mm/rmap, migration: Make rmap_walk_anon() and try_to_unmap_anon() more scalable
[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 #include <linux/uprobes.h>
34
35 #include <asm/uaccess.h>
36 #include <asm/cacheflush.h>
37 #include <asm/tlb.h>
38 #include <asm/mmu_context.h>
39
40 #include "internal.h"
41
42 #ifndef arch_mmap_check
43 #define arch_mmap_check(addr, len, flags)       (0)
44 #endif
45
46 #ifndef arch_rebalance_pgtables
47 #define arch_rebalance_pgtables(addr, len)              (addr)
48 #endif
49
50 static void unmap_region(struct mm_struct *mm,
51                 struct vm_area_struct *vma, struct vm_area_struct *prev,
52                 unsigned long start, unsigned long end);
53
54 /* description of effects of mapping type and prot in current implementation.
55  * this is due to the limited x86 page protection hardware.  The expected
56  * behavior is in parens:
57  *
58  * map_type     prot
59  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
60  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
61  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
62  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
63  *              
64  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
65  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
66  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
67  *
68  */
69 pgprot_t protection_map[16] = {
70         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
71         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
72 };
73
74 pgprot_t vm_get_page_prot(unsigned long vm_flags)
75 {
76         return __pgprot(pgprot_val(protection_map[vm_flags &
77                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
78                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
79 }
80 EXPORT_SYMBOL(vm_get_page_prot);
81
82 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
83 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
84 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
85 /*
86  * Make sure vm_committed_as in one cacheline and not cacheline shared with
87  * other variables. It can be updated by several CPUs frequently.
88  */
89 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
90
91 /*
92  * Check that a process has enough memory to allocate a new virtual
93  * mapping. 0 means there is enough memory for the allocation to
94  * succeed and -ENOMEM implies there is not.
95  *
96  * We currently support three overcommit policies, which are set via the
97  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
98  *
99  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
100  * Additional code 2002 Jul 20 by Robert Love.
101  *
102  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
103  *
104  * Note this is a helper function intended to be used by LSMs which
105  * wish to use this logic.
106  */
107 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
108 {
109         unsigned long free, allowed;
110
111         vm_acct_memory(pages);
112
113         /*
114          * Sometimes we want to use more memory than we have
115          */
116         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
117                 return 0;
118
119         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
120                 free = global_page_state(NR_FREE_PAGES);
121                 free += global_page_state(NR_FILE_PAGES);
122
123                 /*
124                  * shmem pages shouldn't be counted as free in this
125                  * case, they can't be purged, only swapped out, and
126                  * that won't affect the overall amount of available
127                  * memory in the system.
128                  */
129                 free -= global_page_state(NR_SHMEM);
130
131                 free += nr_swap_pages;
132
133                 /*
134                  * Any slabs which are created with the
135                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
136                  * which are reclaimable, under pressure.  The dentry
137                  * cache and most inode caches should fall into this
138                  */
139                 free += global_page_state(NR_SLAB_RECLAIMABLE);
140
141                 /*
142                  * Leave reserved pages. The pages are not for anonymous pages.
143                  */
144                 if (free <= totalreserve_pages)
145                         goto error;
146                 else
147                         free -= totalreserve_pages;
148
149                 /*
150                  * Leave the last 3% for root
151                  */
152                 if (!cap_sys_admin)
153                         free -= free / 32;
154
155                 if (free > pages)
156                         return 0;
157
158                 goto error;
159         }
160
161         allowed = (totalram_pages - hugetlb_total_pages())
162                 * sysctl_overcommit_ratio / 100;
163         /*
164          * Leave the last 3% for root
165          */
166         if (!cap_sys_admin)
167                 allowed -= allowed / 32;
168         allowed += total_swap_pages;
169
170         /* Don't let a single process grow too big:
171            leave 3% of the size of this process for other processes */
172         if (mm)
173                 allowed -= mm->total_vm / 32;
174
175         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
176                 return 0;
177 error:
178         vm_unacct_memory(pages);
179
180         return -ENOMEM;
181 }
182
183 /*
184  * Requires inode->i_mapping->i_mmap_mutex
185  */
186 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
187                 struct file *file, struct address_space *mapping)
188 {
189         if (vma->vm_flags & VM_DENYWRITE)
190                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
191         if (vma->vm_flags & VM_SHARED)
192                 mapping->i_mmap_writable--;
193
194         flush_dcache_mmap_lock(mapping);
195         if (unlikely(vma->vm_flags & VM_NONLINEAR))
196                 list_del_init(&vma->shared.nonlinear);
197         else
198                 vma_interval_tree_remove(vma, &mapping->i_mmap);
199         flush_dcache_mmap_unlock(mapping);
200 }
201
202 /*
203  * Unlink a file-based vm structure from its interval tree, to hide
204  * vma from rmap and vmtruncate before freeing its page tables.
205  */
206 void unlink_file_vma(struct vm_area_struct *vma)
207 {
208         struct file *file = vma->vm_file;
209
210         if (file) {
211                 struct address_space *mapping = file->f_mapping;
212                 mutex_lock(&mapping->i_mmap_mutex);
213                 __remove_shared_vm_struct(vma, file, mapping);
214                 mutex_unlock(&mapping->i_mmap_mutex);
215         }
216 }
217
218 /*
219  * Close a vm structure and free it, returning the next.
220  */
221 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
222 {
223         struct vm_area_struct *next = vma->vm_next;
224
225         might_sleep();
226         if (vma->vm_ops && vma->vm_ops->close)
227                 vma->vm_ops->close(vma);
228         if (vma->vm_file)
229                 fput(vma->vm_file);
230         mpol_put(vma_policy(vma));
231         kmem_cache_free(vm_area_cachep, vma);
232         return next;
233 }
234
235 static unsigned long do_brk(unsigned long addr, unsigned long len);
236
237 SYSCALL_DEFINE1(brk, unsigned long, brk)
238 {
239         unsigned long rlim, retval;
240         unsigned long newbrk, oldbrk;
241         struct mm_struct *mm = current->mm;
242         unsigned long min_brk;
243
244         down_write(&mm->mmap_sem);
245
246 #ifdef CONFIG_COMPAT_BRK
247         /*
248          * CONFIG_COMPAT_BRK can still be overridden by setting
249          * randomize_va_space to 2, which will still cause mm->start_brk
250          * to be arbitrarily shifted
251          */
252         if (current->brk_randomized)
253                 min_brk = mm->start_brk;
254         else
255                 min_brk = mm->end_data;
256 #else
257         min_brk = mm->start_brk;
258 #endif
259         if (brk < min_brk)
260                 goto out;
261
262         /*
263          * Check against rlimit here. If this check is done later after the test
264          * of oldbrk with newbrk then it can escape the test and let the data
265          * segment grow beyond its set limit the in case where the limit is
266          * not page aligned -Ram Gupta
267          */
268         rlim = rlimit(RLIMIT_DATA);
269         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
270                         (mm->end_data - mm->start_data) > rlim)
271                 goto out;
272
273         newbrk = PAGE_ALIGN(brk);
274         oldbrk = PAGE_ALIGN(mm->brk);
275         if (oldbrk == newbrk)
276                 goto set_brk;
277
278         /* Always allow shrinking brk. */
279         if (brk <= mm->brk) {
280                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
281                         goto set_brk;
282                 goto out;
283         }
284
285         /* Check against existing mmap mappings. */
286         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
287                 goto out;
288
289         /* Ok, looks good - let it rip. */
290         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
291                 goto out;
292 set_brk:
293         mm->brk = brk;
294 out:
295         retval = mm->brk;
296         up_write(&mm->mmap_sem);
297         return retval;
298 }
299
300 #ifdef CONFIG_DEBUG_VM_RB
301 static int browse_rb(struct rb_root *root)
302 {
303         int i = 0, j;
304         struct rb_node *nd, *pn = NULL;
305         unsigned long prev = 0, pend = 0;
306
307         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
308                 struct vm_area_struct *vma;
309                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
310                 if (vma->vm_start < prev)
311                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
312                 if (vma->vm_start < pend)
313                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
314                 if (vma->vm_start > vma->vm_end)
315                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
316                 i++;
317                 pn = nd;
318                 prev = vma->vm_start;
319                 pend = vma->vm_end;
320         }
321         j = 0;
322         for (nd = pn; nd; nd = rb_prev(nd)) {
323                 j++;
324         }
325         if (i != j)
326                 printk("backwards %d, forwards %d\n", j, i), i = 0;
327         return i;
328 }
329
330 void validate_mm(struct mm_struct *mm)
331 {
332         int bug = 0;
333         int i = 0;
334         struct vm_area_struct *vma = mm->mmap;
335         while (vma) {
336                 struct anon_vma_chain *avc;
337                 vma_lock_anon_vma(vma);
338                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
339                         anon_vma_interval_tree_verify(avc);
340                 vma_unlock_anon_vma(vma);
341                 vma = vma->vm_next;
342                 i++;
343         }
344         if (i != mm->map_count)
345                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
346         i = browse_rb(&mm->mm_rb);
347         if (i != mm->map_count)
348                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
349         BUG_ON(bug);
350 }
351 #else
352 #define validate_mm(mm) do { } while (0)
353 #endif
354
355 /*
356  * vma has some anon_vma assigned, and is already inserted on that
357  * anon_vma's interval trees.
358  *
359  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
360  * vma must be removed from the anon_vma's interval trees using
361  * anon_vma_interval_tree_pre_update_vma().
362  *
363  * After the update, the vma will be reinserted using
364  * anon_vma_interval_tree_post_update_vma().
365  *
366  * The entire update must be protected by exclusive mmap_sem and by
367  * the root anon_vma's mutex.
