b38b47ef1f776339670aa911cf65bfba22a6e7a5
[linux-3.10.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/export.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33
34 #include <asm/uaccess.h>
35 #include <asm/cacheflush.h>
36 #include <asm/tlb.h>
37 #include <asm/mmu_context.h>
38
39 #include "internal.h"
40
41 #ifndef arch_mmap_check
42 #define arch_mmap_check(addr, len, flags)       (0)
43 #endif
44
45 #ifndef arch_rebalance_pgtables
46 #define arch_rebalance_pgtables(addr, len)              (addr)
47 #endif
48
49 static void unmap_region(struct mm_struct *mm,
50                 struct vm_area_struct *vma, struct vm_area_struct *prev,
51                 unsigned long start, unsigned long end);
52
53 /*
54  * WARNING: the debugging will use recursive algorithms so never enable this
55  * unless you know what you are doing.
56  */
57 #undef DEBUG_MM_RB
58
59 /* description of effects of mapping type and prot in current implementation.
60  * this is due to the limited x86 page protection hardware.  The expected
61  * behavior is in parens:
62  *
63  * map_type     prot
64  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
65  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *              
69  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
70  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
71  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
72  *
73  */
74 pgprot_t protection_map[16] = {
75         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
76         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
77 };
78
79 pgprot_t vm_get_page_prot(unsigned long vm_flags)
80 {
81         return __pgprot(pgprot_val(protection_map[vm_flags &
82                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
83                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
84 }
85 EXPORT_SYMBOL(vm_get_page_prot);
86
87 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
88 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
89 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
90 /*
91  * Make sure vm_committed_as in one cacheline and not cacheline shared with
92  * other variables. It can be updated by several CPUs frequently.
93  */
94 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
95
96 /*
97  * Check that a process has enough memory to allocate a new virtual
98  * mapping. 0 means there is enough memory for the allocation to
99  * succeed and -ENOMEM implies there is not.
100  *
101  * We currently support three overcommit policies, which are set via the
102  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
103  *
104  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
105  * Additional code 2002 Jul 20 by Robert Love.
106  *
107  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
108  *
109  * Note this is a helper function intended to be used by LSMs which
110  * wish to use this logic.
111  */
112 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
113 {
114         unsigned long free, allowed;
115
116         vm_acct_memory(pages);
117
118         /*
119          * Sometimes we want to use more memory than we have
120          */
121         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
122                 return 0;
123
124         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
125                 free = global_page_state(NR_FREE_PAGES);
126                 free += global_page_state(NR_FILE_PAGES);
127
128                 /*
129                  * shmem pages shouldn't be counted as free in this
130                  * case, they can't be purged, only swapped out, and
131                  * that won't affect the overall amount of available
132                  * memory in the system.
133                  */
134                 free -= global_page_state(NR_SHMEM);
135
136                 free += nr_swap_pages;
137
138                 /*
139                  * Any slabs which are created with the
140                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
141                  * which are reclaimable, under pressure.  The dentry
142                  * cache and most inode caches should fall into this
143                  */
144                 free += global_page_state(NR_SLAB_RECLAIMABLE);
145
146                 /*
147                  * Leave reserved pages. The pages are not for anonymous pages.
148                  */
149                 if (free <= totalreserve_pages)
150                         goto error;
151                 else
152                         free -= totalreserve_pages;
153
154                 /*
155                  * Leave the last 3% for root
156                  */
157                 if (!cap_sys_admin)
158                         free -= free / 32;
159
160                 if (free > pages)
161                         return 0;
162
163                 goto error;
164         }
165
166         allowed = (totalram_pages - hugetlb_total_pages())
167                 * sysctl_overcommit_ratio / 100;
168         /*
169          * Leave the last 3% for root
170          */
171         if (!cap_sys_admin)
172                 allowed -= allowed / 32;
173         allowed += total_swap_pages;
174
175         /* Don't let a single process grow too big:
176            leave 3% of the size of this process for other processes */
177         if (mm)
178                 allowed -= mm->total_vm / 32;
179
180         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
181                 return 0;
182 error:
183         vm_unacct_memory(pages);
184
185         return -ENOMEM;
186 }
187
188 /*
189  * Requires inode->i_mapping->i_mmap_mutex
190  */
191 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
192                 struct file *file, struct address_space *mapping)
193 {
194         if (vma->vm_flags & VM_DENYWRITE)
195                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
196         if (vma->vm_flags & VM_SHARED)
197                 mapping->i_mmap_writable--;
198
199         flush_dcache_mmap_lock(mapping);
200         if (unlikely(vma->vm_flags & VM_NONLINEAR))
201                 list_del_init(&vma->shared.vm_set.list);
202         else
203                 vma_prio_tree_remove(vma, &mapping->i_mmap);
204         flush_dcache_mmap_unlock(mapping);
205 }
206
207 /*
208  * Unlink a file-based vm structure from its prio_tree, to hide
209  * vma from rmap and vmtruncate before freeing its page tables.
210  */
211 void unlink_file_vma(struct vm_area_struct *vma)
212 {
213         struct file *file = vma->vm_file;
214
215         if (file) {
216                 struct address_space *mapping = file->f_mapping;
217                 mutex_lock(&mapping->i_mmap_mutex);
218                 __remove_shared_vm_struct(vma, file, mapping);
219                 mutex_unlock(&mapping->i_mmap_mutex);
220         }
221 }
222
223 /*
224  * Close a vm structure and free it, returning the next.
225  */
226 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
227 {
228         struct vm_area_struct *next = vma->vm_next;
229
230         might_sleep();
231         if (vma->vm_ops && vma->vm_ops->close)
232                 vma->vm_ops->close(vma);
233         if (vma->vm_file) {
234                 fput(vma->vm_file);
235                 if (vma->vm_flags & VM_EXECUTABLE)
236                         removed_exe_file_vma(vma->vm_mm);
237         }
238         mpol_put(vma_policy(vma));
239         kmem_cache_free(vm_area_cachep, vma);
240         return next;
241 }
242
243 static unsigned long do_brk(unsigned long addr, unsigned long len);
244
245 SYSCALL_DEFINE1(brk, unsigned long, brk)
246 {
247         unsigned long rlim, retval;
248         unsigned long newbrk, oldbrk;
249         struct mm_struct *mm = current->mm;
250         unsigned long min_brk;
251
252         down_write(&mm->mmap_sem);
253
254 #ifdef CONFIG_COMPAT_BRK
255         /*
256          * CONFIG_COMPAT_BRK can still be overridden by setting
257          * randomize_va_space to 2, which will still cause mm->start_brk
258          * to be arbitrarily shifted
259          */
260         if (current->brk_randomized)
261                 min_brk = mm->start_brk;
262         else
263                 min_brk = mm->end_data;
264 #else
265         min_brk = mm->start_brk;
266 #endif
267         if (brk < min_brk)
268                 goto out;
269
270         /*
271          * Check against rlimit here. If this check is done later after the test
272          * of oldbrk with newbrk then it can escape the test and let the data
273          * segment grow beyond its set limit the in case where the limit is
274          * not page aligned -Ram Gupta
275          */
276         rlim = rlimit(RLIMIT_DATA);
277         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
278                         (mm->end_data - mm->start_data) > rlim)
279                 goto out;
280
281         newbrk = PAGE_ALIGN(brk);
282         oldbrk = PAGE_ALIGN(mm->brk);
283         if (oldbrk == newbrk)
284                 goto set_brk;
285
286         /* Always allow shrinking brk. */
287         if (brk <= mm->brk) {
288                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
289                         goto set_brk;
290                 goto out;
291         }
292
293         /* Check against existing mmap mappings. */
294         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
295                 goto out;
296
297         /* Ok, looks good - let it rip. */
298         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
299                 goto out;
300 set_brk:
301         mm->brk = brk;
302 out:
303         retval = mm->brk;
304         up_write(&mm->mmap_sem);
305         return retval;
306 }
307
308 #ifdef DEBUG_MM_RB
309 static int browse_rb(struct rb_root *root)
310 {
311         int i = 0, j;
312         struct rb_node *nd, *pn = NULL;
313         unsigned long prev = 0, pend = 0;
314
315         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
316                 struct vm_area_struct *vma;
317                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
318                 if (vma->vm_start < prev)
319                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
320                 if (vma->vm_start < pend)
321                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
322                 if (vma->vm_start > vma->vm_end)
323                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
324                 i++;
325                 pn = nd;
326                 prev = vma->vm_start;
327                 pend = vma->vm_end;
328         }
329         j = 0;
330         for (nd = pn; nd; nd = rb_prev(nd)) {
331                 j++;
332         }
333         if (i != j)
334                 printk("backwards %d, forwards %d\n", j, i), i = 0;
335         return i;
336 }
337
338 void validate_mm(struct mm_struct *mm)
339 {
340         int bug = 0;
341         int i = 0;
342         struct vm_area_struct *tmp = mm->mmap;
343         while (tmp) {
344                 tmp = tmp->vm_next;
345                 i++;
346         }
347         if (i != mm->map_count)
348                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
349         i = browse_rb(&mm->mm_rb);
350         if (i != mm->map_count)
351                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
352         BUG_ON(bug);
353 }
354 #else
355 #define validate_mm(mm) do { } while (0)
356 #endif
357
358 static struct vm_area_struct *
359 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
360                 struct vm_area_struct **pprev, struct rb_node ***rb_link,
361                 struct rb_node ** rb_parent)
362 {
363         struct vm_area_struct * vma;
364         struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
365
366         __rb_link = &mm->mm_rb.rb_node;
367         rb_prev = __rb_parent = NULL;
368         vma = NULL;
369
370         while (*__rb_link) {
371                 struct vm_area_struct *vma_tmp;
372
373                 __rb_parent = *__rb_link;
374                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
375
376                 if (vma_tmp->vm_end > addr) {
377                         vma = vma_tmp;
378                         if (vma_tmp->vm_start <= addr)
379                                 break;
380                         __rb_link = &__rb_parent->rb_left;
381                 } else {
382                         rb_prev = __rb_parent;
383                         __rb_link = &__rb_parent->rb_right;
384                 }
385         }
386
387         *pprev = NULL;
388         if (rb_prev)
389                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
390         *rb_link = __rb_link;
391         *rb_parent = __rb_parent;
392         return vma;
393 }
394
395 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
396                 struct rb_node **rb_link, struct rb_node *rb_parent)
397 {
398         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
399         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
400 }
401
402 static void __vma_link_file(struct vm_area_struct *vma)
403 {
404         struct file *file;
405
406         file = vma->vm_file;
407         if (file) {
408                 struct address_space *mapping = file->f_mapping;
409
410                 if (vma->vm_flags & VM_DENYWRITE)
411                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
412                 if (vma->vm_flags & VM_SHARED)
413                         mapping->i_mmap_writable++;
414
415                 flush_dcache_mmap_lock(mapping);
416                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
417                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
418                 else
419                         vma_prio_tree_insert(vma, &mapping->i_mmap);
420                 flush_dcache_mmap_unlock(mapping);
421         }
422 }
423
424 static void
425 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
426         struct vm_area_struct *prev, struct rb_node **rb_link,
427         struct rb_node *rb_parent)
428 {
429         __vma_link_list(mm, vma, prev, rb_parent);
430         __vma_link_rb(mm, vma, rb_link, rb_parent);
431 }
432
433 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
434                         struct vm_area_struct *prev, struct rb_node **rb_link,
435                         struct rb_node *rb_parent)
436 {
437         struct address_space *mapping = NULL;
438
439         if (vma->vm_file)
440                 mapping = vma->vm_file->f_mapping;
441
442         if (mapping)
443                 mutex_lock(&mapping->i_mmap_mutex);
444
445         __vma_link(mm, vma, prev, rb_link, rb_parent);
446         __vma_link_file(vma);
447
448         if (mapping)
449                 mutex_unlock(&mapping->i_mmap_mutex);
450
451         mm->map_count++;
452         validate_mm(mm);
453 }
454
455 /*
456  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
457  * mm's list and rbtree.  It has already been inserted into the prio_tree.
