writeback: remove unused nonblocking and congestion checks
[linux-2.6.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/ima.h>
24 #include <linux/hugetlb.h>
25 #include <linux/profile.h>
26 #include <linux/module.h>
27 #include <linux/mount.h>
28 #include <linux/mempolicy.h>
29 #include <linux/rmap.h>
30 #include <linux/mmu_notifier.h>
31 #include <linux/perf_event.h>
32
33 #include <asm/uaccess.h>
34 #include <asm/cacheflush.h>
35 #include <asm/tlb.h>
36 #include <asm/mmu_context.h>
37
38 #include "internal.h"
39
40 #ifndef arch_mmap_check
41 #define arch_mmap_check(addr, len, flags)       (0)
42 #endif
43
44 #ifndef arch_rebalance_pgtables
45 #define arch_rebalance_pgtables(addr, len)              (addr)
46 #endif
47
48 static void unmap_region(struct mm_struct *mm,
49                 struct vm_area_struct *vma, struct vm_area_struct *prev,
50                 unsigned long start, unsigned long end);
51
52 /*
53  * WARNING: the debugging will use recursive algorithms so never enable this
54  * unless you know what you are doing.
55  */
56 #undef DEBUG_MM_RB
57
58 /* description of effects of mapping type and prot in current implementation.
59  * this is due to the limited x86 page protection hardware.  The expected
60  * behavior is in parens:
61  *
62  * map_type     prot
63  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
64  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
65  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
66  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
67  *              
68  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
69  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
70  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
71  *
72  */
73 pgprot_t protection_map[16] = {
74         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
75         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
76 };
77
78 pgprot_t vm_get_page_prot(unsigned long vm_flags)
79 {
80         return __pgprot(pgprot_val(protection_map[vm_flags &
81                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
82                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
83 }
84 EXPORT_SYMBOL(vm_get_page_prot);
85
86 int sysctl_overcommit_memory = OVERCOMMIT_GUESS;  /* heuristic overcommit */
87 int sysctl_overcommit_ratio = 50;       /* default is 50% */
88 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
89 struct percpu_counter vm_committed_as;
90
91 /*
92  * Check that a process has enough memory to allocate a new virtual
93  * mapping. 0 means there is enough memory for the allocation to
94  * succeed and -ENOMEM implies there is not.
95  *
96  * We currently support three overcommit policies, which are set via the
97  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
98  *
99  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
100  * Additional code 2002 Jul 20 by Robert Love.
101  *
102  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
103  *
104  * Note this is a helper function intended to be used by LSMs which
105  * wish to use this logic.
106  */
107 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
108 {
109         unsigned long free, allowed;
110
111         vm_acct_memory(pages);
112
113         /*
114          * Sometimes we want to use more memory than we have
115          */
116         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
117                 return 0;
118
119         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
120                 unsigned long n;
121
122                 free = global_page_state(NR_FILE_PAGES);
123                 free += nr_swap_pages;
124
125                 /*
126                  * Any slabs which are created with the
127                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
128                  * which are reclaimable, under pressure.  The dentry
129                  * cache and most inode caches should fall into this
130                  */
131                 free += global_page_state(NR_SLAB_RECLAIMABLE);
132
133                 /*
134                  * Leave the last 3% for root
135                  */
136                 if (!cap_sys_admin)
137                         free -= free / 32;
138
139                 if (free > pages)
140                         return 0;
141
142                 /*
143                  * nr_free_pages() is very expensive on large systems,
144                  * only call if we're about to fail.
145                  */
146                 n = nr_free_pages();
147
148                 /*
149                  * Leave reserved pages. The pages are not for anonymous pages.
150                  */
151                 if (n <= totalreserve_pages)
152                         goto error;
153                 else
154                         n -= totalreserve_pages;
155
156                 /*
157                  * Leave the last 3% for root
158                  */
159                 if (!cap_sys_admin)
160                         n -= n / 32;
161                 free += n;
162
163                 if (free > pages)
164                         return 0;
165
166                 goto error;
167         }
168
169         allowed = (totalram_pages - hugetlb_total_pages())
170                 * sysctl_overcommit_ratio / 100;
171         /*
172          * Leave the last 3% for root
173          */
174         if (!cap_sys_admin)
175                 allowed -= allowed / 32;
176         allowed += total_swap_pages;
177
178         /* Don't let a single process grow too big:
179            leave 3% of the size of this process for other processes */
180         if (mm)
181                 allowed -= mm->total_vm / 32;
182
183         if (percpu_counter_read_positive(&vm_committed_as) < 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 SYSCALL_DEFINE1(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                                 break;
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 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 __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
476 {
477         struct vm_area_struct *__vma, *prev;
478         struct rb_node **rb_link, *rb_parent;
479
480         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
481         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
482         __vma_link(mm, vma, prev, rb_link, rb_parent);
483         mm->map_count++;
484 }
485
486 static inline void
487 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
488                 struct vm_area_struct *prev)
489 {
490         prev->vm_next = vma->vm_next;
491         rb_erase(&vma->vm_rb, &mm->mm_rb);
492         if (mm->mmap_cache == vma)
493                 mm->mmap_cache = prev;
494 }
495
496 /*
497  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
498  * is already present in an i_mmap tree without adjusting the tree.
499  * The following helper function should be used when such adjustments
500  * are necessary.  The "insert" vma (if any) is to be inserted
501  * before we drop the necessary locks.
502  */
503 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
504         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
505 {
506         struct mm_struct *mm = vma->vm_mm;
507         struct vm_area_struct *next = vma->vm_next;
508         struct vm_area_struct *importer = NULL;
509         struct address_space *mapping = NULL;
510         struct prio_tree_root *root = NULL;
511         struct file *file = vma->vm_file;
512         struct anon_vma *anon_vma = NULL;
513         long adjust_next = 0;
514         int remove_next = 0;
515
516         if (next && !insert) {
517                 if (end >= next->vm_end) {
518                         /*
519                          * vma expands, overlapping all the next, and
520                          * perhaps the one after too (mprotect case 6).
521                          */
522 again:                  remove_next = 1 + (end > next->vm_end);
523                         end = next->vm_end;
524                         anon_vma = next->anon_vma;
525                         importer = vma;
526                 } else if (end > next->vm_start) {
527                         /*
528                          * vma expands, overlapping part of the next:
529                          * mprotect case 5 shifting the boundary up.
530                          */
531                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
532                         anon_vma = next->anon_vma;
533                         importer = vma;
534                 } else if (end < vma->vm_end) {
535                         /*
536                          * vma shrinks, and !insert tells it's not
537                          * split_vma inserting another: so it must be
538                          * mprotect case 4 shifting the boundary down.
539                          */
540                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
541                         anon_vma = next->anon_vma;
542                         importer = next;
543                 }
544         }
545
546         if (file) {
547                 mapping = file->f_mapping;
548                 if (!(vma->vm_flags & VM_NONLINEAR))
549                         root = &mapping->i_mmap;
550                 spin_lock(&mapping->i_mmap_lock);
551                 if (importer &&
552                     vma->vm_truncate_count != next->vm_truncate_count) {
553                         /*
554                          * unmap_mapping_range might be in progress:
555                          * ensure that the expanding vma is rescanned.
556                          */
557                         importer->vm_truncate_count = 0;
558                 }
559                 if (insert) {
560                         insert->vm_truncate_count = vma->vm_truncate_count;
561                         /*
562                          * Put into prio_tree now, so instantiated pages
563                          * are visible to arm/parisc __flush_dcache_page
564                          * throughout; but we cannot insert into address
565                          * space until vma start or end is updated.
566                          */
567                         __vma_link_file(insert);
568                 }
569         }
570
571         /*
572          * When changing only vma->vm_end, we don't really need
573          * anon_vma lock.
574          */
575         if (vma->anon_vma && (insert || importer || start != vma->vm_start))
576                 anon_vma = vma->anon_vma;
577         if (anon_vma) {
578                 spin_lock(&anon_vma->lock);
579                 /*
580                  * Easily overlooked: when mprotect shifts the boundary,
581                  * make sure the expanding vma has anon_vma set if the
582                  * shrinking vma had, to cover any anon pages imported.
