mm, 0);
1768         update_hiwater_rss(mm);
1769         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1770         vm_unacct_memory(nr_accounted);
1771         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1772                                  next? next->vm_start: 0);
1773         tlb_finish_mmu(tlb, start, end);
1774 }
1775
1776 /*
1777  * Create a list of vma's touched by the unmap, removing them from the mm's
1778  * vma list as we go..
1779  */
1780 static void
1781 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1782         struct vm_area_struct *prev, unsigned long end)
1783 {
1784         struct vm_area_struct **insertion_point;
1785         struct vm_area_struct *tail_vma = NULL;
1786         unsigned long addr;
1787
1788         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1789         do {
1790                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1791                 mm->map_count--;
1792                 tail_vma = vma;
1793                 vma = vma->vm_next;
1794         } while (vma && vma->vm_start < end);
1795         *insertion_point = vma;
1796         tail_vma->vm_next = NULL;
1797         if (mm->unmap_area == arch_unmap_area)
1798                 addr = prev ? prev->vm_end : mm->mmap_base;
1799         else
1800                 addr = vma ?  vma->vm_start : mm->mmap_base;
1801         mm->unmap_area(mm, addr);
1802         mm->mmap_cache = NULL;          /* Kill the cache. */
1803 }
1804
1805 /*
1806  * Split a vma into two pieces at address 'addr', a new vma is allocated
1807  * either for the first part or the tail.
1808  */
1809 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1810               unsigned long addr, int new_below)
1811 {
1812         struct mempolicy *pol;
1813         struct vm_area_struct *new;
1814
1815         if (is_vm_hugetlb_page(vma) && (addr &
1816                                         ~(huge_page_mask(hstate_vma(vma)))))
1817                 return -EINVAL;
1818
1819         if (mm->map_count >= sysctl_max_map_count)
1820                 return -ENOMEM;
1821
1822         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1823         if (!new)
1824                 return -ENOMEM;
1825
1826         /* most fields are the same, copy all, and then fixup */
1827         *new = *vma;
1828
1829         if (new_below)
1830                 new->vm_end = addr;
1831         else {
1832                 new->vm_start = addr;
1833                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1834         }
1835
1836         pol = mpol_dup(vma_policy(vma));
1837         if (IS_ERR(pol)) {
1838                 kmem_cache_free(vm_area_cachep, new);
1839                 return PTR_ERR(pol);
1840         }
1841         vma_set_policy(new, pol);
1842
1843         if (new->vm_file) {
1844                 get_file(new->vm_file);
1845                 if (vma->vm_flags & VM_EXECUTABLE)
1846                         added_exe_file_vma(mm);
1847         }
1848
1849         if (new->vm_ops && new->vm_ops->open)
1850                 new->vm_ops->open(new);
1851
1852         if (new_below)
1853                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1854                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1855         else
1856                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1857
1858         return 0;
1859 }
1860
1861 /* Munmap is split into 2 main parts -- this part which finds
1862  * what needs doing, and the areas themselves, which do the
1863  * work.  This now handles partial unmappings.
1864  * Jeremy Fitzhardinge <jeremy@goop.org>
1865  */
1866 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1867 {
1868         unsigned long end;
1869         struct vm_area_struct *vma, *prev, *last;
1870
1871         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1872                 return -EINVAL;
1873
1874         if ((len = PAGE_ALIGN(len)) == 0)
1875                 return -EINVAL;
1876
1877         /* Find the first overlapping VMA */
1878         vma = find_vma_prev(mm, start, &prev);
1879         if (!vma)
1880                 return 0;
1881         /* we have  start < vma->vm_end  */
1882
1883         /* if it doesn't overlap, we have nothing.. */
1884         end = start + len;
1885         if (vma->vm_start >= end)
1886                 return 0;
1887
1888         /*
1889          * If we need to split any vma, do it now to save pain later.
1890          *
1891          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1892          * unmapped vm_area_struct will remain in use: so lower split_vma
1893          * places tmp vma above, and higher split_vma places tmp vma below.