368  */
369 static inline void
370 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
371 {
372         struct anon_vma_chain *avc;
373
374         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
375                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
376 }
377
378 static inline void
379 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
380 {
381         struct anon_vma_chain *avc;
382
383         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
384                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
385 }
386
387 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
388                 unsigned long end, struct vm_area_struct **pprev,
389                 struct rb_node ***rb_link, struct rb_node **rb_parent)
390 {
391         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
392
393         __rb_link = &mm->mm_rb.rb_node;
394         rb_prev = __rb_parent = NULL;
395
396         while (*__rb_link) {
397                 struct vm_area_struct *vma_tmp;
398
399                 __rb_parent = *__rb_link;
400                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
401
402                 if (vma_tmp->vm_end > addr) {
403                         /* Fail if an existing vma overlaps the area */
404                         if (vma_tmp->vm_start < end)
405                                 return -ENOMEM;
406                         __rb_link = &__rb_parent->rb_left;
407                 } else {
408                         rb_prev = __rb_parent;
409                         __rb_link = &__rb_parent->rb_right;
410                 }
411         }
412
413         *pprev = NULL;
414         if (rb_prev)
415                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
416         *rb_link = __rb_link;
417         *rb_parent = __rb_parent;
418         return 0;
419 }
420
421 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
422                 struct rb_node **rb_link, struct rb_node *rb_parent)
423 {
424         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
425         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
426 }
427
428 static void __vma_link_file(struct vm_area_struct *vma)
429 {
430         struct file *file;
431
432         file = vma->vm_file;
433         if (file) {
434                 struct address_space *mapping = file->f_mapping;
435
436                 if (vma->vm_flags & VM_DENYWRITE)
437                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
438                 if (vma->vm_flags & VM_SHARED)
439                         mapping->i_mmap_writable++;
440
441                 flush_dcache_mmap_lock(mapping);
442                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
443                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
444                 else
445                         vma_interval_tree_insert(vma, &mapping->i_mmap);
446                 flush_dcache_mmap_unlock(mapping);
447         }
448 }
449
450 static void
451 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
452         struct vm_area_struct *prev, struct rb_node **rb_link,
453         struct rb_node *rb_parent)
454 {
455         __vma_link_list(mm, vma, prev, rb_parent);
456         __vma_link_rb(mm, vma, rb_link, rb_parent);
457 }
458
459 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
460                         struct vm_area_struct *prev, struct rb_node **rb_link,
461                         struct rb_node *rb_parent)
462 {
463         struct address_space *mapping = NULL;
464
465         if (vma->vm_file)
466                 mapping = vma->vm_file->f_mapping;
467
468         if (mapping)
469                 mutex_lock(&mapping->i_mmap_mutex);
470
471         __vma_link(mm, vma, prev, rb_link, rb_parent);
472         __vma_link_file(vma);
473
474         if (mapping)
475                 mutex_unlock(&mapping->i_mmap_mutex);
476
477         mm->map_count++;
478         validate_mm(mm);
479 }
480
481 /*
482  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
483  * mm's list and rbtree.  It has already been inserted into the interval tree.
484  */
485 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
486 {
487         struct vm_area_struct *prev;
488         struct rb_node **rb_link, *rb_parent;
489
490         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
491                            &prev, &rb_link, &rb_parent))
492                 BUG();
493         __vma_link(mm, vma, prev, rb_link, rb_parent);
494         mm->map_count++;
495 }
496
497 static inline void
498 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
499                 struct vm_area_struct *prev)
500 {
501         struct vm_area_struct *next = vma->vm_next;
502
503         prev->vm_next = next;
504         if (next)
505                 next->vm_prev = prev;
506         rb_erase(&vma->vm_rb, &mm->mm_rb);
507         if (mm->mmap_cache == vma)
508                 mm->mmap_cache = prev;
509 }
510
511 /*
512  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
513  * is already present in an i_mmap tree without adjusting the tree.
514  * The following helper function should be used when such adjustments
515  * are necessary.  The "insert" vma (if any) is to be inserted
516  * before we drop the necessary locks.
517  */
518 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
519         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
520 {
521         struct mm_struct *mm = vma->vm_mm;
522         struct vm_area_struct *next = vma->vm_next;
523         struct vm_area_struct *importer = NULL;
524         struct address_space *mapping = NULL;
525         struct rb_root *root = NULL;
526         struct anon_vma *anon_vma = NULL;
527         struct file *file = vma->vm_file;
528         long adjust_next = 0;
529         int remove_next = 0;
530
531         if (next && !insert) {
532                 struct vm_area_struct *exporter = NULL;
533
534                 if (end >= next->vm_end) {
535                         /*
536                          * vma expands, overlapping all the next, and
537                          * perhaps the one after too (mprotect case 6).
538                          */
539 again:                  remove_next = 1 + (end > next->vm_end);
540                         end = next->vm_end;
541                         exporter = next;
542                         importer = vma;
543                 } else if (end > next->vm_start) {
544                         /*
545                          * vma expands, overlapping part of the next:
546                          * mprotect case 5 shifting the boundary up.
547                          */
548                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
549                         exporter = next;
550                         importer = vma;
551                 } else if (end < vma->vm_end) {
552                         /*
553                          * vma shrinks, and !insert tells it's not
554                          * split_vma inserting another: so it must be
555                          * mprotect case 4 shifting the boundary down.
556                          */
557                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
558                         exporter = vma;
559                         importer = next;
560                 }
561
562                 /*
563                  * Easily overlooked: when mprotect shifts the boundary,
564                  * make sure the expanding vma has anon_vma set if the
565                  * shrinking vma had, to cover any anon pages imported.
566                  */
567                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
568                         if (anon_vma_clone(importer, exporter))
569                                 return -ENOMEM;
570                         importer->anon_vma = exporter->anon_vma;
571                 }
572         }
573
574         if (file) {
575                 mapping = file->f_mapping;
576                 if (!(vma->vm_flags & VM_NONLINEAR)) {
577                         root = &mapping->i_mmap;
578                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
579
580                         if (adjust_next)
581                                 uprobe_munmap(next, next->vm_start,
582                                                         next->vm_end);
583                 }
584
585                 mutex_lock(&mapping->i_mmap_mutex);
586                 if (insert) {
587                         /*
588                          * Put into interval tree now, so instantiated pages
589                          * are visible to arm/parisc __flush_dcache_page
590                          * throughout; but we cannot insert into address
591                          * space until vma start or end is updated.
592                          */
593                         __vma_link_file(insert);
594                 }
595         }
596
597         vma_adjust_trans_huge(vma, start, end, adjust_next);
598
599         anon_vma = vma->anon_vma;
600         if (!anon_vma && adjust_next)
601                 anon_vma = next->anon_vma;
602         if (anon_vma) {
603                 VM_BUG_ON(adjust_next && next->anon_vma &&
604                           anon_vma != next->anon_vma);
605                 anon_vma_lock_write(anon_vma);
606                 anon_vma_interval_tree_pre_update_vma(vma);
607                 if (adjust_next)
608                         anon_vma_interval_tree_pre_update_vma(next);
609         }
610
611         if (root) {
612                 flush_dcache_mmap_lock(mapping);
613                 vma_interval_tree_remove(vma, root);
614                 if (adjust_next)
615                         vma_interval_tree_remove(next, root);
616         }
617
618         vma->vm_start = start;
619         vma->vm_end = end;
620         vma->vm_pgoff = pgoff;
621         if (adjust_next) {
622                 next->vm_start += adjust_next << PAGE_SHIFT;
623                 next->vm_pgoff += adjust_next;
624         }
625
626         if (root) {
627                 if (adjust_next)
628                         vma_interval_tree_insert(next, root);
629                 vma_interval_tree_insert(vma, root);
630                 flush_dcache_mmap_unlock(mapping);
631         }
632
633         if (remove_next) {
634                 /*
635                  * vma_merge has merged next into vma, and needs
636                  * us to remove next before dropping the locks.
637                  */
638                 __vma_unlink(mm, next, vma);
639                 if (file)
640                         __remove_shared_vm_struct(next, file, mapping);
641         } else if (insert) {
642                 /*
643                  * split_vma has split insert from vma, and needs
644                  * us to insert it before dropping the locks
645                  * (it may either follow vma or precede it).
646                  */
647                 __insert_vm_struct(mm, insert);
648         }
649
650         if (anon_vma) {
651                 anon_vma_interval_tree_post_update_vma(vma);
652                 if (adjust_next)
653                         anon_vma_interval_tree_post_update_vma(next);
654                 anon_vma_unlock(anon_vma);
655         }
656         if (mapping)
657                 mutex_unlock(&mapping->i_mmap_mutex);
658
659         if (root) {
660                 uprobe_mmap(vma);
661
662                 if (adjust_next)
663                         uprobe_mmap(next);
664         }
665
666         if (remove_next) {
667                 if (file) {
668                         uprobe_munmap(next, next->vm_start, next->vm_end);
669                         fput(file);
670                 }
671                 if (next->anon_vma)
672                         anon_vma_merge(vma, next);
673                 mm->map_count--;
674                 mpol_put(vma_policy(next));
675                 kmem_cache_free(vm_area_cachep, next);
676                 /*
677                  * In mprotect's case 6 (see comments on vma_merge),
678                  * we must remove another next too. It would clutter
679                  * up the code too much to do both in one go.
680                  */
681                 if (remove_next == 2) {
682                         next = vma->vm_next;
683                         goto again;
684                 }
685         }
686         if (insert && file)
687                 uprobe_mmap(insert);
688
689         validate_mm(mm);
690
691         return 0;
692 }
693
694 /*
695  * If the vma has a ->close operation then the driver probably needs to release
696  * per-vma resources, so we don't attempt to merge those.
697  */
698 static inline int is_mergeable_vma(struct vm_area_struct *vma,
699                         struct file *file, unsigned long vm_flags)
700 {
701         if (vma->vm_flags ^ vm_flags)
702                 return 0;
703         if (vma->vm_file != file)
704                 return 0;
705         if (vma->vm_ops && vma->vm_ops->close)
706                 return 0;
707         return 1;
708 }
709
710 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
711                                         struct anon_vma *anon_vma2,
712                                         struct vm_area_struct *vma)
713 {
714         /*
715          * The list_is_singular() test is to avoid merging VMA cloned from
716          * parents. This can improve scalability caused by anon_vma lock.
717          */
718         if ((!anon_vma1 || !anon_vma2) && (!vma ||
719                 list_is_singular(&vma->anon_vma_chain)))
720                 return 1;
721         return anon_vma1 == anon_vma2;
722 }
723
724 /*
725  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
726  * in front of (at a lower virtual address and file offset than) the vma.
727  *
728  * We cannot merge two vmas if they have differently assigned (non-NULL)
729  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
730  *
731  * We don't check here for the merged mmap wrapping around the end of pagecache
732  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
733  * wrap, nor mmaps which cover the final page at index -1UL.
