mm: make do_mmap_pgoff return populate as a size in bytes, not as a bool
[linux-3.10.git] / mm / mmap.c
1 /*
2  * mm/mmap.c
3  *
4  * Written by obz.
5  *
6  * Address space accounting code        <alan@lxorguk.ukuu.org.uk>
7  */
8
9 #include <linux/slab.h>
10 #include <linux/backing-dev.h>
11 #include <linux/mm.h>
12 #include <linux/shm.h>
13 #include <linux/mman.h>
14 #include <linux/pagemap.h>
15 #include <linux/swap.h>
16 #include <linux/syscalls.h>
17 #include <linux/capability.h>
18 #include <linux/init.h>
19 #include <linux/file.h>
20 #include <linux/fs.h>
21 #include <linux/personality.h>
22 #include <linux/security.h>
23 #include <linux/hugetlb.h>
24 #include <linux/profile.h>
25 #include <linux/export.h>
26 #include <linux/mount.h>
27 #include <linux/mempolicy.h>
28 #include <linux/rmap.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/perf_event.h>
31 #include <linux/audit.h>
32 #include <linux/khugepaged.h>
33 #include <linux/uprobes.h>
34 #include <linux/rbtree_augmented.h>
35 #include <linux/sched/sysctl.h>
36
37 #include <asm/uaccess.h>
38 #include <asm/cacheflush.h>
39 #include <asm/tlb.h>
40 #include <asm/mmu_context.h>
41
42 #include "internal.h"
43
44 #ifndef arch_mmap_check
45 #define arch_mmap_check(addr, len, flags)       (0)
46 #endif
47
48 #ifndef arch_rebalance_pgtables
49 #define arch_rebalance_pgtables(addr, len)              (addr)
50 #endif
51
52 static void unmap_region(struct mm_struct *mm,
53                 struct vm_area_struct *vma, struct vm_area_struct *prev,
54                 unsigned long start, unsigned long end);
55
56 /* description of effects of mapping type and prot in current implementation.
57  * this is due to the limited x86 page protection hardware.  The expected
58  * behavior is in parens:
59  *
60  * map_type     prot
61  *              PROT_NONE       PROT_READ       PROT_WRITE      PROT_EXEC
62  * MAP_SHARED   r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
63  *              w: (no) no      w: (no) no      w: (yes) yes    w: (no) no
64  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
65  *              
66  * MAP_PRIVATE  r: (no) no      r: (yes) yes    r: (no) yes     r: (no) yes
67  *              w: (no) no      w: (no) no      w: (copy) copy  w: (no) no
68  *              x: (no) no      x: (no) yes     x: (no) yes     x: (yes) yes
69  *
70  */
71 pgprot_t protection_map[16] = {
72         __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
73         __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
74 };
75
76 pgprot_t vm_get_page_prot(unsigned long vm_flags)
77 {
78         return __pgprot(pgprot_val(protection_map[vm_flags &
79                                 (VM_READ|VM_WRITE|VM_EXEC|VM_SHARED)]) |
80                         pgprot_val(arch_vm_get_page_prot(vm_flags)));
81 }
82 EXPORT_SYMBOL(vm_get_page_prot);
83
84 int sysctl_overcommit_memory __read_mostly = OVERCOMMIT_GUESS;  /* heuristic overcommit */
85 int sysctl_overcommit_ratio __read_mostly = 50; /* default is 50% */
86 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
87 /*
88  * Make sure vm_committed_as in one cacheline and not cacheline shared with
89  * other variables. It can be updated by several CPUs frequently.
90  */
91 struct percpu_counter vm_committed_as ____cacheline_aligned_in_smp;
92
93 /*
94  * The global memory commitment made in the system can be a metric
95  * that can be used to drive ballooning decisions when Linux is hosted
96  * as a guest. On Hyper-V, the host implements a policy engine for dynamically
97  * balancing memory across competing virtual machines that are hosted.
98  * Several metrics drive this policy engine including the guest reported
99  * memory commitment.
100  */
101 unsigned long vm_memory_committed(void)
102 {
103         return percpu_counter_read_positive(&vm_committed_as);
104 }
105 EXPORT_SYMBOL_GPL(vm_memory_committed);
106
107 /*
108  * Check that a process has enough memory to allocate a new virtual
109  * mapping. 0 means there is enough memory for the allocation to
110  * succeed and -ENOMEM implies there is not.
111  *
112  * We currently support three overcommit policies, which are set via the
113  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
114  *
115  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
116  * Additional code 2002 Jul 20 by Robert Love.
117  *
118  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
119  *
120  * Note this is a helper function intended to be used by LSMs which
121  * wish to use this logic.
122  */
123 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
124 {
125         unsigned long free, allowed;
126
127         vm_acct_memory(pages);
128
129         /*
130          * Sometimes we want to use more memory than we have
131          */
132         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
133                 return 0;
134
135         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
136                 free = global_page_state(NR_FREE_PAGES);
137                 free += global_page_state(NR_FILE_PAGES);
138
139                 /*
140                  * shmem pages shouldn't be counted as free in this
141                  * case, they can't be purged, only swapped out, and
142                  * that won't affect the overall amount of available
143                  * memory in the system.
144                  */
145                 free -= global_page_state(NR_SHMEM);
146
147                 free += nr_swap_pages;
148
149                 /*
150                  * Any slabs which are created with the
151                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
152                  * which are reclaimable, under pressure.  The dentry
153                  * cache and most inode caches should fall into this
154                  */
155                 free += global_page_state(NR_SLAB_RECLAIMABLE);
156
157                 /*
158                  * Leave reserved pages. The pages are not for anonymous pages.
159                  */
160                 if (free <= totalreserve_pages)
161                         goto error;
162                 else
163                         free -= totalreserve_pages;
164
165                 /*
166                  * Leave the last 3% for root
167                  */
168                 if (!cap_sys_admin)
169                         free -= free / 32;
170
171                 if (free > pages)
172                         return 0;
173
174                 goto error;
175         }
176
177         allowed = (totalram_pages - hugetlb_total_pages())
178                 * sysctl_overcommit_ratio / 100;
179         /*
180          * Leave the last 3% for root
181          */
182         if (!cap_sys_admin)
183                 allowed -= allowed / 32;
184         allowed += total_swap_pages;
185
186         /* Don't let a single process grow too big:
187            leave 3% of the size of this process for other processes */
188         if (mm)
189                 allowed -= mm->total_vm / 32;
190
191         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
192                 return 0;
193 error:
194         vm_unacct_memory(pages);
195
196         return -ENOMEM;
197 }
198
199 /*
200  * Requires inode->i_mapping->i_mmap_mutex
201  */
202 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
203                 struct file *file, struct address_space *mapping)
204 {
205         if (vma->vm_flags & VM_DENYWRITE)
206                 atomic_inc(&file->f_path.dentry->d_inode->i_writecount);
207         if (vma->vm_flags & VM_SHARED)
208                 mapping->i_mmap_writable--;
209
210         flush_dcache_mmap_lock(mapping);
211         if (unlikely(vma->vm_flags & VM_NONLINEAR))
212                 list_del_init(&vma->shared.nonlinear);
213         else
214                 vma_interval_tree_remove(vma, &mapping->i_mmap);
215         flush_dcache_mmap_unlock(mapping);
216 }
217
218 /*
219  * Unlink a file-based vm structure from its interval tree, to hide
220  * vma from rmap and vmtruncate before freeing its page tables.
221  */
222 void unlink_file_vma(struct vm_area_struct *vma)
223 {
224         struct file *file = vma->vm_file;
225
226         if (file) {
227                 struct address_space *mapping = file->f_mapping;
228                 mutex_lock(&mapping->i_mmap_mutex);
229                 __remove_shared_vm_struct(vma, file, mapping);
230                 mutex_unlock(&mapping->i_mmap_mutex);
231         }
232 }
233
234 /*
235  * Close a vm structure and free it, returning the next.
236  */
237 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
238 {
239         struct vm_area_struct *next = vma->vm_next;
240
241         might_sleep();
242         if (vma->vm_ops && vma->vm_ops->close)
243                 vma->vm_ops->close(vma);
244         if (vma->vm_file)
245                 fput(vma->vm_file);
246         mpol_put(vma_policy(vma));
247         kmem_cache_free(vm_area_cachep, vma);
248         return next;
249 }
250
251 static unsigned long do_brk(unsigned long addr, unsigned long len);
252
253 SYSCALL_DEFINE1(brk, unsigned long, brk)
254 {
255         unsigned long rlim, retval;
256         unsigned long newbrk, oldbrk;
257         struct mm_struct *mm = current->mm;
258         unsigned long min_brk;
259         bool populate;
260
261         down_write(&mm->mmap_sem);
262
263 #ifdef CONFIG_COMPAT_BRK
264         /*
265          * CONFIG_COMPAT_BRK can still be overridden by setting
266          * randomize_va_space to 2, which will still cause mm->start_brk
267          * to be arbitrarily shifted
268          */
269         if (current->brk_randomized)
270                 min_brk = mm->start_brk;
271         else
272                 min_brk = mm->end_data;
273 #else
274         min_brk = mm->start_brk;
275 #endif
276         if (brk < min_brk)
277                 goto out;
278
279         /*
280          * Check against rlimit here. If this check is done later after the test
281          * of oldbrk with newbrk then it can escape the test and let the data
282          * segment grow beyond its set limit the in case where the limit is
283          * not page aligned -Ram Gupta
284          */
285         rlim = rlimit(RLIMIT_DATA);
286         if (rlim < RLIM_INFINITY && (brk - mm->start_brk) +
287                         (mm->end_data - mm->start_data) > rlim)
288                 goto out;
289
290         newbrk = PAGE_ALIGN(brk);
291         oldbrk = PAGE_ALIGN(mm->brk);
292         if (oldbrk == newbrk)
293                 goto set_brk;
294
295         /* Always allow shrinking brk. */
296         if (brk <= mm->brk) {
297                 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
298                         goto set_brk;
299                 goto out;
300         }
301
302         /* Check against existing mmap mappings. */
303         if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
304                 goto out;
305
306         /* Ok, looks good - let it rip. */
307         if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
308                 goto out;
309
310 set_brk:
311         mm->brk = brk;
312         populate = newbrk > oldbrk && (mm->def_flags & VM_LOCKED) != 0;
313         up_write(&mm->mmap_sem);
314         if (populate)
315                 mm_populate(oldbrk, newbrk - oldbrk);
316         return brk;
317
318 out:
319         retval = mm->brk;
320         up_write(&mm->mmap_sem);
321         return retval;
322 }
323
324 static long vma_compute_subtree_gap(struct vm_area_struct *vma)
325 {
326         unsigned long max, subtree_gap;
327         max = vma->vm_start;
328         if (vma->vm_prev)
329                 max -= vma->vm_prev->vm_end;
330         if (vma->vm_rb.rb_left) {
331                 subtree_gap = rb_entry(vma->vm_rb.rb_left,
332                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
333                 if (subtree_gap > max)
334                         max = subtree_gap;
335         }
336         if (vma->vm_rb.rb_right) {
337                 subtree_gap = rb_entry(vma->vm_rb.rb_right,
338                                 struct vm_area_struct, vm_rb)->rb_subtree_gap;
339                 if (subtree_gap > max)
340                         max = subtree_gap;
341         }
342         return max;
343 }
344
345 #ifdef CONFIG_DEBUG_VM_RB
346 static int browse_rb(struct rb_root *root)
347 {
348         int i = 0, j, bug = 0;
349         struct rb_node *nd, *pn = NULL;
350         unsigned long prev = 0, pend = 0;
351
352         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
353                 struct vm_area_struct *vma;
354                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
355                 if (vma->vm_start < prev) {
356                         printk("vm_start %lx prev %lx\n", vma->vm_start, prev);
357                         bug = 1;
358                 }
359                 if (vma->vm_start < pend) {
360                         printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
361                         bug = 1;
362                 }
363                 if (vma->vm_start > vma->vm_end) {
364                         printk("vm_end %lx < vm_start %lx\n",
365                                 vma->vm_end, vma->vm_start);
366                         bug = 1;
367                 }
368                 if (vma->rb_subtree_gap != vma_compute_subtree_gap(vma)) {
369                         printk("free gap %lx, correct %lx\n",
370                                vma->rb_subtree_gap,
371                                vma_compute_subtree_gap(vma));
372                         bug = 1;
373                 }
374                 i++;
375                 pn = nd;
376                 prev = vma->vm_start;
377                 pend = vma->vm_end;
378         }
379         j = 0;
380         for (nd = pn; nd; nd = rb_prev(nd))
381                 j++;
382         if (i != j) {
383                 printk("backwards %d, forwards %d\n", j, i);
384                 bug = 1;
385         }
386         return bug ? -1 : i;
387 }
388
389 static void validate_mm_rb(struct rb_root *root, struct vm_area_struct *ignore)
390 {
391         struct rb_node *nd;
392
393         for (nd = rb_first(root); nd; nd = rb_next(nd)) {
394                 struct vm_area_struct *vma;
395                 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
396                 BUG_ON(vma != ignore &&
397                        vma->rb_subtree_gap != vma_compute_subtree_gap(vma));
398         }
399 }
400
401 void validate_mm(struct mm_struct *mm)
402 {
403         int bug = 0;
404         int i = 0;
405         unsigned long highest_address = 0;
406         struct vm_area_struct *vma = mm->mmap;
407         while (vma) {
408                 struct anon_vma_chain *avc;
409                 vma_lock_anon_vma(vma);
410                 list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
411                         anon_vma_interval_tree_verify(avc);
412                 vma_unlock_anon_vma(vma);
413                 highest_address = vma->vm_end;
414                 vma = vma->vm_next;
415                 i++;
416         }
417         if (i != mm->map_count) {
418                 printk("map_count %d vm_next %d\n", mm->map_count, i);
419                 bug = 1;
420         }
421         if (highest_address != mm->highest_vm_end) {
422                 printk("mm->highest_vm_end %lx, found %lx\n",
423                        mm->highest_vm_end, highest_address);
424                 bug = 1;
425         }
426         i = browse_rb(&mm->mm_rb);
427         if (i != mm->map_count) {
428                 printk("map_count %d rb %d\n", mm->map_count, i);
429                 bug = 1;
430         }
431         BUG_ON(bug);
432 }
433 #else
434 #define validate_mm_rb(root, ignore) do { } while (0)
435 #define validate_mm(mm) do { } while (0)
436 #endif
437
438 RB_DECLARE_CALLBACKS(static, vma_gap_callbacks, struct vm_area_struct, vm_rb,
439                      unsigned long, rb_subtree_gap, vma_compute_subtree_gap)
440
441 /*
442  * Update augmented rbtree rb_subtree_gap values after vma->vm_start or
443  * vma->vm_prev->vm_end values changed, without modifying the vma's position
444  * in the rbtree.