458  */
459 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
460 {
461         struct vm_area_struct *__vma, *prev;
462         struct rb_node **rb_link, *rb_parent;
463
464         __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
465         BUG_ON(__vma && __vma->vm_start < vma->vm_end);
466         __vma_link(mm, vma, prev, rb_link, rb_parent);
467         mm->map_count++;
468 }
469
470 static inline void
471 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
472                 struct vm_area_struct *prev)
473 {
474         struct vm_area_struct *next = vma->vm_next;
475
476         prev->vm_next = next;
477         if (next)
478                 next->vm_prev = prev;
479         rb_erase(&vma->vm_rb, &mm->mm_rb);
480         if (mm->mmap_cache == vma)
481                 mm->mmap_cache = prev;
482 }
483
484 /*
485  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
486  * is already present in an i_mmap tree without adjusting the tree.
487  * The following helper function should be used when such adjustments
488  * are necessary.  The "insert" vma (if any) is to be inserted
489  * before we drop the necessary locks.
490  */
491 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
492         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
493 {
494         struct mm_struct *mm = vma->vm_mm;
495         struct vm_area_struct *next = vma->vm_next;
496         struct vm_area_struct *importer = NULL;
497         struct address_space *mapping = NULL;
498         struct prio_tree_root *root = NULL;
499         struct anon_vma *anon_vma = NULL;
500         struct file *file = vma->vm_file;
501         long adjust_next = 0;
502         int remove_next = 0;
503
504         if (next && !insert) {
505                 struct vm_area_struct *exporter = NULL;
506
507                 if (end >= next->vm_end) {
508                         /*
509                          * vma expands, overlapping all the next, and
510                          * perhaps the one after too (mprotect case 6).
511                          */
512 again:                  remove_next = 1 + (end > next->vm_end);
513                         end = next->vm_end;
514                         exporter = next;
515                         importer = vma;
516                 } else if (end > next->vm_start) {
517                         /*
518                          * vma expands, overlapping part of the next:
519                          * mprotect case 5 shifting the boundary up.
520                          */
521                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
522                         exporter = next;
523                         importer = vma;
524                 } else if (end < vma->vm_end) {
525                         /*
526                          * vma shrinks, and !insert tells it's not
527                          * split_vma inserting another: so it must be
528                          * mprotect case 4 shifting the boundary down.
529                          */
530                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
531                         exporter = vma;
532                         importer = next;
533                 }
534
535                 /*
536                  * Easily overlooked: when mprotect shifts the boundary,
537                  * make sure the expanding vma has anon_vma set if the
538                  * shrinking vma had, to cover any anon pages imported.
539                  */
540                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
541                         if (anon_vma_clone(importer, exporter))
542                                 return -ENOMEM;
543                         importer->anon_vma = exporter->anon_vma;
544                 }
545         }
546
547         if (file) {
548                 mapping = file->f_mapping;
549                 if (!(vma->vm_flags & VM_NONLINEAR))
550                         root = &mapping->i_mmap;
551                 mutex_lock(&mapping->i_mmap_mutex);
552                 if (insert) {
553                         /*
554                          * Put into prio_tree now, so instantiated pages
555                          * are visible to arm/parisc __flush_dcache_page
556                          * throughout; but we cannot insert into address
557                          * space until vma start or end is updated.
558                          */
559                         __vma_link_file(insert);
560                 }
561         }
562
563         vma_adjust_trans_huge(vma, start, end, adjust_next);
564
565         /*
566          * When changing only vma->vm_end, we don't really need anon_vma
567          * lock. This is a fairly rare case by itself, but the anon_vma
568          * lock may be shared between many sibling processes.  Skipping
569          * the lock for brk adjustments makes a difference sometimes.
570          */
571         if (vma->anon_vma && (importer || start != vma->vm_start)) {
572                 anon_vma = vma->anon_vma;
573                 anon_vma_lock(anon_vma);
574         }
575
576         if (root) {
577                 flush_dcache_mmap_lock(mapping);
578                 vma_prio_tree_remove(vma, root);
579                 if (adjust_next)
580                         vma_prio_tree_remove(next, root);
581         }
582
583         vma->vm_start = start;
584         vma->vm_end = end;
585         vma->vm_pgoff = pgoff;
586         if (adjust_next) {
587                 next->vm_start += adjust_next << PAGE_SHIFT;
588                 next->vm_pgoff += adjust_next;
589         }
590
591         if (root) {
592                 if (adjust_next)
593                         vma_prio_tree_insert(next, root);
594                 vma_prio_tree_insert(vma, root);
595                 flush_dcache_mmap_unlock(mapping);
596         }
597
598         if (remove_next) {
599                 /*
600                  * vma_merge has merged next into vma, and needs
601                  * us to remove next before dropping the locks.
602                  */
603                 __vma_unlink(mm, next, vma);
604                 if (file)
605                         __remove_shared_vm_struct(next, file, mapping);
606         } else if (insert) {
607                 /*
608                  * split_vma has split insert from vma, and needs
609                  * us to insert it before dropping the locks
610                  * (it may either follow vma or precede it).
611                  */
612                 __insert_vm_struct(mm, insert);
613         }
614
615         if (anon_vma)
616                 anon_vma_unlock(anon_vma);
617         if (mapping)
618                 mutex_unlock(&mapping->i_mmap_mutex);
619
620         if (remove_next) {
621                 if (file) {
622                         fput(file);
623                         if (next->vm_flags & VM_EXECUTABLE)
624                                 removed_exe_file_vma(mm);
625                 }
626                 if (next->anon_vma)
627                         anon_vma_merge(vma, next);
628                 mm->map_count--;
629                 mpol_put(vma_policy(next));
630                 kmem_cache_free(vm_area_cachep, next);
631                 /*
632                  * In mprotect's case 6 (see comments on vma_merge),
633                  * we must remove another next too. It would clutter
634                  * up the code too much to do both in one go.
635                  */
636                 if (remove_next == 2) {
637                         next = vma->vm_next;
638                         goto again;
639                 }
640         }
641
642         validate_mm(mm);
643
644         return 0;
645 }
646
647 /*
648  * If the vma has a ->close operation then the driver probably needs to release
649  * per-vma resources, so we don't attempt to merge those.
650  */
651 static inline int is_mergeable_vma(struct vm_area_struct *vma,
652                         struct file *file, unsigned long vm_flags)
653 {
654         /* VM_CAN_NONLINEAR may get set later by f_op->mmap() */
655         if ((vma->vm_flags ^ vm_flags) & ~VM_CAN_NONLINEAR)
656                 return 0;
657         if (vma->vm_file != file)
658                 return 0;
659         if (vma->vm_ops && vma->vm_ops->close)
660                 return 0;
661         return 1;
662 }
663
664 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
665                                         struct anon_vma *anon_vma2,
666                                         struct vm_area_struct *vma)
667 {
668         /*
669          * The list_is_singular() test is to avoid merging VMA cloned from
670          * parents. This can improve scalability caused by anon_vma lock.
671          */
672         if ((!anon_vma1 || !anon_vma2) && (!vma ||
673                 list_is_singular(&vma->anon_vma_chain)))
674                 return 1;
675         return anon_vma1 == anon_vma2;
676 }
677
678 /*
679  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
680  * in front of (at a lower virtual address and file offset than) the vma.
681  *
682  * We cannot merge two vmas if they have differently assigned (non-NULL)
683  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
684  *
685  * We don't check here for the merged mmap wrapping around the end of pagecache
686  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
687  * wrap, nor mmaps which cover the final page at index -1UL.
688  */
689 static int
690 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
691         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
692 {
693         if (is_mergeable_vma(vma, file, vm_flags) &&
694             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
695                 if (vma->vm_pgoff == vm_pgoff)
696                         return 1;
697         }
698         return 0;
699 }
700
701 /*
702  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
703  * beyond (at a higher virtual address and file offset than) the vma.
704  *
705  * We cannot merge two vmas if they have differently assigned (non-NULL)
706  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
707  */
708 static int
709 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
710         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
711 {
712         if (is_mergeable_vma(vma, file, vm_flags) &&
713             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
714                 pgoff_t vm_pglen;
715                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
716                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
717                         return 1;
718         }
719         return 0;
720 }
721
722 /*
723  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
724  * whether that can be merged with its predecessor or its successor.
725  * Or both (it neatly fills a hole).
726  *
727  * In most cases - when called for mmap, brk or mremap - [addr,end) is
728  * certain not to be mapped by the time vma_merge is called; but when
729  * called for mprotect, it is certain to be already mapped (either at
730  * an offset within prev, or at the start of next), and the flags of
731  * this area are about to be changed to vm_flags - and the no-change
732  * case has already been eliminated.