583                  */
584                 if (importer && !importer->anon_vma) {
585                         importer->anon_vma = anon_vma;
586                         __anon_vma_link(importer);
587                 }
588         }
589
590         if (root) {
591                 flush_dcache_mmap_lock(mapping);
592                 vma_prio_tree_remove(vma, root);
593                 if (adjust_next)
594                         vma_prio_tree_remove(next, root);
595         }
596
597         vma->vm_start = start;
598         vma->vm_end = end;
599         vma->vm_pgoff = pgoff;
600         if (adjust_next) {
601                 next->vm_start += adjust_next << PAGE_SHIFT;
602                 next->vm_pgoff += adjust_next;
603         }
604
605         if (root) {
606                 if (adjust_next)
607                         vma_prio_tree_insert(next, root);
608                 vma_prio_tree_insert(vma, root);
609                 flush_dcache_mmap_unlock(mapping);
610         }
611
612         if (remove_next) {
613                 /*
614                  * vma_merge has merged next into vma, and needs
615                  * us to remove next before dropping the locks.
616                  */
617                 __vma_unlink(mm, next, vma);
618                 if (file)
619                         __remove_shared_vm_struct(next, file, mapping);
620                 if (next->anon_vma)
621                         __anon_vma_merge(vma, next);
622         } else if (insert) {
623                 /*
624                  * split_vma has split insert from vma, and needs
625                  * us to insert it before dropping the locks
626                  * (it may either follow vma or precede it).
627                  */
628                 __insert_vm_struct(mm, insert);
629         }
630
631         if (anon_vma)
632                 spin_unlock(&anon_vma->lock);
633         if (mapping)
634                 spin_unlock(&mapping->i_mmap_lock);
635
636         if (remove_next) {
637                 if (file) {
638                         fput(file);
639                         if (next->vm_flags & VM_EXECUTABLE)
640                                 removed_exe_file_vma(mm);
641                 }
642                 mm->map_count--;
643                 mpol_put(vma_policy(next));
644                 kmem_cache_free(vm_area_cachep, next);
645                 /*
646                  * In mprotect's case 6 (see comments on vma_merge),
647                  * we must remove another next too. It would clutter
648                  * up the code too much to do both in one go.
649                  */
650                 if (remove_next == 2) {
651                         next = vma->vm_next;
652                         goto again;
653                 }
654         }
655
656         validate_mm(mm);
657 }
658
659 /*
660  * If the vma has a ->close operation then the driver probably needs to release
661  * per-vma resources, so we don't attempt to merge those.
662  */
663 static inline int is_mergeable_vma(struct vm_area_struct *vma,
664                         struct file *file, unsigned long vm_flags)
665 {
666         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
667         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
668                 return 0;
669         if (vma->vm_file != file)
670                 return 0;
671         if (vma->vm_ops && vma->vm_ops->close)
672                 return 0;
673         return 1;
674 }
675
676 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
677                                         struct anon_vma *anon_vma2)
678 {
679         return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
680 }
681
682 /*
683  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
684  * in front of (at a lower virtual address and file offset than) the vma.
685  *
686  * We cannot merge two vmas if they have differently assigned (non-NULL)
687  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
688  *
689  * We don't check here for the merged mmap wrapping around the end of pagecache
690  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
691  * wrap, nor mmaps which cover the final page at index -1UL.
692  */
693 static int
694 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
695         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
696 {
697         if (is_mergeable_vma(vma, file, vm_flags) &&
698             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
699                 if (vma->vm_pgoff == vm_pgoff)
700                         return 1;
701         }
702         return 0;
703 }
704
705 /*
706  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
707  * beyond (at a higher virtual address and file offset than) the vma.
708  *
709  * We cannot merge two vmas if they have differently assigned (non-NULL)
710  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
711  */
712 static int
713 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
714         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
715 {
716         if (is_mergeable_vma(vma, file, vm_flags) &&
717             is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
718                 pgoff_t vm_pglen;
719                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
720                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
721                         return 1;
722         }
723         return 0;
724 }
725
726 /*
727  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
728  * whether that can be merged with its predecessor or its successor.
729  * Or both (it neatly fills a hole).
730  *
731  * In most cases - when called for mmap, brk or mremap - [addr,end) is
732  * certain not to be mapped by the time vma_merge is called; but when
733  * called for mprotect, it is certain to be already mapped (either at
734  * an offset within prev, or at the start of next), and the flags of
735  * this area are about to be changed to vm_flags - and the no-change
736  * case has already been eliminated.
737  *
738  * The following mprotect cases have to be considered, where AAAA is
739  * the area passed down from mprotect_fixup, never extending beyond one
740  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
741  *
742  *     AAAA             AAAA                AAAA          AAAA
743  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
744  *    cannot merge    might become    might become    might become
745  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
746  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
747  *    mremap move:                                    PPPPNNNNNNNN 8
748  *        AAAA
749  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
750  *    might become    case 1 below    case 2 below    case 3 below
751  *
752  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
753  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
754  */
755 struct vm_area_struct *vma_merge(struct mm_struct *mm,
756                         struct vm_area_struct *prev, unsigned long addr,
757                         unsigned long end, unsigned long vm_flags,
758                         struct anon_vma *anon_vma, struct file *file,
759                         pgoff_t pgoff, struct mempolicy *policy)
760 {
761         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
762         struct vm_area_struct *area, *next;
763
764         /*
765          * We later require that vma->vm_flags == vm_flags,
766          * so this tests vma->vm_flags & VM_SPECIAL, too.
767          */
768         if (vm_flags & VM_SPECIAL)
769                 return NULL;
770
771         if (prev)
772                 next = prev->vm_next;
773         else
774                 next = mm->mmap;
775         area = next;
776         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
777                 next = next->vm_next;
778
779         /*
780          * Can it merge with the predecessor?
781          */
782         if (prev && prev->vm_end == addr &&
783                         mpol_equal(vma_policy(prev), policy) &&
784                         can_vma_merge_after(prev, vm_flags,
785                                                 anon_vma, file, pgoff)) {
786                 /*
787                  * OK, it can.  Can we now merge in the successor as well?
788                  */
789                 if (next && end == next->vm_start &&
790                                 mpol_equal(policy, vma_policy(next)) &&
791                                 can_vma_merge_before(next, vm_flags,
792                                         anon_vma, file, pgoff+pglen) &&
793                                 is_mergeable_anon_vma(prev->anon_vma,
794                                                       next->anon_vma)) {
795                                                         /* cases 1, 6 */
796                         vma_adjust(prev, prev->vm_start,
797                                 next->vm_end, prev->vm_pgoff, NULL);
798                 } else                                  /* cases 2, 5, 7 */
799                         vma_adjust(prev, prev->vm_start,
800                                 end, prev->vm_pgoff, NULL);
801                 return prev;
802         }
803
804         /*
805          * Can this new request be merged in front of next?
806          */
807         if (next && end == next->vm_start &&
808                         mpol_equal(policy, vma_policy(next)) &&
809                         can_vma_merge_before(next, vm_flags,
810                                         anon_vma, file, pgoff+pglen)) {
811                 if (prev && addr < prev->vm_end)        /* case 4 */
812                         vma_adjust(prev, prev->vm_start,
813                                 addr, prev->vm_pgoff, NULL);
814                 else                                    /* cases 3, 8 */
815                         vma_adjust(area, addr, next->vm_end,
816                                 next->vm_pgoff - pglen, NULL);
817                 return area;
818         }
819
820         return NULL;
821 }
822
823 /*
824  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
825  * neighbouring vmas for a suitable anon_vma, before it goes off
826  * to allocate a new anon_vma.  It checks because a repetitive
827  * sequence of mprotects and faults may otherwise lead to distinct
828  * anon_vmas being allocated, preventing vma merge in subsequent
829  * mprotect.
830  */
831 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
832 {
833         struct vm_area_struct *near;
834         unsigned long vm_flags;
835
836         near = vma->vm_next;
837         if (!near)
838                 goto try_prev;
839
840         /*
841          * Since only mprotect tries to remerge vmas, match flags
842          * which might be mprotected into each other later on.
843          * Neither mlock nor madvise tries to remerge at present,
844          * so leave their flags as obstructing a merge.
845          */
846         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
847         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
848
849         if (near->anon_vma && vma->vm_end == near->vm_start &&
850                         mpol_equal(vma_policy(vma), vma_policy(near)) &&
851                         can_vma_merge_before(near, vm_flags,
852                                 NULL, vma->vm_file, vma->vm_pgoff +
853                                 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
854                 return near->anon_vma;
855 try_prev:
856         /*
857          * It is potentially slow to have to call find_vma_prev here.
858          * But it's only on the first write fault on the vma, not
859          * every time, and we could devise a way to avoid it later
860          * (e.g. stash info in next's anon_vma_node when assigning
861          * an anon_vma, or when trying vma_merge).  Another time.
