]> git.openfabrics.org - ~shefty/rdma-dev.git/blob - mm/mmap.c
mm: augment vma rbtree with rb_subtree_gap
[~shefty/rdma-dev.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 #include <linux/rbtree_augmented.h>
35
36 #include <asm/uaccess.h>
37 #include <asm/cacheflush.h>
38 #include <asm/tlb.h>
39 #include <asm/mmu_context.h>
40
41 #include "internal.h"
42
43 #ifndef arch_mmap_check
44 #define arch_mmap_check(addr, len, flags)       (0)
45 #endif
46
47 #ifndef arch_rebalance_pgtables
48 #define arch_rebalance_pgtables(addr, len)              (addr)
49 #endif
50
51 static void unmap_region(struct mm_struct *mm,
52                 struct vm_area_struct *vma, struct vm_area_struct *prev,
53                 unsigned long start, unsigned long end);
54
55 /* description of effects of mapping type and prot in current implementation.
56  * this is due to the limited x86 page protection hardware.  The expected
57  * behavior is in parens:
58  *
59  * map_type     prot
60  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
61  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
62  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
63  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
64  *              
65  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
66  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
67  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
68  *
69  */
70 pgprot_t protection_map[16] = {
71         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
72         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
73 };
74
75 pgprot_t vm_get_page_prot(unsigned long vm_flags)
76 {
77         return __pgprot(pgprot_val(protection_map[vm_flags &
78                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
79                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
80 }
81 EXPORT_SYMBOL(vm_get_page_prot);
82
83 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
84 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
85 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
86 /*
87  * Make sure vm_committed_as in one cacheline and not cacheline shared with
88  * other variables. It can be updated by several CPUs frequently.
89  */
90 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
91
92 /*
93  * Check that a process has enough memory to allocate a new virtual
94  * mapping. 0 means there is enough memory for the allocation to
95  * succeed and -ENOMEM implies there is not.
96  *
97  * We currently support three overcommit policies, which are set via the
98  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
99  *
100  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
101  * Additional code 2002 Jul 20 by Robert Love.
102  *
103  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
104  *
105  * Note this is a helper function intended to be used by LSMs which
106  * wish to use this logic.
107  */
108 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
109 {
110         unsigned long free, allowed;
111
112         vm_acct_memory(pages);
113
114         /*
115          * Sometimes we want to use more memory than we have
116          */
117         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
118                 return 0;
119
120         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
121                 free = global_page_state(NR_FREE_PAGES);
122                 free += global_page_state(NR_FILE_PAGES);
123
124                 /*
125                  * shmem pages shouldn't be counted as free in this
126                  * case, they can't be purged, only swapped out, and
127                  * that won't affect the overall amount of available
128                  * memory in the system.
129                  */
130                 free -= global_page_state(NR_SHMEM);
131
132                 free += nr_swap_pages;
133
134                 /*
135                  * Any slabs which are created with the
136                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
137                  * which are reclaimable, under pressure.  The dentry
138                  * cache and most inode caches should fall into this
139                  */
140                 free += global_page_state(NR_SLAB_RECLAIMABLE);
141
142                 /*
143                  * Leave reserved pages. The pages are not for anonymous pages.
144                  */
145                 if (free <= totalreserve_pages)
146                         goto error;
147                 else
148                         free -= totalreserve_pages;
149
150                 /*
151                  * Leave the last 3% for root
152                  */
153                 if (!cap_sys_admin)
154                         free -= free / 32;
155
156                 if (free > pages)
157                         return 0;
158
159                 goto error;
160         }
161
162         allowed = (totalram_pages - hugetlb_total_pages())
163                 * sysctl_overcommit_ratio / 100;
164         /*
165          * Leave the last 3% for root
166          */
167         if (!cap_sys_admin)
168                 allowed -= allowed / 32;
169         allowed += total_swap_pages;
170
171         /* Don't let a single process grow too big:
172            leave 3% of the size of this process for other processes */
173         if (mm)
174                 allowed -= mm->total_vm / 32;
175
176         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
177                 return 0;
178 error:
179         vm_unacct_memory(pages);
180
181         return -ENOMEM;
182 }
183
184 /*
185  * Requires inode->i_mapping->i_mmap_mutex
186  */
187 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
188                 struct file *file, struct address_space *mapping)
189 {
190         if (vma->vm_flags & VM_DENYWRITE)
191                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
192         if (vma->vm_flags & VM_SHARED)
193                 mapping->i_mmap_writable--;
194
195         flush_dcache_mmap_lock(mapping);
196         if (unlikely(vma->vm_flags & VM_NONLINEAR))
197                 list_del_init(&vma->shared.nonlinear);
198         else
199                 vma_interval_tree_remove(vma, &mapping->i_mmap);
200         flush_dcache_mmap_unlock(mapping);
201 }
202
203 /*
204  * Unlink a file-based vm structure from its interval tree, to hide
205  * vma from rmap and vmtruncate before freeing its page tables.
206  */
207 void unlink_file_vma(struct vm_area_struct *vma)
208 {
209         struct file *file = vma->vm_file;
210
211         if (file) {
212                 struct address_space *mapping = file->f_mapping;
213                 mutex_lock(&mapping->i_mmap_mutex);
214                 __remove_shared_vm_struct(vma, file, mapping);
215                 mutex_unlock(&mapping->i_mmap_mutex);
216         }
217 }
218
219 /*
220  * Close a vm structure and free it, returning the next.
221  */
222 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
223 {
224         struct vm_area_struct *next = vma->vm_next;
225
226         might_sleep();
227         if (vma->vm_ops && vma->vm_ops->close)
228                 vma->vm_ops->close(vma);
229         if (vma->vm_file)
230                 fput(vma->vm_file);
231         mpol_put(vma_policy(vma));
232         kmem_cache_free(vm_area_cachep, vma);
233         return next;
234 }
235
236 static unsigned long do_brk(unsigned long addr, unsigned long len);
237
238 SYSCALL_DEFINE1(brk, unsigned long, brk)
239 {
240         unsigned long rlim, retval;
241         unsigned long newbrk, oldbrk;
242         struct mm_struct *mm = current->mm;
243         unsigned long min_brk;
244
245         down_write(&mm->mmap_sem);
246
247 #ifdef CONFIG_COMPAT_BRK
248         /*
249          * CONFIG_COMPAT_BRK can still be overridden by setting
250          * randomize_va_space to 2, which will still cause mm->start_brk
251          * to be arbitrarily shifted
252          */
253         if (current->brk_randomized)
254                 min_brk = mm->start_brk;
255         else
256                 min_brk = mm->end_data;
257 #else
258         min_brk = mm->start_brk;
259 #endif
260         if (brk < min_brk)
261                 goto out;
262
263         /*
264          * Check against rlimit here. If this check is done later after the test
265          * of oldbrk with newbrk then it can escape the test and let the data
266          * segment grow beyond its set limit the in case where the limit is
267          * not page aligned -Ram Gupta
268          */
269         rlim = rlimit(RLIMIT_DATA);
270         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
271                         (mm->end_data - mm->start_data) > rlim)
272                 goto out;
273
274         newbrk = PAGE_ALIGN(brk);
275         oldbrk = PAGE_ALIGN(mm->brk);
276         if (oldbrk == newbrk)
277                 goto set_brk;
278
279         /* Always allow shrinking brk. */
280         if (brk <= mm->brk) {
281                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
282                         goto set_brk;
283                 goto out;
284         }
285
286         /* Check against existing mmap mappings. */
287         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
288                 goto out;
289
290         /* Ok, looks good - let it rip. */
291         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
292                 goto out;
293 set_brk:
294         mm->brk = brk;
295 out:
296         retval = mm->brk;
297         up_write(&mm->mmap_sem);
298         return retval;
299 }
300
301 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
302 {
303         unsigned long max, subtree_gap;
304         max = vma->vm_start;
305         if (vma->vm_prev)
306                 max -= vma->vm_prev->vm_end;
307         if (vma->vm_rb.rb_left) {
308                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
309                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
310                 if (subtree_gap > max)
311                         max = subtree_gap;
312         }
313         if (vma->vm_rb.rb_right) {
314                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
315                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
316                 if (subtree_gap > max)
317                         max = subtree_gap;
318         }
319         return max;
320 }
321
322 #ifdef CONFIG_DEBUG_VM_RB
323 static int browse_rb(struct rb_root *root)
324 {
325         int i = 0, j;
326         struct rb_node *nd, *pn = NULL;
327         unsigned long prev = 0, pend = 0;
328
329         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
330                 struct vm_area_struct *vma;
331                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
332                 if (vma->vm_start < prev)
333                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
334                 if (vma->vm_start < pend)
335                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
336                 if (vma->vm_start > vma->vm_end)
337                         printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
338                 i++;
339                 pn = nd;
340                 prev = vma->vm_start;
341                 pend = vma->vm_end;
342         }
343         j = 0;
344         for (nd = pn; nd; nd = rb_prev(nd)) {
345                 j++;
346         }
347         if (i != j)
348                 printk("backwards %d, forwards %d\n", j, i), i = 0;
349         return i;
350 }
351
352 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
353 {
354         struct rb_node *nd;
355
356         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
357                 struct vm_area_struct *vma;
358                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
359                 BUG_ON(vma != ignore &&
360                        vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
361         }
362 }
363
364 void validate_mm(struct mm_struct *mm)
365 {
366         int bug = 0;
367         int i = 0;
368         struct vm_area_struct *vma = mm->mmap;
369         while (vma) {
370                 struct anon_vma_chain *avc;
371                 vma_lock_anon_vma(vma);
372                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
373                         anon_vma_interval_tree_verify(avc);
374                 vma_unlock_anon_vma(vma);
375                 vma = vma->vm_next;
376                 i++;
377         }
378         if (i != mm->map_count)
379                 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
380         i = browse_rb(&mm->mm_rb);
381         if (i != mm->map_count)
382                 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
383         BUG_ON(bug);
384 }
385 #else
386 #define validate_mm_rb(root, ignore) do { } while (0)
387 #define validate_mm(mm) do { } while (0)
388 #endif
389
390 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
391                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
392
393 /*
394  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
395  * vma->vm_prev->vm_end values changed, without modifying the vma's position
396  * in the rbtree.
397  */
398 static void vma_gap_update(struct vm_area_struct *vma)
399 {
400         /*
401          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
402          * function that does exacltly what we want.
403          */
404         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
405 }
406
407 static inline void vma_rb_insert(struct vm_area_struct *vma,
408                                  struct rb_root *root)
409 {
410         /* All rb_subtree_gap values must be consistent prior to insertion */
411         validate_mm_rb(root, NULL);
412
413         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
414 }
415
416 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
417 {
418         /*
419          * All rb_subtree_gap values must be consistent prior to erase,
420          * with the possible exception of the vma being erased.