445  */
446 static void vma_gap_update(struct vm_area_struct *vma)
447 {
448         /*
449          * As it turns out, RB_DECLARE_CALLBACKS() already created a callback
450          * function that does exacltly what we want.
451          */
452         vma_gap_callbacks_propagate(&vma->vm_rb, NULL);
453 }
454
455 static inline void vma_rb_insert(struct vm_area_struct *vma,
456                                  struct rb_root *root)
457 {
458         /* All rb_subtree_gap values must be consistent prior to insertion */
459         validate_mm_rb(root, NULL);
460
461         rb_insert_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
462 }
463
464 static void vma_rb_erase(struct vm_area_struct *vma, struct rb_root *root)
465 {
466         /*
467          * All rb_subtree_gap values must be consistent prior to erase,
468          * with the possible exception of the vma being erased.
469          */
470         validate_mm_rb(root, vma);
471
472         /*
473          * Note rb_erase_augmented is a fairly large inline function,
474          * so make sure we instantiate it only once with our desired
475          * augmented rbtree callbacks.
476          */
477         rb_erase_augmented(&vma->vm_rb, root, &vma_gap_callbacks);
478 }
479
480 /*
481  * vma has some anon_vma assigned, and is already inserted on that
482  * anon_vma's interval trees.
483  *
484  * Before updating the vma's vm_start / vm_end / vm_pgoff fields, the
485  * vma must be removed from the anon_vma's interval trees using
486  * anon_vma_interval_tree_pre_update_vma().
487  *
488  * After the update, the vma will be reinserted using
489  * anon_vma_interval_tree_post_update_vma().
490  *
491  * The entire update must be protected by exclusive mmap_sem and by
492  * the root anon_vma's mutex.
493  */
494 static inline void
495 anon_vma_interval_tree_pre_update_vma(struct vm_area_struct *vma)
496 {
497         struct anon_vma_chain *avc;
498
499         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
500                 anon_vma_interval_tree_remove(avc, &avc->anon_vma->rb_root);
501 }
502
503 static inline void
504 anon_vma_interval_tree_post_update_vma(struct vm_area_struct *vma)
505 {
506         struct anon_vma_chain *avc;
507
508         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
509                 anon_vma_interval_tree_insert(avc, &avc->anon_vma->rb_root);
510 }
511
512 static int find_vma_links(struct mm_struct *mm, unsigned long addr,
513                 unsigned long end, struct vm_area_struct **pprev,
514                 struct rb_node ***rb_link, struct rb_node **rb_parent)
515 {
516         struct rb_node **__rb_link, *__rb_parent, *rb_prev;
517
518         __rb_link = &mm->mm_rb.rb_node;
519         rb_prev = __rb_parent = NULL;
520
521         while (*__rb_link) {
522                 struct vm_area_struct *vma_tmp;
523
524                 __rb_parent = *__rb_link;
525                 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
526
527                 if (vma_tmp->vm_end > addr) {
528                         /* Fail if an existing vma overlaps the area */
529                         if (vma_tmp->vm_start < end)
530                                 return -ENOMEM;
531                         __rb_link = &__rb_parent->rb_left;
532                 } else {
533                         rb_prev = __rb_parent;
534                         __rb_link = &__rb_parent->rb_right;
535                 }
536         }
537
538         *pprev = NULL;
539         if (rb_prev)
540                 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
541         *rb_link = __rb_link;
542         *rb_parent = __rb_parent;
543         return 0;
544 }
545
546 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
547                 struct rb_node **rb_link, struct rb_node *rb_parent)
548 {
549         /* Update tracking information for the gap following the new vma. */
550         if (vma->vm_next)
551                 vma_gap_update(vma->vm_next);
552         else
553                 mm->highest_vm_end = vma->vm_end;
554
555         /*
556          * vma->vm_prev wasn't known when we followed the rbtree to find the
557          * correct insertion point for that vma. As a result, we could not
558          * update the vma vm_rb parents rb_subtree_gap values on the way down.
559          * So, we first insert the vma with a zero rb_subtree_gap value
560          * (to be consistent with what we did on the way down), and then
561          * immediately update the gap to the correct value. Finally we
562          * rebalance the rbtree after all augmented values have been set.
563          */
564         rb_link_node(&vma->vm_rb, rb_parent, rb_link);
565         vma->rb_subtree_gap = 0;
566         vma_gap_update(vma);
567         vma_rb_insert(vma, &mm->mm_rb);
568 }
569
570 static void __vma_link_file(struct vm_area_struct *vma)
571 {
572         struct file *file;
573
574         file = vma->vm_file;
575         if (file) {
576                 struct address_space *mapping = file->f_mapping;
577
578                 if (vma->vm_flags & VM_DENYWRITE)
579                         atomic_dec(&file->f_path.dentry->d_inode->i_writecount);
580                 if (vma->vm_flags & VM_SHARED)
581                         mapping->i_mmap_writable++;
582
583                 flush_dcache_mmap_lock(mapping);
584                 if (unlikely(vma->vm_flags & VM_NONLINEAR))
585                         vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
586                 else
587                         vma_interval_tree_insert(vma, &mapping->i_mmap);
588                 flush_dcache_mmap_unlock(mapping);
589         }
590 }
591
592 static void
593 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
594         struct vm_area_struct *prev, struct rb_node **rb_link,
595         struct rb_node *rb_parent)
596 {
597         __vma_link_list(mm, vma, prev, rb_parent);
598         __vma_link_rb(mm, vma, rb_link, rb_parent);
599 }
600
601 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
602                         struct vm_area_struct *prev, struct rb_node **rb_link,
603                         struct rb_node *rb_parent)
604 {
605         struct address_space *mapping = NULL;
606
607         if (vma->vm_file)
608                 mapping = vma->vm_file->f_mapping;
609
610         if (mapping)
611                 mutex_lock(&mapping->i_mmap_mutex);
612
613         __vma_link(mm, vma, prev, rb_link, rb_parent);
614         __vma_link_file(vma);
615
616         if (mapping)
617                 mutex_unlock(&mapping->i_mmap_mutex);
618
619         mm->map_count++;
620         validate_mm(mm);
621 }
622
623 /*
624  * Helper for vma_adjust() in the split_vma insert case: insert a vma into the
625  * mm's list and rbtree.  It has already been inserted into the interval tree.
626  */
627 static void __insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
628 {
629         struct vm_area_struct *prev;
630         struct rb_node **rb_link, *rb_parent;
631
632         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
633                            &prev, &rb_link, &rb_parent))
634                 BUG();
635         __vma_link(mm, vma, prev, rb_link, rb_parent);
636         mm->map_count++;
637 }
638
639 static inline void
640 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
641                 struct vm_area_struct *prev)
642 {
643         struct vm_area_struct *next;
644
645         vma_rb_erase(vma, &mm->mm_rb);
646         prev->vm_next = next = vma->vm_next;
647         if (next)
648                 next->vm_prev = prev;
649         if (mm->mmap_cache == vma)
650                 mm->mmap_cache = prev;
651 }
652
653 /*
654  * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
655  * is already present in an i_mmap tree without adjusting the tree.
656  * The following helper function should be used when such adjustments
657  * are necessary.  The "insert" vma (if any) is to be inserted
658  * before we drop the necessary locks.
659  */
660 int vma_adjust(struct vm_area_struct *vma, unsigned long start,
661         unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
662 {
663         struct mm_struct *mm = vma->vm_mm;
664         struct vm_area_struct *next = vma->vm_next;
665         struct vm_area_struct *importer = NULL;
666         struct address_space *mapping = NULL;
667         struct rb_root *root = NULL;
668         struct anon_vma *anon_vma = NULL;
669         struct file *file = vma->vm_file;
670         bool start_changed = false, end_changed = false;
671         long adjust_next = 0;
672         int remove_next = 0;
673
674         if (next && !insert) {
675                 struct vm_area_struct *exporter = NULL;
676
677                 if (end >= next->vm_end) {
678                         /*
679                          * vma expands, overlapping all the next, and
680                          * perhaps the one after too (mprotect case 6).
681                          */
682 again:                  remove_next = 1 + (end > next->vm_end);
683                         end = next->vm_end;
684                         exporter = next;
685                         importer = vma;
686                 } else if (end > next->vm_start) {
687                         /*
688                          * vma expands, overlapping part of the next:
689                          * mprotect case 5 shifting the boundary up.
690                          */
691                         adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
692                         exporter = next;
693                         importer = vma;
694                 } else if (end < vma->vm_end) {
695                         /*
696                          * vma shrinks, and !insert tells it's not
697                          * split_vma inserting another: so it must be
698                          * mprotect case 4 shifting the boundary down.
699                          */
700                         adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
701                         exporter = vma;
702                         importer = next;
703                 }
704
705                 /*
706                  * Easily overlooked: when mprotect shifts the boundary,
707                  * make sure the expanding vma has anon_vma set if the
708                  * shrinking vma had, to cover any anon pages imported.
709                  */
710                 if (exporter && exporter->anon_vma && !importer->anon_vma) {
711                         if (anon_vma_clone(importer, exporter))
712                                 return -ENOMEM;
713                         importer->anon_vma = exporter->anon_vma;
714                 }
715         }
716
717         if (file) {
718                 mapping = file->f_mapping;
719                 if (!(vma->vm_flags & VM_NONLINEAR)) {
720                         root = &mapping->i_mmap;
721                         uprobe_munmap(vma, vma->vm_start, vma->vm_end);
722
723                         if (adjust_next)
724                                 uprobe_munmap(next, next->vm_start,
725                                                         next->vm_end);
726                 }
727
728                 mutex_lock(&mapping->i_mmap_mutex);
729                 if (insert) {
730                         /*
731                          * Put into interval tree now, so instantiated pages
732                          * are visible to arm/parisc __flush_dcache_page
733                          * throughout; but we cannot insert into address
734                          * space until vma start or end is updated.
735                          */
736                         __vma_link_file(insert);
737                 }
738         }
739
740         vma_adjust_trans_huge(vma, start, end, adjust_next);
741
742         anon_vma = vma->anon_vma;
743         if (!anon_vma && adjust_next)
744                 anon_vma = next->anon_vma;
745         if (anon_vma) {
746                 VM_BUG_ON(adjust_next && next->anon_vma &&
747                           anon_vma != next->anon_vma);
748                 anon_vma_lock_write(anon_vma);
749                 anon_vma_interval_tree_pre_update_vma(vma);
750                 if (adjust_next)
751                         anon_vma_interval_tree_pre_update_vma(next);
752         }
753
754         if (root) {
755                 flush_dcache_mmap_lock(mapping);
756                 vma_interval_tree_remove(vma, root);
757                 if (adjust_next)
758                         vma_interval_tree_remove(next, root);
759         }
760
761         if (start != vma->vm_start) {
762                 vma->vm_start = start;
763                 start_changed = true;
764         }
765         if (end != vma->vm_end) {
766                 vma->vm_end = end;
767                 end_changed = true;
768         }
769         vma->vm_pgoff = pgoff;
770         if (adjust_next) {
771                 next->vm_start += adjust_next << PAGE_SHIFT;
772                 next->vm_pgoff += adjust_next;
773         }
774
775         if (root) {
776                 if (adjust_next)
777                         vma_interval_tree_insert(next, root);
778                 vma_interval_tree_insert(vma, root);
779                 flush_dcache_mmap_unlock(mapping);
780         }
781
782         if (remove_next) {
783                 /*
784                  * vma_merge has merged next into vma, and needs
785                  * us to remove next before dropping the locks.
786                  */
787                 __vma_unlink(mm, next, vma);
788                 if (file)
789                         __remove_shared_vm_struct(next, file, mapping);
790         } else if (insert) {
791                 /*
792                  * split_vma has split insert from vma, and needs
793                  * us to insert it before dropping the locks
794                  * (it may either follow vma or precede it).