733  *
734  * The following mprotect cases have to be considered, where AAAA is
735  * the area passed down from mprotect_fixup, never extending beyond one
736  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
737  *
738  *     AAAA             AAAA                AAAA          AAAA
739  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
740  *    cannot merge    might become    might become    might become
741  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
742  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
743  *    mremap move:                                    PPPPNNNNNNNN 8
744  *        AAAA
745  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
746  *    might become    case 1 below    case 2 below    case 3 below
747  *
748  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
749  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
750  */
751 struct vm_area_struct *vma_merge(struct mm_struct *mm,
752                         struct vm_area_struct *prev, unsigned long addr,
753                         unsigned long end, unsigned long vm_flags,
754                         struct anon_vma *anon_vma, struct file *file,
755                         pgoff_t pgoff, struct mempolicy *policy)
756 {
757         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
758         struct vm_area_struct *area, *next;
759         int err;
760
761         /*
762          * We later require that vma->vm_flags == vm_flags,
763          * so this tests vma->vm_flags & VM_SPECIAL, too.
764          */
765         if (vm_flags & VM_SPECIAL)
766                 return NULL;
767
768         if (prev)
769                 next = prev->vm_next;
770         else
771                 next = mm->mmap;
772         area = next;
773         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
774                 next = next->vm_next;
775
776         /*
777          * Can it merge with the predecessor?
778          */
779         if (prev && prev->vm_end == addr &&
780                         mpol_equal(vma_policy(prev), policy) &&
781                         can_vma_merge_after(prev, vm_flags,
782                                                 anon_vma, file, pgoff)) {
783                 /*
784                  * OK, it can.  Can we now merge in the successor as well?
785                  */
786                 if (next && end == next->vm_start &&
787                                 mpol_equal(policy, vma_policy(next)) &&
788                                 can_vma_merge_before(next, vm_flags,
789                                         anon_vma, file, pgoff+pglen) &&
790                                 is_mergeable_anon_vma(prev->anon_vma,
791                                                       next->anon_vma, NULL)) {
792                                                         /* cases 1, 6 */
793                         err = vma_adjust(prev, prev->vm_start,
794                                 next->vm_end, prev->vm_pgoff, NULL);
795                 } else                                  /* cases 2, 5, 7 */
796                         err = vma_adjust(prev, prev->vm_start,
797                                 end, prev->vm_pgoff, NULL);
798                 if (err)
799                         return NULL;
800                 khugepaged_enter_vma_merge(prev);
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                         err = vma_adjust(prev, prev->vm_start,
813                                 addr, prev->vm_pgoff, NULL);
814                 else                                    /* cases 3, 8 */
815                         err = vma_adjust(area, addr, next->vm_end,
816                                 next->vm_pgoff - pglen, NULL);
817                 if (err)
818                         return NULL;
819                 khugepaged_enter_vma_merge(area);
820                 return area;
821         }
822
823         return NULL;
824 }
825
826 /*
827  * Rough compatbility check to quickly see if it's even worth looking
828  * at sharing an anon_vma.
829  *
830  * They need to have the same vm_file, and the flags can only differ
831  * in things that mprotect may change.
832  *
833  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
834  * we can merge the two vma's. For example, we refuse to merge a vma if
835  * there is a vm_ops->close() function, because that indicates that the
836  * driver is doing some kind of reference counting. But that doesn't
837  * really matter for the anon_vma sharing case.
838  */
839 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
840 {
841         return a->vm_end == b->vm_start &&
842                 mpol_equal(vma_policy(a), vma_policy(b)) &&
843                 a->vm_file == b->vm_file &&
844                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
845                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
846 }
847
848 /*
849  * Do some basic sanity checking to see if we can re-use the anon_vma
850  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
851  * the same as 'old', the other will be the new one that is trying
852  * to share the anon_vma.
853  *
854  * NOTE! This runs with mm_sem held for reading, so it is possible that
855  * the anon_vma of 'old' is concurrently in the process of being set up
856  * by another page fault trying to merge _that_. But that's ok: if it
857  * is being set up, that automatically means that it will be a singleton
858  * acceptable for merging, so we can do all of this optimistically. But
859  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
860  *
861  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
862  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
863  * is to return an anon_vma that is "complex" due to having gone through
864  * a fork).
865  *
866  * We also make sure that the two vma's are compatible (adjacent,
867  * and with the same memory policies). That's all stable, even with just
868  * a read lock on the mm_sem.
869  */
870 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
871 {
872         if (anon_vma_compatible(a, b)) {
873                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
874
875                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
876                         return anon_vma;
877         }
878         return NULL;
879 }
880
881 /*
882  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
883  * neighbouring vmas for a suitable anon_vma, before it goes off
884  * to allocate a new anon_vma.  It checks because a repetitive
885  * sequence of mprotects and faults may otherwise lead to distinct
886  * anon_vmas being allocated, preventing vma merge in subsequent
887  * mprotect.
888  */
889 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
890 {
891         struct anon_vma *anon_vma;
892         struct vm_area_struct *near;
893
894         near = vma->vm_next;
895         if (!near)
896                 goto try_prev;
897
898         anon_vma = reusable_anon_vma(near, vma, near);
899         if (anon_vma)
900                 return anon_vma;
901 try_prev:
902         near = vma->vm_prev;
903         if (!near)
904                 goto none;
905
906         anon_vma = reusable_anon_vma(near, near, vma);
907         if (anon_vma)
908                 return anon_vma;
909 none:
910         /*
911          * There's no absolute need to look only at touching neighbours:
912          * we could search further afield for "compatible" anon_vmas.
913          * But it would probably just be a waste of time searching,
914          * or lead to too many vmas hanging off the same anon_vma.
915          * We're trying to allow mprotect remerging later on,
916          * not trying to minimize memory used for anon_vmas.
917          */
918         return NULL;
919 }
920
921 #ifdef CONFIG_PROC_FS
922 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
923                                                 struct file *file, long pages)
924 {
925         const unsigned long stack_flags
926                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
927
928         if (file) {
929                 mm->shared_vm += pages;
930                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
931                         mm->exec_vm += pages;
932         } else if (flags & stack_flags)
933                 mm->stack_vm += pages;
934         if (flags & (VM_RESERVED|VM_IO))
935                 mm->reserved_vm += pages;
936 }
937 #endif /* CONFIG_PROC_FS */
938
939 /*
940  * If a hint addr is less than mmap_min_addr change hint to be as
941  * low as possible but still greater than mmap_min_addr
942  */
943 static inline unsigned long round_hint_to_min(unsigned long hint)
944 {
945         hint &= PAGE_MASK;
946         if (((void *)hint != NULL) &&
947             (hint < mmap_min_addr))
948                 return PAGE_ALIGN(mmap_min_addr);
949         return hint;
950 }
951
952 /*
953  * The caller must hold down_write(&current->mm->mmap_sem).
954  */
955
956 static unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
957                         unsigned long len, unsigned long prot,
958                         unsigned long flags, unsigned long pgoff)
959 {
960         struct mm_struct * mm = current->mm;
961         struct inode *inode;
962         vm_flags_t vm_flags;
963         int error;
964         unsigned long reqprot = prot;
965
966         /*
967          * Does the application expect PROT_READ to imply PROT_EXEC?
968          *
969          * (the exception is when the underlying filesystem is noexec
970          *  mounted, in which case we dont add PROT_EXEC.)
971          */
972         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
973                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
974                         prot |= PROT_EXEC;
975
976         if (!len)
977                 return -EINVAL;
978
979         if (!(flags & MAP_FIXED))
980                 addr = round_hint_to_min(addr);
981
982         /* Careful about overflows.. */
983         len = PAGE_ALIGN(len);
984         if (!len)
985                 return -ENOMEM;
986
987         /* offset overflow? */
988         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
989                return -EOVERFLOW;
990
991         /* Too many mappings? */
992         if (mm->map_count > sysctl_max_map_count)
993                 return -ENOMEM;
994
995         /* Obtain the address to map to. we verify (or select) it and ensure
996          * that it represents a valid section of the address space.
997          */
998         addr = get_unmapped_area(file, addr, len, pgoff, flags);
999         if (addr & ~PAGE_MASK)
1000                 return addr;
1001
1002         /* Do simple checking here so the lower-level routines won't have
1003          * to. we assume access permissions have been handled by the open
1004          * of the memory object, so we don't do any here.
1005          */
1006         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1007                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1008
1009         if (flags & MAP_LOCKED)
1010                 if (!can_do_mlock())
1011                         return -EPERM;
1012
1013         /* mlock MCL_FUTURE? */
1014         if (vm_flags & VM_LOCKED) {
1015                 unsigned long locked, lock_limit;
1016                 locked = len >> PAGE_SHIFT;
1017                 locked += mm->locked_vm;
1018                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1019                 lock_limit >>= PAGE_SHIFT;
1020                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1021                         return -EAGAIN;
1022         }
1023
1024         inode = file ? file->f_path.dentry->d_inode : NULL;
1025
1026         if (file) {
1027                 switch (flags & MAP_TYPE) {
1028                 case MAP_SHARED:
1029                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1030                                 return -EACCES;
1031
1032                         /*
1033                          * Make sure we don't allow writing to an append-only
1034                          * file..
1035                          */
1036                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1037                                 return -EACCES;
1038
1039                         /*
1040                          * Make sure there are no mandatory locks on the file.
1041                          */
1042                         if (locks_verify_locked(inode))
1043                                 return -EAGAIN;
1044
1045                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1046                         if (!(file->f_mode & FMODE_WRITE))
1047                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1048
1049                         /* fall through */
1050                 case MAP_PRIVATE:
1051                         if (!(file->f_mode & FMODE_READ))
1052                                 return -EACCES;
1053                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1054                                 if (vm_flags & VM_EXEC)
1055                                         return -EPERM;
1056                                 vm_flags &= ~VM_MAYEXEC;
1057                         }
1058
1059                         if (!file->f_op || !file->f_op->mmap)
1060                                 return -ENODEV;
1061                         break;
1062
1063                 default:
1064                         return -EINVAL;
1065                 }
1066         } else {
1067                 switch (flags & MAP_TYPE) {
1068                 case MAP_SHARED:
1069                         /*
1070                          * Ignore pgoff.