421          */
422         validate_mm_rb(root, vma);
423
424         /*
425          * Note rb_erase_augmented is a fairly large inline function,
426          * so make sure we instantiate it only once with our desired
427          * augmented rbtree callbacks.
428          */
429         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
430 }
431
432 /*
433  * vma has some anon_vma assigned, and is already inserted on that
434  * anon_vma's interval trees.
435  *
436  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
437  * vma must be removed from the anon_vma's interval trees using
438  * anon_vma_interval_tree_pre_update_vma().
439  *
440  * After the update, the vma will be reinserted using
441  * anon_vma_interval_tree_post_update_vma().
442  *
443  * The entire update must be protected by exclusive mmap_sem and by
444  * the root anon_vma's mutex.
445  */
446 static inline void
447 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
448 {
449         struct anon_vma_chain *avc;
450
451         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
452                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
453 }
454
455 static inline void
456 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
457 {
458         struct anon_vma_chain *avc;
459
460         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
461                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
462 }
463
464 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
465                 unsigned long end, struct vm_area_struct **pprev,
466                 struct rb_node ***rb_link, struct rb_node **rb_parent)
467 {
468         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
469
470         __rb_link = &mm->mm_rb.rb_node;
471         rb_prev = __rb_parent = NULL;
472
473         while (*__rb_link) {
474                 struct vm_area_struct *vma_tmp;
475
476                 __rb_parent = *__rb_link;
477                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
478
479                 if (vma_tmp->vm_end > addr) {
480                         /* Fail if an existing vma overlaps the area */
481                         if (vma_tmp->vm_start < end)
482                                 return -ENOMEM;
483                         __rb_link = &__rb_parent->rb_left;
484                 } else {
485                         rb_prev = __rb_parent;
486                         __rb_link = &__rb_parent->rb_right;
487                 }
488         }
489
490         *pprev = NULL;
491         if (rb_prev)
492                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
493         *rb_link = __rb_link;
494         *rb_parent = __rb_parent;
495         return 0;
496 }
497
498 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
499                 struct rb_node **rb_link, struct rb_node *rb_parent)
500 {
501         /* Update tracking information for the gap following the new vma. */
502         if (vma->vm_next)
503                 vma_gap_update(vma->vm_next);
504         else
505                 mm->highest_vm_end = vma->vm_end;
506
507         /*
508          * vma->vm_prev wasn't known when we followed the rbtree to find the
509          * correct insertion point for that vma. As a result, we could not
510          * update the vma vm_rb parents rb_subtree_gap values on the way down.
511          * So, we first insert the vma with a zero rb_subtree_gap value
512          * (to be consistent with what we did on the way down), and then
513          * immediately update the gap to the correct value. Finally we
514          * rebalance the rbtree after all augmented values have been set.
515          */
516         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
517         vma->rb_subtree_gap = 0;
518         vma_gap_update(vma);
519         vma_rb_insert(vma, &mm->mm_rb);
520 }
521
522 static void __vma_link_file(struct vm_area_struct *vma)
523 {
524         struct file *file;
525
526         file = vma->vm_file;
527         if (file) {
528                 struct address_space *mapping = file->f_mapping;
529
530                 if (vma->vm_flags & VM_DENYWRITE)
531                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
532                 if (vma->vm_flags & VM_SHARED)
533                         mapping->i_mmap_writable++;
534
535                 flush_dcache_mmap_lock(mapping);
536                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
537                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
538                 else
539                         vma_interval_tree_insert(vma, &mapping->i_mmap);
540                 flush_dcache_mmap_unlock(mapping);
541         }
542 }
543
544 static void
545 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
546         struct vm_area_struct *prev, struct rb_node **rb_link,
547         struct rb_node *rb_parent)
548 {
549         __vma_link_list(mm, vma, prev, rb_parent);
550         __vma_link_rb(mm, vma, rb_link, rb_parent);
551 }
552
553 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
554                         struct vm_area_struct *prev, struct rb_node **rb_link,
555                         struct rb_node *rb_parent)
556 {
557         struct address_space *mapping = NULL;
558
559         if (vma->vm_file)
560                 mapping = vma->vm_file->f_mapping;
561
562         if (mapping)
563                 mutex_lock(&mapping->i_mmap_mutex);
564
565         __vma_link(mm, vma, prev, rb_link, rb_parent);
566         __vma_link_file(vma);
567
568         if (mapping)
569                 mutex_unlock(&mapping->i_mmap_mutex);
570
571         mm->map_count++;
572         validate_mm(mm);
573 }
574
575 /*
576  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
577  * mm's list and rbtree.  It has already been inserted into the interval tree.
578  */
579 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
580 {
581         struct vm_area_struct *prev;
582         struct rb_node **rb_link, *rb_parent;
583
584         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
585                            &prev, &rb_link, &rb_parent))
586                 BUG();
587         __vma_link(mm, vma, prev, rb_link, rb_parent);
588         mm->map_count++;
589 }
590
591 static inline void
592 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
593                 struct vm_area_struct *prev)
594 {
595         struct vm_area_struct *next;
596
597         vma_rb_erase(vma, &mm->mm_rb);
598         prev->vm_next = next = vma->vm_next;
599         if (next)
600                 next->vm_prev = prev;
601         if (mm->mmap_cache == vma)
602                 mm->mmap_cache = prev;
603 }
604
605 /*
606  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
607  * is already present in an i_mmap tree without adjusting the tree.
608  * The following helper function should be used when such adjustments
609  * are necessary.  The "insert" vma (if any) is to be inserted
610  * before we drop the necessary locks.
611  */
612 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
613         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
614 {
615         struct mm_struct *mm = vma->vm_mm;
616         struct vm_area_struct *next = vma->vm_next;
617         struct vm_area_struct *importer = NULL;
618         struct address_space *mapping = NULL;
619         struct rb_root *root = NULL;
620         struct anon_vma *anon_vma = NULL;
621         struct file *file = vma->vm_file;
622         bool start_changed = false, end_changed = false;
623         long adjust_next = 0;
624         int remove_next = 0;
625
626         if (next && !insert) {
627                 struct vm_area_struct *exporter = NULL;
628
629                 if (end >= next->vm_end) {
630                         /*
631                          * vma expands, overlapping all the next, and
632                          * perhaps the one after too (mprotect case 6).
633                          */
634 again:                  remove_next = 1 + (end > next->vm_end);
635                         end = next->vm_end;
636                         exporter = next;
637                         importer = vma;
638                 } else if (end > next->vm_start) {
639                         /*
640                          * vma expands, overlapping part of the next:
641                          * mprotect case 5 shifting the boundary up.
642                          */
643                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
644                         exporter = next;
645                         importer = vma;
646                 } else if (end < vma->vm_end) {
647                         /*
648                          * vma shrinks, and !insert tells it's not
649                          * split_vma inserting another: so it must be
650                          * mprotect case 4 shifting the boundary down.
651                          */
652                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
653                         exporter = vma;
654                         importer = next;
655                 }
656
657                 /*
658                  * Easily overlooked: when mprotect shifts the boundary,
659                  * make sure the expanding vma has anon_vma set if the
660                  * shrinking vma had, to cover any anon pages imported.
661                  */
662                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
663                         if (anon_vma_clone(importer, exporter))
664                                 return -ENOMEM;
665                         importer->anon_vma = exporter->anon_vma;
666                 }
667         }
668
669         if (file) {
670                 mapping = file->f_mapping;
671                 if (!(vma->vm_flags & VM_NONLINEAR)) {
672                         root = &mapping->i_mmap;
673                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
674
675                         if (adjust_next)
676                                 uprobe_munmap(next, next->vm_start,
677                                                         next->vm_end);
678                 }
679
680                 mutex_lock(&mapping->i_mmap_mutex);
681                 if (insert) {
682                         /*
683                          * Put into interval tree now, so instantiated pages
684                          * are visible to arm/parisc __flush_dcache_page
685                          * throughout; but we cannot insert into address
686                          * space until vma start or end is updated.
687                          */
688                         __vma_link_file(insert);
689                 }
690         }
691
692         vma_adjust_trans_huge(vma, start, end, adjust_next);
693
694         anon_vma = vma->anon_vma;
695         if (!anon_vma && adjust_next)
696                 anon_vma = next->anon_vma;
697         if (anon_vma) {
698                 VM_BUG_ON(adjust_next && next->anon_vma &&
699                           anon_vma != next->anon_vma);
700                 anon_vma_lock(anon_vma);
701                 anon_vma_interval_tree_pre_update_vma(vma);
702                 if (adjust_next)
703                         anon_vma_interval_tree_pre_update_vma(next);
704         }
705
706         if (root) {
707                 flush_dcache_mmap_lock(mapping);
708                 vma_interval_tree_remove(vma, root);
709                 if (adjust_next)
710                         vma_interval_tree_remove(next, root);
711         }
712
713         if (start != vma->vm_start) {
714                 vma->vm_start = start;
715                 start_changed = true;
716         }
717         if (end != vma->vm_end) {
718                 vma->vm_end = end;
719                 end_changed = true;
720         }
721         vma->vm_pgoff = pgoff;
722         if (adjust_next) {
723                 next->vm_start += adjust_next << PAGE_SHIFT;
724                 next->vm_pgoff += adjust_next;
725         }
726
727         if (root) {
728                 if (adjust_next)
729                         vma_interval_tree_insert(next, root);
730                 vma_interval_tree_insert(vma, root);
731                 flush_dcache_mmap_unlock(mapping);
732         }
733
734         if (remove_next) {
735                 /*
736                  * vma_merge has merged next into vma, and needs
737                  * us to remove next before dropping the locks.
738                  */
739                 __vma_unlink(mm, next, vma);
740                 if (file)
741                         __remove_shared_vm_struct(next, file, mapping);
742         } else if (insert) {
743                 /*
744                  * split_vma has split insert from vma, and needs
745                  * us to insert it before dropping the locks
746                  * (it may either follow vma or precede it).
747                  */
748                 __insert_vm_struct(mm, insert);
749         } else {
750                 if (start_changed)
751                         vma_gap_update(vma);
752                 if (end_changed) {
753                         if (!next)
754                                 mm->highest_vm_end = end;
755                         else if (!adjust_next)
756                                 vma_gap_update(next);
757                 }
758         }
759
760         if (anon_vma) {
761                 anon_vma_interval_tree_post_update_vma(vma);
762                 if (adjust_next)
763                         anon_vma_interval_tree_post_update_vma(next);
764                 anon_vma_unlock(anon_vma);
765         }
766         if (mapping)
767                 mutex_unlock(&mapping->i_mmap_mutex);
768
769         if (root) {
770                 uprobe_mmap(vma);
771
772                 if (adjust_next)
773                         uprobe_mmap(next);
774         }
775
776         if (remove_next) {
777                 if (file) {
778                         uprobe_munmap(next, next->vm_start, next->vm_end);
779                         fput(file);
780                 }
781                 if (next->anon_vma)
782                         anon_vma_merge(vma, next);
783                 mm->map_count--;
784                 mpol_put(vma_policy(next));
785                 kmem_cache_free(vm_area_cachep, next);
786                 /*
787                  * In mprotect's case 6 (see comments on vma_merge),
788                  * we must remove another next too. It would clutter
789                  * up the code too much to do both in one go.