1894          */
1895         if (start > vma->vm_start) {
1896                 int error = split_vma(mm, vma, start, 0);
1897                 if (error)
1898                         return error;
1899                 prev = vma;
1900         }
1901
1902         /* Does it split the last one? */
1903         last = find_vma(mm, end);
1904         if (last && end > last->vm_start) {
1905                 int error = split_vma(mm, last, end, 1);
1906                 if (error)
1907                         return error;
1908         }
1909         vma = prev? prev->vm_next: mm->mmap;
1910
1911         /*
1912          * Remove the vma's, and unmap the actual pages
1913          */
1914         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1915         unmap_region(mm, vma, prev, start, end);
1916
1917         /* Fix up all other VM information */
1918         remove_vma_list(mm, vma);
1919
1920         return 0;
1921 }
1922
1923 EXPORT_SYMBOL(do_munmap);
1924
1925 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1926 {
1927         int ret;
1928         struct mm_struct *mm = current->mm;
1929
1930         profile_munmap(addr);
1931
1932         down_write(&mm->mmap_sem);
1933         ret = do_munmap(mm, addr, len);
1934         up_write(&mm->mmap_sem);
1935         return ret;
1936 }
1937
1938 static inline void verify_mm_writelocked(struct mm_struct *mm)
1939 {
1940 #ifdef CONFIG_DEBUG_VM
1941         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1942                 WARN_ON(1);
1943                 up_read(&mm->mmap_sem);
1944         }
1945 #endif
1946 }
1947
1948 /*
1949  *  this is really a simplified "do_mmap".  it only handles
1950  *  anonymous maps.  eventually we may be able to do some
1951  *  brk-specific accounting here.
1952  */
1953 unsigned long do_brk(unsigned long addr, unsigned long len)
1954 {
1955         struct mm_struct * mm = current->mm;
1956         struct vm_area_struct * vma, * prev;
1957         unsigned long flags;
1958         struct rb_node ** rb_link, * rb_parent;
1959         pgoff_t pgoff = addr >> PAGE_SHIFT;
1960         int error;
1961
1962         len = PAGE_ALIGN(len);
1963         if (!len)
1964                 return addr;
1965
1966         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1967                 return -EINVAL;
1968
1969         if (is_hugepage_only_range(mm, addr, len))
1970                 return -EINVAL;
1971
1972         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
1973         if (error)
1974                 return error;
1975
1976         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1977
1978         error = arch_mmap_check(addr, len, flags);
1979         if (error)
1980                 return error;
1981
1982         /*
1983          * mlock MCL_FUTURE?
1984          */
1985         if (mm->def_flags & VM_LOCKED) {
1986                 unsigned long locked, lock_limit;
1987                 locked = len >> PAGE_SHIFT;
1988                 locked += mm->locked_vm;
1989                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1990                 lock_limit >>= PAGE_SHIFT;
1991                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1992                         return -EAGAIN;
1993         }
1994
1995         /*
1996          * mm->mmap_sem is required to protect against another thread
1997          * changing the mappings in case we sleep.
1998          */
1999         verify_mm_writelocked(mm);
2000
2001         /*
2002          * Clear old maps.  this also does some error checking for us
2003          */
2004  munmap_back:
2005         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2006         if (vma && vma->vm_start < addr + len) {
2007                 if (do_munmap(mm, addr, len))
2008                         return -ENOMEM;
2009                 goto munmap_back;
2010         }
2011
2012         /* Check against address space limits *after* clearing old maps... */
2013         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2014                 return -ENOMEM;
2015
2016         if (mm->map_count > sysctl_max_map_count)
2017                 return -ENOMEM;
2018
2019         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2020                 return -ENOMEM;
2021
2022         /* Can we just expand an old private anonymous mapping? */
2023         if (vma_merge(mm, prev, addr, addr + len, flags,
2024                                         NULL, NULL, pgoff, NULL))
2025                 goto out;
2026
2027         /*
2028          * create a vma struct for an anonymous mapping
2029          */
2030         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2031         if (!vma) {
2032                 vm_unacct_memory(len >> PAGE_SHIFT);
2033                 return -ENOMEM;
2034         }
2035
2036         vma->vm_mm = mm;
2037         vma->vm_start = addr;
2038         vma->vm_end = addr + len;
2039         vma->vm_pgoff = pgoff;
2040         vma->vm_flags = flags;
2041         vma->vm_page_prot = vm_get_page_prot(flags);
2042         vma_link(mm, vma, prev, rb_link, rb_parent);
2043 out:
2044         mm->total_vm += len >> PAGE_SHIFT;
2045         if (flags & VM_LOCKED) {
2046                 mm->locked_vm += len >> PAGE_SHIFT;
2047                 make_pages_present(addr, addr + len);
2048         }
2049         return addr;
2050 }
2051
2052 EXPORT_SYMBOL(do_brk);
2053
2054 /* Release all mmaps. */
2055 void exit_mmap(struct mm_struct *mm)
2056 {
2057         struct mmu_gather *tlb;
2058         struct vm_area_struct *vma = mm->mmap;
2059         unsigned long nr_accounted = 0;
2060         unsigned long end;
2061
2062         /* mm's last user has gone, and its about to be pulled down */
2063         arch_exit_mmap(mm);
2064
2065         lru_add_drain();
2066         flush_cache_mm(mm);
2067         tlb = tlb_gather_mmu(mm, 1);
2068         /* Don't update_hiwater_rss(mm) here, do_exit already did */
2069         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2070         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2071         vm_unacct_memory(nr_accounted);
2072         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2073         tlb_finish_mmu(tlb, 0, end);
2074
2075         /*
2076          * Walk the list again, actually closing and freeing it,
2077          * with preemption enabled, without holding any MM locks.