862          */
863         BUG_ON(find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma);
864         if (!near)
865                 goto none;
866
867         vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
868         vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
869
870         if (near->anon_vma && near->vm_end == vma->vm_start &&
871                         mpol_equal(vma_policy(near), vma_policy(vma)) &&
872                         can_vma_merge_after(near, vm_flags,
873                                 NULL, vma->vm_file, vma->vm_pgoff))
874                 return near->anon_vma;
875 none:
876         /*
877          * There's no absolute need to look only at touching neighbours:
878          * we could search further afield for "compatible" anon_vmas.
879          * But it would probably just be a waste of time searching,
880          * or lead to too many vmas hanging off the same anon_vma.
881          * We're trying to allow mprotect remerging later on,
882          * not trying to minimize memory used for anon_vmas.
883          */
884         return NULL;
885 }
886
887 #ifdef CONFIG_PROC_FS
888 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
889                                                 struct file *file, long pages)
890 {
891         const unsigned long stack_flags
892                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
893
894         if (file) {
895                 mm->shared_vm += pages;
896                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
897                         mm->exec_vm += pages;
898         } else if (flags & stack_flags)
899                 mm->stack_vm += pages;
900         if (flags & (VM_RESERVED|VM_IO))
901                 mm->reserved_vm += pages;
902 }
903 #endif /* CONFIG_PROC_FS */
904
905 /*
906  * The caller must hold down_write(&current->mm->mmap_sem).
907  */
908
909 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
910                         unsigned long len, unsigned long prot,
911                         unsigned long flags, unsigned long pgoff)
912 {
913         struct mm_struct * mm = current->mm;
914         struct inode *inode;
915         unsigned int vm_flags;
916         int error;
917         unsigned long reqprot = prot;
918
919         /*
920          * Does the application expect PROT_READ to imply PROT_EXEC?
921          *
922          * (the exception is when the underlying filesystem is noexec
923          *  mounted, in which case we dont add PROT_EXEC.)
924          */
925         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
926                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
927                         prot |= PROT_EXEC;
928
929         if (!len)
930                 return -EINVAL;
931
932         if (!(flags & MAP_FIXED))
933                 addr = round_hint_to_min(addr);
934
935         error = arch_mmap_check(addr, len, flags);
936         if (error)
937                 return error;
938
939         /* Careful about overflows.. */
940         len = PAGE_ALIGN(len);
941         if (!len || len > TASK_SIZE)
942                 return -ENOMEM;
943
944         /* offset overflow? */
945         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
946                return -EOVERFLOW;
947
948         /* Too many mappings? */
949         if (mm->map_count > sysctl_max_map_count)
950                 return -ENOMEM;
951
952         if (flags & MAP_HUGETLB) {
953                 struct user_struct *user = NULL;
954                 if (file)
955                         return -EINVAL;
956
957                 /*
958                  * VM_NORESERVE is used because the reservations will be
959                  * taken when vm_ops->mmap() is called
960                  * A dummy user value is used because we are not locking
961                  * memory so no accounting is necessary
962                  */
963                 len = ALIGN(len, huge_page_size(&default_hstate));
964                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, len, VM_NORESERVE,
965                                                 &user, HUGETLB_ANONHUGE_INODE);
966                 if (IS_ERR(file))
967                         return PTR_ERR(file);
968         }
969
970         /* Obtain the address to map to. we verify (or select) it and ensure
971          * that it represents a valid section of the address space.
972          */
973         addr = get_unmapped_area(file, addr, len, pgoff, flags);
974         if (addr & ~PAGE_MASK)
975                 return addr;
976
977         /* Do simple checking here so the lower-level routines won't have
978          * to. we assume access permissions have been handled by the open
979          * of the memory object, so we don't do any here.
980          */
981         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
982                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
983
984         if (flags & MAP_LOCKED)
985                 if (!can_do_mlock())
986                         return -EPERM;
987
988         /* mlock MCL_FUTURE? */
989         if (vm_flags & VM_LOCKED) {
990                 unsigned long locked, lock_limit;
991                 locked = len >> PAGE_SHIFT;
992                 locked += mm->locked_vm;
993                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
994                 lock_limit >>= PAGE_SHIFT;
995                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
996                         return -EAGAIN;
997         }
998
999         inode = file ? file->f_path.dentry->d_inode : NULL;
1000
1001         if (file) {
1002                 switch (flags & MAP_TYPE) {
1003                 case MAP_SHARED:
1004                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1005                                 return -EACCES;
1006
1007                         /*
1008                          * Make sure we don't allow writing to an append-only
1009                          * file..
1010                          */
1011                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1012                                 return -EACCES;
1013
1014                         /*
1015                          * Make sure there are no mandatory locks on the file.
1016                          */
1017                         if (locks_verify_locked(inode))
1018                                 return -EAGAIN;
1019
1020                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1021                         if (!(file->f_mode & FMODE_WRITE))
1022                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1023
1024                         /* fall through */
1025                 case MAP_PRIVATE:
1026                         if (!(file->f_mode & FMODE_READ))
1027                                 return -EACCES;
1028                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1029                                 if (vm_flags & VM_EXEC)
1030                                         return -EPERM;
1031                                 vm_flags &= ~VM_MAYEXEC;
1032                         }
1033
1034                         if (!file->f_op || !file->f_op->mmap)
1035                                 return -ENODEV;
1036                         break;
1037
1038                 default:
1039                         return -EINVAL;
1040                 }
1041         } else {
1042                 switch (flags & MAP_TYPE) {
1043                 case MAP_SHARED:
1044                         /*
1045                          * Ignore pgoff.
1046                          */
1047                         pgoff = 0;
1048                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1049                         break;
1050                 case MAP_PRIVATE:
1051                         /*
1052                          * Set pgoff according to addr for anon_vma.
1053                          */
1054                         pgoff = addr >> PAGE_SHIFT;
1055                         break;
1056                 default:
1057                         return -EINVAL;
1058                 }
1059         }
1060
1061         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1062         if (error)
1063                 return error;
1064         error = ima_file_mmap(file, prot);
1065         if (error)
1066                 return error;
1067
1068         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1069 }
1070 EXPORT_SYMBOL(do_mmap_pgoff);
1071
1072 /*
1073  * Some shared mappigns will want the pages marked read-only
1074  * to track write events. If so, we'll downgrade vm_page_prot
1075  * to the private version (using protection_map[] without the
1076  * VM_SHARED bit).
1077  */
1078 int vma_wants_writenotify(struct vm_area_struct *vma)
1079 {
1080         unsigned int vm_flags = vma->vm_flags;
1081
1082         /* If it was private or non-writable, the write bit is already clear */
1083         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1084                 return 0;
1085
1086         /* The backer wishes to know when pages are first written to? */
1087         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1088                 return 1;
1089
1090         /* The open routine did something to the protections already? */
1091         if (pgprot_val(vma->vm_page_prot) !=
1092             pgprot_val(vm_get_page_prot(vm_flags)))
1093                 return 0;
1094
1095         /* Specialty mapping? */
1096         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1097                 return 0;
1098
1099         /* Can the mapping track the dirty pages? */
1100         return vma->vm_file && vma->vm_file->f_mapping &&
1101                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1102 }
1103
1104 /*
1105  * We account for memory if it's a private writeable mapping,
1106  * not hugepages and VM_NORESERVE wasn't set.
1107  */
1108 static inline int accountable_mapping(struct file *file, unsigned int vm_flags)
1109 {
1110         /*
1111          * hugetlb has its own accounting separate from the core VM
1112          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1113          */
1114         if (file && is_file_hugepages(file))
1115                 return 0;
1116
1117         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1118 }
1119
1120 unsigned long mmap_region(struct file *file, unsigned long addr,
1121                           unsigned long len, unsigned long flags,
1122                           unsigned int vm_flags, unsigned long pgoff)
1123 {
1124         struct mm_struct *mm = current->mm;
1125         struct vm_area_struct *vma, *prev;
1126         int correct_wcount = 0;
1127         int error;
1128         struct rb_node **rb_link, *rb_parent;
1129         unsigned long charged = 0;
1130         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1131
1132         /* Clear old maps */
1133         error = -ENOMEM;
1134 munmap_back:
1135         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1136         if (vma && vma->vm_start < addr + len) {
1137                 if (do_munmap(mm, addr, len))
1138                         return -ENOMEM;
1139                 goto munmap_back;
1140         }
1141
1142         /* Check against address space limit. */
1143         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1144                 return -ENOMEM;
1145
1146         /*
1147          * Set 'VM_NORESERVE' if we should not account for the
1148          * memory use of this mapping.