790                  */
791                 next = vma->vm_next;
792                 if (remove_next == 2)
793                         goto again;
794                 else if (next)
795                         vma_gap_update(next);
796                 else
797                         mm->highest_vm_end = end;
798         }
799         if (insert && file)
800                 uprobe_mmap(insert);
801
802         validate_mm(mm);
803
804         return 0;
805 }
806
807 /*
808  * If the vma has a ->close operation then the driver probably needs to release
809  * per-vma resources, so we don't attempt to merge those.
810  */
811 static inline int is_mergeable_vma(struct vm_area_struct *vma,
812                         struct file *file, unsigned long vm_flags)
813 {
814         if (vma->vm_flags ^ vm_flags)
815                 return 0;
816         if (vma->vm_file != file)
817                 return 0;
818         if (vma->vm_ops && vma->vm_ops->close)
819                 return 0;
820         return 1;
821 }
822
823 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
824                                         struct anon_vma *anon_vma2,
825                                         struct vm_area_struct *vma)
826 {
827         /*
828          * The list_is_singular() test is to avoid merging VMA cloned from
829          * parents. This can improve scalability caused by anon_vma lock.
830          */
831         if ((!anon_vma1 || !anon_vma2) && (!vma ||
832                 list_is_singular(&vma->anon_vma_chain)))
833                 return 1;
834         return anon_vma1 == anon_vma2;
835 }
836
837 /*
838  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
839  * in front of (at a lower virtual address and file offset than) the vma.
840  *
841  * We cannot merge two vmas if they have differently assigned (non-NULL)
842  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
843  *
844  * We don't check here for the merged mmap wrapping around the end of pagecache
845  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
846  * wrap, nor mmaps which cover the final page at index -1UL.
847  */
848 static int
849 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
850         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
851 {
852         if (is_mergeable_vma(vma, file, vm_flags) &&
853             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
854                 if (vma->vm_pgoff == vm_pgoff)
855                         return 1;
856         }
857         return 0;
858 }
859
860 /*
861  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
862  * beyond (at a higher virtual address and file offset than) the vma.
863  *
864  * We cannot merge two vmas if they have differently assigned (non-NULL)
865  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
866  */
867 static int
868 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
869         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
870 {
871         if (is_mergeable_vma(vma, file, vm_flags) &&
872             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
873                 pgoff_t vm_pglen;
874                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
875                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
876                         return 1;
877         }
878         return 0;
879 }
880
881 /*
882  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
883  * whether that can be merged with its predecessor or its successor.
884  * Or both (it neatly fills a hole).
885  *
886  * In most cases - when called for mmap, brk or mremap - [addr,end) is
887  * certain not to be mapped by the time vma_merge is called; but when
888  * called for mprotect, it is certain to be already mapped (either at
889  * an offset within prev, or at the start of next), and the flags of
890  * this area are about to be changed to vm_flags - and the no-change
891  * case has already been eliminated.
892  *
893  * The following mprotect cases have to be considered, where AAAA is
894  * the area passed down from mprotect_fixup, never extending beyond one
895  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
896  *
897  *     AAAA             AAAA                AAAA          AAAA
898  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
899  *    cannot merge    might become    might become    might become
900  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
901  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
902  *    mremap move:                                    PPPPNNNNNNNN 8
903  *        AAAA
904  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
905  *    might become    case 1 below    case 2 below    case 3 below
906  *
907  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
908  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
909  */
910 struct vm_area_struct *vma_merge(struct mm_struct *mm,
911                         struct vm_area_struct *prev, unsigned long addr,
912                         unsigned long end, unsigned long vm_flags,
913                         struct anon_vma *anon_vma, struct file *file,
914                         pgoff_t pgoff, struct mempolicy *policy)
915 {
916         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
917         struct vm_area_struct *area, *next;
918         int err;
919
920         /*
921          * We later require that vma->vm_flags == vm_flags,
922          * so this tests vma->vm_flags & VM_SPECIAL, too.
923          */
924         if (vm_flags & VM_SPECIAL)
925                 return NULL;
926
927         if (prev)
928                 next = prev->vm_next;
929         else
930                 next = mm->mmap;
931         area = next;
932         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
933                 next = next->vm_next;
934
935         /*
936          * Can it merge with the predecessor?
937          */
938         if (prev && prev->vm_end == addr &&
939                         mpol_equal(vma_policy(prev), policy) &&
940                         can_vma_merge_after(prev, vm_flags,
941                                                 anon_vma, file, pgoff)) {
942                 /*
943                  * OK, it can.  Can we now merge in the successor as well?
944                  */
945                 if (next && end == next->vm_start &&
946                                 mpol_equal(policy, vma_policy(next)) &&
947                                 can_vma_merge_before(next, vm_flags,
948                                         anon_vma, file, pgoff+pglen) &&
949                                 is_mergeable_anon_vma(prev->anon_vma,
950                                                       next->anon_vma, NULL)) {
951                                                         /* cases 1, 6 */
952                         err = vma_adjust(prev, prev->vm_start,
953                                 next->vm_end, prev->vm_pgoff, NULL);
954                 } else                                  /* cases 2, 5, 7 */
955                         err = vma_adjust(prev, prev->vm_start,
956                                 end, prev->vm_pgoff, NULL);
957                 if (err)
958                         return NULL;
959                 khugepaged_enter_vma_merge(prev);
960                 return prev;
961         }
962
963         /*
964          * Can this new request be merged in front of next?
965          */
966         if (next && end == next->vm_start &&
967                         mpol_equal(policy, vma_policy(next)) &&
968                         can_vma_merge_before(next, vm_flags,
969                                         anon_vma, file, pgoff+pglen)) {
970                 if (prev && addr < prev->vm_end)        /* case 4 */
971                         err = vma_adjust(prev, prev->vm_start,
972                                 addr, prev->vm_pgoff, NULL);
973                 else                                    /* cases 3, 8 */
974                         err = vma_adjust(area, addr, next->vm_end,
975                                 next->vm_pgoff - pglen, NULL);
976                 if (err)
977                         return NULL;
978                 khugepaged_enter_vma_merge(area);
979                 return area;
980         }
981
982         return NULL;
983 }
984
985 /*
986  * Rough compatbility check to quickly see if it's even worth looking
987  * at sharing an anon_vma.
988  *
989  * They need to have the same vm_file, and the flags can only differ
990  * in things that mprotect may change.
991  *
992  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
993  * we can merge the two vma's. For example, we refuse to merge a vma if
994  * there is a vm_ops->close() function, because that indicates that the
995  * driver is doing some kind of reference counting. But that doesn't
996  * really matter for the anon_vma sharing case.
997  */
998 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
999 {
1000         return a->vm_end == b->vm_start &&
1001                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1002                 a->vm_file == b->vm_file &&
1003                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
1004                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1005 }
1006
1007 /*
1008  * Do some basic sanity checking to see if we can re-use the anon_vma
1009  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1010  * the same as 'old', the other will be the new one that is trying
1011  * to share the anon_vma.
1012  *
1013  * NOTE! This runs with mm_sem held for reading, so it is possible that
1014  * the anon_vma of 'old' is concurrently in the process of being set up
1015  * by another page fault trying to merge _that_. But that's ok: if it
1016  * is being set up, that automatically means that it will be a singleton
1017  * acceptable for merging, so we can do all of this optimistically. But
1018  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1019  *
1020  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1021  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1022  * is to return an anon_vma that is "complex" due to having gone through
1023  * a fork).
1024  *
1025  * We also make sure that the two vma's are compatible (adjacent,
1026  * and with the same memory policies). That's all stable, even with just
1027  * a read lock on the mm_sem.
1028  */
1029 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1030 {
1031         if (anon_vma_compatible(a, b)) {
1032                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1033
1034                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1035                         return anon_vma;
1036         }
1037         return NULL;
1038 }
1039
1040 /*
1041  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1042  * neighbouring vmas for a suitable anon_vma, before it goes off
1043  * to allocate a new anon_vma.  It checks because a repetitive
1044  * sequence of mprotects and faults may otherwise lead to distinct
1045  * anon_vmas being allocated, preventing vma merge in subsequent
1046  * mprotect.
1047  */
1048 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1049 {
1050         struct anon_vma *anon_vma;
1051         struct vm_area_struct *near;
1052
1053         near = vma->vm_next;
1054         if (!near)
1055                 goto try_prev;
1056
1057         anon_vma = reusable_anon_vma(near, vma, near);
1058         if (anon_vma)
1059                 return anon_vma;
1060 try_prev:
1061         near = vma->vm_prev;
1062         if (!near)
1063                 goto none;
1064
1065         anon_vma = reusable_anon_vma(near, near, vma);
1066         if (anon_vma)
1067                 return anon_vma;
1068 none:
1069         /*
1070          * There's no absolute need to look only at touching neighbours:
1071          * we could search further afield for "compatible" anon_vmas.
1072          * But it would probably just be a waste of time searching,
1073          * or lead to too many vmas hanging off the same anon_vma.
1074          * We're trying to allow mprotect remerging later on,
1075          * not trying to minimize memory used for anon_vmas.
1076          */
1077         return NULL;
1078 }
1079
1080 #ifdef CONFIG_PROC_FS
1081 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1082                                                 struct file *file, long pages)
1083 {
1084         const unsigned long stack_flags
1085                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1086
1087         mm->total_vm += pages;
1088
1089         if (file) {
1090                 mm->shared_vm += pages;
1091                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1092                         mm->exec_vm += pages;
1093         } else if (flags & stack_flags)
1094                 mm->stack_vm += pages;
1095 }
1096 #endif /* CONFIG_PROC_FS */
1097
1098 /*
1099  * If a hint addr is less than mmap_min_addr change hint to be as
1100  * low as possible but still greater than mmap_min_addr
1101  */
1102 static inline unsigned long round_hint_to_min(unsigned long hint)
1103 {
1104         hint &= PAGE_MASK;
1105         if (((void *)hint != NULL) &&
1106             (hint < mmap_min_addr))
1107                 return PAGE_ALIGN(mmap_min_addr);
1108         return hint;
1109 }
1110
1111 /*
1112  * The caller must hold down_write(&current->mm->mmap_sem).
1113  */
1114
1115 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1116                         unsigned long len, unsigned long prot,
1117                         unsigned long flags, unsigned long pgoff)
1118 {
1119         struct mm_struct * mm = current->mm;
1120         struct inode *inode;
1121         vm_flags_t vm_flags;
1122
1123         /*
1124          * Does the application expect PROT_READ to imply PROT_EXEC?