795                  */
796                 __insert_vm_struct(mm, insert);
797         } else {
798                 if (start_changed)
799                         vma_gap_update(vma);
800                 if (end_changed) {
801                         if (!next)
802                                 mm->highest_vm_end = end;
803                         else if (!adjust_next)
804                                 vma_gap_update(next);
805                 }
806         }
807
808         if (anon_vma) {
809                 anon_vma_interval_tree_post_update_vma(vma);
810                 if (adjust_next)
811                         anon_vma_interval_tree_post_update_vma(next);
812                 anon_vma_unlock(anon_vma);
813         }
814         if (mapping)
815                 mutex_unlock(&mapping->i_mmap_mutex);
816
817         if (root) {
818                 uprobe_mmap(vma);
819
820                 if (adjust_next)
821                         uprobe_mmap(next);
822         }
823
824         if (remove_next) {
825                 if (file) {
826                         uprobe_munmap(next, next->vm_start, next->vm_end);
827                         fput(file);
828                 }
829                 if (next->anon_vma)
830                         anon_vma_merge(vma, next);
831                 mm->map_count--;
832                 mpol_put(vma_policy(next));
833                 kmem_cache_free(vm_area_cachep, next);
834                 /*
835                  * In mprotect's case 6 (see comments on vma_merge),
836                  * we must remove another next too. It would clutter
837                  * up the code too much to do both in one go.
838                  */
839                 next = vma->vm_next;
840                 if (remove_next == 2)
841                         goto again;
842                 else if (next)
843                         vma_gap_update(next);
844                 else
845                         mm->highest_vm_end = end;
846         }
847         if (insert && file)
848                 uprobe_mmap(insert);
849
850         validate_mm(mm);
851
852         return 0;
853 }
854
855 /*
856  * If the vma has a ->close operation then the driver probably needs to release
857  * per-vma resources, so we don't attempt to merge those.
858  */
859 static inline int is_mergeable_vma(struct vm_area_struct *vma,
860                         struct file *file, unsigned long vm_flags)
861 {
862         if (vma->vm_flags ^ vm_flags)
863                 return 0;
864         if (vma->vm_file != file)
865                 return 0;
866         if (vma->vm_ops && vma->vm_ops->close)
867                 return 0;
868         return 1;
869 }
870
871 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
872                                         struct anon_vma *anon_vma2,
873                                         struct vm_area_struct *vma)
874 {
875         /*
876          * The list_is_singular() test is to avoid merging VMA cloned from
877          * parents. This can improve scalability caused by anon_vma lock.
878          */
879         if ((!anon_vma1 || !anon_vma2) && (!vma ||
880                 list_is_singular(&vma->anon_vma_chain)))
881                 return 1;
882         return anon_vma1 == anon_vma2;
883 }
884
885 /*
886  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
887  * in front of (at a lower virtual address and file offset than) the vma.
888  *
889  * We cannot merge two vmas if they have differently assigned (non-NULL)
890  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
891  *
892  * We don't check here for the merged mmap wrapping around the end of pagecache
893  * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
894  * wrap, nor mmaps which cover the final page at index -1UL.
895  */
896 static int
897 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
898         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
899 {
900         if (is_mergeable_vma(vma, file, vm_flags) &&
901             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
902                 if (vma->vm_pgoff == vm_pgoff)
903                         return 1;
904         }
905         return 0;
906 }
907
908 /*
909  * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
910  * beyond (at a higher virtual address and file offset than) the vma.
911  *
912  * We cannot merge two vmas if they have differently assigned (non-NULL)
913  * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
914  */
915 static int
916 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
917         struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
918 {
919         if (is_mergeable_vma(vma, file, vm_flags) &&
920             is_mergeable_anon_vma(anon_vma, vma->anon_vma, vma)) {
921                 pgoff_t vm_pglen;
922                 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
923                 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
924                         return 1;
925         }
926         return 0;
927 }
928
929 /*
930  * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
931  * whether that can be merged with its predecessor or its successor.
932  * Or both (it neatly fills a hole).
933  *
934  * In most cases - when called for mmap, brk or mremap - [addr,end) is
935  * certain not to be mapped by the time vma_merge is called; but when
936  * called for mprotect, it is certain to be already mapped (either at
937  * an offset within prev, or at the start of next), and the flags of
938  * this area are about to be changed to vm_flags - and the no-change
939  * case has already been eliminated.
940  *
941  * The following mprotect cases have to be considered, where AAAA is
942  * the area passed down from mprotect_fixup, never extending beyond one
943  * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
944  *
945  *     AAAA             AAAA                AAAA          AAAA
946  *    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPPPNNNNNN    PPPPNNNNXXXX
947  *    cannot merge    might become    might become    might become
948  *                    PPNNNNNNNNNN    PPPPPPPPPPNN    PPPPPPPPPPPP 6 or
949  *    mmap, brk or    case 4 below    case 5 below    PPPPPPPPXXXX 7 or
950  *    mremap move:                                    PPPPNNNNNNNN 8
951  *        AAAA
952  *    PPPP    NNNN    PPPPPPPPPPPP    PPPPPPPPNNNN    PPPPNNNNNNNN
953  *    might become    case 1 below    case 2 below    case 3 below
954  *
955  * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
956  * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
957  */
958 struct vm_area_struct *vma_merge(struct mm_struct *mm,
959                         struct vm_area_struct *prev, unsigned long addr,
960                         unsigned long end, unsigned long vm_flags,
961                         struct anon_vma *anon_vma, struct file *file,
962                         pgoff_t pgoff, struct mempolicy *policy)
963 {
964         pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
965         struct vm_area_struct *area, *next;
966         int err;
967
968         /*
969          * We later require that vma->vm_flags == vm_flags,
970          * so this tests vma->vm_flags & VM_SPECIAL, too.
971          */
972         if (vm_flags & VM_SPECIAL)
973                 return NULL;
974
975         if (prev)
976                 next = prev->vm_next;
977         else
978                 next = mm->mmap;
979         area = next;
980         if (next && next->vm_end == end)                /* cases 6, 7, 8 */
981                 next = next->vm_next;
982
983         /*
984          * Can it merge with the predecessor?
985          */
986         if (prev && prev->vm_end == addr &&
987                         mpol_equal(vma_policy(prev), policy) &&
988                         can_vma_merge_after(prev, vm_flags,
989                                                 anon_vma, file, pgoff)) {
990                 /*
991                  * OK, it can.  Can we now merge in the successor as well?
992                  */
993                 if (next && end == next->vm_start &&
994                                 mpol_equal(policy, vma_policy(next)) &&
995                                 can_vma_merge_before(next, vm_flags,
996                                         anon_vma, file, pgoff+pglen) &&
997                                 is_mergeable_anon_vma(prev->anon_vma,
998                                                       next->anon_vma, NULL)) {
999                                                         /* cases 1, 6 */
1000                         err = vma_adjust(prev, prev->vm_start,
1001                                 next->vm_end, prev->vm_pgoff, NULL);
1002                 } else                                  /* cases 2, 5, 7 */
1003                         err = vma_adjust(prev, prev->vm_start,
1004                                 end, prev->vm_pgoff, NULL);
1005                 if (err)
1006                         return NULL;
1007                 khugepaged_enter_vma_merge(prev);
1008                 return prev;
1009         }
1010
1011         /*
1012          * Can this new request be merged in front of next?
1013          */
1014         if (next && end == next->vm_start &&
1015                         mpol_equal(policy, vma_policy(next)) &&
1016                         can_vma_merge_before(next, vm_flags,
1017                                         anon_vma, file, pgoff+pglen)) {
1018                 if (prev && addr < prev->vm_end)        /* case 4 */
1019                         err = vma_adjust(prev, prev->vm_start,
1020                                 addr, prev->vm_pgoff, NULL);
1021                 else                                    /* cases 3, 8 */
1022                         err = vma_adjust(area, addr, next->vm_end,
1023                                 next->vm_pgoff - pglen, NULL);
1024                 if (err)
1025                         return NULL;
1026                 khugepaged_enter_vma_merge(area);
1027                 return area;
1028         }
1029
1030         return NULL;
1031 }
1032
1033 /*
1034  * Rough compatbility check to quickly see if it's even worth looking
1035  * at sharing an anon_vma.
1036  *
1037  * They need to have the same vm_file, and the flags can only differ
1038  * in things that mprotect may change.
1039  *
1040  * NOTE! The fact that we share an anon_vma doesn't _have_ to mean that
1041  * we can merge the two vma's. For example, we refuse to merge a vma if
1042  * there is a vm_ops->close() function, because that indicates that the
1043  * driver is doing some kind of reference counting. But that doesn't
1044  * really matter for the anon_vma sharing case.
1045  */
1046 static int anon_vma_compatible(struct vm_area_struct *a, struct vm_area_struct *b)
1047 {
1048         return a->vm_end == b->vm_start &&
1049                 mpol_equal(vma_policy(a), vma_policy(b)) &&
1050                 a->vm_file == b->vm_file &&
1051                 !((a->vm_flags ^ b->vm_flags) & ~(VM_READ|VM_WRITE|VM_EXEC)) &&
1052                 b->vm_pgoff == a->vm_pgoff + ((b->vm_start - a->vm_start) >> PAGE_SHIFT);
1053 }
1054
1055 /*
1056  * Do some basic sanity checking to see if we can re-use the anon_vma
1057  * from 'old'. The 'a'/'b' vma's are in VM order - one of them will be
1058  * the same as 'old', the other will be the new one that is trying
1059  * to share the anon_vma.
1060  *
1061  * NOTE! This runs with mm_sem held for reading, so it is possible that
1062  * the anon_vma of 'old' is concurrently in the process of being set up
1063  * by another page fault trying to merge _that_. But that's ok: if it
1064  * is being set up, that automatically means that it will be a singleton
1065  * acceptable for merging, so we can do all of this optimistically. But
1066  * we do that ACCESS_ONCE() to make sure that we never re-load the pointer.
1067  *
1068  * IOW: that the "list_is_singular()" test on the anon_vma_chain only
1069  * matters for the 'stable anon_vma' case (ie the thing we want to avoid
1070  * is to return an anon_vma that is "complex" due to having gone through
1071  * a fork).
1072  *
1073  * We also make sure that the two vma's are compatible (adjacent,
1074  * and with the same memory policies). That's all stable, even with just
1075  * a read lock on the mm_sem.
1076  */
1077 static struct anon_vma *reusable_anon_vma(struct vm_area_struct *old, struct vm_area_struct *a, struct vm_area_struct *b)
1078 {
1079         if (anon_vma_compatible(a, b)) {
1080                 struct anon_vma *anon_vma = ACCESS_ONCE(old->anon_vma);
1081
1082                 if (anon_vma && list_is_singular(&old->anon_vma_chain))
1083                         return anon_vma;
1084         }
1085         return NULL;
1086 }
1087
1088 /*
1089  * find_mergeable_anon_vma is used by anon_vma_prepare, to check
1090  * neighbouring vmas for a suitable anon_vma, before it goes off
1091  * to allocate a new anon_vma.  It checks because a repetitive
1092  * sequence of mprotects and faults may otherwise lead to distinct
1093  * anon_vmas being allocated, preventing vma merge in subsequent
1094  * mprotect.
1095  */
1096 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
1097 {
1098         struct anon_vma *anon_vma;
1099         struct vm_area_struct *near;
1100
1101         near = vma->vm_next;
1102         if (!near)
1103                 goto try_prev;
1104
1105         anon_vma = reusable_anon_vma(near, vma, near);
1106         if (anon_vma)
1107                 return anon_vma;
1108 try_prev:
1109         near = vma->vm_prev;
1110         if (!near)
1111                 goto none;
1112
1113         anon_vma = reusable_anon_vma(near, near, vma);
1114         if (anon_vma)
1115                 return anon_vma;
1116 none:
1117         /*
1118          * There's no absolute need to look only at touching neighbours:
1119          * we could search further afield for "compatible" anon_vmas.
1120          * But it would probably just be a waste of time searching,
1121          * or lead to too many vmas hanging off the same anon_vma.
1122          * We're trying to allow mprotect remerging later on,
1123          * not trying to minimize memory used for anon_vmas.
1124          */
1125         return NULL;
1126 }
1127
1128 #ifdef CONFIG_PROC_FS
1129 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
1130                                                 struct file *file, long pages)
1131 {
1132         const unsigned long stack_flags
1133                 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
1134
1135         mm->total_vm += pages;
1136
1137         if (file) {
1138                 mm->shared_vm += pages;
1139                 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
1140                         mm->exec_vm += pages;
1141         } else if (flags & stack_flags)
1142                 mm->stack_vm += pages;
1143 }
1144 #endif /* CONFIG_PROC_FS */
1145
1146 /*
1147  * If a hint addr is less than mmap_min_addr change hint to be as
1148  * low as possible but still greater than mmap_min_addr
1149  */
1150 static inline unsigned long round_hint_to_min(unsigned long hint)
1151 {
1152         hint &= PAGE_MASK;
1153         if (((void *)hint != NULL) &&
1154             (hint < mmap_min_addr))
1155                 return PAGE_ALIGN(mmap_min_addr);
1156         return hint;
1157 }
1158
1159 /*
1160  * The caller must hold down_write(&current->mm->mmap_sem).