1149          */
1150         if ((flags & MAP_NORESERVE)) {
1151                 /* We honor MAP_NORESERVE if allowed to overcommit */
1152                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1153                         vm_flags |= VM_NORESERVE;
1154
1155                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1156                 if (file && is_file_hugepages(file))
1157                         vm_flags |= VM_NORESERVE;
1158         }
1159
1160         /*
1161          * Private writable mapping: check memory availability
1162          */
1163         if (accountable_mapping(file, vm_flags)) {
1164                 charged = len >> PAGE_SHIFT;
1165                 if (security_vm_enough_memory(charged))
1166                         return -ENOMEM;
1167                 vm_flags |= VM_ACCOUNT;
1168         }
1169
1170         /*
1171          * Can we just expand an old mapping?
1172          */
1173         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1174         if (vma)
1175                 goto out;
1176
1177         /*
1178          * Determine the object being mapped and call the appropriate
1179          * specific mapper. the address has already been validated, but
1180          * not unmapped, but the maps are removed from the list.
1181          */
1182         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1183         if (!vma) {
1184                 error = -ENOMEM;
1185                 goto unacct_error;
1186         }
1187
1188         vma->vm_mm = mm;
1189         vma->vm_start = addr;
1190         vma->vm_end = addr + len;
1191         vma->vm_flags = vm_flags;
1192         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1193         vma->vm_pgoff = pgoff;
1194
1195         if (file) {
1196                 error = -EINVAL;
1197                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1198                         goto free_vma;
1199                 if (vm_flags & VM_DENYWRITE) {
1200                         error = deny_write_access(file);
1201                         if (error)
1202                                 goto free_vma;
1203                         correct_wcount = 1;
1204                 }
1205                 vma->vm_file = file;
1206                 get_file(file);
1207                 error = file->f_op->mmap(file, vma);
1208                 if (error)
1209                         goto unmap_and_free_vma;
1210                 if (vm_flags & VM_EXECUTABLE)
1211                         added_exe_file_vma(mm);
1212
1213                 /* Can addr have changed??
1214                  *
1215                  * Answer: Yes, several device drivers can do it in their
1216                  *         f_op->mmap method. -DaveM
1217                  */
1218                 addr = vma->vm_start;
1219                 pgoff = vma->vm_pgoff;
1220                 vm_flags = vma->vm_flags;
1221         } else if (vm_flags & VM_SHARED) {
1222                 error = shmem_zero_setup(vma);
1223                 if (error)
1224                         goto free_vma;
1225         }
1226
1227         if (vma_wants_writenotify(vma))
1228                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1229
1230         vma_link(mm, vma, prev, rb_link, rb_parent);
1231         file = vma->vm_file;
1232
1233         /* Once vma denies write, undo our temporary denial count */
1234         if (correct_wcount)
1235                 atomic_inc(&inode->i_writecount);
1236 out:
1237         perf_event_mmap(vma);
1238
1239         mm->total_vm += len >> PAGE_SHIFT;
1240         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1241         if (vm_flags & VM_LOCKED) {
1242                 /*
1243                  * makes pages present; downgrades, drops, reacquires mmap_sem
1244                  */
1245                 long nr_pages = mlock_vma_pages_range(vma, addr, addr + len);
1246                 if (nr_pages < 0)
1247                         return nr_pages;        /* vma gone! */
1248                 mm->locked_vm += (len >> PAGE_SHIFT) - nr_pages;
1249         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1250                 make_pages_present(addr, addr + len);
1251         return addr;
1252
1253 unmap_and_free_vma:
1254         if (correct_wcount)
1255                 atomic_inc(&inode->i_writecount);
1256         vma->vm_file = NULL;
1257         fput(file);
1258
1259         /* Undo any partial mapping done by a device driver. */
1260         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1261         charged = 0;
1262 free_vma:
1263         kmem_cache_free(vm_area_cachep, vma);
1264 unacct_error:
1265         if (charged)
1266                 vm_unacct_memory(charged);
1267         return error;
1268 }
1269
1270 /* Get an address range which is currently unmapped.
1271  * For shmat() with addr=0.
1272  *
1273  * Ugly calling convention alert:
1274  * Return value with the low bits set means error value,
1275  * ie
1276  *      if (ret & ~PAGE_MASK)
1277  *              error = ret;
1278  *
1279  * This function "knows" that -ENOMEM has the bits set.
1280  */
1281 #ifndef HAVE_ARCH_UNMAPPED_AREA
1282 unsigned long
1283 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1284                 unsigned long len, unsigned long pgoff, unsigned long flags)
1285 {
1286         struct mm_struct *mm = current->mm;
1287         struct vm_area_struct *vma;
1288         unsigned long start_addr;
1289
1290         if (len > TASK_SIZE)
1291                 return -ENOMEM;
1292
1293         if (flags & MAP_FIXED)
1294                 return addr;
1295
1296         if (addr) {
1297                 addr = PAGE_ALIGN(addr);
1298                 vma = find_vma(mm, addr);
1299                 if (TASK_SIZE - len >= addr &&
1300                     (!vma || addr + len <= vma->vm_start))
1301                         return addr;
1302         }
1303         if (len > mm->cached_hole_size) {
1304                 start_addr = addr = mm->free_area_cache;
1305         } else {
1306                 start_addr = addr = TASK_UNMAPPED_BASE;
1307                 mm->cached_hole_size = 0;
1308         }
1309
1310 full_search:
1311         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1312                 /* At this point:  (!vma || addr < vma->vm_end). */
1313                 if (TASK_SIZE - len < addr) {
1314                         /*
1315                          * Start a new search - just in case we missed
1316                          * some holes.
1317                          */
1318                         if (start_addr != TASK_UNMAPPED_BASE) {
1319                                 addr = TASK_UNMAPPED_BASE;
1320                                 start_addr = addr;
1321                                 mm->cached_hole_size = 0;
1322                                 goto full_search;
1323                         }
1324                         return -ENOMEM;
1325                 }
1326                 if (!vma || addr + len <= vma->vm_start) {
1327                         /*
1328                          * Remember the place where we stopped the search:
1329                          */
1330                         mm->free_area_cache = addr + len;
1331                         return addr;
1332                 }
1333                 if (addr + mm->cached_hole_size < vma->vm_start)
1334                         mm->cached_hole_size = vma->vm_start - addr;
1335                 addr = vma->vm_end;
1336         }
1337 }
1338 #endif  
1339
1340 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1341 {
1342         /*
1343          * Is this a new hole at the lowest possible address?
1344          */
1345         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1346                 mm->free_area_cache = addr;
1347                 mm->cached_hole_size = ~0UL;
1348         }
1349 }
1350
1351 /*
1352  * This mmap-allocator allocates new areas top-down from below the
1353  * stack's low limit (the base):
1354  */
1355 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1356 unsigned long
1357 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1358                           const unsigned long len, const unsigned long pgoff,
1359                           const unsigned long flags)
1360 {
1361         struct vm_area_struct *vma;
1362         struct mm_struct *mm = current->mm;
1363         unsigned long addr = addr0;
1364
1365         /* requested length too big for entire address space */
1366         if (len > TASK_SIZE)
1367                 return -ENOMEM;
1368
1369         if (flags & MAP_FIXED)
1370                 return addr;
1371
1372         /* requesting a specific address */
1373         if (addr) {
1374                 addr = PAGE_ALIGN(addr);
1375                 vma = find_vma(mm, addr);
1376                 if (TASK_SIZE - len >= addr &&
1377                                 (!vma || addr + len <= vma->vm_start))
1378                         return addr;
1379         }
1380
1381         /* check if free_area_cache is useful for us */
1382         if (len <= mm->cached_hole_size) {
1383                 mm->cached_hole_size = 0;
1384                 mm->free_area_cache = mm->mmap_base;
1385         }
1386
1387         /* either no address requested or can't fit in requested address hole */
1388         addr = mm->free_area_cache;
1389
1390         /* make sure it can fit in the remaining address space */
1391         if (addr > len) {
1392                 vma = find_vma(mm, addr-len);
1393                 if (!vma || addr <= vma->vm_start)
1394                         /* remember the address as a hint for next time */
1395                         return (mm->free_area_cache = addr-len);
1396         }
1397
1398         if (mm->mmap_base < len)
1399                 goto bottomup;
1400
1401         addr = mm->mmap_base-len;
1402
1403         do {
1404                 /*
1405                  * Lookup failure means no vma is above this address,
1406                  * else if new region fits below vma->vm_start,
1407                  * return with success:
1408                  */
1409                 vma = find_vma(mm, addr);
1410                 if (!vma || addr+len <= vma->vm_start)
1411                         /* remember the address as a hint for next time */
1412                         return (mm->free_area_cache = addr);
1413
1414                 /* remember the largest hole we saw so far */
1415                 if (addr + mm->cached_hole_size < vma->vm_start)
1416                         mm->cached_hole_size = vma->vm_start - addr;
1417
1418                 /* try just below the current vma->vm_start */
1419                 addr = vma->vm_start-len;
1420         } while (len < vma->vm_start);
1421
1422 bottomup:
1423         /*
1424          * A failed mmap() very likely causes application failure,
1425          * so fall back to the bottom-up function here. This scenario
1426          * can happen with large stack limits and large mmap()
1427          * allocations.