1071                          */
1072                         pgoff = 0;
1073                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1074                         break;
1075                 case MAP_PRIVATE:
1076                         /*
1077                          * Set pgoff according to addr for anon_vma.
1078                          */
1079                         pgoff = addr >> PAGE_SHIFT;
1080                         break;
1081                 default:
1082                         return -EINVAL;
1083                 }
1084         }
1085
1086         error = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1087         if (error)
1088                 return error;
1089
1090         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1091 }
1092
1093 unsigned long do_mmap(struct file *file, unsigned long addr,
1094         unsigned long len, unsigned long prot,
1095         unsigned long flag, unsigned long offset)
1096 {
1097         if (unlikely(offset + PAGE_ALIGN(len) < offset))
1098                 return -EINVAL;
1099         if (unlikely(offset & ~PAGE_MASK))
1100                 return -EINVAL;
1101         return do_mmap_pgoff(file, addr, len, prot, flag, offset >> PAGE_SHIFT);
1102 }
1103 EXPORT_SYMBOL(do_mmap);
1104
1105 unsigned long vm_mmap(struct file *file, unsigned long addr,
1106         unsigned long len, unsigned long prot,
1107         unsigned long flag, unsigned long offset)
1108 {
1109         unsigned long ret;
1110         struct mm_struct *mm = current->mm;
1111
1112         down_write(&mm->mmap_sem);
1113         ret = do_mmap(file, addr, len, prot, flag, offset);
1114         up_write(&mm->mmap_sem);
1115         return ret;
1116 }
1117 EXPORT_SYMBOL(vm_mmap);
1118
1119 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1120                 unsigned long, prot, unsigned long, flags,
1121                 unsigned long, fd, unsigned long, pgoff)
1122 {
1123         struct file *file = NULL;
1124         unsigned long retval = -EBADF;
1125
1126         if (!(flags & MAP_ANONYMOUS)) {
1127                 audit_mmap_fd(fd, flags);
1128                 if (unlikely(flags & MAP_HUGETLB))
1129                         return -EINVAL;
1130                 file = fget(fd);
1131                 if (!file)
1132                         goto out;
1133         } else if (flags & MAP_HUGETLB) {
1134                 struct user_struct *user = NULL;
1135                 /*
1136                  * VM_NORESERVE is used because the reservations will be
1137                  * taken when vm_ops->mmap() is called
1138                  * A dummy user value is used because we are not locking
1139                  * memory so no accounting is necessary
1140                  */
1141                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1142                                                 VM_NORESERVE, &user,
1143                                                 HUGETLB_ANONHUGE_INODE);
1144                 if (IS_ERR(file))
1145                         return PTR_ERR(file);
1146         }
1147
1148         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1149
1150         down_write(&current->mm->mmap_sem);
1151         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1152         up_write(&current->mm->mmap_sem);
1153
1154         if (file)
1155                 fput(file);
1156 out:
1157         return retval;
1158 }
1159
1160 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1161 struct mmap_arg_struct {
1162         unsigned long addr;
1163         unsigned long len;
1164         unsigned long prot;
1165         unsigned long flags;
1166         unsigned long fd;
1167         unsigned long offset;
1168 };
1169
1170 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1171 {
1172         struct mmap_arg_struct a;
1173
1174         if (copy_from_user(&a, arg, sizeof(a)))
1175                 return -EFAULT;
1176         if (a.offset & ~PAGE_MASK)
1177                 return -EINVAL;
1178
1179         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1180                               a.offset >> PAGE_SHIFT);
1181 }
1182 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1183
1184 /*
1185  * Some shared mappigns will want the pages marked read-only
1186  * to track write events. If so, we'll downgrade vm_page_prot
1187  * to the private version (using protection_map[] without the
1188  * VM_SHARED bit).
1189  */
1190 int vma_wants_writenotify(struct vm_area_struct *vma)
1191 {
1192         vm_flags_t vm_flags = vma->vm_flags;
1193
1194         /* If it was private or non-writable, the write bit is already clear */
1195         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1196                 return 0;
1197
1198         /* The backer wishes to know when pages are first written to? */
1199         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1200                 return 1;
1201
1202         /* The open routine did something to the protections already? */
1203         if (pgprot_val(vma->vm_page_prot) !=
1204             pgprot_val(vm_get_page_prot(vm_flags)))
1205                 return 0;
1206
1207         /* Specialty mapping? */
1208         if (vm_flags & (VM_PFNMAP|VM_INSERTPAGE))
1209                 return 0;
1210
1211         /* Can the mapping track the dirty pages? */
1212         return vma->vm_file && vma->vm_file->f_mapping &&
1213                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1214 }
1215
1216 /*
1217  * We account for memory if it's a private writeable mapping,
1218  * not hugepages and VM_NORESERVE wasn't set.
1219  */
1220 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1221 {
1222         /*
1223          * hugetlb has its own accounting separate from the core VM
1224          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1225          */
1226         if (file && is_file_hugepages(file))
1227                 return 0;
1228
1229         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1230 }
1231
1232 unsigned long mmap_region(struct file *file, unsigned long addr,
1233                           unsigned long len, unsigned long flags,
1234                           vm_flags_t vm_flags, unsigned long pgoff)
1235 {
1236         struct mm_struct *mm = current->mm;
1237         struct vm_area_struct *vma, *prev;
1238         int correct_wcount = 0;
1239         int error;
1240         struct rb_node **rb_link, *rb_parent;
1241         unsigned long charged = 0;
1242         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1243
1244         /* Clear old maps */
1245         error = -ENOMEM;
1246 munmap_back:
1247         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1248         if (vma && vma->vm_start < addr + len) {
1249                 if (do_munmap(mm, addr, len))
1250                         return -ENOMEM;
1251                 goto munmap_back;
1252         }
1253
1254         /* Check against address space limit. */
1255         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1256                 return -ENOMEM;
1257
1258         /*
1259          * Set 'VM_NORESERVE' if we should not account for the
1260          * memory use of this mapping.
1261          */
1262         if ((flags & MAP_NORESERVE)) {
1263                 /* We honor MAP_NORESERVE if allowed to overcommit */
1264                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1265                         vm_flags |= VM_NORESERVE;
1266
1267                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1268                 if (file && is_file_hugepages(file))
1269                         vm_flags |= VM_NORESERVE;
1270         }
1271
1272         /*
1273          * Private writable mapping: check memory availability
1274          */
1275         if (accountable_mapping(file, vm_flags)) {
1276                 charged = len >> PAGE_SHIFT;
1277                 if (security_vm_enough_memory_mm(mm, charged))
1278                         return -ENOMEM;
1279                 vm_flags |= VM_ACCOUNT;
1280         }
1281
1282         /*
1283          * Can we just expand an old mapping?
1284          */
1285         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1286         if (vma)
1287                 goto out;
1288
1289         /*
1290          * Determine the object being mapped and call the appropriate
1291          * specific mapper. the address has already been validated, but
1292          * not unmapped, but the maps are removed from the list.
1293          */
1294         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1295         if (!vma) {
1296                 error = -ENOMEM;
1297                 goto unacct_error;
1298         }
1299
1300         vma->vm_mm = mm;
1301         vma->vm_start = addr;
1302         vma->vm_end = addr + len;
1303         vma->vm_flags = vm_flags;
1304         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1305         vma->vm_pgoff = pgoff;
1306         INIT_LIST_HEAD(&vma->anon_vma_chain);
1307
1308         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1309
1310         if (file) {
1311                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1312                         goto free_vma;
1313                 if (vm_flags & VM_DENYWRITE) {
1314                         error = deny_write_access(file);
1315                         if (error)
1316                                 goto free_vma;
1317                         correct_wcount = 1;
1318                 }
1319                 vma->vm_file = file;
1320                 get_file(file);
1321                 error = file->f_op->mmap(file, vma);
1322                 if (error)
1323                         goto unmap_and_free_vma;
1324                 if (vm_flags & VM_EXECUTABLE)
1325                         added_exe_file_vma(mm);
1326
1327                 /* Can addr have changed??
1328                  *
1329                  * Answer: Yes, several device drivers can do it in their
1330                  *         f_op->mmap method. -DaveM
1331                  */
1332                 addr = vma->vm_start;
1333                 pgoff = vma->vm_pgoff;
1334                 vm_flags = vma->vm_flags;
1335         } else if (vm_flags & VM_SHARED) {
1336                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1337                         goto free_vma;
1338                 error = shmem_zero_setup(vma);
1339                 if (error)
1340                         goto free_vma;
1341         }
1342
1343         if (vma_wants_writenotify(vma)) {
1344                 pgprot_t pprot = vma->vm_page_prot;
1345
1346                 /* Can vma->vm_page_prot have changed??
1347                  *
1348                  * Answer: Yes, drivers may have changed it in their
1349                  *         f_op->mmap method.
1350                  *
1351                  * Ensures that vmas marked as uncached stay that way.
1352                  */
1353                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1354                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1355                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1356         }
1357
1358         vma_link(mm, vma, prev, rb_link, rb_parent);
1359         file = vma->vm_file;
1360
1361         /* Once vma denies write, undo our temporary denial count */
1362         if (correct_wcount)
1363                 atomic_inc(&inode->i_writecount);
1364 out:
1365         perf_event_mmap(vma);
1366
1367         mm->total_vm += len >> PAGE_SHIFT;
1368         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1369         if (vm_flags & VM_LOCKED) {
1370                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1371                         mm->locked_vm += (len >> PAGE_SHIFT);
1372         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1373                 make_pages_present(addr, addr + len);
1374         return addr;
1375
1376 unmap_and_free_vma:
1377         if (correct_wcount)
1378                 atomic_inc(&inode->i_writecount);
1379         vma->vm_file = NULL;
1380         fput(file);
1381
1382         /* Undo any partial mapping done by a device driver. */
1383         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1384         charged = 0;
1385 free_vma:
1386         kmem_cache_free(vm_area_cachep, vma);
1387 unacct_error:
1388         if (charged)
1389                 vm_unacct_memory(charged);
1390         return error;
1391 }
1392
1393 /* Get an address range which is currently unmapped.
1394  * For shmat() with addr=0.
1395  *
1396  * Ugly calling convention alert:
1397  * Return value with the low bits set means error value,
1398  * ie
1399  *      if (ret & ~PAGE_MASK)
1400  *              error = ret;
1401  *
1402  * This function "knows" that -ENOMEM has the bits set.