1161  */
1162
1163 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
1164                         unsigned long len, unsigned long prot,
1165                         unsigned long flags, unsigned long pgoff,
1166                         unsigned long *populate)
1167 {
1168         struct mm_struct * mm = current->mm;
1169         struct inode *inode;
1170         vm_flags_t vm_flags;
1171
1172         *populate = 0;
1173
1174         /*
1175          * Does the application expect PROT_READ to imply PROT_EXEC?
1176          *
1177          * (the exception is when the underlying filesystem is noexec
1178          *  mounted, in which case we dont add PROT_EXEC.)
1179          */
1180         if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
1181                 if (!(file && (file->f_path.mnt->mnt_flags & MNT_NOEXEC)))
1182                         prot |= PROT_EXEC;
1183
1184         if (!len)
1185                 return -EINVAL;
1186
1187         if (!(flags & MAP_FIXED))
1188                 addr = round_hint_to_min(addr);
1189
1190         /* Careful about overflows.. */
1191         len = PAGE_ALIGN(len);
1192         if (!len)
1193                 return -ENOMEM;
1194
1195         /* offset overflow? */
1196         if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
1197                return -EOVERFLOW;
1198
1199         /* Too many mappings? */
1200         if (mm->map_count > sysctl_max_map_count)
1201                 return -ENOMEM;
1202
1203         /* Obtain the address to map to. we verify (or select) it and ensure
1204          * that it represents a valid section of the address space.
1205          */
1206         addr = get_unmapped_area(file, addr, len, pgoff, flags);
1207         if (addr & ~PAGE_MASK)
1208                 return addr;
1209
1210         /* Do simple checking here so the lower-level routines won't have
1211          * to. we assume access permissions have been handled by the open
1212          * of the memory object, so we don't do any here.
1213          */
1214         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
1215                         mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1216
1217         if (flags & MAP_LOCKED)
1218                 if (!can_do_mlock())
1219                         return -EPERM;
1220
1221         /* mlock MCL_FUTURE? */
1222         if (vm_flags & VM_LOCKED) {
1223                 unsigned long locked, lock_limit;
1224                 locked = len >> PAGE_SHIFT;
1225                 locked += mm->locked_vm;
1226                 lock_limit = rlimit(RLIMIT_MEMLOCK);
1227                 lock_limit >>= PAGE_SHIFT;
1228                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1229                         return -EAGAIN;
1230         }
1231
1232         inode = file ? file->f_path.dentry->d_inode : NULL;
1233
1234         if (file) {
1235                 switch (flags & MAP_TYPE) {
1236                 case MAP_SHARED:
1237                         if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
1238                                 return -EACCES;
1239
1240                         /*
1241                          * Make sure we don't allow writing to an append-only
1242                          * file..
1243                          */
1244                         if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
1245                                 return -EACCES;
1246
1247                         /*
1248                          * Make sure there are no mandatory locks on the file.
1249                          */
1250                         if (locks_verify_locked(inode))
1251                                 return -EAGAIN;
1252
1253                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1254                         if (!(file->f_mode & FMODE_WRITE))
1255                                 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
1256
1257                         /* fall through */
1258                 case MAP_PRIVATE:
1259                         if (!(file->f_mode & FMODE_READ))
1260                                 return -EACCES;
1261                         if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1262                                 if (vm_flags & VM_EXEC)
1263                                         return -EPERM;
1264                                 vm_flags &= ~VM_MAYEXEC;
1265                         }
1266
1267                         if (!file->f_op || !file->f_op->mmap)
1268                                 return -ENODEV;
1269                         break;
1270
1271                 default:
1272                         return -EINVAL;
1273                 }
1274         } else {
1275                 switch (flags & MAP_TYPE) {
1276                 case MAP_SHARED:
1277                         /*
1278                          * Ignore pgoff.
1279                          */
1280                         pgoff = 0;
1281                         vm_flags |= VM_SHARED | VM_MAYSHARE;
1282                         break;
1283                 case MAP_PRIVATE:
1284                         /*
1285                          * Set pgoff according to addr for anon_vma.
1286                          */
1287                         pgoff = addr >> PAGE_SHIFT;
1288                         break;
1289                 default:
1290                         return -EINVAL;
1291                 }
1292         }
1293
1294         /*
1295          * Set 'VM_NORESERVE' if we should not account for the
1296          * memory use of this mapping.
1297          */
1298         if (flags & MAP_NORESERVE) {
1299                 /* We honor MAP_NORESERVE if allowed to overcommit */
1300                 if (sysctl_overcommit_memory != OVERCOMMIT_NEVER)
1301                         vm_flags |= VM_NORESERVE;
1302
1303                 /* hugetlb applies strict overcommit unless MAP_NORESERVE */
1304                 if (file && is_file_hugepages(file))
1305                         vm_flags |= VM_NORESERVE;
1306         }
1307
1308         addr = mmap_region(file, addr, len, vm_flags, pgoff);
1309         if (!IS_ERR_VALUE(addr) && (vm_flags & VM_POPULATE))
1310                 *populate = len;
1311         return addr;
1312 }
1313
1314 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1315                 unsigned long, prot, unsigned long, flags,
1316                 unsigned long, fd, unsigned long, pgoff)
1317 {
1318         struct file *file = NULL;
1319         unsigned long retval = -EBADF;
1320
1321         if (!(flags & MAP_ANONYMOUS)) {
1322                 audit_mmap_fd(fd, flags);
1323                 if (unlikely(flags & MAP_HUGETLB))
1324                         return -EINVAL;
1325                 file = fget(fd);
1326                 if (!file)
1327                         goto out;
1328         } else if (flags & MAP_HUGETLB) {
1329                 struct user_struct *user = NULL;
1330                 /*
1331                  * VM_NORESERVE is used because the reservations will be
1332                  * taken when vm_ops->mmap() is called
1333                  * A dummy user value is used because we are not locking
1334                  * memory so no accounting is necessary
1335                  */
1336                 file = hugetlb_file_setup(HUGETLB_ANON_FILE, addr, len,
1337                                 VM_NORESERVE,
1338                                 &user, HUGETLB_ANONHUGE_INODE,
1339                                 (flags >> MAP_HUGE_SHIFT) & MAP_HUGE_MASK);
1340                 if (IS_ERR(file))
1341                         return PTR_ERR(file);
1342         }
1343
1344         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1345
1346         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1347         if (file)
1348                 fput(file);
1349 out:
1350         return retval;
1351 }
1352
1353 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1354 struct mmap_arg_struct {
1355         unsigned long addr;
1356         unsigned long len;
1357         unsigned long prot;
1358         unsigned long flags;
1359         unsigned long fd;
1360         unsigned long offset;
1361 };
1362
1363 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1364 {
1365         struct mmap_arg_struct a;
1366
1367         if (copy_from_user(&a, arg, sizeof(a)))
1368                 return -EFAULT;
1369         if (a.offset & ~PAGE_MASK)
1370                 return -EINVAL;
1371
1372         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1373                               a.offset >> PAGE_SHIFT);
1374 }
1375 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1376
1377 /*
1378  * Some shared mappigns will want the pages marked read-only
1379  * to track write events. If so, we'll downgrade vm_page_prot
1380  * to the private version (using protection_map[] without the
1381  * VM_SHARED bit).
1382  */
1383 int vma_wants_writenotify(struct vm_area_struct *vma)
1384 {
1385         vm_flags_t vm_flags = vma->vm_flags;
1386
1387         /* If it was private or non-writable, the write bit is already clear */
1388         if ((vm_flags & (VM_WRITE|VM_SHARED)) != ((VM_WRITE|VM_SHARED)))
1389                 return 0;
1390
1391         /* The backer wishes to know when pages are first written to? */
1392         if (vma->vm_ops && vma->vm_ops->page_mkwrite)
1393                 return 1;
1394
1395         /* The open routine did something to the protections already? */
1396         if (pgprot_val(vma->vm_page_prot) !=
1397             pgprot_val(vm_get_page_prot(vm_flags)))
1398                 return 0;
1399
1400         /* Specialty mapping? */
1401         if (vm_flags & VM_PFNMAP)
1402                 return 0;
1403
1404         /* Can the mapping track the dirty pages? */
1405         return vma->vm_file && vma->vm_file->f_mapping &&
1406                 mapping_cap_account_dirty(vma->vm_file->f_mapping);
1407 }
1408
1409 /*
1410  * We account for memory if it's a private writeable mapping,
1411  * not hugepages and VM_NORESERVE wasn't set.
1412  */
1413 static inline int accountable_mapping(struct file *file, vm_flags_t vm_flags)
1414 {
1415         /*
1416          * hugetlb has its own accounting separate from the core VM
1417          * VM_HUGETLB may not be set yet so we cannot check for that flag.
1418          */
1419         if (file && is_file_hugepages(file))
1420                 return 0;
1421
1422         return (vm_flags & (VM_NORESERVE | VM_SHARED | VM_WRITE)) == VM_WRITE;
1423 }
1424
1425 unsigned long mmap_region(struct file *file, unsigned long addr,
1426                 unsigned long len, vm_flags_t vm_flags, unsigned long pgoff)
1427 {
1428         struct mm_struct *mm = current->mm;
1429         struct vm_area_struct *vma, *prev;
1430         int correct_wcount = 0;
1431         int error;
1432         struct rb_node **rb_link, *rb_parent;
1433         unsigned long charged = 0;
1434         struct inode *inode =  file ? file->f_path.dentry->d_inode : NULL;
1435
1436         /* Clear old maps */
1437         error = -ENOMEM;
1438 munmap_back:
1439         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
1440                 if (do_munmap(mm, addr, len))
1441                         return -ENOMEM;
1442                 goto munmap_back;
1443         }
1444
1445         /* Check against address space limit. */
1446         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1447                 return -ENOMEM;
1448
1449         /*
1450          * Private writable mapping: check memory availability
1451          */
1452         if (accountable_mapping(file, vm_flags)) {
1453                 charged = len >> PAGE_SHIFT;
1454                 if (security_vm_enough_memory_mm(mm, charged))
1455                         return -ENOMEM;
1456                 vm_flags |= VM_ACCOUNT;
1457         }
1458
1459         /*
1460          * Can we just expand an old mapping?
1461          */
1462         vma = vma_merge(mm, prev, addr, addr + len, vm_flags, NULL, file, pgoff, NULL);
1463         if (vma)
1464                 goto out;
1465
1466         /*
1467          * Determine the object being mapped and call the appropriate
1468          * specific mapper. the address has already been validated, but
1469          * not unmapped, but the maps are removed from the list.
1470          */
1471         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1472         if (!vma) {
1473                 error = -ENOMEM;
1474                 goto unacct_error;
1475         }
1476
1477         vma->vm_mm = mm;
1478         vma->vm_start = addr;
1479         vma->vm_end = addr + len;
1480         vma->vm_flags = vm_flags;
1481         vma->vm_page_prot = vm_get_page_prot(vm_flags);
1482         vma->vm_pgoff = pgoff;
1483         INIT_LIST_HEAD(&vma->anon_vma_chain);
1484
1485         error = -EINVAL;        /* when rejecting VM_GROWSDOWN|VM_GROWSUP */
1486
1487         if (file) {
1488                 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1489                         goto free_vma;
1490                 if (vm_flags & VM_DENYWRITE) {
1491                         error = deny_write_access(file);
1492                         if (error)
1493                                 goto free_vma;
1494                         correct_wcount = 1;
1495                 }
1496                 vma->vm_file = get_file(file);
1497                 error = file->f_op->mmap(file, vma);
1498                 if (error)
1499                         goto unmap_and_free_vma;
1500
1501                 /* Can addr have changed??
1502                  *
1503                  * Answer: Yes, several device drivers can do it in their
1504                  *         f_op->mmap method. -DaveM
1505                  * Bug: If addr is changed, prev, rb_link, rb_parent should
1506                  *      be updated for vma_link()
1507                  */
1508                 WARN_ON_ONCE(addr != vma->vm_start);
1509
1510                 addr = vma->vm_start;
1511                 pgoff = vma->vm_pgoff;
1512                 vm_flags = vma->vm_flags;
1513         } else if (vm_flags & VM_SHARED) {
1514                 if (unlikely(vm_flags & (VM_GROWSDOWN|VM_GROWSUP)))
1515                         goto free_vma;
1516                 error = shmem_zero_setup(vma);
1517                 if (error)
1518                         goto free_vma;
1519         }
1520
1521         if (vma_wants_writenotify(vma)) {
1522                 pgprot_t pprot = vma->vm_page_prot;
1523
1524                 /* Can vma->vm_page_prot have changed??
1525                  *
1526                  * Answer: Yes, drivers may have changed it in their
1527                  *         f_op->mmap method.
1528                  *
1529                  * Ensures that vmas marked as uncached stay that way.