1428          */
1429         mm->cached_hole_size = ~0UL;
1430         mm->free_area_cache = TASK_UNMAPPED_BASE;
1431         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1432         /*
1433          * Restore the topdown base:
1434          */
1435         mm->free_area_cache = mm->mmap_base;
1436         mm->cached_hole_size = ~0UL;
1437
1438         return addr;
1439 }
1440 #endif
1441
1442 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1443 {
1444         /*
1445          * Is this a new hole at the highest possible address?
1446          */
1447         if (addr > mm->free_area_cache)
1448                 mm->free_area_cache = addr;
1449
1450         /* dont allow allocations above current base */
1451         if (mm->free_area_cache > mm->mmap_base)
1452                 mm->free_area_cache = mm->mmap_base;
1453 }
1454
1455 unsigned long
1456 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1457                 unsigned long pgoff, unsigned long flags)
1458 {
1459         unsigned long (*get_area)(struct file *, unsigned long,
1460                                   unsigned long, unsigned long, unsigned long);
1461
1462         get_area = current->mm->get_unmapped_area;
1463         if (file && file->f_op && file->f_op->get_unmapped_area)
1464                 get_area = file->f_op->get_unmapped_area;
1465         addr = get_area(file, addr, len, pgoff, flags);
1466         if (IS_ERR_VALUE(addr))
1467                 return addr;
1468
1469         if (addr > TASK_SIZE - len)
1470                 return -ENOMEM;
1471         if (addr & ~PAGE_MASK)
1472                 return -EINVAL;
1473
1474         return arch_rebalance_pgtables(addr, len);
1475 }
1476
1477 EXPORT_SYMBOL(get_unmapped_area);
1478
1479 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1480 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1481 {
1482         struct vm_area_struct *vma = NULL;
1483
1484         if (mm) {
1485                 /* Check the cache first. */
1486                 /* (Cache hit rate is typically around 35%.) */
1487                 vma = mm->mmap_cache;
1488                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1489                         struct rb_node * rb_node;
1490
1491                         rb_node = mm->mm_rb.rb_node;
1492                         vma = NULL;
1493
1494                         while (rb_node) {
1495                                 struct vm_area_struct * vma_tmp;
1496
1497                                 vma_tmp = rb_entry(rb_node,
1498                                                 struct vm_area_struct, vm_rb);
1499
1500                                 if (vma_tmp->vm_end > addr) {
1501                                         vma = vma_tmp;
1502                                         if (vma_tmp->vm_start <= addr)
1503                                                 break;
1504                                         rb_node = rb_node->rb_left;
1505                                 } else
1506                                         rb_node = rb_node->rb_right;
1507                         }
1508                         if (vma)
1509                                 mm->mmap_cache = vma;
1510                 }
1511         }
1512         return vma;
1513 }
1514
1515 EXPORT_SYMBOL(find_vma);
1516
1517 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1518 struct vm_area_struct *
1519 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1520                         struct vm_area_struct **pprev)
1521 {
1522         struct vm_area_struct *vma = NULL, *prev = NULL;
1523         struct rb_node *rb_node;
1524         if (!mm)
1525                 goto out;
1526
1527         /* Guard against addr being lower than the first VMA */
1528         vma = mm->mmap;
1529
1530         /* Go through the RB tree quickly. */
1531         rb_node = mm->mm_rb.rb_node;
1532
1533         while (rb_node) {
1534                 struct vm_area_struct *vma_tmp;
1535                 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1536
1537                 if (addr < vma_tmp->vm_end) {
1538                         rb_node = rb_node->rb_left;
1539                 } else {
1540                         prev = vma_tmp;
1541                         if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1542                                 break;
1543                         rb_node = rb_node->rb_right;
1544                 }
1545         }
1546
1547 out:
1548         *pprev = prev;
1549         return prev ? prev->vm_next : vma;
1550 }
1551
1552 /*
1553  * Verify that the stack growth is acceptable and
1554  * update accounting. This is shared with both the
1555  * grow-up and grow-down cases.
1556  */
1557 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1558 {
1559         struct mm_struct *mm = vma->vm_mm;
1560         struct rlimit *rlim = current->signal->rlim;
1561         unsigned long new_start;
1562
1563         /* address space limit tests */
1564         if (!may_expand_vm(mm, grow))
1565                 return -ENOMEM;
1566
1567         /* Stack limit test */
1568         if (size > rlim[RLIMIT_STACK].rlim_cur)
1569                 return -ENOMEM;
1570
1571         /* mlock limit tests */
1572         if (vma->vm_flags & VM_LOCKED) {
1573                 unsigned long locked;
1574                 unsigned long limit;
1575                 locked = mm->locked_vm + grow;
1576                 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1577                 if (locked > limit && !capable(CAP_IPC_LOCK))
1578                         return -ENOMEM;
1579         }
1580
1581         /* Check to ensure the stack will not grow into a hugetlb-only region */
1582         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1583                         vma->vm_end - size;
1584         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1585                 return -EFAULT;
1586
1587         /*
1588          * Overcommit..  This must be the final test, as it will
1589          * update security statistics.
1590          */
1591         if (security_vm_enough_memory_mm(mm, grow))
1592                 return -ENOMEM;
1593
1594         /* Ok, everything looks good - let it rip */
1595         mm->total_vm += grow;
1596         if (vma->vm_flags & VM_LOCKED)
1597                 mm->locked_vm += grow;
1598         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1599         return 0;
1600 }
1601
1602 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1603 /*
1604  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1605  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1606  */
1607 #ifndef CONFIG_IA64
1608 static
1609 #endif
1610 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1611 {
1612         int error;
1613
1614         if (!(vma->vm_flags & VM_GROWSUP))
1615                 return -EFAULT;
1616
1617         /*
1618          * We must make sure the anon_vma is allocated
1619          * so that the anon_vma locking is not a noop.
1620          */
1621         if (unlikely(anon_vma_prepare(vma)))
1622                 return -ENOMEM;
1623         anon_vma_lock(vma);
1624
1625         /*
1626          * vma->vm_start/vm_end cannot change under us because the caller
1627          * is required to hold the mmap_sem in read mode.  We need the
1628          * anon_vma lock to serialize against concurrent expand_stacks.
1629          * Also guard against wrapping around to address 0.
1630          */
1631         if (address < PAGE_ALIGN(address+4))
1632                 address = PAGE_ALIGN(address+4);
1633         else {
1634                 anon_vma_unlock(vma);
1635                 return -ENOMEM;
1636         }
1637         error = 0;
1638
1639         /* Somebody else might have raced and expanded it already */
1640         if (address > vma->vm_end) {
1641                 unsigned long size, grow;
1642
1643                 size = address - vma->vm_start;
1644                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1645
1646                 error = acct_stack_growth(vma, size, grow);
1647                 if (!error)
1648                         vma->vm_end = address;
1649         }
1650         anon_vma_unlock(vma);
1651         return error;
1652 }
1653 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1654
1655 /*
1656  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1657  */
1658 static int expand_downwards(struct vm_area_struct *vma,
1659                                    unsigned long address)
1660 {
1661         int error;
1662
1663         /*
1664          * We must make sure the anon_vma is allocated
1665          * so that the anon_vma locking is not a noop.
1666          */
1667         if (unlikely(anon_vma_prepare(vma)))
1668                 return -ENOMEM;
1669
1670         address &= PAGE_MASK;
1671         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1672         if (error)
1673                 return error;
1674
1675         anon_vma_lock(vma);
1676
1677         /*
1678          * vma->vm_start/vm_end cannot change under us because the caller
1679          * is required to hold the mmap_sem in read mode.  We need the
1680          * anon_vma lock to serialize against concurrent expand_stacks.