1403  */
1404 #ifndef HAVE_ARCH_UNMAPPED_AREA
1405 unsigned long
1406 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1407                 unsigned long len, unsigned long pgoff, unsigned long flags)
1408 {
1409         struct mm_struct *mm = current->mm;
1410         struct vm_area_struct *vma;
1411         unsigned long start_addr;
1412
1413         if (len > TASK_SIZE)
1414                 return -ENOMEM;
1415
1416         if (flags & MAP_FIXED)
1417                 return addr;
1418
1419         if (addr) {
1420                 addr = PAGE_ALIGN(addr);
1421                 vma = find_vma(mm, addr);
1422                 if (TASK_SIZE - len >= addr &&
1423                     (!vma || addr + len <= vma->vm_start))
1424                         return addr;
1425         }
1426         if (len > mm->cached_hole_size) {
1427                 start_addr = addr = mm->free_area_cache;
1428         } else {
1429                 start_addr = addr = TASK_UNMAPPED_BASE;
1430                 mm->cached_hole_size = 0;
1431         }
1432
1433 full_search:
1434         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1435                 /* At this point:  (!vma || addr < vma->vm_end). */
1436                 if (TASK_SIZE - len < addr) {
1437                         /*
1438                          * Start a new search - just in case we missed
1439                          * some holes.
1440                          */
1441                         if (start_addr != TASK_UNMAPPED_BASE) {
1442                                 addr = TASK_UNMAPPED_BASE;
1443                                 start_addr = addr;
1444                                 mm->cached_hole_size = 0;
1445                                 goto full_search;
1446                         }
1447                         return -ENOMEM;
1448                 }
1449                 if (!vma || addr + len <= vma->vm_start) {
1450                         /*
1451                          * Remember the place where we stopped the search:
1452                          */
1453                         mm->free_area_cache = addr + len;
1454                         return addr;
1455                 }
1456                 if (addr + mm->cached_hole_size < vma->vm_start)
1457                         mm->cached_hole_size = vma->vm_start - addr;
1458                 addr = vma->vm_end;
1459         }
1460 }
1461 #endif  
1462
1463 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1464 {
1465         /*
1466          * Is this a new hole at the lowest possible address?
1467          */
1468         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1469                 mm->free_area_cache = addr;
1470 }
1471
1472 /*
1473  * This mmap-allocator allocates new areas top-down from below the
1474  * stack's low limit (the base):
1475  */
1476 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1477 unsigned long
1478 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1479                           const unsigned long len, const unsigned long pgoff,
1480                           const unsigned long flags)
1481 {
1482         struct vm_area_struct *vma;
1483         struct mm_struct *mm = current->mm;
1484         unsigned long addr = addr0, start_addr;
1485
1486         /* requested length too big for entire address space */
1487         if (len > TASK_SIZE)
1488                 return -ENOMEM;
1489
1490         if (flags & MAP_FIXED)
1491                 return addr;
1492
1493         /* requesting a specific address */
1494         if (addr) {
1495                 addr = PAGE_ALIGN(addr);
1496                 vma = find_vma(mm, addr);
1497                 if (TASK_SIZE - len >= addr &&
1498                                 (!vma || addr + len <= vma->vm_start))
1499                         return addr;
1500         }
1501
1502         /* check if free_area_cache is useful for us */
1503         if (len <= mm->cached_hole_size) {
1504                 mm->cached_hole_size = 0;
1505                 mm->free_area_cache = mm->mmap_base;
1506         }
1507
1508 try_again:
1509         /* either no address requested or can't fit in requested address hole */
1510         start_addr = addr = mm->free_area_cache;
1511
1512         if (addr < len)
1513                 goto fail;
1514
1515         addr -= len;
1516         do {
1517                 /*
1518                  * Lookup failure means no vma is above this address,
1519                  * else if new region fits below vma->vm_start,
1520                  * return with success:
1521                  */
1522                 vma = find_vma(mm, addr);
1523                 if (!vma || addr+len <= vma->vm_start)
1524                         /* remember the address as a hint for next time */
1525                         return (mm->free_area_cache = addr);
1526
1527                 /* remember the largest hole we saw so far */
1528                 if (addr + mm->cached_hole_size < vma->vm_start)
1529                         mm->cached_hole_size = vma->vm_start - addr;
1530
1531                 /* try just below the current vma->vm_start */
1532                 addr = vma->vm_start-len;
1533         } while (len < vma->vm_start);
1534
1535 fail:
1536         /*
1537          * if hint left us with no space for the requested
1538          * mapping then try again:
1539          *
1540          * Note: this is different with the case of bottomup
1541          * which does the fully line-search, but we use find_vma
1542          * here that causes some holes skipped.
1543          */
1544         if (start_addr != mm->mmap_base) {
1545                 mm->free_area_cache = mm->mmap_base;
1546                 mm->cached_hole_size = 0;
1547                 goto try_again;
1548         }
1549
1550         /*
1551          * A failed mmap() very likely causes application failure,
1552          * so fall back to the bottom-up function here. This scenario
1553          * can happen with large stack limits and large mmap()
1554          * allocations.
1555          */
1556         mm->cached_hole_size = ~0UL;
1557         mm->free_area_cache = TASK_UNMAPPED_BASE;
1558         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1559         /*
1560          * Restore the topdown base:
1561          */
1562         mm->free_area_cache = mm->mmap_base;
1563         mm->cached_hole_size = ~0UL;
1564
1565         return addr;
1566 }
1567 #endif
1568
1569 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1570 {
1571         /*
1572          * Is this a new hole at the highest possible address?
1573          */
1574         if (addr > mm->free_area_cache)
1575                 mm->free_area_cache = addr;
1576
1577         /* dont allow allocations above current base */
1578         if (mm->free_area_cache > mm->mmap_base)
1579                 mm->free_area_cache = mm->mmap_base;
1580 }
1581
1582 unsigned long
1583 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1584                 unsigned long pgoff, unsigned long flags)
1585 {
1586         unsigned long (*get_area)(struct file *, unsigned long,
1587                                   unsigned long, unsigned long, unsigned long);
1588
1589         unsigned long error = arch_mmap_check(addr, len, flags);
1590         if (error)
1591                 return error;
1592
1593         /* Careful about overflows.. */
1594         if (len > TASK_SIZE)
1595                 return -ENOMEM;
1596
1597         get_area = current->mm->get_unmapped_area;
1598         if (file && file->f_op && file->f_op->get_unmapped_area)
1599                 get_area = file->f_op->get_unmapped_area;
1600         addr = get_area(file, addr, len, pgoff, flags);
1601         if (IS_ERR_VALUE(addr))
1602                 return addr;
1603
1604         if (addr > TASK_SIZE - len)
1605                 return -ENOMEM;
1606         if (addr & ~PAGE_MASK)
1607                 return -EINVAL;
1608
1609         return arch_rebalance_pgtables(addr, len);
1610 }
1611
1612 EXPORT_SYMBOL(get_unmapped_area);
1613
1614 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1615 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1616 {
1617         struct vm_area_struct *vma = NULL;
1618
1619         if (mm) {
1620                 /* Check the cache first. */
1621                 /* (Cache hit rate is typically around 35%.) */
1622                 vma = mm->mmap_cache;
1623                 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1624                         struct rb_node * rb_node;
1625
1626                         rb_node = mm->mm_rb.rb_node;
1627                         vma = NULL;
1628
1629                         while (rb_node) {
1630                                 struct vm_area_struct * vma_tmp;
1631
1632                                 vma_tmp = rb_entry(rb_node,
1633                                                 struct vm_area_struct, vm_rb);
1634
1635                                 if (vma_tmp->vm_end > addr) {
1636                                         vma = vma_tmp;
1637                                         if (vma_tmp->vm_start <= addr)
1638                                                 break;
1639                                         rb_node = rb_node->rb_left;
1640                                 } else
1641                                         rb_node = rb_node->rb_right;
1642                         }
1643                         if (vma)
1644                                 mm->mmap_cache = vma;
1645                 }
1646         }
1647         return vma;
1648 }
1649
1650 EXPORT_SYMBOL(find_vma);
1651
1652 /*
1653  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1654  */
1655 struct vm_area_struct *
1656 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1657                         struct vm_area_struct **pprev)
1658 {
1659         struct vm_area_struct *vma;
1660
1661         vma = find_vma(mm, addr);
1662         if (vma) {
1663                 *pprev = vma->vm_prev;
1664         } else {
1665                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1666                 *pprev = NULL;
1667                 while (rb_node) {
1668                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1669                         rb_node = rb_node->rb_right;
1670                 }
1671         }
1672         return vma;
1673 }
1674
1675 /*
1676  * Verify that the stack growth is acceptable and
1677  * update accounting. This is shared with both the
1678  * grow-up and grow-down cases.
1679  */
1680 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1681 {
1682         struct mm_struct *mm = vma->vm_mm;
1683         struct rlimit *rlim = current->signal->rlim;
1684         unsigned long new_start;
1685
1686         /* address space limit tests */
1687         if (!may_expand_vm(mm, grow))
1688                 return -ENOMEM;
1689
1690         /* Stack limit test */
1691         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1692                 return -ENOMEM;
1693
1694         /* mlock limit tests */
1695         if (vma->vm_flags & VM_LOCKED) {
1696                 unsigned long locked;
1697                 unsigned long limit;
1698                 locked = mm->locked_vm + grow;
1699                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1700                 limit >>= PAGE_SHIFT;
1701                 if (locked > limit && !capable(CAP_IPC_LOCK))
1702                         return -ENOMEM;
1703         }
1704
1705         /* Check to ensure the stack will not grow into a hugetlb-only region */
1706         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1707                         vma->vm_end - size;
1708         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1709                 return -EFAULT;
1710
1711         /*
1712          * Overcommit..  This must be the final test, as it will
1713          * update security statistics.
1714          */
1715         if (security_vm_enough_memory_mm(mm, grow))
1716                 return -ENOMEM;
1717
1718         /* Ok, everything looks good - let it rip */
1719         mm->total_vm += grow;
1720         if (vma->vm_flags & VM_LOCKED)
1721                 mm->locked_vm += grow;
1722         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1723         return 0;
1724 }
1725
1726 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1727 /*
1728  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1729  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1730  */
1731 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1732 {
1733         int error;
1734
1735         if (!(vma->vm_flags & VM_GROWSUP))
1736                 return -EFAULT;
1737
1738         /*
1739          * We must make sure the anon_vma is allocated
1740          * so that the anon_vma locking is not a noop.