2078          */
2079         while (vma)
2080                 vma = remove_vma(vma);
2081
2082         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2083 }
2084
2085 /* Insert vm structure into process list sorted by address
2086  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2087  * then i_mmap_lock is taken here.
2088  */
2089 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2090 {
2091         struct vm_area_struct * __vma, * prev;
2092         struct rb_node ** rb_link, * rb_parent;
2093
2094         /*
2095          * The vm_pgoff of a purely anonymous vma should be irrelevant
2096          * until its first write fault, when page's anon_vma and index
2097          * are set.  But now set the vm_pgoff it will almost certainly
2098          * end up with (unless mremap moves it elsewhere before that
2099          * first wfault), so /proc/pid/maps tells a consistent story.
2100          *
2101          * By setting it to reflect the virtual start address of the
2102          * vma, merges and splits can happen in a seamless way, just
2103          * using the existing file pgoff checks and manipulations.
2104          * Similarly in do_mmap_pgoff and in do_brk.
2105          */
2106         if (!vma->vm_file) {
2107                 BUG_ON(vma->anon_vma);
2108                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2109         }
2110         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2111         if (__vma && __vma->vm_start < vma->vm_end)
2112                 return -ENOMEM;
2113         if ((vma->vm_flags & VM_ACCOUNT) &&
2114              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2115                 return -ENOMEM;
2116         vma_link(mm, vma, prev, rb_link, rb_parent);
2117         return 0;
2118 }
2119
2120 /*
2121  * Copy the vma structure to a new location in the same mm,
2122  * prior to moving page table entries, to effect an mremap move.
2123  */
2124 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2125         unsigned long addr, unsigned long len, pgoff_t pgoff)
2126 {
2127         struct vm_area_struct *vma = *vmap;
2128         unsigned long vma_start = vma->vm_start;
2129         struct mm_struct *mm = vma->vm_mm;
2130         struct vm_area_struct *new_vma, *prev;
2131         struct rb_node **rb_link, *rb_parent;
2132         struct mempolicy *pol;
2133
2134         /*
2135          * If anonymous vma has not yet been faulted, update new pgoff
2136          * to match new location, to increase its chance of merging.