1681          */
1682
1683         /* Somebody else might have raced and expanded it already */
1684         if (address < vma->vm_start) {
1685                 unsigned long size, grow;
1686
1687                 size = vma->vm_end - address;
1688                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1689
1690                 error = acct_stack_growth(vma, size, grow);
1691                 if (!error) {
1692                         vma->vm_start = address;
1693                         vma->vm_pgoff -= grow;
1694                 }
1695         }
1696         anon_vma_unlock(vma);
1697         return error;
1698 }
1699
1700 int expand_stack_downwards(struct vm_area_struct *vma, unsigned long address)
1701 {
1702         return expand_downwards(vma, address);
1703 }
1704
1705 #ifdef CONFIG_STACK_GROWSUP
1706 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1707 {
1708         return expand_upwards(vma, address);
1709 }
1710
1711 struct vm_area_struct *
1712 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1713 {
1714         struct vm_area_struct *vma, *prev;
1715
1716         addr &= PAGE_MASK;
1717         vma = find_vma_prev(mm, addr, &prev);
1718         if (vma && (vma->vm_start <= addr))
1719                 return vma;
1720         if (!prev || expand_stack(prev, addr))
1721                 return NULL;
1722         if (prev->vm_flags & VM_LOCKED) {
1723                 if (mlock_vma_pages_range(prev, addr, prev->vm_end) < 0)
1724                         return NULL;    /* vma gone! */
1725         }
1726         return prev;
1727 }
1728 #else
1729 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1730 {
1731         return expand_downwards(vma, address);
1732 }
1733
1734 struct vm_area_struct *
1735 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1736 {
1737         struct vm_area_struct * vma;
1738         unsigned long start;
1739
1740         addr &= PAGE_MASK;
1741         vma = find_vma(mm,addr);
1742         if (!vma)
1743                 return NULL;
1744         if (vma->vm_start <= addr)
1745                 return vma;
1746         if (!(vma->vm_flags & VM_GROWSDOWN))
1747                 return NULL;
1748         start = vma->vm_start;
1749         if (expand_stack(vma, addr))
1750                 return NULL;
1751         if (vma->vm_flags & VM_LOCKED) {
1752                 if (mlock_vma_pages_range(vma, addr, start) < 0)
1753                         return NULL;    /* vma gone! */
1754         }
1755         return vma;
1756 }
1757 #endif
1758
1759 /*
1760  * Ok - we have the memory areas we should free on the vma list,
1761  * so release them, and do the vma updates.
1762  *
1763  * Called with the mm semaphore held.
1764  */
1765 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1766 {
1767         /* Update high watermark before we lower total_vm */
1768         update_hiwater_vm(mm);
1769         do {
1770                 long nrpages = vma_pages(vma);
1771
1772                 mm->total_vm -= nrpages;
1773                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1774                 vma = remove_vma(vma);
1775         } while (vma);
1776         validate_mm(mm);
1777 }
1778
1779 /*
1780  * Get rid of page table information in the indicated region.
1781  *
1782  * Called with the mm semaphore held.
1783  */
1784 static void unmap_region(struct mm_struct *mm,
1785                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1786                 unsigned long start, unsigned long end)
1787 {
1788         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1789         struct mmu_gather *tlb;
1790         unsigned long nr_accounted = 0;
1791
1792         lru_add_drain();
1793         tlb = tlb_gather_mmu(mm, 0);
1794         update_hiwater_rss(mm);
1795         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1796         vm_unacct_memory(nr_accounted);
1797         free_pgtables(tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1798                                  next? next->vm_start: 0);
1799         tlb_finish_mmu(tlb, start, end);
1800 }
1801
1802 /*
1803  * Create a list of vma's touched by the unmap, removing them from the mm's
1804  * vma list as we go..
1805  */
1806 static void
1807 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1808         struct vm_area_struct *prev, unsigned long end)
1809 {
1810         struct vm_area_struct **insertion_point;
1811         struct vm_area_struct *tail_vma = NULL;
1812         unsigned long addr;
1813
1814         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1815         do {
1816                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1817                 mm->map_count--;
1818                 tail_vma = vma;
1819                 vma = vma->vm_next;
1820         } while (vma && vma->vm_start < end);
1821         *insertion_point = vma;
1822         tail_vma->vm_next = NULL;
1823         if (mm->unmap_area == arch_unmap_area)
1824                 addr = prev ? prev->vm_end : mm->mmap_base;
1825         else
1826                 addr = vma ?  vma->vm_start : mm->mmap_base;
1827         mm->unmap_area(mm, addr);
1828         mm->mmap_cache = NULL;          /* Kill the cache. */
1829 }
1830
1831 /*
1832  * Split a vma into two pieces at address 'addr', a new vma is allocated
1833  * either for the first part or the tail.
1834  */
1835 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1836               unsigned long addr, int new_below)
1837 {
1838         struct mempolicy *pol;
1839         struct vm_area_struct *new;
1840
1841         if (is_vm_hugetlb_page(vma) && (addr &
1842                                         ~(huge_page_mask(hstate_vma(vma)))))
1843                 return -EINVAL;
1844
1845         if (mm->map_count >= sysctl_max_map_count)
1846                 return -ENOMEM;
1847
1848         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1849         if (!new)
1850                 return -ENOMEM;
1851
1852         /* most fields are the same, copy all, and then fixup */
1853         *new = *vma;
1854
1855         if (new_below)
1856                 new->vm_end = addr;
1857         else {
1858                 new->vm_start = addr;
1859                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1860         }
1861
1862         pol = mpol_dup(vma_policy(vma));
1863         if (IS_ERR(pol)) {
1864                 kmem_cache_free(vm_area_cachep, new);
1865                 return PTR_ERR(pol);
1866         }
1867         vma_set_policy(new, pol);
1868
1869         if (new->vm_file) {
1870                 get_file(new->vm_file);
1871                 if (vma->vm_flags & VM_EXECUTABLE)
1872                         added_exe_file_vma(mm);
1873         }
1874
1875         if (new->vm_ops && new->vm_ops->open)
1876                 new->vm_ops->open(new);
1877
1878         if (new_below)
1879                 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1880                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
1881         else
1882                 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1883
1884         return 0;
1885 }
1886
1887 /* Munmap is split into 2 main parts -- this part which finds
1888  * what needs doing, and the areas themselves, which do the
1889  * work.  This now handles partial unmappings.
1890  * Jeremy Fitzhardinge <jeremy@goop.org>
1891  */
1892 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1893 {
1894         unsigned long end;
1895         struct vm_area_struct *vma, *prev, *last;
1896
1897         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1898                 return -EINVAL;
1899
1900         if ((len = PAGE_ALIGN(len)) == 0)
1901                 return -EINVAL;
1902
1903         /* Find the first overlapping VMA */
1904         vma = find_vma_prev(mm, start, &prev);
1905         if (!vma)
1906                 return 0;
1907         /* we have  start < vma->vm_end  */
1908
1909         /* if it doesn't overlap, we have nothing.. */
1910         end = start + len;
1911         if (vma->vm_start >= end)
1912                 return 0;
1913
1914         /*
1915          * If we need to split any vma, do it now to save pain later.
1916          *
1917          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1918          * unmapped vm_area_struct will remain in use: so lower split_vma
1919          * places tmp vma above, and higher split_vma places tmp vma below.
1920          */
1921         if (start > vma->vm_start) {
1922                 int error = split_vma(mm, vma, start, 0);
1923                 if (error)
1924                         return error;
1925                 prev = vma;
1926         }
1927
1928         /* Does it split the last one? */
1929         last = find_vma(mm, end);
1930         if (last && end > last->vm_start) {
1931                 int error = split_vma(mm, last, end, 1);
1932                 if (error)
1933                         return error;
1934         }
1935         vma = prev? prev->vm_next: mm->mmap;
1936
1937         /*
1938          * unlock any mlock()ed ranges before detaching vmas
1939          */
1940         if (mm->locked_vm) {
1941                 struct vm_area_struct *tmp = vma;
1942                 while (tmp && tmp->vm_start < end) {
1943                         if (tmp->vm_flags & VM_LOCKED) {
1944                                 mm->locked_vm -= vma_pages(tmp);
1945                                 munlock_vma_pages_all(tmp);
1946                         }
1947                         tmp = tmp->vm_next;
1948                 }
1949         }
1950
1951         /*
1952          * Remove the vma's, and unmap the actual pages
1953          */
1954         detach_vmas_to_be_unmapped(mm, vma, prev, end);
1955         unmap_region(mm, vma, prev, start, end);
1956
1957         /* Fix up all other VM information */
1958         remove_vma_list(mm, vma);
1959
1960         return 0;
1961 }
1962
1963 EXPORT_SYMBOL(do_munmap);
1964
1965 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1966 {
1967         int ret;
1968         struct mm_struct *mm = current->mm;
1969
1970         profile_munmap(addr);
1971
1972         down_write(&mm->mmap_sem);
1973         ret = do_munmap(mm, addr, len);
1974         up_write(&mm->mmap_sem);
1975         return ret;
1976 }
1977
1978 static inline void verify_mm_writelocked(struct mm_struct *mm)
1979 {
1980 #ifdef CONFIG_DEBUG_VM
1981         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1982                 WARN_ON(1);
1983                 up_read(&mm->mmap_sem);
1984         }
1985 #endif
1986 }
1987
1988 /*
1989  *  this is really a simplified "do_mmap".  it only handles
1990  *  anonymous maps.  eventually we may be able to do some
1991  *  brk-specific accounting here.