734  */
735 static int
736 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
737         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
738 {
739         if (is_mergeable_vma(vma, file, vm_flags) &&
740             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
741                 if (vma->vm_pgoff == vm_pgoff)
742                         return 1;
743         }
744         return 0;
745 }
746
747 /*
748  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
749  * beyond (at a higher virtual address and file offset than) the vma.
750  *
751  * We cannot merge two vmas if they have differently assigned (non-NULL)
752  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
753  */
754 static int
755 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
756         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
757 {
758         if (is_mergeable_vma(vma, file, vm_flags) &&
759             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
760                 pgoff_t vm_pglen;
761                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
762                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
763                         return 1;
764         }
765         return 0;
766 }
767
768 /*
769  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
770  * whether that can be merged with its predecessor or its successor.
771  * Or both (it neatly fills a hole).
772  *
773  * In most cases - when called for mmap, brk or mremap - [addr,end) is
774  * certain not to be mapped by the time vma_merge is called; but when
775  * called for mprotect, it is certain to be already mapped (either at
776  * an offset within prev, or at the start of next), and the flags of
777  * this area are about to be changed to vm_flags - and the no-change
778  * case has already been eliminated.
779  *
780  * The following mprotect cases have to be considered, where AAAA is
781  * the area passed down from mprotect_fixup, never extending beyond one
782  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
783  *
784  *     AAAA             AAAA                AAAA          AAAA
785  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
786  *    cannot merge    might become    might become    might become
787  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
788  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
789  *    mremap move:                                    PPPPNNNNNNNN 8
790  *        AAAA
791  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
792  *    might become    case 1 below    case 2 below    case 3 below
793  *
794  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
795  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
796  */
797 struct vm_area_struct *vma_merge(struct mm_struct *mm,
798                         struct vm_area_struct *prev, unsigned long addr,
799                         unsigned long end, unsigned long vm_flags,
800                         struct anon_vma *anon_vma, struct file *file,
801                         pgoff_t pgoff, struct mempolicy *policy)
802 {
803         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
804         struct vm_area_struct *area, *next;
805         int err;
806
807         /*
808          * We later require that vma->vm_flags == vm_flags,
809          * so this tests vma->vm_flags & VM_SPECIAL, too.
810          */
811         if (vm_flags & VM_SPECIAL)
812                 return NULL;
813
814         if (prev)
815                 next = prev->vm_next;
816         else
817                 next = mm->mmap;
818         area = next;
819         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
820                 next = next->vm_next;
821
822         /*
823          * Can it merge with the predecessor?
824          */
825         if (prev && prev->vm_end == addr &&
826                         mpol_equal(vma_policy(prev), policy) &&
827                         can_vma_merge_after(prev, vm_flags,
828                                                 anon_vma, file, pgoff)) {
829                 /*
830                  * OK, it can.  Can we now merge in the successor as well?
831                  */
832                 if (next && end == next->vm_start &&
833                                 mpol_equal(policy, vma_policy(next)) &&
834                                 can_vma_merge_before(next, vm_flags,
835                                         anon_vma, file, pgoff+pglen) &&
836                                 is_mergeable_anon_vma(prev->anon_vma,
837                                                       next->anon_vma, NULL)) {
838                                                         /* cases 1, 6 */
839                         err = vma_adjust(prev, prev->vm_start,
840                                 next->vm_end, prev->vm_pgoff, NULL);
841                 } else                                  /* cases 2, 5, 7 */
842                         err = vma_adjust(prev, prev->vm_start,
843                                 end, prev->vm_pgoff, NULL);
844                 if (err)
845                         return NULL;
846                 khugepaged_enter_vma_merge(prev);
847                 return prev;
848         }
849
850         /*
851          * Can this new request be merged in front of next?
852          */
853         if (next && end == next->vm_start &&
854                         mpol_equal(policy, vma_policy(next)) &&
855                         can_vma_merge_before(next, vm_flags,
856                                         anon_vma, file, pgoff+pglen)) {
857                 if (prev && addr < prev->vm_end)        /* case 4 */
858                         err = vma_adjust(prev, prev->vm_start,
859                                 addr, prev->vm_pgoff, NULL);
860                 else                                    /* cases 3, 8 */
861                         err = vma_adjust(area, addr, next->vm_end,
862                                 next->vm_pgoff - pglen, NULL);
863                 if (err)
864                         return NULL;
865                 khugepaged_enter_vma_merge(area);
866                 return area;
867         }
868
869         return NULL;
870 }
871
872 /*
873  * Rough compatbility check to quickly see if it's even worth looking
874  * at sharing an anon_vma.
875  *
876  * They need to have the same vm_file, and the flags can only differ
877  * in things that mprotect may change.
878  *
879  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
880  * we can merge the two vma's. For example, we refuse to merge a vma if
881  * there is a vm_ops->close() function, because that indicates that the
882  * driver is doing some kind of reference counting. But that doesn't
883  * really matter for the anon_vma sharing case.
884  */
885 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
886 {
887         return a->vm_end == b->vm_start &&
888                 mpol_equal(vma_policy(a), vma_policy(b)) &&
889                 a->vm_file == b->vm_file &&
890                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
891                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
892 }
893
894 /*
895  * Do some basic sanity checking to see if we can re-use the anon_vma
896  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
897  * the same as 'old', the other will be the new one that is trying
898  * to share the anon_vma.
899  *
900  * NOTE! This runs with mm_sem held for reading, so it is possible that
901  * the anon_vma of 'old' is concurrently in the process of being set up
902  * by another page fault trying to merge _that_. But that's ok: if it
903  * is being set up, that automatically means that it will be a singleton
904  * acceptable for merging, so we can do all of this optimistically. But
905  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
906  *
907  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
908  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
909  * is to return an anon_vma that is "complex" due to having gone through
910  * a fork).
911  *
912  * We also make sure that the two vma's are compatible (adjacent,
913  * and with the same memory policies). That's all stable, even with just
914  * a read lock on the mm_sem.
915  */
916 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
917 {
918         if (anon_vma_compatible(a, b)) {
919                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
920
921                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
922                         return anon_vma;
923         }
924         return NULL;
925 }
926
927 /*
928  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
929  * neighbouring vmas for a suitable anon_vma, before it goes off
930  * to allocate a new anon_vma.  It checks because a repetitive
931  * sequence of mprotects and faults may otherwise lead to distinct
932  * anon_vmas being allocated, preventing vma merge in subsequent
933  * mprotect.
934  */
935 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
936 {
937         struct anon_vma *anon_vma;
938         struct vm_area_struct *near;
939
940         near = vma->vm_next;
941         if (!near)
942                 goto try_prev;
943
944         anon_vma = reusable_anon_vma(near, vma, near);
945         if (anon_vma)
946                 return anon_vma;
947 try_prev:
948         near = vma->vm_prev;
949         if (!near)
950                 goto none;
951
952         anon_vma = reusable_anon_vma(near, near, vma);
953         if (anon_vma)
954                 return anon_vma;
955 none:
956         /*
957          * There's no absolute need to look only at touching neighbours:
958          * we could search further afield for "compatible" anon_vmas.
959          * But it would probably just be a waste of time searching,
960          * or lead to too many vmas hanging off the same anon_vma.
961          * We're trying to allow mprotect remerging later on,
962          * not trying to minimize memory used for anon_vmas.
963          */
964         return NULL;
965 }
966
967 #ifdef CONFIG_PROC_FS
968 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
969                                                 struct file *file, long pages)
970 {
971         const unsigned long stack_flags
972                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
973
974         mm->total_vm += pages;
975
976         if (file) {
977                 mm->shared_vm += pages;
978                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
979                         mm->exec_vm += pages;
980         } else if (flags & stack_flags)
981                 mm->stack_vm += pages;
982 }
983 #endif /* CONFIG_PROC_FS */
984
985 /*
986  * If a hint addr is less than mmap_min_addr change hint to be as
987  * low as possible but still greater than mmap_min_addr
988  */
989 static inline unsigned long round_hint_to_min(unsigned long hint)
990 {
991         hint &= PAGE_MASK;
992         if (((void *)hint != NULL) &&
993             (hint < mmap_min_addr))
994                 return PAGE_ALIGN(mmap_min_addr);
995         return hint;
996 }
997
998 /*
999  * The caller must hold down_write(&current->mm->mmap_sem).
1000  */
1001
1002 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1003                         unsigned long len, unsigned long prot,
1004                         unsigned long flags, unsigned long pgoff)
1005 {
1006         struct mm_struct * mm = current->mm;
1007         struct inode *inode;
1008         vm_flags_t vm_flags;
1009
1010         /*
1011          * Does the application expect PROT_READ to imply PROT_EXEC?
1012          *
1013          * (the exception is when the underlying filesystem is noexec
1014          *  mounted, in which case we dont add PROT_EXEC.)
1015          */
1016         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1017                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1018                         prot |= PROT_EXEC;
1019
1020         if (!len)
1021                 return -EINVAL;
1022
1023         if (!(flags & MAP_FIXED))
1024                 addr = round_hint_to_min(addr);
1025
1026         /* Careful about overflows.. */
1027         len = PAGE_ALIGN(len);
1028         if (!len)
1029                 return -ENOMEM;
1030
1031         /* offset overflow? */
1032         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1033                return -EOVERFLOW;
1034
1035         /* Too many mappings? */
1036         if (mm->map_count > sysctl_max_map_count)
1037                 return -ENOMEM;
1038
1039         /* Obtain the address to map to. we verify (or select) it and ensure
1040          * that it represents a valid section of the address space.
1041          */
1042         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1043         if (addr & ~PAGE_MASK)
1044                 return addr;
1045
1046         /* Do simple checking here so the lower-level routines won't have
1047          * to. we assume access permissions have been handled by the open
1048          * of the memory object, so we don't do any here.
1049          */
1050         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1051                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1052
1053         if (flags & MAP_LOCKED)
1054                 if (!can_do_mlock())
1055                         return -EPERM;
1056
1057         /* mlock MCL_FUTURE? */
1058         if (vm_flags & VM_LOCKED) {
1059                 unsigned long locked, lock_limit;
1060                 locked = len >> PAGE_SHIFT;
1061                 locked += mm->locked_vm;
1062                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1063                 lock_limit >>= PAGE_SHIFT;
1064                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1065                         return -EAGAIN;
1066         }
1067
1068         inode = file ? file->f_path.dentry->d_inode : NULL;
1069
1070         if (file) {
1071                 switch (flags & MAP_TYPE) {
1072                 case MAP_SHARED:
1073                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1074                                 return -EACCES;
1075
1076                         /*
1077                          * Make sure we don't allow writing to an append-only
1078                          * file..