1125          *
1126          * (the exception is when the underlying filesystem is noexec
1127          *  mounted, in which case we dont add PROT_EXEC.)
1128          */
1129         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1130                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1131                         prot |= PROT_EXEC;
1132
1133         if (!len)
1134                 return -EINVAL;
1135
1136         if (!(flags & MAP_FIXED))
1137                 addr = round_hint_to_min(addr);
1138
1139         /* Careful about overflows.. */
1140         len = PAGE_ALIGN(len);
1141         if (!len)
1142                 return -ENOMEM;
1143
1144         /* offset overflow? */
1145         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1146                return -EOVERFLOW;
1147
1148         /* Too many mappings? */
1149         if (mm->map_count > sysctl_max_map_count)
1150                 return -ENOMEM;
1151
1152         /* Obtain the address to map to. we verify (or select) it and ensure
1153          * that it represents a valid section of the address space.
1154          */
1155         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1156         if (addr & ~PAGE_MASK)
1157                 return addr;
1158
1159         /* Do simple checking here so the lower-level routines won't have
1160          * to. we assume access permissions have been handled by the open
1161          * of the memory object, so we don't do any here.
1162          */
1163         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1164                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1165
1166         if (flags & MAP_LOCKED)
1167                 if (!can_do_mlock())
1168                         return -EPERM;
1169
1170         /* mlock MCL_FUTURE? */
1171         if (vm_flags & VM_LOCKED) {
1172                 unsigned long locked, lock_limit;
1173                 locked = len >> PAGE_SHIFT;
1174                 locked += mm->locked_vm;
1175                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1176                 lock_limit >>= PAGE_SHIFT;
1177                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1178                         return -EAGAIN;
1179         }
1180
1181         inode = file ? file->f_path.dentry->d_inode : NULL;
1182
1183         if (file) {
1184                 switch (flags & MAP_TYPE) {
1185                 case MAP_SHARED:
1186                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1187                                 return -EACCES;
1188
1189                         /*
1190                          * Make sure we don't allow writing to an append-only
1191                          * file..
1192                          */
1193                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1194                                 return -EACCES;
1195
1196                         /*
1197                          * Make sure there are no mandatory locks on the file.
1198                          */
1199                         if (locks_verify_locked(inode))
1200                                 return -EAGAIN;
1201
1202                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1203                         if (!(file->f_mode & FMODE_WRITE))
1204                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1205
1206                         /* fall through */
1207                 case MAP_PRIVATE:
1208                         if (!(file->f_mode & FMODE_READ))
1209                                 return -EACCES;
1210                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1211                                 if (vm_flags & VM_EXEC)
1212                                         return -EPERM;
1213                                 vm_flags &= ~VM_MAYEXEC;
1214                         }
1215
1216                         if (!file->f_op || !file->f_op->mmap)
1217                                 return -ENODEV;
1218                         break;
1219
1220                 default:
1221                         return -EINVAL;
1222                 }
1223         } else {
1224                 switch (flags & MAP_TYPE) {
1225                 case MAP_SHARED:
1226                         /*
1227                          * Ignore pgoff.
1228                          */
1229                         pgoff = 0;
1230                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1231                         break;
1232                 case MAP_PRIVATE:
1233                         /*
1234                          * Set pgoff according to addr for anon_vma.
1235                          */
1236                         pgoff = addr >> PAGE_SHIFT;
1237                         break;
1238                 default:
1239                         return -EINVAL;
1240                 }
1241         }
1242
1243         return mmap_region(file, addr, len, flags, vm_flags, pgoff);
1244 }
1245
1246 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1247                 unsigned long, prot, unsigned long, flags,
1248                 unsigned long, fd, unsigned long, pgoff)
1249 {
1250         struct file *file = NULL;
1251         unsigned long retval = -EBADF;
1252
1253         if (!(flags & MAP_ANONYMOUS)) {
1254                 audit_mmap_fd(fd, flags);
1255                 if (unlikely(flags & MAP_HUGETLB))
1256                         return -EINVAL;
1257                 file = fget(fd);
1258                 if (!file)
1259                         goto out;
1260         } else if (flags & MAP_HUGETLB) {
1261                 struct user_struct *user = NULL;
1262                 /*
1263                  * VM_NORESERVE is used because the reservations will be
1264                  * taken when vm_ops->mmap() is called
1265                  * A dummy user value is used because we are not locking
1266                  * memory so no accounting is necessary
1267                  */
1268                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1269                                 VM_NORESERVE,
1270                                 &user, HUGETLB_ANONHUGE_INODE,
1271                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1272                 if (IS_ERR(file))
1273                         return PTR_ERR(file);
1274         }
1275
1276         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1277
1278         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1279         if (file)
1280                 fput(file);
1281 out:
1282         return retval;
1283 }
1284
1285 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1286 struct mmap_arg_struct {
1287         unsigned long addr;
1288         unsigned long len;
1289         unsigned long prot;
1290         unsigned long flags;
1291         unsigned long fd;
1292         unsigned long offset;
1293 };
1294
1295 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1296 {
1297         struct mmap_arg_struct a;
1298
1299         if (copy_from_user(&a, arg, sizeof(a)))
1300                 return -EFAULT;
1301         if (a.offset & ~PAGE_MASK)
1302                 return -EINVAL;
1303
1304         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1305                               a.offset >> PAGE_SHIFT);
1306 }
1307 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1308
1309 /*
1310  * Some shared mappigns will want the pages marked read-only
1311  * to track write events. If so, we'll downgrade vm_page_prot
1312  * to the private version (using protection_map[] without the
1313  * VM_SHARED bit).
1314  */
1315 int vma_wants_writenotify(struct vm_area_struct *vma)
1316 {
1317         vm_flags_t vm_flags = vma->vm_flags;
1318
1319         /* If it was private or non-writable, the write bit is already clear */
1320         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1321                 return 0;
1322
1323         /* The backer wishes to know when pages are first written to? */
1324         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1325                 return 1;
1326
1327         /* The open routine did something to the protections already? */
1328         if (pgprot_val(vma->vm_page_prot) !=
1329             pgprot_val(vm_get_page_prot(vm_flags)))
1330                 return 0;
1331
1332         /* Specialty mapping? */
1333         if (vm_flags & VM_PFNMAP)
1334                 return 0;
1335
1336         /* Can the mapping track the dirty pages? */
1337         return vma->vm_file && vma->vm_file->f_mapping &&
1338                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1339 }
1340
1341 /*
1342  * We account for memory if it's a private writeable mapping,
1343  * not hugepages and VM_NORESERVE wasn't set.
1344  */
1345 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1346 {
1347         /*
1348          * hugetlb has its own accounting separate from the core VM
1349          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1350          */
1351         if (file && is_file_hugepages(file))
1352                 return 0;
1353
1354         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1355 }
1356
1357 unsigned long mmap_region(struct file *file, unsigned long addr,
1358                           unsigned long len, unsigned long flags,
1359                           vm_flags_t vm_flags, unsigned long pgoff)
1360 {
1361         struct mm_struct *mm = current->mm;
1362         struct vm_area_struct *vma, *prev;
1363         int correct_wcount = 0;
1364         int error;
1365         struct rb_node **rb_link, *rb_parent;
1366         unsigned long charged = 0;
1367         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1368
1369         /* Clear old maps */
1370         error = -ENOMEM;
1371 munmap_back:
1372         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1373                 if (do_munmap(mm, addr, len))
1374                         return -ENOMEM;
1375                 goto munmap_back;
1376         }
1377
1378         /* Check against address space limit. */
1379         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1380                 return -ENOMEM;
1381
1382         /*
1383          * Set 'VM_NORESERVE' if we should not account for the
1384          * memory use of this mapping.
1385          */
1386         if ((flags & MAP_NORESERVE)) {
1387                 /* We honor MAP_NORESERVE if allowed to overcommit */
1388                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1389                         vm_flags |= VM_NORESERVE;
1390
1391                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1392                 if (file && is_file_hugepages(file))
1393                         vm_flags |= VM_NORESERVE;
1394         }
1395
1396         /*
1397          * Private writable mapping: check memory availability
1398          */
1399         if (accountable_mapping(file, vm_flags)) {
1400                 charged = len >> PAGE_SHIFT;
1401                 if (security_vm_enough_memory_mm(mm, charged))
1402                         return -ENOMEM;
1403                 vm_flags |= VM_ACCOUNT;
1404         }
1405
1406         /*
1407          * Can we just expand an old mapping?
1408          */
1409         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1410         if (vma)
1411                 goto out;
1412
1413         /*
1414          * Determine the object being mapped and call the appropriate
1415          * specific mapper. the address has already been validated, but
1416          * not unmapped, but the maps are removed from the list.
1417          */
1418         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1419         if (!vma) {
1420                 error = -ENOMEM;
1421                 goto unacct_error;
1422         }
1423
1424         vma->vm_mm = mm;
1425         vma->vm_start = addr;
1426         vma->vm_end = addr + len;
1427         vma->vm_flags = vm_flags;
1428         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1429         vma->vm_pgoff = pgoff;
1430         INIT_LIST_HEAD(&vma->anon_vma_chain);
1431
1432         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1433
1434         if (file) {
1435                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1436                         goto free_vma;
1437                 if (vm_flags & VM_DENYWRITE) {
1438                         error = deny_write_access(file);
1439                         if (error)
1440                                 goto free_vma;
1441                         correct_wcount = 1;
1442                 }
1443                 vma->vm_file = get_file(file);
1444                 error = file->f_op->mmap(file, vma);
1445                 if (error)
1446                         goto unmap_and_free_vma;
1447
1448                 /* Can addr have changed??
1449                  *
1450                  * Answer: Yes, several device drivers can do it in their
1451                  *         f_op->mmap method. -DaveM
1452                  */
1453                 addr = vma->vm_start;
1454                 pgoff = vma->vm_pgoff;
1455                 vm_flags = vma->vm_flags;
1456         } else if (vm_flags & VM_SHARED) {
1457                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1458                         goto free_vma;
1459                 error = shmem_zero_setup(vma);
1460                 if (error)
1461                         goto free_vma;
1462         }
1463
1464         if (vma_wants_writenotify(vma)) {
1465                 pgprot_t pprot = vma->vm_page_prot;
1466
1467                 /* Can vma->vm_page_prot have changed??
1468                  *
1469                  * Answer: Yes, drivers may have changed it in their
1470                  *         f_op->mmap method.
1471                  *
1472                  * Ensures that vmas marked as uncached stay that way.