1992  */
1993 unsigned long do_brk(unsigned long addr, unsigned long len)
1994 {
1995         struct mm_struct * mm = current->mm;
1996         struct vm_area_struct * vma, * prev;
1997         unsigned long flags;
1998         struct rb_node ** rb_link, * rb_parent;
1999         pgoff_t pgoff = addr >> PAGE_SHIFT;
2000         int error;
2001
2002         len = PAGE_ALIGN(len);
2003         if (!len)
2004                 return addr;
2005
2006         if ((addr + len) > TASK_SIZE || (addr + len) < addr)
2007                 return -EINVAL;
2008
2009         if (is_hugepage_only_range(mm, addr, len))
2010                 return -EINVAL;
2011
2012         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2013         if (error)
2014                 return error;
2015
2016         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2017
2018         error = arch_mmap_check(addr, len, flags);
2019         if (error)
2020                 return error;
2021
2022         /*
2023          * mlock MCL_FUTURE?
2024          */
2025         if (mm->def_flags & VM_LOCKED) {
2026                 unsigned long locked, lock_limit;
2027                 locked = len >> PAGE_SHIFT;
2028                 locked += mm->locked_vm;
2029                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
2030                 lock_limit >>= PAGE_SHIFT;
2031                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2032                         return -EAGAIN;
2033         }
2034
2035         /*
2036          * mm->mmap_sem is required to protect against another thread
2037          * changing the mappings in case we sleep.
2038          */
2039         verify_mm_writelocked(mm);
2040
2041         /*
2042          * Clear old maps.  this also does some error checking for us
2043          */
2044  munmap_back:
2045         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2046         if (vma && vma->vm_start < addr + len) {
2047                 if (do_munmap(mm, addr, len))
2048                         return -ENOMEM;
2049                 goto munmap_back;
2050         }
2051
2052         /* Check against address space limits *after* clearing old maps... */
2053         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2054                 return -ENOMEM;
2055
2056         if (mm->map_count > sysctl_max_map_count)
2057                 return -ENOMEM;
2058
2059         if (security_vm_enough_memory(len >> PAGE_SHIFT))
2060                 return -ENOMEM;
2061
2062         /* Can we just expand an old private anonymous mapping? */
2063         vma = vma_merge(mm, prev, addr, addr + len, flags,
2064                                         NULL, NULL, pgoff, NULL);
2065         if (vma)
2066                 goto out;
2067
2068         /*
2069          * create a vma struct for an anonymous mapping
2070          */
2071         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2072         if (!vma) {
2073                 vm_unacct_memory(len >> PAGE_SHIFT);
2074                 return -ENOMEM;
2075         }
2076
2077         vma->vm_mm = mm;
2078         vma->vm_start = addr;
2079         vma->vm_end = addr + len;
2080         vma->vm_pgoff = pgoff;
2081         vma->vm_flags = flags;
2082         vma->vm_page_prot = vm_get_page_prot(flags);
2083         vma_link(mm, vma, prev, rb_link, rb_parent);
2084 out:
2085         mm->total_vm += len >> PAGE_SHIFT;
2086         if (flags & VM_LOCKED) {
2087                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2088                         mm->locked_vm += (len >> PAGE_SHIFT);
2089         }
2090         return addr;
2091 }
2092
2093 EXPORT_SYMBOL(do_brk);
2094
2095 /* Release all mmaps. */
2096 void exit_mmap(struct mm_struct *mm)
2097 {
2098         struct mmu_gather *tlb;
2099         struct vm_area_struct *vma;
2100         unsigned long nr_accounted = 0;
2101         unsigned long end;
2102
2103         /* mm's last user has gone, and its about to be pulled down */
2104         mmu_notifier_release(mm);
2105
2106         if (mm->locked_vm) {
2107                 vma = mm->mmap;
2108                 while (vma) {
2109                         if (vma->vm_flags & VM_LOCKED)
2110                                 munlock_vma_pages_all(vma);
2111                         vma = vma->vm_next;
2112                 }
2113         }
2114
2115         arch_exit_mmap(mm);
2116
2117         vma = mm->mmap;
2118         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2119                 return;
2120
2121         lru_add_drain();
2122         flush_cache_mm(mm);
2123         tlb = tlb_gather_mmu(mm, 1);
2124         /* update_hiwater_rss(mm) here? but nobody should be looking */
2125         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2126         end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2127         vm_unacct_memory(nr_accounted);
2128
2129         free_pgtables(tlb, vma, FIRST_USER_ADDRESS, 0);
2130         tlb_finish_mmu(tlb, 0, end);
2131
2132         /*
2133          * Walk the list again, actually closing and freeing it,
2134          * with preemption enabled, without holding any MM locks.
2135          */
2136         while (vma)
2137                 vma = remove_vma(vma);
2138
2139         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2140 }
2141
2142 /* Insert vm structure into process list sorted by address
2143  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2144  * then i_mmap_lock is taken here.
2145  */
2146 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2147 {
2148         struct vm_area_struct * __vma, * prev;
2149         struct rb_node ** rb_link, * rb_parent;
2150
2151         /*
2152          * The vm_pgoff of a purely anonymous vma should be irrelevant
2153          * until its first write fault, when page's anon_vma and index
2154          * are set.  But now set the vm_pgoff it will almost certainly
2155          * end up with (unless mremap moves it elsewhere before that
2156          * first wfault), so /proc/pid/maps tells a consistent story.
2157          *
2158          * By setting it to reflect the virtual start address of the
2159          * vma, merges and splits can happen in a seamless way, just
2160          * using the existing file pgoff checks and manipulations.
2161          * Similarly in do_mmap_pgoff and in do_brk.
2162          */
2163         if (!vma->vm_file) {
2164                 BUG_ON(vma->anon_vma);
2165                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2166         }
2167         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2168         if (__vma && __vma->vm_start < vma->vm_end)
2169                 return -ENOMEM;
2170         if ((vma->vm_flags & VM_ACCOUNT) &&
2171              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2172                 return -ENOMEM;
2173         vma_link(mm, vma, prev, rb_link, rb_parent);
2174         return 0;
2175 }
2176
2177 /*
2178  * Copy the vma structure to a new location in the same mm,
2179  * prior to moving page table entries, to effect an mremap move.
2180  */
2181 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2182         unsigned long addr, unsigned long len, pgoff_t pgoff)
2183 {
2184         struct vm_area_struct *vma = *vmap;
2185         unsigned long vma_start = vma->vm_start;
2186         struct mm_struct *mm = vma->vm_mm;
2187         struct vm_area_struct *new_vma, *prev;
2188         struct rb_node **rb_link, *rb_parent;
2189         struct mempolicy *pol;
2190
2191         /*
2192          * If anonymous vma has not yet been faulted, update new pgoff
2193          * to match new location, to increase its chance of merging.
2194          */
2195         if (!vma->vm_file && !vma->anon_vma)
2196                 pgoff = addr >> PAGE_SHIFT;
2197
2198         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2199         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2200                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2201         if (new_vma) {
2202                 /*
2203                  * Source vma may have been merged into new_vma
2204                  */
2205                 if (vma_start >= new_vma->vm_start &&
2206                     vma_start < new_vma->vm_end)
2207                         *vmap = new_vma;
2208         } else {
2209                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2210                 if (new_vma) {
2211                         *new_vma = *vma;
2212                         pol = mpol_dup(vma_policy(vma));
2213                         if (IS_ERR(pol)) {
2214                                 kmem_cache_free(vm_area_cachep, new_vma);
2215                                 return NULL;
2216                         }
2217                         vma_set_policy(new_vma, pol);
2218                         new_vma->vm_start = addr;
2219                         new_vma->vm_end = addr + len;
2220                         new_vma->vm_pgoff = pgoff;
2221                         if (new_vma->vm_file) {
2222                                 get_file(new_vma->vm_file);
2223                                 if (vma->vm_flags & VM_EXECUTABLE)
2224                                         added_exe_file_vma(mm);
2225                         }
2226                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2227                                 new_vma->vm_ops->open(new_vma);
2228                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2229                 }
2230         }
2231         return new_vma;
2232 }
2233
2234 /*
2235  * Return true if the calling process may expand its vm space by the passed
2236  * number of pages
2237  */
2238 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2239 {
2240         unsigned long cur = mm->total_vm;       /* pages */
2241         unsigned long lim;
2242
2243         lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2244
2245         if (cur + npages > lim)
2246                 return 0;
2247         return 1;
2248 }
2249
2250
2251 static int special_mapping_fault(struct vm_area_struct *vma,
2252                                 struct vm_fault *vmf)
2253 {
2254         pgoff_t pgoff;
2255         struct page **pages;
2256
2257         /*
2258          * special mappings have no vm_file, and in that case, the mm
2259          * uses vm_pgoff internally. So we have to subtract it from here.