1741          */
1742         if (unlikely(anon_vma_prepare(vma)))
1743                 return -ENOMEM;
1744         vma_lock_anon_vma(vma);
1745
1746         /*
1747          * vma->vm_start/vm_end cannot change under us because the caller
1748          * is required to hold the mmap_sem in read mode.  We need the
1749          * anon_vma lock to serialize against concurrent expand_stacks.
1750          * Also guard against wrapping around to address 0.
1751          */
1752         if (address < PAGE_ALIGN(address+4))
1753                 address = PAGE_ALIGN(address+4);
1754         else {
1755                 vma_unlock_anon_vma(vma);
1756                 return -ENOMEM;
1757         }
1758         error = 0;
1759
1760         /* Somebody else might have raced and expanded it already */
1761         if (address > vma->vm_end) {
1762                 unsigned long size, grow;
1763
1764                 size = address - vma->vm_start;
1765                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1766
1767                 error = -ENOMEM;
1768                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1769                         error = acct_stack_growth(vma, size, grow);
1770                         if (!error) {
1771                                 vma->vm_end = address;
1772                                 perf_event_mmap(vma);
1773                         }
1774                 }
1775         }
1776         vma_unlock_anon_vma(vma);
1777         khugepaged_enter_vma_merge(vma);
1778         return error;
1779 }
1780 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1781
1782 /*
1783  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1784  */
1785 int expand_downwards(struct vm_area_struct *vma,
1786                                    unsigned long address)
1787 {
1788         int error;
1789
1790         /*
1791          * We must make sure the anon_vma is allocated
1792          * so that the anon_vma locking is not a noop.
1793          */
1794         if (unlikely(anon_vma_prepare(vma)))
1795                 return -ENOMEM;
1796
1797         address &= PAGE_MASK;
1798         error = security_file_mmap(NULL, 0, 0, 0, address, 1);
1799         if (error)
1800                 return error;
1801
1802         vma_lock_anon_vma(vma);
1803
1804         /*
1805          * vma->vm_start/vm_end cannot change under us because the caller
1806          * is required to hold the mmap_sem in read mode.  We need the
1807          * anon_vma lock to serialize against concurrent expand_stacks.
1808          */
1809
1810         /* Somebody else might have raced and expanded it already */
1811         if (address < vma->vm_start) {
1812                 unsigned long size, grow;
1813
1814                 size = vma->vm_end - address;
1815                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1816
1817                 error = -ENOMEM;
1818                 if (grow <= vma->vm_pgoff) {
1819                         error = acct_stack_growth(vma, size, grow);
1820                         if (!error) {
1821                                 vma->vm_start = address;
1822                                 vma->vm_pgoff -= grow;
1823                                 perf_event_mmap(vma);
1824                         }
1825                 }
1826         }
1827         vma_unlock_anon_vma(vma);
1828         khugepaged_enter_vma_merge(vma);
1829         return error;
1830 }
1831
1832 #ifdef CONFIG_STACK_GROWSUP
1833 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1834 {
1835         return expand_upwards(vma, address);
1836 }
1837
1838 struct vm_area_struct *
1839 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1840 {
1841         struct vm_area_struct *vma, *prev;
1842
1843         addr &= PAGE_MASK;
1844         vma = find_vma_prev(mm, addr, &prev);
1845         if (vma && (vma->vm_start <= addr))
1846                 return vma;
1847         if (!prev || expand_stack(prev, addr))
1848                 return NULL;
1849         if (prev->vm_flags & VM_LOCKED) {
1850                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1851         }
1852         return prev;
1853 }
1854 #else
1855 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1856 {
1857         return expand_downwards(vma, address);
1858 }
1859
1860 struct vm_area_struct *
1861 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1862 {
1863         struct vm_area_struct * vma;
1864         unsigned long start;
1865
1866         addr &= PAGE_MASK;
1867         vma = find_vma(mm,addr);
1868         if (!vma)
1869                 return NULL;
1870         if (vma->vm_start <= addr)
1871                 return vma;
1872         if (!(vma->vm_flags & VM_GROWSDOWN))
1873                 return NULL;
1874         start = vma->vm_start;
1875         if (expand_stack(vma, addr))
1876                 return NULL;
1877         if (vma->vm_flags & VM_LOCKED) {
1878                 mlock_vma_pages_range(vma, addr, start);
1879         }
1880         return vma;
1881 }
1882 #endif
1883
1884 /*
1885  * Ok - we have the memory areas we should free on the vma list,
1886  * so release them, and do the vma updates.
1887  *
1888  * Called with the mm semaphore held.
1889  */
1890 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1891 {
1892         /* Update high watermark before we lower total_vm */
1893         update_hiwater_vm(mm);
1894         do {
1895                 long nrpages = vma_pages(vma);
1896
1897                 mm->total_vm -= nrpages;
1898                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1899                 vma = remove_vma(vma);
1900         } while (vma);
1901         validate_mm(mm);
1902 }
1903
1904 /*
1905  * Get rid of page table information in the indicated region.
1906  *
1907  * Called with the mm semaphore held.
1908  */
1909 static void unmap_region(struct mm_struct *mm,
1910                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1911                 unsigned long start, unsigned long end)
1912 {
1913         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1914         struct mmu_gather tlb;
1915         unsigned long nr_accounted = 0;
1916
1917         lru_add_drain();
1918         tlb_gather_mmu(&tlb, mm, 0);
1919         update_hiwater_rss(mm);
1920         unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1921         vm_unacct_memory(nr_accounted);
1922         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1923                                  next ? next->vm_start : 0);
1924         tlb_finish_mmu(&tlb, start, end);
1925 }
1926
1927 /*
1928  * Create a list of vma's touched by the unmap, removing them from the mm's
1929  * vma list as we go..
1930  */
1931 static void
1932 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1933         struct vm_area_struct *prev, unsigned long end)
1934 {
1935         struct vm_area_struct **insertion_point;
1936         struct vm_area_struct *tail_vma = NULL;
1937         unsigned long addr;
1938
1939         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1940         vma->vm_prev = NULL;
1941         do {
1942                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1943                 mm->map_count--;
1944                 tail_vma = vma;
1945                 vma = vma->vm_next;
1946         } while (vma && vma->vm_start < end);
1947         *insertion_point = vma;
1948         if (vma)
1949                 vma->vm_prev = prev;
1950         tail_vma->vm_next = NULL;
1951         if (mm->unmap_area == arch_unmap_area)
1952                 addr = prev ? prev->vm_end : mm->mmap_base;
1953         else
1954                 addr = vma ?  vma->vm_start : mm->mmap_base;
1955         mm->unmap_area(mm, addr);
1956         mm->mmap_cache = NULL;          /* Kill the cache. */
1957 }
1958
1959 /*
1960  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1961  * munmap path where it doesn't make sense to fail.
1962  */
1963 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1964               unsigned long addr, int new_below)
1965 {
1966         struct mempolicy *pol;
1967         struct vm_area_struct *new;
1968         int err = -ENOMEM;
1969
1970         if (is_vm_hugetlb_page(vma) && (addr &
1971                                         ~(huge_page_mask(hstate_vma(vma)))))
1972                 return -EINVAL;
1973
1974         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1975         if (!new)
1976                 goto out_err;
1977
1978         /* most fields are the same, copy all, and then fixup */
1979         *new = *vma;
1980
1981         INIT_LIST_HEAD(&new->anon_vma_chain);
1982
1983         if (new_below)
1984                 new->vm_end = addr;
1985         else {
1986                 new->vm_start = addr;
1987                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1988         }
1989
1990         pol = mpol_dup(vma_policy(vma));
1991         if (IS_ERR(pol)) {
1992                 err = PTR_ERR(pol);
1993                 goto out_free_vma;
1994         }
1995         vma_set_policy(new, pol);
1996
1997         if (anon_vma_clone(new, vma))
1998                 goto out_free_mpol;
1999
2000         if (new->vm_file) {
2001                 get_file(new->vm_file);
2002                 if (vma->vm_flags & VM_EXECUTABLE)
2003                         added_exe_file_vma(mm);
2004         }
2005
2006         if (new->vm_ops && new->vm_ops->open)
2007                 new->vm_ops->open(new);
2008
2009         if (new_below)
2010                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2011                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2012         else
2013                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2014
2015         /* Success. */
2016         if (!err)
2017                 return 0;
2018
2019         /* Clean everything up if vma_adjust failed. */
2020         if (new->vm_ops && new->vm_ops->close)
2021                 new->vm_ops->close(new);
2022         if (new->vm_file) {
2023                 if (vma->vm_flags & VM_EXECUTABLE)
2024                         removed_exe_file_vma(mm);
2025                 fput(new->vm_file);
2026         }
2027         unlink_anon_vmas(new);
2028  out_free_mpol:
2029         mpol_put(pol);
2030  out_free_vma:
2031         kmem_cache_free(vm_area_cachep, new);
2032  out_err:
2033         return err;
2034 }
2035
2036 /*
2037  * Split a vma into two pieces at address 'addr', a new vma is allocated
2038  * either for the first part or the tail.
2039  */
2040 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2041               unsigned long addr, int new_below)
2042 {
2043         if (mm->map_count >= sysctl_max_map_count)
2044                 return -ENOMEM;
2045
2046         return __split_vma(mm, vma, addr, new_below);
2047 }
2048
2049 /* Munmap is split into 2 main parts -- this part which finds
2050  * what needs doing, and the areas themselves, which do the
2051  * work.  This now handles partial unmappings.
2052  * Jeremy Fitzhardinge <jeremy@goop.org>
2053  */
2054 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2055 {
2056         unsigned long end;
2057         struct vm_area_struct *vma, *prev, *last;
2058
2059         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2060                 return -EINVAL;
2061
2062         if ((len = PAGE_ALIGN(len)) == 0)
2063                 return -EINVAL;
2064
2065         /* Find the first overlapping VMA */
2066         vma = find_vma(mm, start);
2067         if (!vma)
2068                 return 0;
2069         prev = vma->vm_prev;
2070         /* we have  start < vma->vm_end  */
2071
2072         /* if it doesn't overlap, we have nothing.. */
2073         end = start + len;
2074         if (vma->vm_start >= end)
2075                 return 0;
2076
2077         /*
2078          * If we need to split any vma, do it now to save pain later.