1473                  */
1474                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1475                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1476                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1477         }
1478
1479         vma_link(mm, vma, prev, rb_link, rb_parent);
1480         file = vma->vm_file;
1481
1482         /* Once vma denies write, undo our temporary denial count */
1483         if (correct_wcount)
1484                 atomic_inc(&inode->i_writecount);
1485 out:
1486         perf_event_mmap(vma);
1487
1488         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1489         if (vm_flags & VM_LOCKED) {
1490                 if (!mlock_vma_pages_range(vma, addr, addr + len))
1491                         mm->locked_vm += (len >> PAGE_SHIFT);
1492         } else if ((flags & MAP_POPULATE) && !(flags & MAP_NONBLOCK))
1493                 make_pages_present(addr, addr + len);
1494
1495         if (file)
1496                 uprobe_mmap(vma);
1497
1498         return addr;
1499
1500 unmap_and_free_vma:
1501         if (correct_wcount)
1502                 atomic_inc(&inode->i_writecount);
1503         vma->vm_file = NULL;
1504         fput(file);
1505
1506         /* Undo any partial mapping done by a device driver. */
1507         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1508         charged = 0;
1509 free_vma:
1510         kmem_cache_free(vm_area_cachep, vma);
1511 unacct_error:
1512         if (charged)
1513                 vm_unacct_memory(charged);
1514         return error;
1515 }
1516
1517 /* Get an address range which is currently unmapped.
1518  * For shmat() with addr=0.
1519  *
1520  * Ugly calling convention alert:
1521  * Return value with the low bits set means error value,
1522  * ie
1523  *      if (ret & ~PAGE_MASK)
1524  *              error = ret;
1525  *
1526  * This function "knows" that -ENOMEM has the bits set.
1527  */
1528 #ifndef HAVE_ARCH_UNMAPPED_AREA
1529 unsigned long
1530 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1531                 unsigned long len, unsigned long pgoff, unsigned long flags)
1532 {
1533         struct mm_struct *mm = current->mm;
1534         struct vm_area_struct *vma;
1535         unsigned long start_addr;
1536
1537         if (len > TASK_SIZE)
1538                 return -ENOMEM;
1539
1540         if (flags & MAP_FIXED)
1541                 return addr;
1542
1543         if (addr) {
1544                 addr = PAGE_ALIGN(addr);
1545                 vma = find_vma(mm, addr);
1546                 if (TASK_SIZE - len >= addr &&
1547                     (!vma || addr + len <= vma->vm_start))
1548                         return addr;
1549         }
1550         if (len > mm->cached_hole_size) {
1551                 start_addr = addr = mm->free_area_cache;
1552         } else {
1553                 start_addr = addr = TASK_UNMAPPED_BASE;
1554                 mm->cached_hole_size = 0;
1555         }
1556
1557 full_search:
1558         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1559                 /* At this point:  (!vma || addr < vma->vm_end). */
1560                 if (TASK_SIZE - len < addr) {
1561                         /*
1562                          * Start a new search - just in case we missed
1563                          * some holes.
1564                          */
1565                         if (start_addr != TASK_UNMAPPED_BASE) {
1566                                 addr = TASK_UNMAPPED_BASE;
1567                                 start_addr = addr;
1568                                 mm->cached_hole_size = 0;
1569                                 goto full_search;
1570                         }
1571                         return -ENOMEM;
1572                 }
1573                 if (!vma || addr + len <= vma->vm_start) {
1574                         /*
1575                          * Remember the place where we stopped the search:
1576                          */
1577                         mm->free_area_cache = addr + len;
1578                         return addr;
1579                 }
1580                 if (addr + mm->cached_hole_size < vma->vm_start)
1581                         mm->cached_hole_size = vma->vm_start - addr;
1582                 addr = vma->vm_end;
1583         }
1584 }
1585 #endif  
1586
1587 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1588 {
1589         /*
1590          * Is this a new hole at the lowest possible address?
1591          */
1592         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1593                 mm->free_area_cache = addr;
1594 }
1595
1596 /*
1597  * This mmap-allocator allocates new areas top-down from below the
1598  * stack's low limit (the base):
1599  */
1600 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1601 unsigned long
1602 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1603                           const unsigned long len, const unsigned long pgoff,
1604                           const unsigned long flags)
1605 {
1606         struct vm_area_struct *vma;
1607         struct mm_struct *mm = current->mm;
1608         unsigned long addr = addr0, start_addr;
1609
1610         /* requested length too big for entire address space */
1611         if (len > TASK_SIZE)
1612                 return -ENOMEM;
1613
1614         if (flags & MAP_FIXED)
1615                 return addr;
1616
1617         /* requesting a specific address */
1618         if (addr) {
1619                 addr = PAGE_ALIGN(addr);
1620                 vma = find_vma(mm, addr);
1621                 if (TASK_SIZE - len >= addr &&
1622                                 (!vma || addr + len <= vma->vm_start))
1623                         return addr;
1624         }
1625
1626         /* check if free_area_cache is useful for us */
1627         if (len <= mm->cached_hole_size) {
1628                 mm->cached_hole_size = 0;
1629                 mm->free_area_cache = mm->mmap_base;
1630         }
1631
1632 try_again:
1633         /* either no address requested or can't fit in requested address hole */
1634         start_addr = addr = mm->free_area_cache;
1635
1636         if (addr < len)
1637                 goto fail;
1638
1639         addr -= len;
1640         do {
1641                 /*
1642                  * Lookup failure means no vma is above this address,
1643                  * else if new region fits below vma->vm_start,
1644                  * return with success:
1645                  */
1646                 vma = find_vma(mm, addr);
1647                 if (!vma || addr+len <= vma->vm_start)
1648                         /* remember the address as a hint for next time */
1649                         return (mm->free_area_cache = addr);
1650
1651                 /* remember the largest hole we saw so far */
1652                 if (addr + mm->cached_hole_size < vma->vm_start)
1653                         mm->cached_hole_size = vma->vm_start - addr;
1654
1655                 /* try just below the current vma->vm_start */
1656                 addr = vma->vm_start-len;
1657         } while (len < vma->vm_start);
1658
1659 fail:
1660         /*
1661          * if hint left us with no space for the requested
1662          * mapping then try again:
1663          *
1664          * Note: this is different with the case of bottomup
1665          * which does the fully line-search, but we use find_vma
1666          * here that causes some holes skipped.
1667          */
1668         if (start_addr != mm->mmap_base) {
1669                 mm->free_area_cache = mm->mmap_base;
1670                 mm->cached_hole_size = 0;
1671                 goto try_again;
1672         }
1673
1674         /*
1675          * A failed mmap() very likely causes application failure,
1676          * so fall back to the bottom-up function here. This scenario
1677          * can happen with large stack limits and large mmap()
1678          * allocations.
1679          */
1680         mm->cached_hole_size = ~0UL;
1681         mm->free_area_cache = TASK_UNMAPPED_BASE;
1682         addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1683         /*
1684          * Restore the topdown base:
1685          */
1686         mm->free_area_cache = mm->mmap_base;
1687         mm->cached_hole_size = ~0UL;
1688
1689         return addr;
1690 }
1691 #endif
1692
1693 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1694 {
1695         /*
1696          * Is this a new hole at the highest possible address?
1697          */
1698         if (addr > mm->free_area_cache)
1699                 mm->free_area_cache = addr;
1700
1701         /* dont allow allocations above current base */
1702         if (mm->free_area_cache > mm->mmap_base)
1703                 mm->free_area_cache = mm->mmap_base;
1704 }
1705
1706 unsigned long
1707 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1708                 unsigned long pgoff, unsigned long flags)
1709 {
1710         unsigned long (*get_area)(struct file *, unsigned long,
1711                                   unsigned long, unsigned long, unsigned long);
1712
1713         unsigned long error = arch_mmap_check(addr, len, flags);
1714         if (error)
1715                 return error;
1716
1717         /* Careful about overflows.. */
1718         if (len > TASK_SIZE)
1719                 return -ENOMEM;
1720
1721         get_area = current->mm->get_unmapped_area;
1722         if (file && file->f_op && file->f_op->get_unmapped_area)
1723                 get_area = file->f_op->get_unmapped_area;
1724         addr = get_area(file, addr, len, pgoff, flags);
1725         if (IS_ERR_VALUE(addr))
1726                 return addr;
1727
1728         if (addr > TASK_SIZE - len)
1729                 return -ENOMEM;
1730         if (addr & ~PAGE_MASK)
1731                 return -EINVAL;
1732
1733         addr = arch_rebalance_pgtables(addr, len);
1734         error = security_mmap_addr(addr);
1735         return error ? error : addr;
1736 }
1737
1738 EXPORT_SYMBOL(get_unmapped_area);
1739
1740 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1741 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1742 {
1743         struct vm_area_struct *vma = NULL;
1744
1745         if (WARN_ON_ONCE(!mm))          /* Remove this in linux-3.6 */
1746                 return NULL;
1747
1748         /* Check the cache first. */
1749         /* (Cache hit rate is typically around 35%.) */
1750         vma = mm->mmap_cache;
1751         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1752                 struct rb_node *rb_node;
1753
1754                 rb_node = mm->mm_rb.rb_node;
1755                 vma = NULL;
1756
1757                 while (rb_node) {
1758                         struct vm_area_struct *vma_tmp;
1759
1760                         vma_tmp = rb_entry(rb_node,
1761                                            struct vm_area_struct, vm_rb);
1762
1763                         if (vma_tmp->vm_end > addr) {
1764                                 vma = vma_tmp;
1765                                 if (vma_tmp->vm_start <= addr)
1766                                         break;
1767                                 rb_node = rb_node->rb_left;
1768                         } else
1769                                 rb_node = rb_node->rb_right;
1770                 }
1771                 if (vma)
1772                         mm->mmap_cache = vma;
1773         }
1774         return vma;
1775 }
1776
1777 EXPORT_SYMBOL(find_vma);
1778
1779 /*
1780  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1781  */
1782 struct vm_area_struct *
1783 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1784                         struct vm_area_struct **pprev)
1785 {
1786         struct vm_area_struct *vma;
1787
1788         vma = find_vma(mm, addr);
1789         if (vma) {
1790                 *pprev = vma->vm_prev;
1791         } else {
1792                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1793                 *pprev = NULL;
1794                 while (rb_node) {
1795                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1796                         rb_node = rb_node->rb_right;
1797                 }
1798         }
1799         return vma;
1800 }
1801
1802 /*
1803  * Verify that the stack growth is acceptable and
1804  * update accounting. This is shared with both the
1805  * grow-up and grow-down cases.
1806  */
1807 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1808 {
1809         struct mm_struct *mm = vma->vm_mm;
1810         struct rlimit *rlim = current->signal->rlim;
1811         unsigned long new_start;
1812
1813         /* address space limit tests */
1814         if (!may_expand_vm(mm, grow))
1815                 return -ENOMEM;
1816
1817         /* Stack limit test */
1818         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
1819                 return -ENOMEM;
1820
1821         /* mlock limit tests */
1822         if (vma->vm_flags & VM_LOCKED) {
1823                 unsigned long locked;
1824                 unsigned long limit;
1825                 locked = mm->locked_vm + grow;
1826                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
1827                 limit >>= PAGE_SHIFT;
1828                 if (locked > limit && !capable(CAP_IPC_LOCK))
1829                         return -ENOMEM;
1830         }
1831
1832         /* Check to ensure the stack will not grow into a hugetlb-only region */
1833         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
1834                         vma->vm_end - size;
1835         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
1836                 return -EFAULT;
1837
1838         /*
1839          * Overcommit..  This must be the final test, as it will
1840          * update security statistics.