1530                  */
1531                 vma->vm_page_prot = vm_get_page_prot(vm_flags & ~VM_SHARED);
1532                 if (pgprot_val(pprot) == pgprot_val(pgprot_noncached(pprot)))
1533                         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1534         }
1535
1536         vma_link(mm, vma, prev, rb_link, rb_parent);
1537         file = vma->vm_file;
1538
1539         /* Once vma denies write, undo our temporary denial count */
1540         if (correct_wcount)
1541                 atomic_inc(&inode->i_writecount);
1542 out:
1543         perf_event_mmap(vma);
1544
1545         vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1546         if (vm_flags & VM_LOCKED) {
1547                 if (!((vm_flags & VM_SPECIAL) || is_vm_hugetlb_page(vma) ||
1548                                         vma == get_gate_vma(current->mm)))
1549                         mm->locked_vm += (len >> PAGE_SHIFT);
1550                 else
1551                         vma->vm_flags &= ~VM_LOCKED;
1552         }
1553
1554         if (file)
1555                 uprobe_mmap(vma);
1556
1557         return addr;
1558
1559 unmap_and_free_vma:
1560         if (correct_wcount)
1561                 atomic_inc(&inode->i_writecount);
1562         vma->vm_file = NULL;
1563         fput(file);
1564
1565         /* Undo any partial mapping done by a device driver. */
1566         unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1567         charged = 0;
1568 free_vma:
1569         kmem_cache_free(vm_area_cachep, vma);
1570 unacct_error:
1571         if (charged)
1572                 vm_unacct_memory(charged);
1573         return error;
1574 }
1575
1576 unsigned long unmapped_area(struct vm_unmapped_area_info *info)
1577 {
1578         /*
1579          * We implement the search by looking for an rbtree node that
1580          * immediately follows a suitable gap. That is,
1581          * - gap_start = vma->vm_prev->vm_end <= info->high_limit - length;
1582          * - gap_end   = vma->vm_start        >= info->low_limit  + length;
1583          * - gap_end - gap_start >= length
1584          */
1585
1586         struct mm_struct *mm = current->mm;
1587         struct vm_area_struct *vma;
1588         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1589
1590         /* Adjust search length to account for worst case alignment overhead */
1591         length = info->length + info->align_mask;
1592         if (length < info->length)
1593                 return -ENOMEM;
1594
1595         /* Adjust search limits by the desired length */
1596         if (info->high_limit < length)
1597                 return -ENOMEM;
1598         high_limit = info->high_limit - length;
1599
1600         if (info->low_limit > high_limit)
1601                 return -ENOMEM;
1602         low_limit = info->low_limit + length;
1603
1604         /* Check if rbtree root looks promising */
1605         if (RB_EMPTY_ROOT(&mm->mm_rb))
1606                 goto check_highest;
1607         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1608         if (vma->rb_subtree_gap < length)
1609                 goto check_highest;
1610
1611         while (true) {
1612                 /* Visit left subtree if it looks promising */
1613                 gap_end = vma->vm_start;
1614                 if (gap_end >= low_limit && vma->vm_rb.rb_left) {
1615                         struct vm_area_struct *left =
1616                                 rb_entry(vma->vm_rb.rb_left,
1617                                          struct vm_area_struct, vm_rb);
1618                         if (left->rb_subtree_gap >= length) {
1619                                 vma = left;
1620                                 continue;
1621                         }
1622                 }
1623
1624                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1625 check_current:
1626                 /* Check if current node has a suitable gap */
1627                 if (gap_start > high_limit)
1628                         return -ENOMEM;
1629                 if (gap_end >= low_limit && gap_end - gap_start >= length)
1630                         goto found;
1631
1632                 /* Visit right subtree if it looks promising */
1633                 if (vma->vm_rb.rb_right) {
1634                         struct vm_area_struct *right =
1635                                 rb_entry(vma->vm_rb.rb_right,
1636                                          struct vm_area_struct, vm_rb);
1637                         if (right->rb_subtree_gap >= length) {
1638                                 vma = right;
1639                                 continue;
1640                         }
1641                 }
1642
1643                 /* Go back up the rbtree to find next candidate node */
1644                 while (true) {
1645                         struct rb_node *prev = &vma->vm_rb;
1646                         if (!rb_parent(prev))
1647                                 goto check_highest;
1648                         vma = rb_entry(rb_parent(prev),
1649                                        struct vm_area_struct, vm_rb);
1650                         if (prev == vma->vm_rb.rb_left) {
1651                                 gap_start = vma->vm_prev->vm_end;
1652                                 gap_end = vma->vm_start;
1653                                 goto check_current;
1654                         }
1655                 }
1656         }
1657
1658 check_highest:
1659         /* Check highest gap, which does not precede any rbtree node */
1660         gap_start = mm->highest_vm_end;
1661         gap_end = ULONG_MAX;  /* Only for VM_BUG_ON below */
1662         if (gap_start > high_limit)
1663                 return -ENOMEM;
1664
1665 found:
1666         /* We found a suitable gap. Clip it with the original low_limit. */
1667         if (gap_start < info->low_limit)
1668                 gap_start = info->low_limit;
1669
1670         /* Adjust gap address to the desired alignment */
1671         gap_start += (info->align_offset - gap_start) & info->align_mask;
1672
1673         VM_BUG_ON(gap_start + info->length > info->high_limit);
1674         VM_BUG_ON(gap_start + info->length > gap_end);
1675         return gap_start;
1676 }
1677
1678 unsigned long unmapped_area_topdown(struct vm_unmapped_area_info *info)
1679 {
1680         struct mm_struct *mm = current->mm;
1681         struct vm_area_struct *vma;
1682         unsigned long length, low_limit, high_limit, gap_start, gap_end;
1683
1684         /* Adjust search length to account for worst case alignment overhead */
1685         length = info->length + info->align_mask;
1686         if (length < info->length)
1687                 return -ENOMEM;
1688
1689         /*
1690          * Adjust search limits by the desired length.
1691          * See implementation comment at top of unmapped_area().
1692          */
1693         gap_end = info->high_limit;
1694         if (gap_end < length)
1695                 return -ENOMEM;
1696         high_limit = gap_end - length;
1697
1698         if (info->low_limit > high_limit)
1699                 return -ENOMEM;
1700         low_limit = info->low_limit + length;
1701
1702         /* Check highest gap, which does not precede any rbtree node */
1703         gap_start = mm->highest_vm_end;
1704         if (gap_start <= high_limit)
1705                 goto found_highest;
1706
1707         /* Check if rbtree root looks promising */
1708         if (RB_EMPTY_ROOT(&mm->mm_rb))
1709                 return -ENOMEM;
1710         vma = rb_entry(mm->mm_rb.rb_node, struct vm_area_struct, vm_rb);
1711         if (vma->rb_subtree_gap < length)
1712                 return -ENOMEM;
1713
1714         while (true) {
1715                 /* Visit right subtree if it looks promising */
1716                 gap_start = vma->vm_prev ? vma->vm_prev->vm_end : 0;
1717                 if (gap_start <= high_limit && vma->vm_rb.rb_right) {
1718                         struct vm_area_struct *right =
1719                                 rb_entry(vma->vm_rb.rb_right,
1720                                          struct vm_area_struct, vm_rb);
1721                         if (right->rb_subtree_gap >= length) {
1722                                 vma = right;
1723                                 continue;
1724                         }
1725                 }
1726
1727 check_current:
1728                 /* Check if current node has a suitable gap */
1729                 gap_end = vma->vm_start;
1730                 if (gap_end < low_limit)
1731                         return -ENOMEM;
1732                 if (gap_start <= high_limit && gap_end - gap_start >= length)
1733                         goto found;
1734
1735                 /* Visit left subtree if it looks promising */
1736                 if (vma->vm_rb.rb_left) {
1737                         struct vm_area_struct *left =
1738                                 rb_entry(vma->vm_rb.rb_left,
1739                                          struct vm_area_struct, vm_rb);
1740                         if (left->rb_subtree_gap >= length) {
1741                                 vma = left;
1742                                 continue;
1743                         }
1744                 }
1745
1746                 /* Go back up the rbtree to find next candidate node */
1747                 while (true) {
1748                         struct rb_node *prev = &vma->vm_rb;
1749                         if (!rb_parent(prev))
1750                                 return -ENOMEM;
1751                         vma = rb_entry(rb_parent(prev),
1752                                        struct vm_area_struct, vm_rb);
1753                         if (prev == vma->vm_rb.rb_right) {
1754                                 gap_start = vma->vm_prev ?
1755                                         vma->vm_prev->vm_end : 0;
1756                                 goto check_current;
1757                         }
1758                 }
1759         }
1760
1761 found:
1762         /* We found a suitable gap. Clip it with the original high_limit. */
1763         if (gap_end > info->high_limit)
1764                 gap_end = info->high_limit;
1765
1766 found_highest:
1767         /* Compute highest gap address at the desired alignment */
1768         gap_end -= info->length;
1769         gap_end -= (gap_end - info->align_offset) & info->align_mask;
1770
1771         VM_BUG_ON(gap_end < info->low_limit);
1772         VM_BUG_ON(gap_end < gap_start);
1773         return gap_end;
1774 }
1775
1776 /* Get an address range which is currently unmapped.
1777  * For shmat() with addr=0.
1778  *
1779  * Ugly calling convention alert:
1780  * Return value with the low bits set means error value,
1781  * ie
1782  *      if (ret & ~PAGE_MASK)
1783  *              error = ret;
1784  *
1785  * This function "knows" that -ENOMEM has the bits set.
1786  */
1787 #ifndef HAVE_ARCH_UNMAPPED_AREA
1788 unsigned long
1789 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1790                 unsigned long len, unsigned long pgoff, unsigned long flags)
1791 {
1792         struct mm_struct *mm = current->mm;
1793         struct vm_area_struct *vma;
1794         struct vm_unmapped_area_info info;
1795
1796         if (len > TASK_SIZE)
1797                 return -ENOMEM;
1798
1799         if (flags & MAP_FIXED)
1800                 return addr;
1801
1802         if (addr) {
1803                 addr = PAGE_ALIGN(addr);
1804                 vma = find_vma(mm, addr);
1805                 if (TASK_SIZE - len >= addr &&
1806                     (!vma || addr + len <= vma->vm_start))
1807                         return addr;
1808         }
1809
1810         info.flags = 0;
1811         info.length = len;
1812         info.low_limit = TASK_UNMAPPED_BASE;
1813         info.high_limit = TASK_SIZE;
1814         info.align_mask = 0;
1815         return vm_unmapped_area(&info);
1816 }
1817 #endif  
1818
1819 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1820 {
1821         /*
1822          * Is this a new hole at the lowest possible address?
1823          */
1824         if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache)
1825                 mm->free_area_cache = addr;
1826 }
1827
1828 /*
1829  * This mmap-allocator allocates new areas top-down from below the
1830  * stack's low limit (the base):
1831  */
1832 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1833 unsigned long
1834 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1835                           const unsigned long len, const unsigned long pgoff,
1836                           const unsigned long flags)
1837 {
1838         struct vm_area_struct *vma;
1839         struct mm_struct *mm = current->mm;
1840         unsigned long addr = addr0;
1841         struct vm_unmapped_area_info info;
1842
1843         /* requested length too big for entire address space */
1844         if (len > TASK_SIZE)
1845                 return -ENOMEM;
1846
1847         if (flags & MAP_FIXED)
1848                 return addr;
1849
1850         /* requesting a specific address */
1851         if (addr) {
1852                 addr = PAGE_ALIGN(addr);
1853                 vma = find_vma(mm, addr);
1854                 if (TASK_SIZE - len >= addr &&
1855                                 (!vma || addr + len <= vma->vm_start))
1856                         return addr;
1857         }
1858
1859         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
1860         info.length = len;
1861         info.low_limit = PAGE_SIZE;
1862         info.high_limit = mm->mmap_base;
1863         info.align_mask = 0;
1864         addr = vm_unmapped_area(&info);
1865
1866         /*
1867          * A failed mmap() very likely causes application failure,
1868          * so fall back to the bottom-up function here. This scenario
1869          * can happen with large stack limits and large mmap()
1870          * allocations.
1871          */
1872         if (addr & ~PAGE_MASK) {
1873                 VM_BUG_ON(addr != -ENOMEM);
1874                 info.flags = 0;
1875                 info.low_limit = TASK_UNMAPPED_BASE;
1876                 info.high_limit = TASK_SIZE;
1877                 addr = vm_unmapped_area(&info);
1878         }
1879
1880         return addr;
1881 }
1882 #endif
1883
1884 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1885 {
1886         /*
1887          * Is this a new hole at the highest possible address?