2260          * We are allowed to do this because we are the mm; do not copy
2261          * this code into drivers!
2262          */
2263         pgoff = vmf->pgoff - vma->vm_pgoff;
2264
2265         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2266                 pgoff--;
2267
2268         if (*pages) {
2269                 struct page *page = *pages;
2270                 get_page(page);
2271                 vmf->page = page;
2272                 return 0;
2273         }
2274
2275         return VM_FAULT_SIGBUS;
2276 }
2277
2278 /*
2279  * Having a close hook prevents vma merging regardless of flags.
2280  */
2281 static void special_mapping_close(struct vm_area_struct *vma)
2282 {
2283 }
2284
2285 static const struct vm_operations_struct special_mapping_vmops = {
2286         .close = special_mapping_close,
2287         .fault = special_mapping_fault,
2288 };
2289
2290 /*
2291  * Called with mm->mmap_sem held for writing.
2292  * Insert a new vma covering the given region, with the given flags.
2293  * Its pages are supplied by the given array of struct page *.
2294  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2295  * The region past the last page supplied will always produce SIGBUS.
2296  * The array pointer and the pages it points to are assumed to stay alive
2297  * for as long as this mapping might exist.
2298  */
2299 int install_special_mapping(struct mm_struct *mm,
2300                             unsigned long addr, unsigned long len,
2301                             unsigned long vm_flags, struct page **pages)
2302 {
2303         struct vm_area_struct *vma;
2304
2305         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2306         if (unlikely(vma == NULL))
2307                 return -ENOMEM;
2308
2309         vma->vm_mm = mm;
2310         vma->vm_start = addr;
2311         vma->vm_end = addr + len;
2312
2313         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2314         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2315
2316         vma->vm_ops = &special_mapping_vmops;
2317         vma->vm_private_data = pages;
2318
2319         if (unlikely(insert_vm_struct(mm, vma))) {
2320                 kmem_cache_free(vm_area_cachep, vma);
2321                 return -ENOMEM;
2322         }
2323
2324         mm->total_vm += len >> PAGE_SHIFT;
2325
2326         perf_event_mmap(vma);
2327
2328         return 0;
2329 }
2330
2331 static DEFINE_MUTEX(mm_all_locks_mutex);
2332
2333 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2334 {
2335         if (!test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2336                 /*
2337                  * The LSB of head.next can't change from under us
2338                  * because we hold the mm_all_locks_mutex.
2339                  */
2340                 spin_lock_nest_lock(&anon_vma->lock, &mm->mmap_sem);
2341                 /*
2342                  * We can safely modify head.next after taking the
2343                  * anon_vma->lock. If some other vma in this mm shares
2344                  * the same anon_vma we won't take it again.
2345                  *
2346                  * No need of atomic instructions here, head.next
2347                  * can't change from under us thanks to the
2348                  * anon_vma->lock.
2349                  */
2350                 if (__test_and_set_bit(0, (unsigned long *)
2351                                        &anon_vma->head.next))
2352                         BUG();
2353         }
2354 }
2355
2356 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2357 {
2358         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2359                 /*
2360                  * AS_MM_ALL_LOCKS can't change from under us because
2361                  * we hold the mm_all_locks_mutex.
2362                  *
2363                  * Operations on ->flags have to be atomic because
2364                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2365                  * mm_all_locks_mutex, there may be other cpus
2366                  * changing other bitflags in parallel to us.
2367                  */
2368                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2369                         BUG();
2370                 spin_lock_nest_lock(&mapping->i_mmap_lock, &mm->mmap_sem);
2371         }
2372 }
2373
2374 /*
2375  * This operation locks against the VM for all pte/vma/mm related
2376  * operations that could ever happen on a certain mm. This includes
2377  * vmtruncate, try_to_unmap, and all page faults.
2378  *
2379  * The caller must take the mmap_sem in write mode before calling
2380  * mm_take_all_locks(). The caller isn't allowed to release the
2381  * mmap_sem until mm_drop_all_locks() returns.
2382  *
2383  * mmap_sem in write mode is required in order to block all operations
2384  * that could modify pagetables and free pages without need of
2385  * altering the vma layout (for example populate_range() with
2386  * nonlinear vmas). It's also needed in write mode to avoid new
2387  * anon_vmas to be associated with existing vmas.
2388  *
2389  * A single task can't take more than one mm_take_all_locks() in a row
2390  * or it would deadlock.
2391  *
2392  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2393  * mapping->flags avoid to take the same lock twice, if more than one
2394  * vma in this mm is backed by the same anon_vma or address_space.
2395  *
2396  * We can take all the locks in random order because the VM code
2397  * taking i_mmap_lock or anon_vma->lock outside the mmap_sem never
2398  * takes more than one of them in a row. Secondly we're protected
2399  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2400  *
2401  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2402  * that may have to take thousand of locks.
2403  *
2404  * mm_take_all_locks() can fail if it's interrupted by signals.
2405  */
2406 int mm_take_all_locks(struct mm_struct *mm)
2407 {
2408         struct vm_area_struct *vma;
2409         int ret = -EINTR;
2410
2411         BUG_ON(down_read_trylock(&mm->mmap_sem));
2412
2413         mutex_lock(&mm_all_locks_mutex);
2414
2415         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2416                 if (signal_pending(current))
2417                         goto out_unlock;
2418                 if (vma->vm_file && vma->vm_file->f_mapping)
2419                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2420         }
2421
2422         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2423                 if (signal_pending(current))
2424                         goto out_unlock;
2425                 if (vma->anon_vma)
2426                         vm_lock_anon_vma(mm, vma->anon_vma);
2427         }
2428
2429         ret = 0;
2430
2431 out_unlock:
2432         if (ret)
2433                 mm_drop_all_locks(mm);
2434
2435         return ret;
2436 }
2437
2438 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2439 {
2440         if (test_bit(0, (unsigned long *) &anon_vma->head.next)) {
2441                 /*
2442                  * The LSB of head.next can't change to 0 from under
2443                  * us because we hold the mm_all_locks_mutex.
2444                  *
2445                  * We must however clear the bitflag before unlocking
2446                  * the vma so the users using the anon_vma->head will
2447                  * never see our bitflag.
2448                  *
2449                  * No need of atomic instructions here, head.next
2450                  * can't change from under us until we release the
2451                  * anon_vma->lock.
2452                  */
2453                 if (!__test_and_clear_bit(0, (unsigned long *)
2454                                           &anon_vma->head.next))
2455                         BUG();
2456                 spin_unlock(&anon_vma->lock);
2457         }
2458 }
2459
2460 static void vm_unlock_mapping(struct address_space *mapping)
2461 {
2462         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2463                 /*
2464                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2465                  * because we hold the mm_all_locks_mutex.
2466                  */
2467                 spin_unlock(&mapping->i_mmap_lock);
2468                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2469                                         &mapping->flags))
2470                         BUG();
2471         }
2472 }
2473
2474 /*
2475  * The mmap_sem cannot be released by the caller until
2476  * mm_drop_all_locks() returns.
2477  */
2478 void mm_drop_all_locks(struct mm_struct *mm)
2479 {
2480         struct vm_area_struct *vma;
2481
2482         BUG_ON(down_read_trylock(&mm->mmap_sem));
2483         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2484
2485         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2486                 if (vma->anon_vma)
2487                         vm_unlock_anon_vma(vma->anon_vma);
2488                 if (vma->vm_file && vma->vm_file->f_mapping)
2489                         vm_unlock_mapping(vma->vm_file->f_mapping);
2490         }
2491
2492         mutex_unlock(&mm_all_locks_mutex);
2493 }
2494
2495 /*
2496  * initialise the VMA slab
2497  */
2498 void __init mmap_init(void)
2499 {
2500         int ret;
2501
2502         ret = percpu_counter_init(&vm_committed_as, 0);
2503         VM_BUG_ON(ret);
2504 }