1079                          */
1080                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1081                                 return -EACCES;
1082
1083                         /*
1084                          * Make sure there are no mandatory locks on the file.
1085                          */
1086                         if (locks_verify_locked(inode))
1087                                 return -EAGAIN;
1088
1089                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1090                         if (!(file->f_mode & FMODE_WRITE))
1091                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1092
1093                         /* fall through */
1094                 case MAP_PRIVATE:
1095                         if (!(file->f_mode & FMODE_READ))
1096                                 return -EACCES;
1097                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1098                                 if (vm_flags & VM_EXEC)
1099                                         return -EPERM;
1100                                 vm_flags &= ~VM_MAYEXEC;
1101                         }
1102
1103                         if (!file->f_op || !file->f_op->mmap)
1104                                 return -ENODEV;
1105                         break;
1106
1107                 default:
1108                         return -EINVAL;
1109                 }
1110         } else {
1111                 switch (flags & MAP_TYPE) {
1112                 case MAP_SHARED:
1113                         /*
1114                          * Ignore pgoff.
1115                          */
1116                         pgoff = 0;
1117                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1118                         break;
1119                 case MAP_PRIVATE:
1120                         /*
1121                          * Set pgoff according to addr for anon_vma.
1122                          */
1123                         pgoff = addr >> PAGE_SHIFT;
1124                         break;
1125                 default:
1126                         return -EINVAL;
1127                 }
1128         }
1129
1130         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1131 }
1132
1133 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1134                 unsigned long, prot, unsigned long, flags,
1135                 unsigned long, fd, unsigned long, pgoff)
1136 {
1137         struct file *file = NULL;
1138         unsigned long retval = -EBADF;
1139
1140         if (!(flags & MAP_ANONYMOUS)) {
1141                 audit_mmap_fd(fd, flags);
1142                 if (unlikely(flags & MAP_HUGETLB))
1143                         return -EINVAL;
1144                 file = fget(fd);
1145                 if (!file)
1146                         goto out;
1147         } else if (flags & MAP_HUGETLB) {
1148                 struct user_struct *user = NULL;
1149                 /*
1150                  * VM_NORESERVE is used because the reservations will be
1151                  * taken when vm_ops->mmap() is called
1152                  * A dummy user value is used because we are not locking
1153                  * memory so no accounting is necessary
1154                  */
1155                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1156                                                 VM_NORESERVE, &user,
1157                                                 HUGETLB_ANONHUGE_INODE);
1158                 if (IS_ERR(file))
1159                         return PTR_ERR(file);
1160         }
1161
1162         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1163
1164         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1165         if (file)
1166                 fput(file);
1167 out:
1168         return retval;
1169 }
1170
1171 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1172 struct mmap_arg_struct {
1173         unsigned long addr;
1174         unsigned long len;
1175         unsigned long prot;
1176         unsigned long flags;
1177         unsigned long fd;
1178         unsigned long offset;
1179 };
1180
1181 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1182 {
1183         struct mmap_arg_struct a;
1184
1185         if (copy_from_user(&a, arg, sizeof(a)))
1186                 return -EFAULT;
1187         if (a.offset & ~PAGE_MASK)
1188                 return -EINVAL;
1189
1190         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1191                               a.offset >> PAGE_SHIFT);
1192 }
1193 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1194
1195 /*
1196  * Some shared mappigns will want the pages marked read-only
1197  * to track write events. If so, we'll downgrade vm_page_prot
1198  * to the private version (using protection_map[] without the
1199  * VM_SHARED bit).
1200  */
1201 int vma_wants_writenotify(struct vm_area_struct *vma)
1202 {
1203         vm_flags_t vm_flags = vma->vm_flags;
1204
1205         /* If it was private or non-writable, the write bit is already clear */
1206         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1207                 return 0;
1208
1209         /* The backer wishes to know when pages are first written to? */
1210         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1211                 return 1;
1212
1213         /* The open routine did something to the protections already? */
1214         if (pgprot_val(vma->vm_page_prot) !=
1215             pgprot_val(vm_get_page_prot(vm_flags)))
1216                 return 0;
1217
1218         /* Specialty mapping? */
1219         if (vm_flags & VM_PFNMAP)
1220                 return 0;
1221
1222         /* Can the mapping track the dirty pages? */
1223         return vma->vm_file && vma->vm_file->f_mapping &&
1224                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1225 }
1226
1227 /*
1228  * We account for memory if it's a private writeable mapping,
1229  * not hugepages and VM_NORESERVE wasn't set.
1230  */
1231 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1232 {
1233         /*
1234          * hugetlb has its own accounting separate from the core VM
1235          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1236          */
1237         if (file && is_file_hugepages(file))
1238                 return 0;
1239
1240         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1241 }
1242
1243 unsigned long mmap_region(struct file *file, unsigned long addr,
1244                           unsigned long len, unsigned long flags,
1245                           vm_flags_t vm_flags, unsigned long pgoff)
1246 {
1247         struct mm_struct *mm = current->mm;
1248         struct vm_area_struct *vma, *prev;
1249         int correct_wcount = 0;
1250         int error;
1251         struct rb_node **rb_link, *rb_parent;
1252         unsigned long charged = 0;
1253         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1254
1255         /* Clear old maps */
1256         error = -ENOMEM;
1257 munmap_back:
1258         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1259                 if (do_munmap(mm, addr, len))
1260                         return -ENOMEM;
1261                 goto munmap_back;
1262         }
1263
1264         /* Check against address space limit. */
1265         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1266                 return -ENOMEM;
1267
1268         /*
1269          * Set 'VM_NORESERVE' if we should not account for the
1270          * memory use of this mapping.
1271          */
1272         if ((flags & MAP_NORESERVE)) {
1273                 /* We honor MAP_NORESERVE if allowed to overcommit */
1274                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1275                         vm_flags |= VM_NORESERVE;
1276
1277                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1278                 if (file && is_file_hugepages(file))
1279                         vm_flags |= VM_NORESERVE;
1280         }
1281
1282         /*
1283          * Private writable mapping: check memory availability
1284          */
1285         if (accountable_mapping(file, vm_flags)) {
1286                 charged = len >> PAGE_SHIFT;
1287                 if (security_vm_enough_memory_mm(mm, charged))
1288                         return -ENOMEM;
1289                 vm_flags |= VM_ACCOUNT;
1290         }
1291
1292         /*
1293          * Can we just expand an old mapping?
1294          */
1295         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1296         if (vma)
1297                 goto out;
1298
1299         /*
1300          * Determine the object being mapped and call the appropriate
1301          * specific mapper. the address has already been validated, but
1302          * not unmapped, but the maps are removed from the list.
1303          */
1304         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1305         if (!vma) {
1306                 error = -ENOMEM;
1307                 goto unacct_error;
1308         }
1309
1310         vma->vm_mm = mm;
1311         vma->vm_start = addr;
1312         vma->vm_end = addr + len;
1313         vma->vm_flags = vm_flags;
1314         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1315         vma->vm_pgoff = pgoff;
1316         INIT_LIST_HEAD(&vma->anon_vma_chain);
1317
1318         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1319
1320         if (file) {
1321                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1322                         goto free_vma;
1323                 if (vm_flags & VM_DENYWRITE) {
1324                         error = deny_write_access(file);
1325                         if (error)
1326                                 goto free_vma;
1327                         correct_wcount = 1;
1328                 }
1329                 vma->vm_file = get_file(file);
1330                 error = file->f_op->mmap(file, vma);
1331                 if (error)
1332                         goto unmap_and_free_vma;
1333
1334                 /* Can addr have changed??
1335                  *
1336                  * Answer: Yes, several device drivers can do it in their
1337                  *         f_op->mmap method. -DaveM
1338                  */
1339                 addr = vma->vm_start;
1340                 pgoff = vma->vm_pgoff;
1341                 vm_flags = vma->vm_flags;
1342         } else if (vm_flags & VM_SHARED) {
1343                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1344                         goto free_vma;
1345                 error = shmem_zero_setup(vma);
1346                 if (error)
1347                         goto free_vma;
1348         }
1349
1350         if (vma_wants_writenotify(vma)) {
1351                 pgprot_t pprot = vma->vm_page_prot;
1352
1353                 /* Can vma->vm_page_prot have changed??
1354                  *
1355                  * Answer: Yes, drivers may have changed it in their
1356                  *         f_op->mmap method.
1357                  *
1358                  * Ensures that vmas marked as uncached stay that way.
1359                  */
1360                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1361                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1362                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1363         }
1364
1365         vma_link(mm, vma, prev, rb_link, rb_parent);
1366         file = vma->vm_file;
1367
1368         /* Once vma denies write, undo our temporary denial count */
1369         if (correct_wcount)
1370                 atomic_inc(&inode->i_writecount);
1371 out:
1372         perf_event_mmap(vma);
1373
1374         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1375         if (vm_flags & VM_LOCKED) {
1376                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1377                         mm->locked_vm += (len >> PAGE_SHIFT);
1378         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1379                 make_pages_present(addr, addr + len);
1380
1381         if (file)
1382                 uprobe_mmap(vma);
1383
1384         return addr;
1385
1386 unmap_and_free_vma:
1387         if (correct_wcount)
1388                 atomic_inc(&inode->i_writecount);
1389         vma->vm_file = NULL;
1390         fput(file);
1391
1392         /* Undo any partial mapping done by a device driver. */
1393         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1394         charged = 0;
1395 free_vma:
1396         kmem_cache_free(vm_area_cachep, vma);
1397 unacct_error:
1398         if (charged)
1399                 vm_unacct_memory(charged);
1400         return error;
1401 }
1402
1403 /* Get an address range which is currently unmapped.
1404  * For shmat() with addr=0.
1405  *
1406  * Ugly calling convention alert:
1407  * Return value with the low bits set means error value,
1408  * ie
1409  *      if (ret & ~PAGE_MASK)
1410  *              error = ret;
1411  *
1412  * This function "knows" that -ENOMEM has the bits set.