1888          */
1889         if (addr > mm->free_area_cache)
1890                 mm->free_area_cache = addr;
1891
1892         /* dont allow allocations above current base */
1893         if (mm->free_area_cache > mm->mmap_base)
1894                 mm->free_area_cache = mm->mmap_base;
1895 }
1896
1897 unsigned long
1898 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1899                 unsigned long pgoff, unsigned long flags)
1900 {
1901         unsigned long (*get_area)(struct file *, unsigned long,
1902                                   unsigned long, unsigned long, unsigned long);
1903
1904         unsigned long error = arch_mmap_check(addr, len, flags);
1905         if (error)
1906                 return error;
1907
1908         /* Careful about overflows.. */
1909         if (len > TASK_SIZE)
1910                 return -ENOMEM;
1911
1912         get_area = current->mm->get_unmapped_area;
1913         if (file && file->f_op && file->f_op->get_unmapped_area)
1914                 get_area = file->f_op->get_unmapped_area;
1915         addr = get_area(file, addr, len, pgoff, flags);
1916         if (IS_ERR_VALUE(addr))
1917                 return addr;
1918
1919         if (addr > TASK_SIZE - len)
1920                 return -ENOMEM;
1921         if (addr & ~PAGE_MASK)
1922                 return -EINVAL;
1923
1924         addr = arch_rebalance_pgtables(addr, len);
1925         error = security_mmap_addr(addr);
1926         return error ? error : addr;
1927 }
1928
1929 EXPORT_SYMBOL(get_unmapped_area);
1930
1931 /* Look up the first VMA which satisfies  addr < vm_end,  NULL if none. */
1932 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
1933 {
1934         struct vm_area_struct *vma = NULL;
1935
1936         if (WARN_ON_ONCE(!mm))          /* Remove this in linux-3.6 */
1937                 return NULL;
1938
1939         /* Check the cache first. */
1940         /* (Cache hit rate is typically around 35%.) */
1941         vma = mm->mmap_cache;
1942         if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1943                 struct rb_node *rb_node;
1944
1945                 rb_node = mm->mm_rb.rb_node;
1946                 vma = NULL;
1947
1948                 while (rb_node) {
1949                         struct vm_area_struct *vma_tmp;
1950
1951                         vma_tmp = rb_entry(rb_node,
1952                                            struct vm_area_struct, vm_rb);
1953
1954                         if (vma_tmp->vm_end > addr) {
1955                                 vma = vma_tmp;
1956                                 if (vma_tmp->vm_start <= addr)
1957                                         break;
1958                                 rb_node = rb_node->rb_left;
1959                         } else
1960                                 rb_node = rb_node->rb_right;
1961                 }
1962                 if (vma)
1963                         mm->mmap_cache = vma;
1964         }
1965         return vma;
1966 }
1967
1968 EXPORT_SYMBOL(find_vma);
1969
1970 /*
1971  * Same as find_vma, but also return a pointer to the previous VMA in *pprev.
1972  */
1973 struct vm_area_struct *
1974 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1975                         struct vm_area_struct **pprev)
1976 {
1977         struct vm_area_struct *vma;
1978
1979         vma = find_vma(mm, addr);
1980         if (vma) {
1981                 *pprev = vma->vm_prev;
1982         } else {
1983                 struct rb_node *rb_node = mm->mm_rb.rb_node;
1984                 *pprev = NULL;
1985                 while (rb_node) {
1986                         *pprev = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1987                         rb_node = rb_node->rb_right;
1988                 }
1989         }
1990         return vma;
1991 }
1992
1993 /*
1994  * Verify that the stack growth is acceptable and
1995  * update accounting. This is shared with both the
1996  * grow-up and grow-down cases.
1997  */
1998 static int acct_stack_growth(struct vm_area_struct *vma, unsigned long size, unsigned long grow)
1999 {
2000         struct mm_struct *mm = vma->vm_mm;
2001         struct rlimit *rlim = current->signal->rlim;
2002         unsigned long new_start;
2003
2004         /* address space limit tests */
2005         if (!may_expand_vm(mm, grow))
2006                 return -ENOMEM;
2007
2008         /* Stack limit test */
2009         if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur))
2010                 return -ENOMEM;
2011
2012         /* mlock limit tests */
2013         if (vma->vm_flags & VM_LOCKED) {
2014                 unsigned long locked;
2015                 unsigned long limit;
2016                 locked = mm->locked_vm + grow;
2017                 limit = ACCESS_ONCE(rlim[RLIMIT_MEMLOCK].rlim_cur);
2018                 limit >>= PAGE_SHIFT;
2019                 if (locked > limit && !capable(CAP_IPC_LOCK))
2020                         return -ENOMEM;
2021         }
2022
2023         /* Check to ensure the stack will not grow into a hugetlb-only region */
2024         new_start = (vma->vm_flags & VM_GROWSUP) ? vma->vm_start :
2025                         vma->vm_end - size;
2026         if (is_hugepage_only_range(vma->vm_mm, new_start, size))
2027                 return -EFAULT;
2028
2029         /*
2030          * Overcommit..  This must be the final test, as it will
2031          * update security statistics.
2032          */
2033         if (security_vm_enough_memory_mm(mm, grow))
2034                 return -ENOMEM;
2035
2036         /* Ok, everything looks good - let it rip */
2037         if (vma->vm_flags & VM_LOCKED)
2038                 mm->locked_vm += grow;
2039         vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
2040         return 0;
2041 }
2042
2043 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
2044 /*
2045  * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
2046  * vma is the last one with address > vma->vm_end.  Have to extend vma.
2047  */
2048 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
2049 {
2050         int error;
2051
2052         if (!(vma->vm_flags & VM_GROWSUP))
2053                 return -EFAULT;
2054
2055         /*
2056          * We must make sure the anon_vma is allocated
2057          * so that the anon_vma locking is not a noop.
2058          */
2059         if (unlikely(anon_vma_prepare(vma)))
2060                 return -ENOMEM;
2061         vma_lock_anon_vma(vma);
2062
2063         /*
2064          * vma->vm_start/vm_end cannot change under us because the caller
2065          * is required to hold the mmap_sem in read mode.  We need the
2066          * anon_vma lock to serialize against concurrent expand_stacks.
2067          * Also guard against wrapping around to address 0.
2068          */
2069         if (address < PAGE_ALIGN(address+4))
2070                 address = PAGE_ALIGN(address+4);
2071         else {
2072                 vma_unlock_anon_vma(vma);
2073                 return -ENOMEM;
2074         }
2075         error = 0;
2076
2077         /* Somebody else might have raced and expanded it already */
2078         if (address > vma->vm_end) {
2079                 unsigned long size, grow;
2080
2081                 size = address - vma->vm_start;
2082                 grow = (address - vma->vm_end) >> PAGE_SHIFT;
2083
2084                 error = -ENOMEM;
2085                 if (vma->vm_pgoff + (size >> PAGE_SHIFT) >= vma->vm_pgoff) {
2086                         error = acct_stack_growth(vma, size, grow);
2087                         if (!error) {
2088                                 /*
2089                                  * vma_gap_update() doesn't support concurrent
2090                                  * updates, but we only hold a shared mmap_sem
2091                                  * lock here, so we need to protect against
2092                                  * concurrent vma expansions.
2093                                  * vma_lock_anon_vma() doesn't help here, as
2094                                  * we don't guarantee that all growable vmas
2095                                  * in a mm share the same root anon vma.
2096                                  * So, we reuse mm->page_table_lock to guard
2097                                  * against concurrent vma expansions.
2098                                  */
2099                                 spin_lock(&vma->vm_mm->page_table_lock);
2100                                 anon_vma_interval_tree_pre_update_vma(vma);
2101                                 vma->vm_end = address;
2102                                 anon_vma_interval_tree_post_update_vma(vma);
2103                                 if (vma->vm_next)
2104                                         vma_gap_update(vma->vm_next);
2105                                 else
2106                                         vma->vm_mm->highest_vm_end = address;
2107                                 spin_unlock(&vma->vm_mm->page_table_lock);
2108
2109                                 perf_event_mmap(vma);
2110                         }
2111                 }
2112         }
2113         vma_unlock_anon_vma(vma);
2114         khugepaged_enter_vma_merge(vma);
2115         validate_mm(vma->vm_mm);
2116         return error;
2117 }
2118 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
2119
2120 /*
2121  * vma is the first one with address < vma->vm_start.  Have to extend vma.
2122  */
2123 int expand_downwards(struct vm_area_struct *vma,
2124                                    unsigned long address)
2125 {
2126         int error;
2127
2128         /*
2129          * We must make sure the anon_vma is allocated
2130          * so that the anon_vma locking is not a noop.
2131          */
2132         if (unlikely(anon_vma_prepare(vma)))
2133                 return -ENOMEM;
2134
2135         address &= PAGE_MASK;
2136         error = security_mmap_addr(address);
2137         if (error)
2138                 return error;
2139
2140         vma_lock_anon_vma(vma);
2141
2142         /*
2143          * vma->vm_start/vm_end cannot change under us because the caller
2144          * is required to hold the mmap_sem in read mode.  We need the
2145          * anon_vma lock to serialize against concurrent expand_stacks.
2146          */
2147
2148         /* Somebody else might have raced and expanded it already */
2149         if (address < vma->vm_start) {
2150                 unsigned long size, grow;
2151
2152                 size = vma->vm_end - address;
2153                 grow = (vma->vm_start - address) >> PAGE_SHIFT;
2154
2155                 error = -ENOMEM;
2156                 if (grow <= vma->vm_pgoff) {
2157                         error = acct_stack_growth(vma, size, grow);
2158                         if (!error) {
2159                                 /*
2160                                  * vma_gap_update() doesn't support concurrent
2161                                  * updates, but we only hold a shared mmap_sem
2162                                  * lock here, so we need to protect against
2163                                  * concurrent vma expansions.
2164                                  * vma_lock_anon_vma() doesn't help here, as
2165                                  * we don't guarantee that all growable vmas
2166                                  * in a mm share the same root anon vma.
2167                                  * So, we reuse mm->page_table_lock to guard
2168                                  * against concurrent vma expansions.
2169                                  */
2170                                 spin_lock(&vma->vm_mm->page_table_lock);
2171                                 anon_vma_interval_tree_pre_update_vma(vma);
2172                                 vma->vm_start = address;
2173                                 vma->vm_pgoff -= grow;
2174                                 anon_vma_interval_tree_post_update_vma(vma);
2175                                 vma_gap_update(vma);
2176                                 spin_unlock(&vma->vm_mm->page_table_lock);
2177
2178                                 perf_event_mmap(vma);
2179                         }
2180                 }
2181         }
2182         vma_unlock_anon_vma(vma);
2183         khugepaged_enter_vma_merge(vma);
2184         validate_mm(vma->vm_mm);
2185         return error;
2186 }
2187
2188 #ifdef CONFIG_STACK_GROWSUP
2189 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2190 {
2191         return expand_upwards(vma, address);
2192 }
2193
2194 struct vm_area_struct *
2195 find_extend_vma(struct mm_struct *mm, unsigned long addr)
2196 {
2197         struct vm_area_struct *vma, *prev;
2198
2199         addr &= PAGE_MASK;
2200         vma = find_vma_prev(mm, addr, &prev);
2201         if (vma && (vma->vm_start <= addr))
2202                 return vma;
2203         if (!prev || expand_stack(prev, addr))
2204                 return NULL;
2205         if (prev->vm_flags & VM_LOCKED)
2206                 __mlock_vma_pages_range(prev, addr, prev->vm_end, NULL);
2207         return prev;
2208 }
2209 #else
2210 int expand_stack(struct vm_area_struct *vma, unsigned long address)
2211 {
2212         return expand_downwards(vma, address);
2213 }
2214
2215 struct vm_area_struct *
2216 find_extend_vma(struct mm_struct * mm, unsigned long addr)
2217 {
2218         struct vm_area_struct * vma;
2219         unsigned long start;
2220
2221         addr &= PAGE_MASK;
2222         vma = find_vma(mm,addr);
2223         if (!vma)
2224                 return NULL;
2225         if (vma->vm_start <= addr)
2226                 return vma;
2227         if (!(vma->vm_flags & VM_GROWSDOWN))
2228                 return NULL;
2229         start = vma->vm_start;
2230         if (expand_stack(vma, addr))
2231                 return NULL;
2232         if (vma->vm_flags & VM_LOCKED)
2233                 __mlock_vma_pages_range(vma, addr, start, NULL);
2234         return vma;
2235 }
2236 #endif
2237
2238 /*
2239  * Ok - we have the memory areas we should free on the vma list,
2240  * so release them, and do the vma updates.
2241  *
2242  * Called with the mm semaphore held.
2243  */
2244 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
2245 {
2246         unsigned long nr_accounted = 0;
2247
2248         /* Update high watermark before we lower total_vm */
2249         update_hiwater_vm(mm);
2250         do {
2251                 long nrpages = vma_pages(vma);
2252
2253                 if (vma->vm_flags & VM_ACCOUNT)
2254                         nr_accounted += nrpages;
2255                 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
2256                 vma = remove_vma(vma);
2257         } while (vma);
2258         vm_unacct_memory(nr_accounted);
2259         validate_mm(mm);
2260 }
2261
2262 /*
2263  * Get rid of page table information in the indicated region.
2264  *
2265  * Called with the mm semaphore held.
2266  */
2267 static void unmap_region(struct mm_struct *mm,
2268                 struct vm_area_struct *vma, struct vm_area_struct *prev,
2269                 unsigned long start, unsigned long end)
2270 {
2271         struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
2272         struct mmu_gather tlb;
2273
2274         lru_add_drain();
2275         tlb_gather_mmu(&tlb, mm, 0);
2276         update_hiwater_rss(mm);
2277         unmap_vmas(&tlb, vma, start, end);
2278         free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
2279                                  next ? next->vm_start : 0);
2280         tlb_finish_mmu(&tlb, start, end);
2281 }
2282
2283 /*
2284  * Create a list of vma's touched by the unmap, removing them from the mm's
2285  * vma list as we go..