2079          *
2080          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2081          * unmapped vm_area_struct will remain in use: so lower split_vma
2082          * places tmp vma above, and higher split_vma places tmp vma below.
2083          */
2084         if (start > vma->vm_start) {
2085                 int error;
2086
2087                 /*
2088                  * Make sure that map_count on return from munmap() will
2089                  * not exceed its limit; but let map_count go just above
2090                  * its limit temporarily, to help free resources as expected.
2091                  */
2092                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2093                         return -ENOMEM;
2094
2095                 error = __split_vma(mm, vma, start, 0);
2096                 if (error)
2097                         return error;
2098                 prev = vma;
2099         }
2100
2101         /* Does it split the last one? */
2102         last = find_vma(mm, end);
2103         if (last && end > last->vm_start) {
2104                 int error = __split_vma(mm, last, end, 1);
2105                 if (error)
2106                         return error;
2107         }
2108         vma = prev? prev->vm_next: mm->mmap;
2109
2110         /*
2111          * unlock any mlock()ed ranges before detaching vmas
2112          */
2113         if (mm->locked_vm) {
2114                 struct vm_area_struct *tmp = vma;
2115                 while (tmp && tmp->vm_start < end) {
2116                         if (tmp->vm_flags & VM_LOCKED) {
2117                                 mm->locked_vm -= vma_pages(tmp);
2118                                 munlock_vma_pages_all(tmp);
2119                         }
2120                         tmp = tmp->vm_next;
2121                 }
2122         }
2123
2124         /*
2125          * Remove the vma's, and unmap the actual pages
2126          */
2127         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2128         unmap_region(mm, vma, prev, start, end);
2129
2130         /* Fix up all other VM information */
2131         remove_vma_list(mm, vma);
2132
2133         return 0;
2134 }
2135 EXPORT_SYMBOL(do_munmap);
2136
2137 int vm_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2138 {
2139         int ret;
2140
2141         down_write(&mm->mmap_sem);
2142         ret = do_munmap(mm, start, len);
2143         up_write(&mm->mmap_sem);
2144         return ret;
2145 }
2146 EXPORT_SYMBOL(vm_munmap);
2147
2148 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2149 {
2150         profile_munmap(addr);
2151         return vm_munmap(current->mm, addr, len);
2152 }
2153
2154 static inline void verify_mm_writelocked(struct mm_struct *mm)
2155 {
2156 #ifdef CONFIG_DEBUG_VM
2157         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2158                 WARN_ON(1);
2159                 up_read(&mm->mmap_sem);
2160         }
2161 #endif
2162 }
2163
2164 /*
2165  *  this is really a simplified "do_mmap".  it only handles
2166  *  anonymous maps.  eventually we may be able to do some
2167  *  brk-specific accounting here.
2168  */
2169 static unsigned long do_brk(unsigned long addr, unsigned long len)
2170 {
2171         struct mm_struct * mm = current->mm;
2172         struct vm_area_struct * vma, * prev;
2173         unsigned long flags;
2174         struct rb_node ** rb_link, * rb_parent;
2175         pgoff_t pgoff = addr >> PAGE_SHIFT;
2176         int error;
2177
2178         len = PAGE_ALIGN(len);
2179         if (!len)
2180                 return addr;
2181
2182         error = security_file_mmap(NULL, 0, 0, 0, addr, 1);
2183         if (error)
2184                 return error;
2185
2186         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2187
2188         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2189         if (error & ~PAGE_MASK)
2190                 return error;
2191
2192         /*
2193          * mlock MCL_FUTURE?
2194          */
2195         if (mm->def_flags & VM_LOCKED) {
2196                 unsigned long locked, lock_limit;
2197                 locked = len >> PAGE_SHIFT;
2198                 locked += mm->locked_vm;
2199                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2200                 lock_limit >>= PAGE_SHIFT;
2201                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2202                         return -EAGAIN;
2203         }
2204
2205         /*
2206          * mm->mmap_sem is required to protect against another thread
2207          * changing the mappings in case we sleep.
2208          */
2209         verify_mm_writelocked(mm);
2210
2211         /*
2212          * Clear old maps.  this also does some error checking for us
2213          */
2214  munmap_back:
2215         vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2216         if (vma && vma->vm_start < addr + len) {
2217                 if (do_munmap(mm, addr, len))
2218                         return -ENOMEM;
2219                 goto munmap_back;
2220         }
2221
2222         /* Check against address space limits *after* clearing old maps... */
2223         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2224                 return -ENOMEM;
2225
2226         if (mm->map_count > sysctl_max_map_count)
2227                 return -ENOMEM;
2228
2229         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2230                 return -ENOMEM;
2231
2232         /* Can we just expand an old private anonymous mapping? */
2233         vma = vma_merge(mm, prev, addr, addr + len, flags,
2234                                         NULL, NULL, pgoff, NULL);
2235         if (vma)
2236                 goto out;
2237
2238         /*
2239          * create a vma struct for an anonymous mapping
2240          */
2241         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2242         if (!vma) {
2243                 vm_unacct_memory(len >> PAGE_SHIFT);
2244                 return -ENOMEM;
2245         }
2246
2247         INIT_LIST_HEAD(&vma->anon_vma_chain);
2248         vma->vm_mm = mm;
2249         vma->vm_start = addr;
2250         vma->vm_end = addr + len;
2251         vma->vm_pgoff = pgoff;
2252         vma->vm_flags = flags;
2253         vma->vm_page_prot = vm_get_page_prot(flags);
2254         vma_link(mm, vma, prev, rb_link, rb_parent);
2255 out:
2256         perf_event_mmap(vma);
2257         mm->total_vm += len >> PAGE_SHIFT;
2258         if (flags & VM_LOCKED) {
2259                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2260                         mm->locked_vm += (len >> PAGE_SHIFT);
2261         }
2262         return addr;
2263 }
2264
2265 unsigned long vm_brk(unsigned long addr, unsigned long len)
2266 {
2267         struct mm_struct *mm = current->mm;
2268         unsigned long ret;
2269
2270         down_write(&mm->mmap_sem);
2271         ret = do_brk(addr, len);
2272         up_write(&mm->mmap_sem);
2273         return ret;
2274 }
2275 EXPORT_SYMBOL(vm_brk);
2276
2277 /* Release all mmaps. */
2278 void exit_mmap(struct mm_struct *mm)
2279 {
2280         struct mmu_gather tlb;
2281         struct vm_area_struct *vma;
2282         unsigned long nr_accounted = 0;
2283
2284         /* mm's last user has gone, and its about to be pulled down */
2285         mmu_notifier_release(mm);
2286
2287         if (mm->locked_vm) {
2288                 vma = mm->mmap;
2289                 while (vma) {
2290                         if (vma->vm_flags & VM_LOCKED)
2291                                 munlock_vma_pages_all(vma);
2292                         vma = vma->vm_next;
2293                 }
2294         }
2295
2296         arch_exit_mmap(mm);
2297
2298         vma = mm->mmap;
2299         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2300                 return;
2301
2302         lru_add_drain();
2303         flush_cache_mm(mm);
2304         tlb_gather_mmu(&tlb, mm, 1);
2305         /* update_hiwater_rss(mm) here? but nobody should be looking */
2306         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2307         unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
2308         vm_unacct_memory(nr_accounted);
2309
2310         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2311         tlb_finish_mmu(&tlb, 0, -1);
2312
2313         /*
2314          * Walk the list again, actually closing and freeing it,
2315          * with preemption enabled, without holding any MM locks.
2316          */
2317         while (vma)
2318                 vma = remove_vma(vma);
2319
2320         BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2321 }
2322
2323 /* Insert vm structure into process list sorted by address
2324  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2325  * then i_mmap_mutex is taken here.
2326  */
2327 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
2328 {
2329         struct vm_area_struct * __vma, * prev;
2330         struct rb_node ** rb_link, * rb_parent;
2331
2332         /*
2333          * The vm_pgoff of a purely anonymous vma should be irrelevant
2334          * until its first write fault, when page's anon_vma and index
2335          * are set.  But now set the vm_pgoff it will almost certainly
2336          * end up with (unless mremap moves it elsewhere before that
2337          * first wfault), so /proc/pid/maps tells a consistent story.
2338          *
2339          * By setting it to reflect the virtual start address of the
2340          * vma, merges and splits can happen in a seamless way, just
2341          * using the existing file pgoff checks and manipulations.
2342          * Similarly in do_mmap_pgoff and in do_brk.
2343          */
2344         if (!vma->vm_file) {
2345                 BUG_ON(vma->anon_vma);
2346                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2347         }
2348         __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2349         if (__vma && __vma->vm_start < vma->vm_end)
2350                 return -ENOMEM;
2351         if ((vma->vm_flags & VM_ACCOUNT) &&
2352              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2353                 return -ENOMEM;
2354         vma_link(mm, vma, prev, rb_link, rb_parent);
2355         return 0;
2356 }
2357
2358 /*
2359  * Copy the vma structure to a new location in the same mm,
2360  * prior to moving page table entries, to effect an mremap move.
2361  */
2362 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2363         unsigned long addr, unsigned long len, pgoff_t pgoff)
2364 {
2365         struct vm_area_struct *vma = *vmap;
2366         unsigned long vma_start = vma->vm_start;
2367         struct mm_struct *mm = vma->vm_mm;
2368         struct vm_area_struct *new_vma, *prev;
2369         struct rb_node **rb_link, *rb_parent;
2370         struct mempolicy *pol;
2371         bool faulted_in_anon_vma = true;
2372
2373         /*
2374          * If anonymous vma has not yet been faulted, update new pgoff
2375          * to match new location, to increase its chance of merging.
2376          */
2377         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2378                 pgoff = addr >> PAGE_SHIFT;
2379                 faulted_in_anon_vma = false;
2380         }
2381
2382         find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2383         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2384                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2385         if (new_vma) {
2386                 /*
2387                  * Source vma may have been merged into new_vma
2388                  */
2389                 if (unlikely(vma_start >= new_vma->vm_start &&
2390                              vma_start < new_vma->vm_end)) {
2391                         /*
2392                          * The only way we can get a vma_merge with
2393                          * self during an mremap is if the vma hasn't
2394                          * been faulted in yet and we were allowed to
2395                          * reset the dst vma->vm_pgoff to the
2396                          * destination address of the mremap to allow
2397                          * the merge to happen. mremap must change the
2398                          * vm_pgoff linearity between src and dst vmas
2399                          * (in turn preventing a vma_merge) to be
2400                          * safe. It is only safe to keep the vm_pgoff
2401                          * linear if there are no pages mapped yet.