1413  */
1414 #ifndef HAVE_ARCH_UNMAPPED_AREA
1415 unsigned long
1416 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1417                 unsigned long len, unsigned long pgoff, unsigned long flags)
1418 {
1419         struct mm_struct *mm = current->mm;
1420         struct vm_area_struct *vma;
1421         unsigned long start_addr;
1422
1423         if (len > TASK_SIZE)
1424                 return -ENOMEM;
1425
1426         if (flags & MAP_FIXED)
1427                 return addr;
1428
1429         if (addr) {
1430                 addr = PAGE_ALIGN(addr);
1431                 vma = find_vma(mm, addr);
1432                 if (TASK_SIZE - len >= addr &&
1433                     (!vma || addr + len <= vma->vm_start))
1434                         return addr;
1435         }
1436         if (len > mm->cached_hole_size) {
1437                 start_addr = addr = mm->free_area_cache;
1438         } else {
1439                 start_addr = addr = TASK_UNMAPPED_BASE;
1440                 mm->cached_hole_size = 0;
1441         }
1442
1443 full_search:
1444         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1445                 /* At this point:  (!vma || addr < vma->vm_end). */
1446                 if (TASK_SIZE - len < addr) {
1447                         /*
1448                          * Start a new search - just in case we missed
1449                          * some holes.
1450                          */
1451                         if (start_addr != TASK_UNMAPPED_BASE) {
1452                                 addr = TASK_UNMAPPED_BASE;
1453                                 start_addr = addr;
1454                                 mm->cached_hole_size = 0;
1455                                 goto full_search;
1456                         }
1457                         return -ENOMEM;
1458                 }
1459                 if (!vma || addr + len <= vma->vm_start) {
1460                         /*
1461                          * Remember the place where we stopped the search:
1462                          */
1463                         mm->free_area_cache = addr + len;
1464                         return addr;
1465                 }
1466                 if (addr + mm->cached_hole_size < vma->vm_start)
1467                         mm->cached_hole_size = vma->vm_start - addr;
1468                 addr = vma->vm_end;
1469         }
1470 }
1471 #endif  
1472
1473 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1474 {
1475         /*
1476          * Is this a new hole at the lowest possible address?
1477          */
1478         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1479                 mm->free_area_cache = addr;
1480 }
1481
1482 /*
1483  * This mmap-allocator allocates new areas top-down from below the
1484  * stack's low limit (the base):
1485  */
1486 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1487 unsigned long
1488 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1489                           const unsigned long len, const unsigned long pgoff,
1490                           const unsigned long flags)
1491 {
1492         struct vm_area_struct *vma;
1493         struct mm_struct *mm = current->mm;
1494         unsigned long addr = addr0, start_addr;
1495
1496         /* requested length too big for entire address space */
1497         if (len > TASK_SIZE)
1498                 return -ENOMEM;
1499
1500         if (flags & MAP_FIXED)
1501                 return addr;
1502
1503         /* requesting a specific address */
1504         if (addr) {
1505                 addr = PAGE_ALIGN(addr);
1506                 vma = find_vma(mm, addr);
1507                 if (TASK_SIZE - len >= addr &&
1508                                 (!vma || addr + len <= vma->vm_start))
1509                         return addr;
1510         }
1511
1512         /* check if free_area_cache is useful for us */
1513         if (len <= mm->cached_hole_size) {
1514                 mm->cached_hole_size = 0;
1515                 mm->free_area_cache = mm->mmap_base;
1516         }
1517
1518 try_again:
1519         /* either no address requested or can't fit in requested address hole */
1520         start_addr = addr = mm->free_area_cache;
1521
1522         if (addr < len)
1523                 goto fail;
1524
1525         addr -= len;
1526         do {
1527                 /*
1528                  * Lookup failure means no vma is above this address,
1529                  * else if new region fits below vma->vm_start,
1530                  * return with success:
1531                  */
1532                 vma = find_vma(mm, addr);
1533                 if (!vma || addr+len <= vma->vm_start)
1534                         /* remember the address as a hint for next time */
1535                         return (mm->free_area_cache = addr);
1536
1537                 /* remember the largest hole we saw so far */
1538                 if (addr + mm->cached_hole_size < vma->vm_start)
1539                         mm->cached_hole_size = vma->vm_start - addr;
1540
1541                 /* try just below the current vma->vm_start */
1542                 addr = vma->vm_start-len;
1543         } while (len < vma->vm_start);
1544
1545 fail:
1546         /*
1547          * if hint left us with no space for the requested
1548          * mapping then try again:
1549          *
1550          * Note: this is different with the case of bottomup
1551          * which does the fully line-search, but we use find_vma
1552          * here that causes some holes skipped.
1553          */
1554         if (start_addr != mm->mmap_base) {
1555                 mm->free_area_cache = mm->mmap_base;
1556                 mm->cached_hole_size = 0;
1557                 goto try_again;
1558         }
1559
1560         /*
1561          * A failed mmap() very likely causes application failure,
1562          * so fall back to the bottom-up function here. This scenario
1563          * can happen with large stack limits and large mmap()
1564          * allocations.
1565          */
1566         mm->cached_hole_size = ~0UL;
1567         mm->free_area_cache = TASK_UNMAPPED_BASE;
1568         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1569         /*
1570          * Restore the topdown base:
1571          */
1572         mm->free_area_cache = mm->mmap_base;
1573         mm->cached_hole_size = ~0UL;
1574
1575         return addr;
1576 }
1577 #endif
1578
1579 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1580 {
1581         /*
1582          * Is this a new hole at the highest possible address?
1583          */
1584         if (addr > mm->free_area_cache)
1585                 mm->free_area_cache = addr;
1586
1587         /* dont allow allocations above current base */
1588         if (mm->free_area_cache > mm->mmap_base)
1589                 mm->free_area_cache = mm->mmap_base;
1590 }
1591
1592 unsigned long
1593 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1594                 unsigned long pgoff, unsigned long flags)
1595 {
1596         unsigned long (*get_area)(struct file *, unsigned long,
1597                                   unsigned long, unsigned long, unsigned long);
1598
1599         unsigned long error = arch_mmap_check(addr, len, flags);
1600         if (error)
1601                 return error;
1602
1603         /* Careful about overflows.. */
1604         if (len > TASK_SIZE)
1605                 return -ENOMEM;
1606
1607         get_area = current->mm->get_unmapped_area;
1608         if (file && file->f_op && file->f_op->get_unmapped_area)
1609                 get_area = file->f_op->get_unmapped_area;
1610         addr = get_area(file, addr, len, pgoff, flags);
1611         if (IS_ERR_VALUE(addr))
1612                 return addr;
1613
1614         if (addr > TASK_SIZE - len)
1615                 return -ENOMEM;
1616         if (addr & ~PAGE_MASK)
1617                 return -EINVAL;
1618
1619         addr = arch_rebalance_pgtables(addr, len);
1620         error = security_mmap_addr(addr);
1621         return error ? error : addr;
1622 }
1623
1624 EXPORT_SYMBOL(get_unmapped_area);
1625
1626 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1627 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1628 {
1629         struct vm_area_struct *vma = NULL;
1630
1631         if (WARN_ON_ONCE(!mm))          /* Remove this in linux-3.6 */
1632                 return NULL;
1633
1634         /* Check the cache first. */
1635         /* (Cache hit rate is typically around 35%.) */
1636         vma = mm->mmap_cache;
1637         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1638                 struct rb_node *rb_node;
1639
1640                 rb_node = mm->mm_rb.rb_node;
1641                 vma = NULL;
1642
1643                 while (rb_node) {
1644                         struct vm_area_struct *vma_tmp;
1645
1646                         vma_tmp = rb_entry(rb_node,
1647                                            struct vm_area_struct, vm_rb);
1648
1649                         if (vma_tmp->vm_end > addr) {
1650                                 vma = vma_tmp;
1651                                 if (vma_tmp->vm_start <= addr)
1652                                         break;
1653                                 rb_node = rb_node->rb_left;
1654                         } else
1655                                 rb_node = rb_node->rb_right;
1656                 }
1657                 if (vma)
1658                         mm->mmap_cache = vma;
1659         }
1660         return vma;
1661 }
1662
1663 EXPORT_SYMBOL(find_vma);
1664
1665 /*
1666  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1667  */
1668 struct vm_area_struct *
1669 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1670                         struct vm_area_struct **pprev)
1671 {
1672         struct vm_area_struct *vma;
1673
1674         vma = find_vma(mm, addr);
1675         if (vma) {
1676                 *pprev = vma->vm_prev;
1677         } else {
1678                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1679                 *pprev = NULL;
1680                 while (rb_node) {
1681                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1682                         rb_node = rb_node->rb_right;
1683                 }
1684         }
1685         return vma;
1686 }
1687
1688 /*
1689  * Verify that the stack growth is acceptable and
1690  * update accounting. This is shared with both the
1691  * grow-up and grow-down cases.
1692  */
1693 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1694 {
1695         struct mm_struct *mm = vma->vm_mm;
1696         struct rlimit *rlim = current->signal->rlim;
1697         unsigned long new_start;
1698
1699         /* address space limit tests */
1700         if (!may_expand_vm(mm, grow))
1701                 return -ENOMEM;
1702
1703         /* Stack limit test */
1704         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1705                 return -ENOMEM;
1706
1707         /* mlock limit tests */
1708         if (vma->vm_flags & VM_LOCKED) {
1709                 unsigned long locked;
1710                 unsigned long limit;
1711                 locked = mm->locked_vm + grow;
1712                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1713                 limit >>= PAGE_SHIFT;
1714                 if (locked > limit && !capable(CAP_IPC_LOCK))
1715                         return -ENOMEM;
1716         }
1717
1718         /* Check to ensure the stack will not grow into a hugetlb-only region */
1719         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1720                         vma->vm_end - size;
1721         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1722                 return -EFAULT;
1723
1724         /*
1725          * Overcommit..  This must be the final test, as it will
1726          * update security statistics.
1727          */
1728         if (security_vm_enough_memory_mm(mm, grow))
1729                 return -ENOMEM;
1730
1731         /* Ok, everything looks good - let it rip */
1732         if (vma->vm_flags & VM_LOCKED)
1733                 mm->locked_vm += grow;
1734         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1735         return 0;
1736 }
1737
1738 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1739 /*
1740  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1741  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1742  */
1743 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1744 {
1745         int error;
1746
1747         if (!(vma->vm_flags & VM_GROWSUP))
1748                 return -EFAULT;
1749
1750         /*
1751          * We must make sure the anon_vma is allocated
1752          * so that the anon_vma locking is not a noop.