1841          */
1842         if (security_vm_enough_memory_mm(mm, grow))
1843                 return -ENOMEM;
1844
1845         /* Ok, everything looks good - let it rip */
1846         if (vma->vm_flags & VM_LOCKED)
1847                 mm->locked_vm += grow;
1848         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1849         return 0;
1850 }
1851
1852 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1853 /*
1854  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1855  * vma is the last one with address > vma->vm_end.  Have to extend vma.
1856  */
1857 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1858 {
1859         int error;
1860
1861         if (!(vma->vm_flags & VM_GROWSUP))
1862                 return -EFAULT;
1863
1864         /*
1865          * We must make sure the anon_vma is allocated
1866          * so that the anon_vma locking is not a noop.
1867          */
1868         if (unlikely(anon_vma_prepare(vma)))
1869                 return -ENOMEM;
1870         vma_lock_anon_vma(vma);
1871
1872         /*
1873          * vma->vm_start/vm_end cannot change under us because the caller
1874          * is required to hold the mmap_sem in read mode.  We need the
1875          * anon_vma lock to serialize against concurrent expand_stacks.
1876          * Also guard against wrapping around to address 0.
1877          */
1878         if (address < PAGE_ALIGN(address+4))
1879                 address = PAGE_ALIGN(address+4);
1880         else {
1881                 vma_unlock_anon_vma(vma);
1882                 return -ENOMEM;
1883         }
1884         error = 0;
1885
1886         /* Somebody else might have raced and expanded it already */
1887         if (address > vma->vm_end) {
1888                 unsigned long size, grow;
1889
1890                 size = address - vma->vm_start;
1891                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1892
1893                 error = -ENOMEM;
1894                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
1895                         error = acct_stack_growth(vma, size, grow);
1896                         if (!error) {
1897                                 anon_vma_interval_tree_pre_update_vma(vma);
1898                                 vma->vm_end = address;
1899                                 anon_vma_interval_tree_post_update_vma(vma);
1900                                 if (vma->vm_next)
1901                                         vma_gap_update(vma->vm_next);
1902                                 else
1903                                         vma->vm_mm->highest_vm_end = address;
1904                                 perf_event_mmap(vma);
1905                         }
1906                 }
1907         }
1908         vma_unlock_anon_vma(vma);
1909         khugepaged_enter_vma_merge(vma);
1910         validate_mm(vma->vm_mm);
1911         return error;
1912 }
1913 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1914
1915 /*
1916  * vma is the first one with address < vma->vm_start.  Have to extend vma.
1917  */
1918 int expand_downwards(struct vm_area_struct *vma,
1919                                    unsigned long address)
1920 {
1921         int error;
1922
1923         /*
1924          * We must make sure the anon_vma is allocated
1925          * so that the anon_vma locking is not a noop.
1926          */
1927         if (unlikely(anon_vma_prepare(vma)))
1928                 return -ENOMEM;
1929
1930         address &= PAGE_MASK;
1931         error = security_mmap_addr(address);
1932         if (error)
1933                 return error;
1934
1935         vma_lock_anon_vma(vma);
1936
1937         /*
1938          * vma->vm_start/vm_end cannot change under us because the caller
1939          * is required to hold the mmap_sem in read mode.  We need the
1940          * anon_vma lock to serialize against concurrent expand_stacks.
1941          */
1942
1943         /* Somebody else might have raced and expanded it already */
1944         if (address < vma->vm_start) {
1945                 unsigned long size, grow;
1946
1947                 size = vma->vm_end - address;
1948                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1949
1950                 error = -ENOMEM;
1951                 if (grow <= vma->vm_pgoff) {
1952                         error = acct_stack_growth(vma, size, grow);
1953                         if (!error) {
1954                                 anon_vma_interval_tree_pre_update_vma(vma);
1955                                 vma->vm_start = address;
1956                                 vma->vm_pgoff -= grow;
1957                                 anon_vma_interval_tree_post_update_vma(vma);
1958                                 vma_gap_update(vma);
1959                                 perf_event_mmap(vma);
1960                         }
1961                 }
1962         }
1963         vma_unlock_anon_vma(vma);
1964         khugepaged_enter_vma_merge(vma);
1965         validate_mm(vma->vm_mm);
1966         return error;
1967 }
1968
1969 #ifdef CONFIG_STACK_GROWSUP
1970 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1971 {
1972         return expand_upwards(vma, address);
1973 }
1974
1975 struct vm_area_struct *
1976 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1977 {
1978         struct vm_area_struct *vma, *prev;
1979
1980         addr &= PAGE_MASK;
1981         vma = find_vma_prev(mm, addr, &prev);
1982         if (vma && (vma->vm_start <= addr))
1983                 return vma;
1984         if (!prev || expand_stack(prev, addr))
1985                 return NULL;
1986         if (prev->vm_flags & VM_LOCKED) {
1987                 mlock_vma_pages_range(prev, addr, prev->vm_end);
1988         }
1989         return prev;
1990 }
1991 #else
1992 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1993 {
1994         return expand_downwards(vma, address);
1995 }
1996
1997 struct vm_area_struct *
1998 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1999 {
2000         struct vm_area_struct * vma;
2001         unsigned long start;
2002
2003         addr &= PAGE_MASK;
2004         vma = find_vma(mm,addr);
2005         if (!vma)
2006                 return NULL;
2007         if (vma->vm_start <= addr)
2008                 return vma;
2009         if (!(vma->vm_flags & VM_GROWSDOWN))
2010                 return NULL;
2011         start = vma->vm_start;
2012         if (expand_stack(vma, addr))
2013                 return NULL;
2014         if (vma->vm_flags & VM_LOCKED) {
2015                 mlock_vma_pages_range(vma, addr, start);
2016         }
2017         return vma;
2018 }
2019 #endif
2020
2021 /*
2022  * Ok - we have the memory areas we should free on the vma list,
2023  * so release them, and do the vma updates.
2024  *
2025  * Called with the mm semaphore held.
2026  */
2027 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2028 {
2029         unsigned long nr_accounted = 0;
2030
2031         /* Update high watermark before we lower total_vm */
2032         update_hiwater_vm(mm);
2033         do {
2034                 long nrpages = vma_pages(vma);
2035
2036                 if (vma->vm_flags & VM_ACCOUNT)
2037                         nr_accounted += nrpages;
2038                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2039                 vma = remove_vma(vma);
2040         } while (vma);
2041         vm_unacct_memory(nr_accounted);
2042         validate_mm(mm);
2043 }
2044
2045 /*
2046  * Get rid of page table information in the indicated region.
2047  *
2048  * Called with the mm semaphore held.
2049  */
2050 static void unmap_region(struct mm_struct *mm,
2051                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2052                 unsigned long start, unsigned long end)
2053 {
2054         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2055         struct mmu_gather tlb;
2056
2057         lru_add_drain();
2058         tlb_gather_mmu(&tlb, mm, 0);
2059         update_hiwater_rss(mm);
2060         unmap_vmas(&tlb, vma, start, end);
2061         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2062                                  next ? next->vm_start : 0);
2063         tlb_finish_mmu(&tlb, start, end);
2064 }
2065
2066 /*
2067  * Create a list of vma's touched by the unmap, removing them from the mm's
2068  * vma list as we go..
2069  */
2070 static void
2071 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2072         struct vm_area_struct *prev, unsigned long end)
2073 {
2074         struct vm_area_struct **insertion_point;
2075         struct vm_area_struct *tail_vma = NULL;
2076         unsigned long addr;
2077
2078         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2079         vma->vm_prev = NULL;
2080         do {
2081                 vma_rb_erase(vma, &mm->mm_rb);
2082                 mm->map_count--;
2083                 tail_vma = vma;
2084                 vma = vma->vm_next;
2085         } while (vma && vma->vm_start < end);
2086         *insertion_point = vma;
2087         if (vma) {
2088                 vma->vm_prev = prev;
2089                 vma_gap_update(vma);
2090         } else
2091                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2092         tail_vma->vm_next = NULL;
2093         if (mm->unmap_area == arch_unmap_area)
2094                 addr = prev ? prev->vm_end : mm->mmap_base;
2095         else
2096                 addr = vma ?  vma->vm_start : mm->mmap_base;
2097         mm->unmap_area(mm, addr);
2098         mm->mmap_cache = NULL;          /* Kill the cache. */
2099 }
2100
2101 /*
2102  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2103  * munmap path where it doesn't make sense to fail.
2104  */
2105 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2106               unsigned long addr, int new_below)
2107 {
2108         struct mempolicy *pol;
2109         struct vm_area_struct *new;
2110         int err = -ENOMEM;
2111
2112         if (is_vm_hugetlb_page(vma) && (addr &
2113                                         ~(huge_page_mask(hstate_vma(vma)))))
2114                 return -EINVAL;
2115
2116         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2117         if (!new)
2118                 goto out_err;
2119
2120         /* most fields are the same, copy all, and then fixup */
2121         *new = *vma;
2122
2123         INIT_LIST_HEAD(&new->anon_vma_chain);
2124
2125         if (new_below)
2126                 new->vm_end = addr;
2127         else {
2128                 new->vm_start = addr;
2129                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2130         }
2131
2132         pol = mpol_dup(vma_policy(vma));
2133         if (IS_ERR(pol)) {
2134                 err = PTR_ERR(pol);
2135                 goto out_free_vma;
2136         }
2137         vma_set_policy(new, pol);
2138
2139         if (anon_vma_clone(new, vma))
2140                 goto out_free_mpol;
2141
2142         if (new->vm_file)
2143                 get_file(new->vm_file);
2144
2145         if (new->vm_ops && new->vm_ops->open)
2146                 new->vm_ops->open(new);
2147
2148         if (new_below)
2149                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2150                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2151         else
2152                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2153
2154         /* Success. */
2155         if (!err)
2156                 return 0;
2157
2158         /* Clean everything up if vma_adjust failed. */
2159         if (new->vm_ops && new->vm_ops->close)
2160                 new->vm_ops->close(new);
2161         if (new->vm_file)
2162                 fput(new->vm_file);
2163         unlink_anon_vmas(new);
2164  out_free_mpol:
2165         mpol_put(pol);
2166  out_free_vma:
2167         kmem_cache_free(vm_area_cachep, new);
2168  out_err:
2169         return err;
2170 }
2171
2172 /*
2173  * Split a vma into two pieces at address 'addr', a new vma is allocated
2174  * either for the first part or the tail.