2402                          */
2403                         VM_BUG_ON(faulted_in_anon_vma);
2404                         *vmap = new_vma;
2405                 } else
2406                         anon_vma_moveto_tail(new_vma);
2407         } else {
2408                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2409                 if (new_vma) {
2410                         *new_vma = *vma;
2411                         pol = mpol_dup(vma_policy(vma));
2412                         if (IS_ERR(pol))
2413                                 goto out_free_vma;
2414                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2415                         if (anon_vma_clone(new_vma, vma))
2416                                 goto out_free_mempol;
2417                         vma_set_policy(new_vma, pol);
2418                         new_vma->vm_start = addr;
2419                         new_vma->vm_end = addr + len;
2420                         new_vma->vm_pgoff = pgoff;
2421                         if (new_vma->vm_file) {
2422                                 get_file(new_vma->vm_file);
2423                                 if (vma->vm_flags & VM_EXECUTABLE)
2424                                         added_exe_file_vma(mm);
2425                         }
2426                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2427                                 new_vma->vm_ops->open(new_vma);
2428                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2429                 }
2430         }
2431         return new_vma;
2432
2433  out_free_mempol:
2434         mpol_put(pol);
2435  out_free_vma:
2436         kmem_cache_free(vm_area_cachep, new_vma);
2437         return NULL;
2438 }
2439
2440 /*
2441  * Return true if the calling process may expand its vm space by the passed
2442  * number of pages
2443  */
2444 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2445 {
2446         unsigned long cur = mm->total_vm;       /* pages */
2447         unsigned long lim;
2448
2449         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2450
2451         if (cur + npages > lim)
2452                 return 0;
2453         return 1;
2454 }
2455
2456
2457 static int special_mapping_fault(struct vm_area_struct *vma,
2458                                 struct vm_fault *vmf)
2459 {
2460         pgoff_t pgoff;
2461         struct page **pages;
2462
2463         /*
2464          * special mappings have no vm_file, and in that case, the mm
2465          * uses vm_pgoff internally. So we have to subtract it from here.
2466          * We are allowed to do this because we are the mm; do not copy
2467          * this code into drivers!
2468          */
2469         pgoff = vmf->pgoff - vma->vm_pgoff;
2470
2471         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2472                 pgoff--;
2473
2474         if (*pages) {
2475                 struct page *page = *pages;
2476                 get_page(page);
2477                 vmf->page = page;
2478                 return 0;
2479         }
2480
2481         return VM_FAULT_SIGBUS;
2482 }
2483
2484 /*
2485  * Having a close hook prevents vma merging regardless of flags.
2486  */
2487 static void special_mapping_close(struct vm_area_struct *vma)
2488 {
2489 }
2490
2491 static const struct vm_operations_struct special_mapping_vmops = {
2492         .close = special_mapping_close,
2493         .fault = special_mapping_fault,
2494 };
2495
2496 /*
2497  * Called with mm->mmap_sem held for writing.
2498  * Insert a new vma covering the given region, with the given flags.
2499  * Its pages are supplied by the given array of struct page *.
2500  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2501  * The region past the last page supplied will always produce SIGBUS.
2502  * The array pointer and the pages it points to are assumed to stay alive
2503  * for as long as this mapping might exist.
2504  */
2505 int install_special_mapping(struct mm_struct *mm,
2506                             unsigned long addr, unsigned long len,
2507                             unsigned long vm_flags, struct page **pages)
2508 {
2509         int ret;
2510         struct vm_area_struct *vma;
2511
2512         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2513         if (unlikely(vma == NULL))
2514                 return -ENOMEM;
2515
2516         INIT_LIST_HEAD(&vma->anon_vma_chain);
2517         vma->vm_mm = mm;
2518         vma->vm_start = addr;
2519         vma->vm_end = addr + len;
2520
2521         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2522         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2523
2524         vma->vm_ops = &special_mapping_vmops;
2525         vma->vm_private_data = pages;
2526
2527         ret = security_file_mmap(NULL, 0, 0, 0, vma->vm_start, 1);
2528         if (ret)
2529                 goto out;
2530
2531         ret = insert_vm_struct(mm, vma);
2532         if (ret)
2533                 goto out;
2534
2535         mm->total_vm += len >> PAGE_SHIFT;
2536
2537         perf_event_mmap(vma);
2538
2539         return 0;
2540
2541 out:
2542         kmem_cache_free(vm_area_cachep, vma);
2543         return ret;
2544 }
2545
2546 static DEFINE_MUTEX(mm_all_locks_mutex);
2547
2548 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2549 {
2550         if (!test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2551                 /*
2552                  * The LSB of head.next can't change from under us
2553                  * because we hold the mm_all_locks_mutex.
2554                  */
2555                 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2556                 /*
2557                  * We can safely modify head.next after taking the
2558                  * anon_vma->root->mutex. If some other vma in this mm shares
2559                  * the same anon_vma we won't take it again.
2560                  *
2561                  * No need of atomic instructions here, head.next
2562                  * can't change from under us thanks to the
2563                  * anon_vma->root->mutex.
2564                  */
2565                 if (__test_and_set_bit(0, (unsigned long *)
2566                                        &anon_vma->root->head.next))
2567                         BUG();
2568         }
2569 }
2570
2571 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2572 {
2573         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2574                 /*
2575                  * AS_MM_ALL_LOCKS can't change from under us because
2576                  * we hold the mm_all_locks_mutex.
2577                  *
2578                  * Operations on ->flags have to be atomic because
2579                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2580                  * mm_all_locks_mutex, there may be other cpus
2581                  * changing other bitflags in parallel to us.
2582                  */
2583                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2584                         BUG();
2585                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2586         }
2587 }
2588
2589 /*
2590  * This operation locks against the VM for all pte/vma/mm related
2591  * operations that could ever happen on a certain mm. This includes
2592  * vmtruncate, try_to_unmap, and all page faults.
2593  *
2594  * The caller must take the mmap_sem in write mode before calling
2595  * mm_take_all_locks(). The caller isn't allowed to release the
2596  * mmap_sem until mm_drop_all_locks() returns.
2597  *
2598  * mmap_sem in write mode is required in order to block all operations
2599  * that could modify pagetables and free pages without need of
2600  * altering the vma layout (for example populate_range() with
2601  * nonlinear vmas). It's also needed in write mode to avoid new
2602  * anon_vmas to be associated with existing vmas.
2603  *
2604  * A single task can't take more than one mm_take_all_locks() in a row
2605  * or it would deadlock.
2606  *
2607  * The LSB in anon_vma->head.next and the AS_MM_ALL_LOCKS bitflag in
2608  * mapping->flags avoid to take the same lock twice, if more than one
2609  * vma in this mm is backed by the same anon_vma or address_space.
2610  *
2611  * We can take all the locks in random order because the VM code
2612  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2613  * takes more than one of them in a row. Secondly we're protected
2614  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2615  *
2616  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2617  * that may have to take thousand of locks.
2618  *
2619  * mm_take_all_locks() can fail if it's interrupted by signals.
2620  */
2621 int mm_take_all_locks(struct mm_struct *mm)
2622 {
2623         struct vm_area_struct *vma;
2624         struct anon_vma_chain *avc;
2625
2626         BUG_ON(down_read_trylock(&mm->mmap_sem));
2627
2628         mutex_lock(&mm_all_locks_mutex);
2629
2630         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2631                 if (signal_pending(current))
2632                         goto out_unlock;
2633                 if (vma->vm_file && vma->vm_file->f_mapping)
2634                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2635         }
2636
2637         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2638                 if (signal_pending(current))
2639                         goto out_unlock;
2640                 if (vma->anon_vma)
2641                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2642                                 vm_lock_anon_vma(mm, avc->anon_vma);
2643         }
2644
2645         return 0;
2646
2647 out_unlock:
2648         mm_drop_all_locks(mm);
2649         return -EINTR;
2650 }
2651
2652 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2653 {
2654         if (test_bit(0, (unsigned long *) &anon_vma->root->head.next)) {
2655                 /*
2656                  * The LSB of head.next can't change to 0 from under
2657                  * us because we hold the mm_all_locks_mutex.
2658                  *
2659                  * We must however clear the bitflag before unlocking
2660                  * the vma so the users using the anon_vma->head will
2661                  * never see our bitflag.
2662                  *
2663                  * No need of atomic instructions here, head.next
2664                  * can't change from under us until we release the
2665                  * anon_vma->root->mutex.
2666                  */
2667                 if (!__test_and_clear_bit(0, (unsigned long *)
2668                                           &anon_vma->root->head.next))
2669                         BUG();
2670                 anon_vma_unlock(anon_vma);
2671         }
2672 }
2673
2674 static void vm_unlock_mapping(struct address_space *mapping)
2675 {
2676         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2677                 /*
2678                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2679                  * because we hold the mm_all_locks_mutex.
2680                  */
2681                 mutex_unlock(&mapping->i_mmap_mutex);
2682                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2683                                         &mapping->flags))
2684                         BUG();
2685         }
2686 }
2687
2688 /*
2689  * The mmap_sem cannot be released by the caller until
2690  * mm_drop_all_locks() returns.
2691  */
2692 void mm_drop_all_locks(struct mm_struct *mm)
2693 {
2694         struct vm_area_struct *vma;
2695         struct anon_vma_chain *avc;
2696
2697         BUG_ON(down_read_trylock(&mm->mmap_sem));
2698         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2699
2700         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2701                 if (vma->anon_vma)
2702                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2703                                 vm_unlock_anon_vma(avc->anon_vma);
2704                 if (vma->vm_file && vma->vm_file->f_mapping)
2705                         vm_unlock_mapping(vma->vm_file->f_mapping);
2706         }
2707
2708         mutex_unlock(&mm_all_locks_mutex);
2709 }
2710
2711 /*
2712  * initialise the VMA slab
2713  */
2714 void __init mmap_init(void)
2715 {
2716         int ret;
2717
2718         ret = percpu_counter_init(&vm_committed_as, 0);
2719         VM_BUG_ON(ret);
2720 }