1753          */
1754         if (unlikely(anon_vma_prepare(vma)))
1755                 return -ENOMEM;
1756         vma_lock_anon_vma(vma);
1757
1758         /*
1759          * vma->vm_start/vm_end cannot change under us because the caller
1760          * is required to hold the mmap_sem in read mode.  We need the
1761          * anon_vma lock to serialize against concurrent expand_stacks.
1762          * Also guard against wrapping around to address 0.
1763          */
1764         if (address < PAGE_ALIGN(address+4))
1765                 address = PAGE_ALIGN(address+4);
1766         else {
1767                 vma_unlock_anon_vma(vma);
1768                 return -ENOMEM;
1769         }
1770         error = 0;
1771
1772         /* Somebody else might have raced and expanded it already */
1773         if (address > vma->vm_end) {
1774                 unsigned long size, grow;
1775
1776                 size = address - vma->vm_start;
1777                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1778
1779                 error = -ENOMEM;
1780                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1781                         error = acct_stack_growth(vma, size, grow);
1782                         if (!error) {
1783                                 anon_vma_interval_tree_pre_update_vma(vma);
1784                                 vma->vm_end = address;
1785                                 anon_vma_interval_tree_post_update_vma(vma);
1786                                 perf_event_mmap(vma);
1787                         }
1788                 }
1789         }
1790         vma_unlock_anon_vma(vma);
1791         khugepaged_enter_vma_merge(vma);
1792         validate_mm(vma->vm_mm);
1793         return error;
1794 }
1795 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1796
1797 /*
1798  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1799  */
1800 int expand_downwards(struct vm_area_struct *vma,
1801                                    unsigned long address)
1802 {
1803         int error;
1804
1805         /*
1806          * We must make sure the anon_vma is allocated
1807          * so that the anon_vma locking is not a noop.
1808          */
1809         if (unlikely(anon_vma_prepare(vma)))
1810                 return -ENOMEM;
1811
1812         address &= PAGE_MASK;
1813         error = security_mmap_addr(address);
1814         if (error)
1815                 return error;
1816
1817         vma_lock_anon_vma(vma);
1818
1819         /*
1820          * vma->vm_start/vm_end cannot change under us because the caller
1821          * is required to hold the mmap_sem in read mode.  We need the
1822          * anon_vma lock to serialize against concurrent expand_stacks.
1823          */
1824
1825         /* Somebody else might have raced and expanded it already */
1826         if (address < vma->vm_start) {
1827                 unsigned long size, grow;
1828
1829                 size = vma->vm_end - address;
1830                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1831
1832                 error = -ENOMEM;
1833                 if (grow <= vma->vm_pgoff) {
1834                         error = acct_stack_growth(vma, size, grow);
1835                         if (!error) {
1836                                 anon_vma_interval_tree_pre_update_vma(vma);
1837                                 vma->vm_start = address;
1838                                 vma->vm_pgoff -= grow;
1839                                 anon_vma_interval_tree_post_update_vma(vma);
1840                                 perf_event_mmap(vma);
1841                         }
1842                 }
1843         }
1844         vma_unlock_anon_vma(vma);
1845         khugepaged_enter_vma_merge(vma);
1846         validate_mm(vma->vm_mm);
1847         return error;
1848 }
1849
1850 #ifdef CONFIG_STACK_GROWSUP
1851 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1852 {
1853         return expand_upwards(vma, address);
1854 }
1855
1856 struct vm_area_struct *
1857 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1858 {
1859         struct vm_area_struct *vma, *prev;
1860
1861         addr &= PAGE_MASK;
1862         vma = find_vma_prev(mm, addr, &prev);
1863         if (vma && (vma->vm_start <= addr))
1864                 return vma;
1865         if (!prev || expand_stack(prev, addr))
1866                 return NULL;
1867         if (prev->vm_flags & VM_LOCKED) {
1868                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1869         }
1870         return prev;
1871 }
1872 #else
1873 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1874 {
1875         return expand_downwards(vma, address);
1876 }
1877
1878 struct vm_area_struct *
1879 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1880 {
1881         struct vm_area_struct * vma;
1882         unsigned long start;
1883
1884         addr &= PAGE_MASK;
1885         vma = find_vma(mm,addr);
1886         if (!vma)
1887                 return NULL;
1888         if (vma->vm_start <= addr)
1889                 return vma;
1890         if (!(vma->vm_flags & VM_GROWSDOWN))
1891                 return NULL;
1892         start = vma->vm_start;
1893         if (expand_stack(vma, addr))
1894                 return NULL;
1895         if (vma->vm_flags & VM_LOCKED) {
1896                 mlock_vma_pages_range(vma, addr, start);
1897         }
1898         return vma;
1899 }
1900 #endif
1901
1902 /*
1903  * Ok - we have the memory areas we should free on the vma list,
1904  * so release them, and do the vma updates.
1905  *
1906  * Called with the mm semaphore held.
1907  */
1908 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1909 {
1910         unsigned long nr_accounted = 0;
1911
1912         /* Update high watermark before we lower total_vm */
1913         update_hiwater_vm(mm);
1914         do {
1915                 long nrpages = vma_pages(vma);
1916
1917                 if (vma->vm_flags & VM_ACCOUNT)
1918                         nr_accounted += nrpages;
1919                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1920                 vma = remove_vma(vma);
1921         } while (vma);
1922         vm_unacct_memory(nr_accounted);
1923         validate_mm(mm);
1924 }
1925
1926 /*
1927  * Get rid of page table information in the indicated region.
1928  *
1929  * Called with the mm semaphore held.
1930  */
1931 static void unmap_region(struct mm_struct *mm,
1932                 struct vm_area_struct *vma, struct vm_area_struct *prev,
1933                 unsigned long start, unsigned long end)
1934 {
1935         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1936         struct mmu_gather tlb;
1937
1938         lru_add_drain();
1939         tlb_gather_mmu(&tlb, mm, 0);
1940         update_hiwater_rss(mm);
1941         unmap_vmas(&tlb, vma, start, end);
1942         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
1943                                  next ? next->vm_start : 0);
1944         tlb_finish_mmu(&tlb, start, end);
1945 }
1946
1947 /*
1948  * Create a list of vma's touched by the unmap, removing them from the mm's
1949  * vma list as we go..
1950  */
1951 static void
1952 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1953         struct vm_area_struct *prev, unsigned long end)
1954 {
1955         struct vm_area_struct **insertion_point;
1956         struct vm_area_struct *tail_vma = NULL;
1957         unsigned long addr;
1958
1959         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1960         vma->vm_prev = NULL;
1961         do {
1962                 rb_erase(&vma->vm_rb, &mm->mm_rb);
1963                 mm->map_count--;
1964                 tail_vma = vma;
1965                 vma = vma->vm_next;
1966         } while (vma && vma->vm_start < end);
1967         *insertion_point = vma;
1968         if (vma)
1969                 vma->vm_prev = prev;
1970         tail_vma->vm_next = NULL;
1971         if (mm->unmap_area == arch_unmap_area)
1972                 addr = prev ? prev->vm_end : mm->mmap_base;
1973         else
1974                 addr = vma ?  vma->vm_start : mm->mmap_base;
1975         mm->unmap_area(mm, addr);
1976         mm->mmap_cache = NULL;          /* Kill the cache. */
1977 }
1978
1979 /*
1980  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
1981  * munmap path where it doesn't make sense to fail.
1982  */
1983 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1984               unsigned long addr, int new_below)
1985 {
1986         struct mempolicy *pol;
1987         struct vm_area_struct *new;
1988         int err = -ENOMEM;
1989
1990         if (is_vm_hugetlb_page(vma) && (addr &
1991                                         ~(huge_page_mask(hstate_vma(vma)))))
1992                 return -EINVAL;
1993
1994         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1995         if (!new)
1996                 goto out_err;
1997
1998         /* most fields are the same, copy all, and then fixup */
1999         *new = *vma;
2000
2001         INIT_LIST_HEAD(&new->anon_vma_chain);
2002
2003         if (new_below)
2004                 new->vm_end = addr;
2005         else {
2006                 new->vm_start = addr;
2007                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2008         }
2009
2010         pol = mpol_dup(vma_policy(vma));
2011         if (IS_ERR(pol)) {
2012                 err = PTR_ERR(pol);
2013                 goto out_free_vma;
2014         }
2015         vma_set_policy(new, pol);
2016
2017         if (anon_vma_clone(new, vma))
2018                 goto out_free_mpol;
2019
2020         if (new->vm_file)
2021                 get_file(new->vm_file);
2022
2023         if (new->vm_ops && new->vm_ops->open)
2024                 new->vm_ops->open(new);
2025
2026         if (new_below)
2027                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2028                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2029         else
2030                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2031
2032         /* Success. */
2033         if (!err)
2034                 return 0;
2035
2036         /* Clean everything up if vma_adjust failed. */
2037         if (new->vm_ops && new->vm_ops->close)
2038                 new->vm_ops->close(new);
2039         if (new->vm_file)
2040                 fput(new->vm_file);
2041         unlink_anon_vmas(new);
2042  out_free_mpol:
2043         mpol_put(pol);
2044  out_free_vma:
2045         kmem_cache_free(vm_area_cachep, new);
2046  out_err:
2047         return err;
2048 }
2049
2050 /*
2051  * Split a vma into two pieces at address 'addr', a new vma is allocated
2052  * either for the first part or the tail.
2053  */
2054 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2055               unsigned long addr, int new_below)
2056 {
2057         if (mm->map_count >= sysctl_max_map_count)
2058                 return -ENOMEM;
2059
2060         return __split_vma(mm, vma, addr, new_below);
2061 }
2062
2063 /* Munmap is split into 2 main parts -- this part which finds
2064  * what needs doing, and the areas themselves, which do the
2065  * work.  This now handles partial unmappings.
2066  * Jeremy Fitzhardinge <jeremy@goop.org>
2067  */
2068 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2069 {
2070         unsigned long end;
2071         struct vm_area_struct *vma, *prev, *last;
2072
2073         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2074                 return -EINVAL;
2075
2076         if ((len = PAGE_ALIGN(len)) == 0)
2077                 return -EINVAL;
2078
2079         /* Find the first overlapping VMA */
2080         vma = find_vma(mm, start);
2081         if (!vma)
2082                 return 0;
2083         prev = vma->vm_prev;
2084         /* we have  start < vma->vm_end  */
2085
2086         /* if it doesn't overlap, we have nothing.. */
2087         end = start + len;
2088         if (vma->vm_start >= end)
2089                 return 0;
2090
2091         /*
2092          * If we need to split any vma, do it now to save pain later.