2286  */
2287 static void
2288 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
2289         struct vm_area_struct *prev, unsigned long end)
2290 {
2291         struct vm_area_struct **insertion_point;
2292         struct vm_area_struct *tail_vma = NULL;
2293         unsigned long addr;
2294
2295         insertion_point = (prev ? &prev->vm_next : &mm->mmap);
2296         vma->vm_prev = NULL;
2297         do {
2298                 vma_rb_erase(vma, &mm->mm_rb);
2299                 mm->map_count--;
2300                 tail_vma = vma;
2301                 vma = vma->vm_next;
2302         } while (vma && vma->vm_start < end);
2303         *insertion_point = vma;
2304         if (vma) {
2305                 vma->vm_prev = prev;
2306                 vma_gap_update(vma);
2307         } else
2308                 mm->highest_vm_end = prev ? prev->vm_end : 0;
2309         tail_vma->vm_next = NULL;
2310         if (mm->unmap_area == arch_unmap_area)
2311                 addr = prev ? prev->vm_end : mm->mmap_base;
2312         else
2313                 addr = vma ?  vma->vm_start : mm->mmap_base;
2314         mm->unmap_area(mm, addr);
2315         mm->mmap_cache = NULL;          /* Kill the cache. */
2316 }
2317
2318 /*
2319  * __split_vma() bypasses sysctl_max_map_count checking.  We use this on the
2320  * munmap path where it doesn't make sense to fail.
2321  */
2322 static int __split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
2323               unsigned long addr, int new_below)
2324 {
2325         struct mempolicy *pol;
2326         struct vm_area_struct *new;
2327         int err = -ENOMEM;
2328
2329         if (is_vm_hugetlb_page(vma) && (addr &
2330                                         ~(huge_page_mask(hstate_vma(vma)))))
2331                 return -EINVAL;
2332
2333         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2334         if (!new)
2335                 goto out_err;
2336
2337         /* most fields are the same, copy all, and then fixup */
2338         *new = *vma;
2339
2340         INIT_LIST_HEAD(&new->anon_vma_chain);
2341
2342         if (new_below)
2343                 new->vm_end = addr;
2344         else {
2345                 new->vm_start = addr;
2346                 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
2347         }
2348
2349         pol = mpol_dup(vma_policy(vma));
2350         if (IS_ERR(pol)) {
2351                 err = PTR_ERR(pol);
2352                 goto out_free_vma;
2353         }
2354         vma_set_policy(new, pol);
2355
2356         if (anon_vma_clone(new, vma))
2357                 goto out_free_mpol;
2358
2359         if (new->vm_file)
2360                 get_file(new->vm_file);
2361
2362         if (new->vm_ops && new->vm_ops->open)
2363                 new->vm_ops->open(new);
2364
2365         if (new_below)
2366                 err = vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
2367                         ((addr - new->vm_start) >> PAGE_SHIFT), new);
2368         else
2369                 err = vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
2370
2371         /* Success. */
2372         if (!err)
2373                 return 0;
2374
2375         /* Clean everything up if vma_adjust failed. */
2376         if (new->vm_ops && new->vm_ops->close)
2377                 new->vm_ops->close(new);
2378         if (new->vm_file)
2379                 fput(new->vm_file);
2380         unlink_anon_vmas(new);
2381  out_free_mpol:
2382         mpol_put(pol);
2383  out_free_vma:
2384         kmem_cache_free(vm_area_cachep, new);
2385  out_err:
2386         return err;
2387 }
2388
2389 /*
2390  * Split a vma into two pieces at address 'addr', a new vma is allocated
2391  * either for the first part or the tail.
2392  */
2393 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
2394               unsigned long addr, int new_below)
2395 {
2396         if (mm->map_count >= sysctl_max_map_count)
2397                 return -ENOMEM;
2398
2399         return __split_vma(mm, vma, addr, new_below);
2400 }
2401
2402 /* Munmap is split into 2 main parts -- this part which finds
2403  * what needs doing, and the areas themselves, which do the
2404  * work.  This now handles partial unmappings.
2405  * Jeremy Fitzhardinge <jeremy@goop.org>
2406  */
2407 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
2408 {
2409         unsigned long end;
2410         struct vm_area_struct *vma, *prev, *last;
2411
2412         if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
2413                 return -EINVAL;
2414
2415         if ((len = PAGE_ALIGN(len)) == 0)
2416                 return -EINVAL;
2417
2418         /* Find the first overlapping VMA */
2419         vma = find_vma(mm, start);
2420         if (!vma)
2421                 return 0;
2422         prev = vma->vm_prev;
2423         /* we have  start < vma->vm_end  */
2424
2425         /* if it doesn't overlap, we have nothing.. */
2426         end = start + len;
2427         if (vma->vm_start >= end)
2428                 return 0;
2429
2430         /*
2431          * If we need to split any vma, do it now to save pain later.
2432          *
2433          * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
2434          * unmapped vm_area_struct will remain in use: so lower split_vma
2435          * places tmp vma above, and higher split_vma places tmp vma below.
2436          */
2437         if (start > vma->vm_start) {
2438                 int error;
2439
2440                 /*
2441                  * Make sure that map_count on return from munmap() will
2442                  * not exceed its limit; but let map_count go just above
2443                  * its limit temporarily, to help free resources as expected.
2444                  */
2445                 if (end < vma->vm_end && mm->map_count >= sysctl_max_map_count)
2446                         return -ENOMEM;
2447
2448                 error = __split_vma(mm, vma, start, 0);
2449                 if (error)
2450                         return error;
2451                 prev = vma;
2452         }
2453
2454         /* Does it split the last one? */
2455         last = find_vma(mm, end);
2456         if (last && end > last->vm_start) {
2457                 int error = __split_vma(mm, last, end, 1);
2458                 if (error)
2459                         return error;
2460         }
2461         vma = prev? prev->vm_next: mm->mmap;
2462
2463         /*
2464          * unlock any mlock()ed ranges before detaching vmas
2465          */
2466         if (mm->locked_vm) {
2467                 struct vm_area_struct *tmp = vma;
2468                 while (tmp && tmp->vm_start < end) {
2469                         if (tmp->vm_flags & VM_LOCKED) {
2470                                 mm->locked_vm -= vma_pages(tmp);
2471                                 munlock_vma_pages_all(tmp);
2472                         }
2473                         tmp = tmp->vm_next;
2474                 }
2475         }
2476
2477         /*
2478          * Remove the vma's, and unmap the actual pages
2479          */
2480         detach_vmas_to_be_unmapped(mm, vma, prev, end);
2481         unmap_region(mm, vma, prev, start, end);
2482
2483         /* Fix up all other VM information */
2484         remove_vma_list(mm, vma);
2485
2486         return 0;
2487 }
2488
2489 int vm_munmap(unsigned long start, size_t len)
2490 {
2491         int ret;
2492         struct mm_struct *mm = current->mm;
2493
2494         down_write(&mm->mmap_sem);
2495         ret = do_munmap(mm, start, len);
2496         up_write(&mm->mmap_sem);
2497         return ret;
2498 }
2499 EXPORT_SYMBOL(vm_munmap);
2500
2501 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
2502 {
2503         profile_munmap(addr);
2504         return vm_munmap(addr, len);
2505 }
2506
2507 static inline void verify_mm_writelocked(struct mm_struct *mm)
2508 {
2509 #ifdef CONFIG_DEBUG_VM
2510         if (unlikely(down_read_trylock(&mm->mmap_sem))) {
2511                 WARN_ON(1);
2512                 up_read(&mm->mmap_sem);
2513         }
2514 #endif
2515 }
2516
2517 /*
2518  *  this is really a simplified "do_mmap".  it only handles
2519  *  anonymous maps.  eventually we may be able to do some
2520  *  brk-specific accounting here.
2521  */
2522 static unsigned long do_brk(unsigned long addr, unsigned long len)
2523 {
2524         struct mm_struct * mm = current->mm;
2525         struct vm_area_struct * vma, * prev;
2526         unsigned long flags;
2527         struct rb_node ** rb_link, * rb_parent;
2528         pgoff_t pgoff = addr >> PAGE_SHIFT;
2529         int error;
2530
2531         len = PAGE_ALIGN(len);
2532         if (!len)
2533                 return addr;
2534
2535         flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
2536
2537         error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
2538         if (error & ~PAGE_MASK)
2539                 return error;
2540
2541         /*
2542          * mlock MCL_FUTURE?
2543          */
2544         if (mm->def_flags & VM_LOCKED) {
2545                 unsigned long locked, lock_limit;
2546                 locked = len >> PAGE_SHIFT;
2547                 locked += mm->locked_vm;
2548                 lock_limit = rlimit(RLIMIT_MEMLOCK);
2549                 lock_limit >>= PAGE_SHIFT;
2550                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
2551                         return -EAGAIN;
2552         }
2553
2554         /*
2555          * mm->mmap_sem is required to protect against another thread
2556          * changing the mappings in case we sleep.
2557          */
2558         verify_mm_writelocked(mm);
2559
2560         /*
2561          * Clear old maps.  this also does some error checking for us
2562          */
2563  munmap_back:
2564         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent)) {
2565                 if (do_munmap(mm, addr, len))
2566                         return -ENOMEM;
2567                 goto munmap_back;
2568         }
2569
2570         /* Check against address space limits *after* clearing old maps... */
2571         if (!may_expand_vm(mm, len >> PAGE_SHIFT))
2572                 return -ENOMEM;
2573
2574         if (mm->map_count > sysctl_max_map_count)
2575                 return -ENOMEM;
2576
2577         if (security_vm_enough_memory_mm(mm, len >> PAGE_SHIFT))
2578                 return -ENOMEM;
2579
2580         /* Can we just expand an old private anonymous mapping? */
2581         vma = vma_merge(mm, prev, addr, addr + len, flags,
2582                                         NULL, NULL, pgoff, NULL);
2583         if (vma)
2584                 goto out;
2585
2586         /*
2587          * create a vma struct for an anonymous mapping
2588          */
2589         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2590         if (!vma) {
2591                 vm_unacct_memory(len >> PAGE_SHIFT);
2592                 return -ENOMEM;
2593         }
2594
2595         INIT_LIST_HEAD(&vma->anon_vma_chain);
2596         vma->vm_mm = mm;
2597         vma->vm_start = addr;
2598         vma->vm_end = addr + len;
2599         vma->vm_pgoff = pgoff;
2600         vma->vm_flags = flags;
2601         vma->vm_page_prot = vm_get_page_prot(flags);
2602         vma_link(mm, vma, prev, rb_link, rb_parent);
2603 out:
2604         perf_event_mmap(vma);
2605         mm->total_vm += len >> PAGE_SHIFT;
2606         if (flags & VM_LOCKED)
2607                 mm->locked_vm += (len >> PAGE_SHIFT);
2608         return addr;
2609 }
2610
2611 unsigned long vm_brk(unsigned long addr, unsigned long len)
2612 {
2613         struct mm_struct *mm = current->mm;
2614         unsigned long ret;
2615         bool populate;
2616
2617         down_write(&mm->mmap_sem);
2618         ret = do_brk(addr, len);
2619         populate = ((mm->def_flags & VM_LOCKED) != 0);
2620         up_write(&mm->mmap_sem);
2621         if (populate)
2622                 mm_populate(addr, len);
2623         return ret;
2624 }
2625 EXPORT_SYMBOL(vm_brk);
2626
2627 /* Release all mmaps. */
2628 void exit_mmap(struct mm_struct *mm)
2629 {
2630         struct mmu_gather tlb;
2631         struct vm_area_struct *vma;
2632         unsigned long nr_accounted = 0;
2633
2634         /* mm's last user has gone, and its about to be pulled down */
2635         mmu_notifier_release(mm);
2636
2637         if (mm->locked_vm) {
2638                 vma = mm->mmap;
2639                 while (vma) {
2640                         if (vma->vm_flags & VM_LOCKED)
2641                                 munlock_vma_pages_all(vma);
2642                         vma = vma->vm_next;
2643                 }
2644         }
2645
2646         arch_exit_mmap(mm);
2647
2648         vma = mm->mmap;
2649         if (!vma)       /* Can happen if dup_mmap() received an OOM */
2650                 return;
2651
2652         lru_add_drain();
2653         flush_cache_mm(mm);
2654         tlb_gather_mmu(&tlb, mm, 1);
2655         /* update_hiwater_rss(mm) here? but nobody should be looking */
2656         /* Use -1 here to ensure all VMAs in the mm are unmapped */
2657         unmap_vmas(&tlb, vma, 0, -1);
2658
2659         free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
2660         tlb_finish_mmu(&tlb, 0, -1);
2661
2662         /*
2663          * Walk the list again, actually closing and freeing it,
2664          * with preemption enabled, without holding any MM locks.
2665          */
2666         while (vma) {
2667                 if (vma->vm_flags & VM_ACCOUNT)
2668                         nr_accounted += vma_pages(vma);
2669                 vma = remove_vma(vma);
2670         }
2671         vm_unacct_memory(nr_accounted);
2672
2673         WARN_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
2674 }
2675
2676 /* Insert vm structure into process list sorted by address
2677  * and into the inode's i_mmap tree.  If vm_file is non-NULL
2678  * then i_mmap_mutex is taken here.
2679  */
2680 int insert_vm_struct(struct mm_struct *mm, struct vm_area_struct *vma)
2681 {
2682         struct vm_area_struct *prev;
2683         struct rb_node **rb_link, *rb_parent;
2684
2685         /*
2686          * The vm_pgoff of a purely anonymous vma should be irrelevant
2687          * until its first write fault, when page's anon_vma and index
2688          * are set.  But now set the vm_pgoff it will almost certainly
2689          * end up with (unless mremap moves it elsewhere before that
2690          * first wfault), so /proc/pid/maps tells a consistent story.