2137          */
2138         if (!vma->vm_file && !vma->anon_vma)
2139                 pgoff = addr >> PAGE_SHIFT;
2140
2141         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2142         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2143                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2144         if (new_vma) {
2145                 /*
2146                  * Source vma may have been merged into new_vma
2147                  */
2148                 if (vma_start >= new_vma->vm_start &&
2149                     vma_start < new_vma->vm_end)
2150                         *vmap = new_vma;
2151         } else {
2152                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2153                 if (new_vma) {
2154                         *new_vma = *vma;
2155                         pol = mpol_dup(vma_policy(vma));
2156                         if (IS_ERR(pol)) {
2157                                 kmem_cache_free(vm_area_cachep, new_vma);
2158                                 return NULL;
2159                         }
2160                         vma_set_policy(new_vma, pol);
2161                         new_vma->vm_start = addr;
2162                         new_vma->vm_end = addr + len;
2163                         new_vma->vm_pgoff = pgoff;
2164                         if (new_vma->vm_file) {
2165                                 get_file(new_vma->vm_file);
2166                                 if (vma->vm_flags & VM_EXECUTABLE)
2167                                         added_exe_file_vma(mm);
2168                         }
2169                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2170                                 new_vma->vm_ops->open(new_vma);
2171                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2172                 }
2173         }
2174         return new_vma;
2175 }
2176
2177 /*
2178  * Return true if the calling process may expand its vm space by the passed
2179  * number of pages
2180  */
2181 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2182 {
2183         unsigned long cur = mm->total_vm;       /* pages */
2184         unsigned long lim;
2185
2186         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2187
2188         if (cur + npages > lim)
2189                 return 0;
2190         return 1;
2191 }
2192
2193
2194 static int special_mapping_fault(struct vm_area_struct *vma,
2195                                 struct vm_fault *vmf)
2196 {
2197         pgoff_t pgoff;
2198         struct page **pages;
2199
2200         /*
2201          * special mappings have no vm_file, and in that case, the mm
2202          * uses vm_pgoff internally. So we have to subtract it from here.
2203          * We are allowed to do this because we are the mm; do not copy
2204          * this code into drivers!
2205          */
2206         pgoff = vmf->pgoff - vma->vm_pgoff;
2207
2208         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2209                 pgoff--;
2210
2211         if (*pages) {
2212                 struct page *page = *pages;
2213                 get_page(page);
2214                 vmf->page = page;
2215                 return 0;
2216         }
2217
2218         return VM_FAULT_SIGBUS;
2219 }
2220
2221 /*
2222  * Having a close hook prevents vma merging regardless of flags.
2223  */
2224 static void special_mapping_close(struct vm_area_struct *vma)
2225 {
2226 }
2227
2228 static struct vm_operations_struct special_mapping_vmops = {
2229         .close = special_mapping_close,
2230         .fault = special_mapping_fault,
2231 };
2232
2233 /*
2234  * Called with mm->mmap_sem held for writing.
2235  * Insert a new vma covering the given region, with the given flags.
2236  * Its pages are supplied by the given array of struct page *.
2237  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2238  * The region past the last page supplied will always produce SIGBUS.
2239  * The array pointer and the pages it points to are assumed to stay alive
2240  * for as long as this mapping might exist.
2241  */
2242 int install_special_mapping(struct mm_struct *mm,
2243                             unsigned long addr, unsigned long len,
2244                             unsigned long vm_flags, struct page **pages)
2245 {
2246         struct vm_area_struct *vma;
2247
2248         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2249         if (unlikely(vma == NULL))
2250                 return -ENOMEM;
2251
2252         vma->vm_mm = mm;
2253         vma->vm_start = addr;
2254         vma->vm_end = addr + len;
2255
2256         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2257         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2258
2259         vma->vm_ops = &special_mapping_vmops;
2260         vma->vm_private_data = pages;
2261
2262         if (unlikely(insert_vm_struct(mm, vma))) {
2263                 kmem_cache_free(vm_area_cachep, vma);
2264                 return -ENOMEM;
2265         }
2266
2267         mm->total_vm += len >> PAGE_SHIFT;
2268
2269         return 0;
2270 }
2271
2272 static DEFINE_MUTEX(mm_all_locks_mutex);
2273
2274 static void vm_lock_anon_vma(struct anon_vma *anon_vma)
2275 {
2276         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2277                 /*
2278                  * The LSB of head.next can't change from under us
2279                  * because we hold the mm_all_locks_mutex.
2280                  */
2281                 spin_lock(&anon_vma->lock);
2282                 /*
2283                  * We can safely modify head.next after taking the
2284                  * anon_vma->lock. If some other vma in this mm shares
2285                  * the same anon_vma we won't take it again.
2286                  *
2287                  * No need of atomic instructions here, head.next
2288                  * can't change from under us thanks to the
2289                  * anon_vma->lock.
2290                  */
2291                 if (__test_and_set_bit(0, (unsigned long *)
2292                                        &anon_vma->head.next))
2293                         BUG();
2294         }
2295 }
2296
2297 static void vm_lock_mapping(struct address_space *mapping)
2298 {
2299         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2300                 /*
2301                  * AS_MM_ALL_LOCKS can't change from under us because
2302                  * we hold the mm_all_locks_mutex.