2175  */
2176 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2177               unsigned long addr, int new_below)
2178 {
2179         if (mm->map_count >= sysctl_max_map_count)
2180                 return -ENOMEM;
2181
2182         return __split_vma(mm, vma, addr, new_below);
2183 }
2184
2185 /* Munmap is split into 2 main parts -- this part which finds
2186  * what needs doing, and the areas themselves, which do the
2187  * work.  This now handles partial unmappings.
2188  * Jeremy Fitzhardinge <jeremy@goop.org>
2189  */
2190 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2191 {
2192         unsigned long end;
2193         struct vm_area_struct *vma, *prev, *last;
2194
2195         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2196                 return -EINVAL;
2197
2198         if ((len = PAGE_ALIGN(len)) == 0)
2199                 return -EINVAL;
2200
2201         /* Find the first overlapping VMA */
2202         vma = find_vma(mm, start);
2203         if (!vma)
2204                 return 0;
2205         prev = vma->vm_prev;
2206         /* we have  start < vma->vm_end  */
2207
2208         /* if it doesn't overlap, we have nothing.. */
2209         end = start + len;
2210         if (vma->vm_start >= end)
2211                 return 0;
2212
2213         /*
2214          * If we need to split any vma, do it now to save pain later.
2215          *
2216          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2217          * unmapped vm_area_struct will remain in use: so lower split_vma
2218          * places tmp vma above, and higher split_vma places tmp vma below.
2219          */
2220         if (start > vma->vm_start) {
2221                 int error;
2222
2223                 /*
2224                  * Make sure that map_count on return from munmap() will
2225                  * not exceed its limit; but let map_count go just above
2226                  * its limit temporarily, to help free resources as expected.
2227                  */
2228                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2229                         return -ENOMEM;
2230
2231                 error = __split_vma(mm, vma, start, 0);
2232                 if (error)
2233                         return error;
2234                 prev = vma;
2235         }
2236
2237         /* Does it split the last one? */
2238         last = find_vma(mm, end);
2239         if (last && end > last->vm_start) {
2240                 int error = __split_vma(mm, last, end, 1);
2241                 if (error)
2242                         return error;
2243         }
2244         vma = prev? prev->vm_next: mm->mmap;
2245
2246         /*
2247          * unlock any mlock()ed ranges before detaching vmas
2248          */
2249         if (mm->locked_vm) {
2250                 struct vm_area_struct *tmp = vma;
2251                 while (tmp && tmp->vm_start < end) {
2252                         if (tmp->vm_flags & VM_LOCKED) {
2253                                 mm->locked_vm -= vma_pages(tmp);
2254                                 munlock_vma_pages_all(tmp);
2255                         }
2256                         tmp = tmp->vm_next;
2257                 }
2258         }
2259
2260         /*
2261          * Remove the vma's, and unmap the actual pages
2262          */
2263         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2264         unmap_region(mm, vma, prev, start, end);
2265
2266         /* Fix up all other VM information */
2267         remove_vma_list(mm, vma);
2268
2269         return 0;
2270 }
2271
2272 int vm_munmap(unsigned long start, size_t len)
2273 {
2274         int ret;
2275         struct mm_struct *mm = current->mm;
2276
2277         down_write(&mm->mmap_sem);
2278         ret = do_munmap(mm, start, len);
2279         up_write(&mm->mmap_sem);
2280         return ret;
2281 }
2282 EXPORT_SYMBOL(vm_munmap);
2283
2284 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2285 {
2286         profile_munmap(addr);
2287         return vm_munmap(addr, len);
2288 }
2289
2290 static inline void verify_mm_writelocked(struct mm_struct *mm)
2291 {
2292 #ifdef CONFIG_DEBUG_VM
2293         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2294                 WARN_ON(1);
2295                 up_read(&mm->mmap_sem);
2296         }
2297 #endif
2298 }
2299
2300 /*
2301  *  this is really a simplified "do_mmap".  it only handles
2302  *  anonymous maps.  eventually we may be able to do some
2303  *  brk-specific accounting here.
2304  */
2305 static unsigned long do_brk(unsigned long addr, unsigned long len)
2306 {
2307         struct mm_struct * mm = current->mm;
2308         struct vm_area_struct * vma, * prev;
2309         unsigned long flags;
2310         struct rb_node ** rb_link, * rb_parent;
2311         pgoff_t pgoff = addr >> PAGE_SHIFT;
2312         int error;
2313
2314         len = PAGE_ALIGN(len);
2315         if (!len)
2316                 return addr;
2317
2318         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2319
2320         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2321         if (error & ~PAGE_MASK)
2322                 return error;
2323
2324         /*
2325          * mlock MCL_FUTURE?
2326          */
2327         if (mm->def_flags & VM_LOCKED) {
2328                 unsigned long locked, lock_limit;
2329                 locked = len >> PAGE_SHIFT;
2330                 locked += mm->locked_vm;
2331                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2332                 lock_limit >>= PAGE_SHIFT;
2333                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2334                         return -EAGAIN;
2335         }
2336
2337         /*
2338          * mm->mmap_sem is required to protect against another thread
2339          * changing the mappings in case we sleep.
2340          */
2341         verify_mm_writelocked(mm);
2342
2343         /*
2344          * Clear old maps.  this also does some error checking for us
2345          */
2346  munmap_back:
2347         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2348                 if (do_munmap(mm, addr, len))
2349                         return -ENOMEM;
2350                 goto munmap_back;
2351         }
2352
2353         /* Check against address space limits *after* clearing old maps... */
2354         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2355                 return -ENOMEM;
2356
2357         if (mm->map_count > sysctl_max_map_count)
2358                 return -ENOMEM;
2359
2360         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2361                 return -ENOMEM;
2362
2363         /* Can we just expand an old private anonymous mapping? */
2364         vma = vma_merge(mm, prev, addr, addr + len, flags,
2365                                         NULL, NULL, pgoff, NULL);
2366         if (vma)
2367                 goto out;
2368
2369         /*
2370          * create a vma struct for an anonymous mapping
2371          */
2372         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2373         if (!vma) {
2374                 vm_unacct_memory(len >> PAGE_SHIFT);
2375                 return -ENOMEM;
2376         }
2377
2378         INIT_LIST_HEAD(&vma->anon_vma_chain);
2379         vma->vm_mm = mm;
2380         vma->vm_start = addr;
2381         vma->vm_end = addr + len;
2382         vma->vm_pgoff = pgoff;
2383         vma->vm_flags = flags;
2384         vma->vm_page_prot = vm_get_page_prot(flags);
2385         vma_link(mm, vma, prev, rb_link, rb_parent);
2386 out:
2387         perf_event_mmap(vma);
2388         mm->total_vm += len >> PAGE_SHIFT;
2389         if (flags & VM_LOCKED) {
2390                 if (!mlock_vma_pages_range(vma, addr, addr + len))
2391                         mm->locked_vm += (len >> PAGE_SHIFT);
2392         }
2393         return addr;
2394 }
2395
2396 unsigned long vm_brk(unsigned long addr, unsigned long len)
2397 {
2398         struct mm_struct *mm = current->mm;
2399         unsigned long ret;
2400
2401         down_write(&mm->mmap_sem);
2402         ret = do_brk(addr, len);
2403         up_write(&mm->mmap_sem);
2404         return ret;
2405 }
2406 EXPORT_SYMBOL(vm_brk);
2407
2408 /* Release all mmaps. */
2409 void exit_mmap(struct mm_struct *mm)
2410 {
2411         struct mmu_gather tlb;
2412         struct vm_area_struct *vma;
2413         unsigned long nr_accounted = 0;
2414
2415         /* mm's last user has gone, and its about to be pulled down */
2416         mmu_notifier_release(mm);
2417
2418         if (mm->locked_vm) {
2419                 vma = mm->mmap;
2420                 while (vma) {
2421                         if (vma->vm_flags & VM_LOCKED)
2422                                 munlock_vma_pages_all(vma);
2423                         vma = vma->vm_next;
2424                 }
2425         }
2426
2427         arch_exit_mmap(mm);
2428
2429         vma = mm->mmap;
2430         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2431                 return;
2432
2433         lru_add_drain();
2434         flush_cache_mm(mm);
2435         tlb_gather_mmu(&tlb, mm, 1);
2436         /* update_hiwater_rss(mm) here? but nobody should be looking */
2437         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2438         unmap_vmas(&tlb, vma, 0, -1);
2439
2440         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2441         tlb_finish_mmu(&tlb, 0, -1);
2442
2443         /*
2444          * Walk the list again, actually closing and freeing it,
2445          * with preemption enabled, without holding any MM locks.
2446          */
2447         while (vma) {
2448                 if (vma->vm_flags & VM_ACCOUNT)
2449                         nr_accounted += vma_pages(vma);
2450                 vma = remove_vma(vma);
2451         }
2452         vm_unacct_memory(nr_accounted);
2453
2454         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2455 }
2456
2457 /* Insert vm structure into process list sorted by address
2458  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2459  * then i_mmap_mutex is taken here.
2460  */
2461 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2462 {
2463         struct vm_area_struct *prev;
2464         struct rb_node **rb_link, *rb_parent;
2465
2466         /*
2467          * The vm_pgoff of a purely anonymous vma should be irrelevant
2468          * until its first write fault, when page's anon_vma and index
2469          * are set.  But now set the vm_pgoff it will almost certainly
2470          * end up with (unless mremap moves it elsewhere before that
2471          * first wfault), so /proc/pid/maps tells a consistent story.
2472          *
2473          * By setting it to reflect the virtual start address of the
2474          * vma, merges and splits can happen in a seamless way, just
2475          * using the existing file pgoff checks and manipulations.
2476          * Similarly in do_mmap_pgoff and in do_brk.
2477          */
2478         if (!vma->vm_file) {
2479                 BUG_ON(vma->anon_vma);
2480                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2481         }
2482         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2483                            &prev, &rb_link, &rb_parent))
2484                 return -ENOMEM;
2485         if ((vma->vm_flags & VM_ACCOUNT) &&
2486              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2487                 return -ENOMEM;
2488
2489         vma_link(mm, vma, prev, rb_link, rb_parent);
2490         return 0;
2491 }
2492
2493 /*
2494  * Copy the vma structure to a new location in the same mm,
2495  * prior to moving page table entries, to effect an mremap move.
2496  */
2497 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2498         unsigned long addr, unsigned long len, pgoff_t pgoff,
2499         bool *need_rmap_locks)
2500 {
2501         struct vm_area_struct *vma = *vmap;
2502         unsigned long vma_start = vma->vm_start;
2503         struct mm_struct *mm = vma->vm_mm;
2504         struct vm_area_struct *new_vma, *prev;
2505         struct rb_node **rb_link, *rb_parent;
2506         struct mempolicy *pol;
2507         bool faulted_in_anon_vma = true;
2508
2509         /*
2510          * If anonymous vma has not yet been faulted, update new pgoff
2511          * to match new location, to increase its chance of merging.