2093          *
2094          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2095          * unmapped vm_area_struct will remain in use: so lower split_vma
2096          * places tmp vma above, and higher split_vma places tmp vma below.
2097          */
2098         if (start > vma->vm_start) {
2099                 int error;
2100
2101                 /*
2102                  * Make sure that map_count on return from munmap() will
2103                  * not exceed its limit; but let map_count go just above
2104                  * its limit temporarily, to help free resources as expected.
2105                  */
2106                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2107                         return -ENOMEM;
2108
2109                 error = __split_vma(mm, vma, start, 0);
2110                 if (error)
2111                         return error;
2112                 prev = vma;
2113         }
2114
2115         /* Does it split the last one? */
2116         last = find_vma(mm, end);
2117         if (last && end > last->vm_start) {
2118                 int error = __split_vma(mm, last, end, 1);
2119                 if (error)
2120                         return error;
2121         }
2122         vma = prev? prev->vm_next: mm->mmap;
2123
2124         /*
2125          * unlock any mlock()ed ranges before detaching vmas
2126          */
2127         if (mm->locked_vm) {
2128                 struct vm_area_struct *tmp = vma;
2129                 while (tmp && tmp->vm_start < end) {
2130                         if (tmp->vm_flags & VM_LOCKED) {
2131                                 mm->locked_vm -= vma_pages(tmp);
2132                                 munlock_vma_pages_all(tmp);
2133                         }
2134                         tmp = tmp->vm_next;
2135                 }
2136         }
2137
2138         /*
2139          * Remove the vma's, and unmap the actual pages
2140          */
2141         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2142         unmap_region(mm, vma, prev, start, end);
2143
2144         /* Fix up all other VM information */
2145         remove_vma_list(mm, vma);
2146
2147         return 0;
2148 }
2149
2150 int vm_munmap(unsigned long start, size_t len)
2151 {
2152         int ret;
2153         struct mm_struct *mm = current->mm;
2154
2155         down_write(&mm->mmap_sem);
2156         ret = do_munmap(mm, start, len);
2157         up_write(&mm->mmap_sem);
2158         return ret;
2159 }
2160 EXPORT_SYMBOL(vm_munmap);
2161
2162 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2163 {
2164         profile_munmap(addr);
2165         return vm_munmap(addr, len);
2166 }
2167
2168 static inline void verify_mm_writelocked(struct mm_struct *mm)
2169 {
2170 #ifdef CONFIG_DEBUG_VM
2171         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2172                 WARN_ON(1);
2173                 up_read(&mm->mmap_sem);
2174         }
2175 #endif
2176 }
2177
2178 /*
2179  *  this is really a simplified "do_mmap".  it only handles
2180  *  anonymous maps.  eventually we may be able to do some
2181  *  brk-specific accounting here.
2182  */
2183 static unsigned long do_brk(unsigned long addr, unsigned long len)
2184 {
2185         struct mm_struct * mm = current->mm;
2186         struct vm_area_struct * vma, * prev;
2187         unsigned long flags;
2188         struct rb_node ** rb_link, * rb_parent;
2189         pgoff_t pgoff = addr >> PAGE_SHIFT;
2190         int error;
2191
2192         len = PAGE_ALIGN(len);
2193         if (!len)
2194                 return addr;
2195
2196         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2197
2198         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2199         if (error & ~PAGE_MASK)
2200                 return error;
2201
2202         /*
2203          * mlock MCL_FUTURE?
2204          */
2205         if (mm->def_flags & VM_LOCKED) {
2206                 unsigned long locked, lock_limit;
2207                 locked = len >> PAGE_SHIFT;
2208                 locked += mm->locked_vm;
2209                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2210                 lock_limit >>= PAGE_SHIFT;
2211                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2212                         return -EAGAIN;
2213         }
2214
2215         /*
2216          * mm->mmap_sem is required to protect against another thread
2217          * changing the mappings in case we sleep.
2218          */
2219         verify_mm_writelocked(mm);
2220
2221         /*
2222          * Clear old maps.  this also does some error checking for us
2223          */
2224  munmap_back:
2225         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2226                 if (do_munmap(mm, addr, len))
2227                         return -ENOMEM;
2228                 goto munmap_back;
2229         }
2230
2231         /* Check against address space limits *after* clearing old maps... */
2232         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2233                 return -ENOMEM;
2234
2235         if (mm->map_count > sysctl_max_map_count)
2236                 return -ENOMEM;
2237
2238         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2239                 return -ENOMEM;
2240
2241         /* Can we just expand an old private anonymous mapping? */
2242         vma = vma_merge(mm, prev, addr, addr + len, flags,
2243                                         NULL, NULL, pgoff, NULL);
2244         if (vma)
2245                 goto out;
2246
2247         /*
2248          * create a vma struct for an anonymous mapping
2249          */
2250         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2251         if (!vma) {
2252                 vm_unacct_memory(len >> PAGE_SHIFT);
2253                 return -ENOMEM;
2254         }
2255
2256         INIT_LIST_HEAD(&vma->anon_vma_chain);
2257         vma->vm_mm = mm;
2258         vma->vm_start = addr;
2259         vma->vm_end = addr + len;
2260         vma->vm_pgoff = pgoff;
2261         vma->vm_flags = flags;
2262         vma->vm_page_prot = vm_get_page_prot(flags);
2263         vma_link(mm, vma, prev, rb_link, rb_parent);
2264 out:
2265         perf_event_mmap(vma);
2266         mm->total_vm += len >> PAGE_SHIFT;
2267         if (flags & VM_LOCKED) {
2268                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2269                         mm->locked_vm += (len >> PAGE_SHIFT);
2270         }
2271         return addr;
2272 }
2273
2274 unsigned long vm_brk(unsigned long addr, unsigned long len)
2275 {
2276         struct mm_struct *mm = current->mm;
2277         unsigned long ret;
2278
2279         down_write(&mm->mmap_sem);
2280         ret = do_brk(addr, len);
2281         up_write(&mm->mmap_sem);
2282         return ret;
2283 }
2284 EXPORT_SYMBOL(vm_brk);
2285
2286 /* Release all mmaps. */
2287 void exit_mmap(struct mm_struct *mm)
2288 {
2289         struct mmu_gather tlb;
2290         struct vm_area_struct *vma;
2291         unsigned long nr_accounted = 0;
2292
2293         /* mm's last user has gone, and its about to be pulled down */
2294         mmu_notifier_release(mm);
2295
2296         if (mm->locked_vm) {
2297                 vma = mm->mmap;
2298                 while (vma) {
2299                         if (vma->vm_flags & VM_LOCKED)
2300                                 munlock_vma_pages_all(vma);
2301                         vma = vma->vm_next;
2302                 }
2303         }
2304
2305         arch_exit_mmap(mm);
2306
2307         vma = mm->mmap;
2308         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2309                 return;
2310
2311         lru_add_drain();
2312         flush_cache_mm(mm);
2313         tlb_gather_mmu(&tlb, mm, 1);
2314         /* update_hiwater_rss(mm) here? but nobody should be looking */
2315         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2316         unmap_vmas(&tlb, vma, 0, -1);
2317
2318         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2319         tlb_finish_mmu(&tlb, 0, -1);
2320
2321         /*
2322          * Walk the list again, actually closing and freeing it,
2323          * with preemption enabled, without holding any MM locks.
2324          */
2325         while (vma) {
2326                 if (vma->vm_flags & VM_ACCOUNT)
2327                         nr_accounted += vma_pages(vma);
2328                 vma = remove_vma(vma);
2329         }
2330         vm_unacct_memory(nr_accounted);
2331
2332         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2333 }
2334
2335 /* Insert vm structure into process list sorted by address
2336  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2337  * then i_mmap_mutex is taken here.
2338  */
2339 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2340 {
2341         struct vm_area_struct *prev;
2342         struct rb_node **rb_link, *rb_parent;
2343
2344         /*
2345          * The vm_pgoff of a purely anonymous vma should be irrelevant
2346          * until its first write fault, when page's anon_vma and index
2347          * are set.  But now set the vm_pgoff it will almost certainly
2348          * end up with (unless mremap moves it elsewhere before that
2349          * first wfault), so /proc/pid/maps tells a consistent story.
2350          *
2351          * By setting it to reflect the virtual start address of the
2352          * vma, merges and splits can happen in a seamless way, just
2353          * using the existing file pgoff checks and manipulations.
2354          * Similarly in do_mmap_pgoff and in do_brk.
2355          */
2356         if (!vma->vm_file) {
2357                 BUG_ON(vma->anon_vma);
2358                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2359         }
2360         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2361                            &prev, &rb_link, &rb_parent))
2362                 return -ENOMEM;
2363         if ((vma->vm_flags & VM_ACCOUNT) &&
2364              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2365                 return -ENOMEM;
2366
2367         vma_link(mm, vma, prev, rb_link, rb_parent);
2368         return 0;
2369 }
2370
2371 /*
2372  * Copy the vma structure to a new location in the same mm,
2373  * prior to moving page table entries, to effect an mremap move.
2374  */
2375 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2376         unsigned long addr, unsigned long len, pgoff_t pgoff,
2377         bool *need_rmap_locks)
2378 {
2379         struct vm_area_struct *vma = *vmap;
2380         unsigned long vma_start = vma->vm_start;
2381         struct mm_struct *mm = vma->vm_mm;
2382         struct vm_area_struct *new_vma, *prev;
2383         struct rb_node **rb_link, *rb_parent;
2384         struct mempolicy *pol;
2385         bool faulted_in_anon_vma = true;
2386
2387         /*
2388          * If anonymous vma has not yet been faulted, update new pgoff
2389          * to match new location, to increase its chance of merging.