2303                  *
2304                  * Operations on ->flags have to be atomic because
2305                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2306                  * mm_all_locks_mutex, there may be other cpus
2307                  * changing other bitflags in parallel to us.
2308                  */
2309                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2310                         BUG();
2311                 spin_lock(&mapping->i_mmap_lock);
2312         }
2313 }
2314
2315 /*
2316  * This operation locks against the VM for all pte/vma/mm related
2317  * operations that could ever happen on a certain mm. This includes
2318  * vmtruncate, try_to_unmap, and all page faults.
2319  *
2320  * The caller must take the mmap_sem in write mode before calling
2321  * mm_take_all_locks(). The caller isn't allowed to release the
2322  * mmap_sem until mm_drop_all_locks() returns.
2323  *
2324  * mmap_sem in write mode is required in order to block all operations
2325  * that could modify pagetables and free pages without need of
2326  * altering the vma layout (for example populate_range() with
2327  * nonlinear vmas). It's also needed in write mode to avoid new
2328  * anon_vmas to be associated with existing vmas.
2329  *
2330  * A single task can't take more than one mm_take_all_locks() in a row
2331  * or it would deadlock.
2332  *
2333  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2334  * mapping->flags avoid to take the same lock twice, if more than one
2335  * vma in this mm is backed by the same anon_vma or address_space.
2336  *
2337  * We can take all the locks in random order because the VM code
2338  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2339  * takes more than one of them in a row. Secondly we're protected
2340  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2341  *
2342  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2343  * that may have to take thousand of locks.
2344  *
2345  * mm_take_all_locks() can fail if it's interrupted by signals.
2346  */
2347 int mm_take_all_locks(struct mm_struct *mm)
2348 {
2349         struct vm_area_struct *vma;
2350         int ret = -EINTR;
2351
2352         BUG_ON(down_read_trylock(&mm->mmap_sem));
2353
2354         mutex_lock(&mm_all_locks_mutex);
2355
2356         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2357                 if (signal_pending(current))
2358                         goto out_unlock;
2359                 if (vma->anon_vma)
2360                         vm_lock_anon_vma(vma->anon_vma);
2361                 if (vma->vm_file && vma->vm_file->f_mapping)
2362                         vm_lock_mapping(vma->vm_file->f_mapping);
2363         }
2364         ret = 0;
2365
2366 out_unlock:
2367         if (ret)
2368                 mm_drop_all_locks(mm);
2369
2370         return ret;
2371 }
2372
2373 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2374 {
2375         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2376                 /*
2377                  * The LSB of head.next can't change to 0 from under
2378                  * us because we hold the mm_all_locks_mutex.
2379                  *
2380                  * We must however clear the bitflag before unlocking
2381                  * the vma so the users using the anon_vma->head will
2382                  * never see our bitflag.
2383                  *
2384                  * No need of atomic instructions here, head.next
2385                  * can't change from under us until we release the
2386                  * anon_vma->lock.
2387                  */
2388                 if (!__test_and_clear_bit(0, (unsigned long *)
2389                                           &anon_vma->head.next))
2390                         BUG();
2391                 spin_unlock(&anon_vma->lock);
2392         }
2393 }
2394
2395 static void vm_unlock_mapping(struct address_space *mapping)
2396 {
2397         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2398                 /*
2399                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2400                  * because we hold the mm_all_locks_mutex.
2401                  */
2402                 spin_unlock(&mapping->i_mmap_lock);
2403                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2404                                         &mapping->flags))
2405                         BUG();
2406         }
2407 }
2408
2409 /*
2410  * The mmap_sem cannot be released by the caller until
2411  * mm_drop_all_locks() returns.
2412  */
2413 void mm_drop_all_locks(struct mm_struct *mm)
2414 {
2415         struct vm_area_struct *vma;
2416
2417         BUG_ON(down_read_trylock(&mm->mmap_sem));
2418         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2419
2420         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2421                 if (vma->anon_vma)
2422                         vm_unlock_anon_vma(vma->anon_vma);
2423                 if (vma->vm_file && vma->vm_file->f_mapping)
2424                         vm_unlock_mapping(vma->vm_file->f_mapping);
2425         }
2426
2427         mutex_unlock(&mm_all_locks_mutex);
2428 }