2512          */
2513         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2514                 pgoff = addr >> PAGE_SHIFT;
2515                 faulted_in_anon_vma = false;
2516         }
2517
2518         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2519                 return NULL;    /* should never get here */
2520         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2521                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2522         if (new_vma) {
2523                 /*
2524                  * Source vma may have been merged into new_vma
2525                  */
2526                 if (unlikely(vma_start >= new_vma->vm_start &&
2527                              vma_start < new_vma->vm_end)) {
2528                         /*
2529                          * The only way we can get a vma_merge with
2530                          * self during an mremap is if the vma hasn't
2531                          * been faulted in yet and we were allowed to
2532                          * reset the dst vma->vm_pgoff to the
2533                          * destination address of the mremap to allow
2534                          * the merge to happen. mremap must change the
2535                          * vm_pgoff linearity between src and dst vmas
2536                          * (in turn preventing a vma_merge) to be
2537                          * safe. It is only safe to keep the vm_pgoff
2538                          * linear if there are no pages mapped yet.
2539                          */
2540                         VM_BUG_ON(faulted_in_anon_vma);
2541                         *vmap = vma = new_vma;
2542                 }
2543                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2544         } else {
2545                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2546                 if (new_vma) {
2547                         *new_vma = *vma;
2548                         new_vma->vm_start = addr;
2549                         new_vma->vm_end = addr + len;
2550                         new_vma->vm_pgoff = pgoff;
2551                         pol = mpol_dup(vma_policy(vma));
2552                         if (IS_ERR(pol))
2553                                 goto out_free_vma;
2554                         vma_set_policy(new_vma, pol);
2555                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2556                         if (anon_vma_clone(new_vma, vma))
2557                                 goto out_free_mempol;
2558                         if (new_vma->vm_file)
2559                                 get_file(new_vma->vm_file);
2560                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2561                                 new_vma->vm_ops->open(new_vma);
2562                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2563                         *need_rmap_locks = false;
2564                 }
2565         }
2566         return new_vma;
2567
2568  out_free_mempol:
2569         mpol_put(pol);
2570  out_free_vma:
2571         kmem_cache_free(vm_area_cachep, new_vma);
2572         return NULL;
2573 }
2574
2575 /*
2576  * Return true if the calling process may expand its vm space by the passed
2577  * number of pages
2578  */
2579 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2580 {
2581         unsigned long cur = mm->total_vm;       /* pages */
2582         unsigned long lim;
2583
2584         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2585
2586         if (cur + npages > lim)
2587                 return 0;
2588         return 1;
2589 }
2590
2591
2592 static int special_mapping_fault(struct vm_area_struct *vma,
2593                                 struct vm_fault *vmf)
2594 {
2595         pgoff_t pgoff;
2596         struct page **pages;
2597
2598         /*
2599          * special mappings have no vm_file, and in that case, the mm
2600          * uses vm_pgoff internally. So we have to subtract it from here.
2601          * We are allowed to do this because we are the mm; do not copy
2602          * this code into drivers!
2603          */
2604         pgoff = vmf->pgoff - vma->vm_pgoff;
2605
2606         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2607                 pgoff--;
2608
2609         if (*pages) {
2610                 struct page *page = *pages;
2611                 get_page(page);
2612                 vmf->page = page;
2613                 return 0;
2614         }
2615
2616         return VM_FAULT_SIGBUS;
2617 }
2618
2619 /*
2620  * Having a close hook prevents vma merging regardless of flags.
2621  */
2622 static void special_mapping_close(struct vm_area_struct *vma)
2623 {
2624 }
2625
2626 static const struct vm_operations_struct special_mapping_vmops = {
2627         .close = special_mapping_close,
2628         .fault = special_mapping_fault,
2629 };
2630
2631 /*
2632  * Called with mm->mmap_sem held for writing.
2633  * Insert a new vma covering the given region, with the given flags.
2634  * Its pages are supplied by the given array of struct page *.
2635  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2636  * The region past the last page supplied will always produce SIGBUS.
2637  * The array pointer and the pages it points to are assumed to stay alive
2638  * for as long as this mapping might exist.
2639  */
2640 int install_special_mapping(struct mm_struct *mm,
2641                             unsigned long addr, unsigned long len,
2642                             unsigned long vm_flags, struct page **pages)
2643 {
2644         int ret;
2645         struct vm_area_struct *vma;
2646
2647         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2648         if (unlikely(vma == NULL))
2649                 return -ENOMEM;
2650
2651         INIT_LIST_HEAD(&vma->anon_vma_chain);
2652         vma->vm_mm = mm;
2653         vma->vm_start = addr;
2654         vma->vm_end = addr + len;
2655
2656         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2657         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2658
2659         vma->vm_ops = &special_mapping_vmops;
2660         vma->vm_private_data = pages;
2661
2662         ret = insert_vm_struct(mm, vma);
2663         if (ret)
2664                 goto out;
2665
2666         mm->total_vm += len >> PAGE_SHIFT;
2667
2668         perf_event_mmap(vma);
2669
2670         return 0;
2671
2672 out:
2673         kmem_cache_free(vm_area_cachep, vma);
2674         return ret;
2675 }
2676
2677 static DEFINE_MUTEX(mm_all_locks_mutex);
2678
2679 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2680 {
2681         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2682                 /*
2683                  * The LSB of head.next can't change from under us
2684                  * because we hold the mm_all_locks_mutex.
2685                  */
2686                 mutex_lock_nest_lock(&anon_vma->root->mutex, &mm->mmap_sem);
2687                 /*
2688                  * We can safely modify head.next after taking the
2689                  * anon_vma->root->mutex. If some other vma in this mm shares
2690                  * the same anon_vma we won't take it again.
2691                  *
2692                  * No need of atomic instructions here, head.next
2693                  * can't change from under us thanks to the
2694                  * anon_vma->root->mutex.
2695                  */
2696                 if (__test_and_set_bit(0, (unsigned long *)
2697                                        &anon_vma->root->rb_root.rb_node))
2698                         BUG();
2699         }
2700 }
2701
2702 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2703 {
2704         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2705                 /*
2706                  * AS_MM_ALL_LOCKS can't change from under us because
2707                  * we hold the mm_all_locks_mutex.
2708                  *
2709                  * Operations on ->flags have to be atomic because
2710                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2711                  * mm_all_locks_mutex, there may be other cpus
2712                  * changing other bitflags in parallel to us.
2713                  */
2714                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2715                         BUG();
2716                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2717         }
2718 }
2719
2720 /*
2721  * This operation locks against the VM for all pte/vma/mm related
2722  * operations that could ever happen on a certain mm. This includes
2723  * vmtruncate, try_to_unmap, and all page faults.
2724  *
2725  * The caller must take the mmap_sem in write mode before calling
2726  * mm_take_all_locks(). The caller isn't allowed to release the
2727  * mmap_sem until mm_drop_all_locks() returns.
2728  *
2729  * mmap_sem in write mode is required in order to block all operations
2730  * that could modify pagetables and free pages without need of
2731  * altering the vma layout (for example populate_range() with
2732  * nonlinear vmas). It's also needed in write mode to avoid new
2733  * anon_vmas to be associated with existing vmas.
2734  *
2735  * A single task can't take more than one mm_take_all_locks() in a row
2736  * or it would deadlock.
2737  *
2738  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2739  * mapping->flags avoid to take the same lock twice, if more than one
2740  * vma in this mm is backed by the same anon_vma or address_space.
2741  *
2742  * We can take all the locks in random order because the VM code
2743  * taking i_mmap_mutex or anon_vma->mutex outside the mmap_sem never
2744  * takes more than one of them in a row. Secondly we're protected
2745  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2746  *
2747  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2748  * that may have to take thousand of locks.
2749  *
2750  * mm_take_all_locks() can fail if it's interrupted by signals.
2751  */
2752 int mm_take_all_locks(struct mm_struct *mm)
2753 {
2754         struct vm_area_struct *vma;
2755         struct anon_vma_chain *avc;
2756
2757         BUG_ON(down_read_trylock(&mm->mmap_sem));
2758
2759         mutex_lock(&mm_all_locks_mutex);
2760
2761         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2762                 if (signal_pending(current))
2763                         goto out_unlock;
2764                 if (vma->vm_file && vma->vm_file->f_mapping)
2765                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2766         }
2767
2768         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2769                 if (signal_pending(current))
2770                         goto out_unlock;
2771                 if (vma->anon_vma)
2772                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2773                                 vm_lock_anon_vma(mm, avc->anon_vma);
2774         }
2775
2776         return 0;
2777
2778 out_unlock:
2779         mm_drop_all_locks(mm);
2780         return -EINTR;
2781 }
2782
2783 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
2784 {
2785         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2786                 /*
2787                  * The LSB of head.next can't change to 0 from under
2788                  * us because we hold the mm_all_locks_mutex.
2789                  *
2790                  * We must however clear the bitflag before unlocking
2791                  * the vma so the users using the anon_vma->rb_root will
2792                  * never see our bitflag.
2793                  *
2794                  * No need of atomic instructions here, head.next
2795                  * can't change from under us until we release the
2796                  * anon_vma->root->mutex.
2797                  */
2798                 if (!__test_and_clear_bit(0, (unsigned long *)
2799                                           &anon_vma->root->rb_root.rb_node))
2800                         BUG();
2801                 anon_vma_unlock(anon_vma);
2802         }
2803 }
2804
2805 static void vm_unlock_mapping(struct address_space *mapping)
2806 {
2807         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2808                 /*
2809                  * AS_MM_ALL_LOCKS can't change to 0 from under us
2810                  * because we hold the mm_all_locks_mutex.
2811                  */
2812                 mutex_unlock(&mapping->i_mmap_mutex);
2813                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
2814                                         &mapping->flags))
2815                         BUG();
2816         }
2817 }
2818
2819 /*
2820  * The mmap_sem cannot be released by the caller until
2821  * mm_drop_all_locks() returns.
2822  */
2823 void mm_drop_all_locks(struct mm_struct *mm)
2824 {
2825         struct vm_area_struct *vma;
2826         struct anon_vma_chain *avc;
2827
2828         BUG_ON(down_read_trylock(&mm->mmap_sem));
2829         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
2830
2831         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2832                 if (vma->anon_vma)
2833                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2834                                 vm_unlock_anon_vma(avc->anon_vma);
2835                 if (vma->vm_file && vma->vm_file->f_mapping)
2836                         vm_unlock_mapping(vma->vm_file->f_mapping);
2837         }
2838
2839         mutex_unlock(&mm_all_locks_mutex);
2840 }
2841
2842 /*
2843  * initialise the VMA slab
2844  */
2845 void __init mmap_init(void)
2846 {
2847         int ret;
2848
2849         ret = percpu_counter_init(&vm_committed_as, 0);
2850         VM_BUG_ON(ret);
2851 }