2390          */
2391         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2392                 pgoff = addr >> PAGE_SHIFT;
2393                 faulted_in_anon_vma = false;
2394         }
2395
2396         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2397                 return NULL;    /* should never get here */
2398         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2399                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2400         if (new_vma) {
2401                 /*
2402                  * Source vma may have been merged into new_vma
2403                  */
2404                 if (unlikely(vma_start >= new_vma->vm_start &&
2405                              vma_start < new_vma->vm_end)) {
2406                         /*
2407                          * The only way we can get a vma_merge with
2408                          * self during an mremap is if the vma hasn't
2409                          * been faulted in yet and we were allowed to
2410                          * reset the dst vma->vm_pgoff to the
2411                          * destination address of the mremap to allow
2412                          * the merge to happen. mremap must change the
2413                          * vm_pgoff linearity between src and dst vmas
2414                          * (in turn preventing a vma_merge) to be
2415                          * safe. It is only safe to keep the vm_pgoff
2416                          * linear if there are no pages mapped yet.
2417                          */
2418                         VM_BUG_ON(faulted_in_anon_vma);
2419                         *vmap = vma = new_vma;
2420                 }
2421                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2422         } else {
2423                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2424                 if (new_vma) {
2425                         *new_vma = *vma;
2426                         new_vma->vm_start = addr;
2427                         new_vma->vm_end = addr + len;
2428                         new_vma->vm_pgoff = pgoff;
2429                         pol = mpol_dup(vma_policy(vma));
2430                         if (IS_ERR(pol))
2431                                 goto out_free_vma;
2432                         vma_set_policy(new_vma, pol);
2433                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2434                         if (anon_vma_clone(new_vma, vma))
2435                                 goto out_free_mempol;
2436                         if (new_vma->vm_file)
2437                                 get_file(new_vma->vm_file);
2438                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2439                                 new_vma->vm_ops->open(new_vma);
2440                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2441                         *need_rmap_locks = false;
2442                 }
2443         }
2444         return new_vma;
2445
2446  out_free_mempol:
2447         mpol_put(pol);
2448  out_free_vma:
2449         kmem_cache_free(vm_area_cachep, new_vma);
2450         return NULL;
2451 }
2452
2453 /*
2454  * Return true if the calling process may expand its vm space by the passed
2455  * number of pages
2456  */
2457 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2458 {
2459         unsigned long cur = mm->total_vm;       /* pages */
2460         unsigned long lim;
2461
2462         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2463
2464         if (cur + npages > lim)
2465                 return 0;
2466         return 1;
2467 }
2468
2469
2470 static int special_mapping_fault(struct vm_area_struct *vma,
2471                                 struct vm_fault *vmf)
2472 {
2473         pgoff_t pgoff;
2474         struct page **pages;
2475
2476         /*
2477          * special mappings have no vm_file, and in that case, the mm
2478          * uses vm_pgoff internally. So we have to subtract it from here.
2479          * We are allowed to do this because we are the mm; do not copy
2480          * this code into drivers!
2481          */
2482         pgoff = vmf->pgoff - vma->vm_pgoff;
2483
2484         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2485                 pgoff--;
2486
2487         if (*pages) {
2488                 struct page *page = *pages;
2489                 get_page(page);
2490                 vmf->page = page;
2491                 return 0;
2492         }
2493
2494         return VM_FAULT_SIGBUS;
2495 }
2496
2497 /*
2498  * Having a close hook prevents vma merging regardless of flags.
2499  */
2500 static void special_mapping_close(struct vm_area_struct *vma)
2501 {
2502 }
2503
2504 static const struct vm_operations_struct special_mapping_vmops = {
2505         .close = special_mapping_close,
2506         .fault = special_mapping_fault,
2507 };
2508
2509 /*
2510  * Called with mm->mmap_sem held for writing.
2511  * Insert a new vma covering the given region, with the given flags.
2512  * Its pages are supplied by the given array of struct page *.
2513  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2514  * The region past the last page supplied will always produce SIGBUS.
2515  * The array pointer and the pages it points to are assumed to stay alive
2516  * for as long as this mapping might exist.
2517  */
2518 int install_special_mapping(struct mm_struct *mm,
2519                             unsigned long addr, unsigned long len,
2520                             unsigned long vm_flags, struct page **pages)
2521 {
2522         int ret;
2523         struct vm_area_struct *vma;
2524
2525         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2526         if (unlikely(vma == NULL))
2527                 return -ENOMEM;
2528
2529         INIT_LIST_HEAD(&vma->anon_vma_chain);
2530         vma->vm_mm = mm;
2531         vma->vm_start = addr;
2532         vma->vm_end = addr + len;
2533
2534         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2535         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2536
2537         vma->vm_ops = &special_mapping_vmops;
2538         vma->vm_private_data = pages;
2539
2540         ret = insert_vm_struct(mm, vma);
2541         if (ret)
2542                 goto out;
2543
2544         mm->total_vm += len >> PAGE_SHIFT;
2545
2546         perf_event_mmap(vma);
2547
2548         return 0;
2549
2550 out:
2551         kmem_cache_free(vm_area_cachep, vma);
2552         return ret;
2553 }
2554
2555 static DEFINE_MUTEX(mm_all_locks_mutex);
2556
2557 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2558 {
2559         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2560                 /*
2561                  * The LSB of head.next can't change from under us
2562                  * because we hold the mm_all_locks_mutex.
2563                  */
2564                 down_write(&anon_vma->root->rwsem);
2565                 /*
2566                  * We can safely modify head.next after taking the
2567                  * anon_vma->root->rwsem. If some other vma in this mm shares
2568                  * the same anon_vma we won't take it again.
2569                  *
2570                  * No need of atomic instructions here, head.next
2571                  * can't change from under us thanks to the
2572                  * anon_vma->root->rwsem.
2573                  */
2574                 if (__test_and_set_bit(0, (unsigned long *)
2575                                        &anon_vma->root->rb_root.rb_node))
2576                         BUG();
2577         }
2578 }
2579
2580 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2581 {
2582         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2583                 /*
2584                  * AS_MM_ALL_LOCKS can't change from under us because
2585                  * we hold the mm_all_locks_mutex.
2586                  *
2587                  * Operations on ->flags have to be atomic because
2588                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2589                  * mm_all_locks_mutex, there may be other cpus
2590                  * changing other bitflags in parallel to us.
2591                  */
2592                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2593                         BUG();
2594                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2595         }
2596 }
2597
2598 /*
2599  * This operation locks against the VM for all pte/vma/mm related
2600  * operations that could ever happen on a certain mm. This includes
2601  * vmtruncate, try_to_unmap, and all page faults.
2602  *
2603  * The caller must take the mmap_sem in write mode before calling
2604  * mm_take_all_locks(). The caller isn't allowed to release the
2605  * mmap_sem until mm_drop_all_locks() returns.
2606  *
2607  * mmap_sem in write mode is required in order to block all operations
2608  * that could modify pagetables and free pages without need of
2609  * altering the vma layout (for example populate_range() with
2610  * nonlinear vmas). It's also needed in write mode to avoid new
2611  * anon_vmas to be associated with existing vmas.
2612  *
2613  * A single task can't take more than one mm_take_all_locks() in a row
2614  * or it would deadlock.
2615  *
2616  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2617  * mapping->flags avoid to take the same lock twice, if more than one
2618  * vma in this mm is backed by the same anon_vma or address_space.
2619  *
2620  * We can take all the locks in random order because the VM code
2621  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2622  * takes more than one of them in a row. Secondly we're protected
2623  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2624  *
2625  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2626  * that may have to take thousand of locks.
2627  *
2628  * mm_take_all_locks() can fail if it's interrupted by signals.
2629  */
2630 int mm_take_all_locks(struct mm_struct *mm)
2631 {
2632         struct vm_area_struct *vma;
2633         struct anon_vma_chain *avc;
2634
2635         BUG_ON(down_read_trylock(&mm->mmap_sem));
2636
2637         mutex_lock(&mm_all_locks_mutex);
2638
2639         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2640                 if (signal_pending(current))
2641                         goto out_unlock;
2642                 if (vma->vm_file && vma->vm_file->f_mapping)
2643                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2644         }
2645
2646         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2647                 if (signal_pending(current))
2648                         goto out_unlock;
2649                 if (vma->anon_vma)
2650                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2651                                 vm_lock_anon_vma(mm, avc->anon_vma);
2652         }
2653
2654         return 0;
2655
2656 out_unlock:
2657         mm_drop_all_locks(mm);
2658         return -EINTR;
2659 }
2660
2661 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2662 {
2663         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2664                 /*
2665                  * The LSB of head.next can't change to 0 from under
2666                  * us because we hold the mm_all_locks_mutex.
2667                  *
2668                  * We must however clear the bitflag before unlocking
2669                  * the vma so the users using the anon_vma->rb_root will
2670                  * never see our bitflag.
2671                  *
2672                  * No need of atomic instructions here, head.next
2673                  * can't change from under us until we release the
2674                  * anon_vma->root->rwsem.
2675                  */
2676                 if (!__test_and_clear_bit(0, (unsigned long *)
2677                                           &anon_vma->root->rb_root.rb_node))
2678                         BUG();
2679                 anon_vma_unlock(anon_vma);
2680         }
2681 }
2682
2683 static void vm_unlock_mapping(struct address_space *mapping)
2684 {
2685         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2686                 /*
2687                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2688                  * because we hold the mm_all_locks_mutex.
2689                  */
2690                 mutex_unlock(&mapping->i_mmap_mutex);
2691                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2692                                         &mapping->flags))
2693                         BUG();
2694         }
2695 }
2696
2697 /*
2698  * The mmap_sem cannot be released by the caller until
2699  * mm_drop_all_locks() returns.
2700  */
2701 void mm_drop_all_locks(struct mm_struct *mm)
2702 {
2703         struct vm_area_struct *vma;
2704         struct anon_vma_chain *avc;
2705
2706         BUG_ON(down_read_trylock(&mm->mmap_sem));
2707         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2708
2709         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2710                 if (vma->anon_vma)
2711                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2712                                 vm_unlock_anon_vma(avc->anon_vma);
2713                 if (vma->vm_file && vma->vm_file->f_mapping)
2714                         vm_unlock_mapping(vma->vm_file->f_mapping);
2715         }
2716
2717         mutex_unlock(&mm_all_locks_mutex);
2718 }
2719
2720 /*
2721  * initialise the VMA slab
2722  */
2723 void __init mmap_init(void)
2724 {
2725         int ret;
2726
2727         ret = percpu_counter_init(&vm_committed_as, 0);
2728         VM_BUG_ON(ret);
2729 }