2691          *
2692          * By setting it to reflect the virtual start address of the
2693          * vma, merges and splits can happen in a seamless way, just
2694          * using the existing file pgoff checks and manipulations.
2695          * Similarly in do_mmap_pgoff and in do_brk.
2696          */
2697         if (!vma->vm_file) {
2698                 BUG_ON(vma->anon_vma);
2699                 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2700         }
2701         if (find_vma_links(mm, vma->vm_start, vma->vm_end,
2702                            &prev, &rb_link, &rb_parent))
2703                 return -ENOMEM;
2704         if ((vma->vm_flags & VM_ACCOUNT) &&
2705              security_vm_enough_memory_mm(mm, vma_pages(vma)))
2706                 return -ENOMEM;
2707
2708         vma_link(mm, vma, prev, rb_link, rb_parent);
2709         return 0;
2710 }
2711
2712 /*
2713  * Copy the vma structure to a new location in the same mm,
2714  * prior to moving page table entries, to effect an mremap move.
2715  */
2716 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2717         unsigned long addr, unsigned long len, pgoff_t pgoff,
2718         bool *need_rmap_locks)
2719 {
2720         struct vm_area_struct *vma = *vmap;
2721         unsigned long vma_start = vma->vm_start;
2722         struct mm_struct *mm = vma->vm_mm;
2723         struct vm_area_struct *new_vma, *prev;
2724         struct rb_node **rb_link, *rb_parent;
2725         struct mempolicy *pol;
2726         bool faulted_in_anon_vma = true;
2727
2728         /*
2729          * If anonymous vma has not yet been faulted, update new pgoff
2730          * to match new location, to increase its chance of merging.
2731          */
2732         if (unlikely(!vma->vm_file && !vma->anon_vma)) {
2733                 pgoff = addr >> PAGE_SHIFT;
2734                 faulted_in_anon_vma = false;
2735         }
2736
2737         if (find_vma_links(mm, addr, addr + len, &prev, &rb_link, &rb_parent))
2738                 return NULL;    /* should never get here */
2739         new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2740                         vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2741         if (new_vma) {
2742                 /*
2743                  * Source vma may have been merged into new_vma
2744                  */
2745                 if (unlikely(vma_start >= new_vma->vm_start &&
2746                              vma_start < new_vma->vm_end)) {
2747                         /*
2748                          * The only way we can get a vma_merge with
2749                          * self during an mremap is if the vma hasn't
2750                          * been faulted in yet and we were allowed to
2751                          * reset the dst vma->vm_pgoff to the
2752                          * destination address of the mremap to allow
2753                          * the merge to happen. mremap must change the
2754                          * vm_pgoff linearity between src and dst vmas
2755                          * (in turn preventing a vma_merge) to be
2756                          * safe. It is only safe to keep the vm_pgoff
2757                          * linear if there are no pages mapped yet.
2758                          */
2759                         VM_BUG_ON(faulted_in_anon_vma);
2760                         *vmap = vma = new_vma;
2761                 }
2762                 *need_rmap_locks = (new_vma->vm_pgoff <= vma->vm_pgoff);
2763         } else {
2764                 new_vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
2765                 if (new_vma) {
2766                         *new_vma = *vma;
2767                         new_vma->vm_start = addr;
2768                         new_vma->vm_end = addr + len;
2769                         new_vma->vm_pgoff = pgoff;
2770                         pol = mpol_dup(vma_policy(vma));
2771                         if (IS_ERR(pol))
2772                                 goto out_free_vma;
2773                         vma_set_policy(new_vma, pol);
2774                         INIT_LIST_HEAD(&new_vma->anon_vma_chain);
2775                         if (anon_vma_clone(new_vma, vma))
2776                                 goto out_free_mempol;
2777                         if (new_vma->vm_file)
2778                                 get_file(new_vma->vm_file);
2779                         if (new_vma->vm_ops && new_vma->vm_ops->open)
2780                                 new_vma->vm_ops->open(new_vma);
2781                         vma_link(mm, new_vma, prev, rb_link, rb_parent);
2782                         *need_rmap_locks = false;
2783                 }
2784         }
2785         return new_vma;
2786
2787  out_free_mempol:
2788         mpol_put(pol);
2789  out_free_vma:
2790         kmem_cache_free(vm_area_cachep, new_vma);
2791         return NULL;
2792 }
2793
2794 /*
2795  * Return true if the calling process may expand its vm space by the passed
2796  * number of pages
2797  */
2798 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2799 {
2800         unsigned long cur = mm->total_vm;       /* pages */
2801         unsigned long lim;
2802
2803         lim = rlimit(RLIMIT_AS) >> PAGE_SHIFT;
2804
2805         if (cur + npages > lim)
2806                 return 0;
2807         return 1;
2808 }
2809
2810
2811 static int special_mapping_fault(struct vm_area_struct *vma,
2812                                 struct vm_fault *vmf)
2813 {
2814         pgoff_t pgoff;
2815         struct page **pages;
2816
2817         /*
2818          * special mappings have no vm_file, and in that case, the mm
2819          * uses vm_pgoff internally. So we have to subtract it from here.
2820          * We are allowed to do this because we are the mm; do not copy
2821          * this code into drivers!
2822          */
2823         pgoff = vmf->pgoff - vma->vm_pgoff;
2824
2825         for (pages = vma->vm_private_data; pgoff && *pages; ++pages)
2826                 pgoff--;
2827
2828         if (*pages) {
2829                 struct page *page = *pages;
2830                 get_page(page);
2831                 vmf->page = page;
2832                 return 0;
2833         }
2834
2835         return VM_FAULT_SIGBUS;
2836 }
2837
2838 /*
2839  * Having a close hook prevents vma merging regardless of flags.
2840  */
2841 static void special_mapping_close(struct vm_area_struct *vma)
2842 {
2843 }
2844
2845 static const struct vm_operations_struct special_mapping_vmops = {
2846         .close = special_mapping_close,
2847         .fault = special_mapping_fault,
2848 };
2849
2850 /*
2851  * Called with mm->mmap_sem held for writing.
2852  * Insert a new vma covering the given region, with the given flags.
2853  * Its pages are supplied by the given array of struct page *.
2854  * The array can be shorter than len >> PAGE_SHIFT if it's null-terminated.
2855  * The region past the last page supplied will always produce SIGBUS.
2856  * The array pointer and the pages it points to are assumed to stay alive
2857  * for as long as this mapping might exist.
2858  */
2859 int install_special_mapping(struct mm_struct *mm,
2860                             unsigned long addr, unsigned long len,
2861                             unsigned long vm_flags, struct page **pages)
2862 {
2863         int ret;
2864         struct vm_area_struct *vma;
2865
2866         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2867         if (unlikely(vma == NULL))
2868                 return -ENOMEM;
2869
2870         INIT_LIST_HEAD(&vma->anon_vma_chain);
2871         vma->vm_mm = mm;
2872         vma->vm_start = addr;
2873         vma->vm_end = addr + len;
2874
2875         vma->vm_flags = vm_flags | mm->def_flags | VM_DONTEXPAND;
2876         vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
2877
2878         vma->vm_ops = &special_mapping_vmops;
2879         vma->vm_private_data = pages;
2880
2881         ret = insert_vm_struct(mm, vma);
2882         if (ret)
2883                 goto out;
2884
2885         mm->total_vm += len >> PAGE_SHIFT;
2886
2887         perf_event_mmap(vma);
2888
2889         return 0;
2890
2891 out:
2892         kmem_cache_free(vm_area_cachep, vma);
2893         return ret;
2894 }
2895
2896 static DEFINE_MUTEX(mm_all_locks_mutex);
2897
2898 static void vm_lock_anon_vma(struct mm_struct *mm, struct anon_vma *anon_vma)
2899 {
2900         if (!test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
2901                 /*
2902                  * The LSB of head.next can't change from under us
2903                  * because we hold the mm_all_locks_mutex.
2904                  */
2905                 down_write_nest_lock(&anon_vma->root->rwsem, &mm->mmap_sem);
2906                 /*
2907                  * We can safely modify head.next after taking the
2908                  * anon_vma->root->rwsem. If some other vma in this mm shares
2909                  * the same anon_vma we won't take it again.
2910                  *
2911                  * No need of atomic instructions here, head.next
2912                  * can't change from under us thanks to the
2913                  * anon_vma->root->rwsem.
2914                  */
2915                 if (__test_and_set_bit(0, (unsigned long *)
2916                                        &anon_vma->root->rb_root.rb_node))
2917                         BUG();
2918         }
2919 }
2920
2921 static void vm_lock_mapping(struct mm_struct *mm, struct address_space *mapping)
2922 {
2923         if (!test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
2924                 /*
2925                  * AS_MM_ALL_LOCKS can't change from under us because
2926                  * we hold the mm_all_locks_mutex.
2927                  *
2928                  * Operations on ->flags have to be atomic because
2929                  * even if AS_MM_ALL_LOCKS is stable thanks to the
2930                  * mm_all_locks_mutex, there may be other cpus
2931                  * changing other bitflags in parallel to us.
2932                  */
2933                 if (test_and_set_bit(AS_MM_ALL_LOCKS, &mapping->flags))
2934                         BUG();
2935                 mutex_lock_nest_lock(&mapping->i_mmap_mutex, &mm->mmap_sem);
2936         }
2937 }
2938
2939 /*
2940  * This operation locks against the VM for all pte/vma/mm related
2941  * operations that could ever happen on a certain mm. This includes
2942  * vmtruncate, try_to_unmap, and all page faults.
2943  *
2944  * The caller must take the mmap_sem in write mode before calling
2945  * mm_take_all_locks(). The caller isn't allowed to release the
2946  * mmap_sem until mm_drop_all_locks() returns.
2947  *
2948  * mmap_sem in write mode is required in order to block all operations
2949  * that could modify pagetables and free pages without need of
2950  * altering the vma layout (for example populate_range() with
2951  * nonlinear vmas). It's also needed in write mode to avoid new
2952  * anon_vmas to be associated with existing vmas.
2953  *
2954  * A single task can't take more than one mm_take_all_locks() in a row
2955  * or it would deadlock.
2956  *
2957  * The LSB in anon_vma->rb_root.rb_node and the AS_MM_ALL_LOCKS bitflag in
2958  * mapping->flags avoid to take the same lock twice, if more than one
2959  * vma in this mm is backed by the same anon_vma or address_space.
2960  *
2961  * We can take all the locks in random order because the VM code
2962  * taking i_mmap_mutex or anon_vma->rwsem outside the mmap_sem never
2963  * takes more than one of them in a row. Secondly we're protected
2964  * against a concurrent mm_take_all_locks() by the mm_all_locks_mutex.
2965  *
2966  * mm_take_all_locks() and mm_drop_all_locks are expensive operations
2967  * that may have to take thousand of locks.
2968  *
2969  * mm_take_all_locks() can fail if it's interrupted by signals.
2970  */
2971 int mm_take_all_locks(struct mm_struct *mm)
2972 {
2973         struct vm_area_struct *vma;
2974         struct anon_vma_chain *avc;
2975
2976         BUG_ON(down_read_trylock(&mm->mmap_sem));
2977
2978         mutex_lock(&mm_all_locks_mutex);
2979
2980         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2981                 if (signal_pending(current))
2982                         goto out_unlock;
2983                 if (vma->vm_file && vma->vm_file->f_mapping)
2984                         vm_lock_mapping(mm, vma->vm_file->f_mapping);
2985         }
2986
2987         for (vma = mm->mmap; vma; vma = vma->vm_next) {
2988                 if (signal_pending(current))
2989                         goto out_unlock;
2990                 if (vma->anon_vma)
2991                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
2992                                 vm_lock_anon_vma(mm, avc->anon_vma);
2993         }
2994
2995         return 0;
2996
2997 out_unlock:
2998         mm_drop_all_locks(mm);
2999         return -EINTR;
3000 }
3001
3002 static void vm_unlock_anon_vma(struct anon_vma *anon_vma)
3003 {
3004         if (test_bit(0, (unsigned long *) &anon_vma->root->rb_root.rb_node)) {
3005                 /*
3006                  * The LSB of head.next can't change to 0 from under
3007                  * us because we hold the mm_all_locks_mutex.
3008                  *
3009                  * We must however clear the bitflag before unlocking
3010                  * the vma so the users using the anon_vma->rb_root will
3011                  * never see our bitflag.
3012                  *
3013                  * No need of atomic instructions here, head.next
3014                  * can't change from under us until we release the
3015                  * anon_vma->root->rwsem.
3016                  */
3017                 if (!__test_and_clear_bit(0, (unsigned long *)
3018                                           &anon_vma->root->rb_root.rb_node))
3019                         BUG();
3020                 anon_vma_unlock(anon_vma);
3021         }
3022 }
3023
3024 static void vm_unlock_mapping(struct address_space *mapping)
3025 {
3026         if (test_bit(AS_MM_ALL_LOCKS, &mapping->flags)) {
3027                 /*
3028                  * AS_MM_ALL_LOCKS can't change to 0 from under us
3029                  * because we hold the mm_all_locks_mutex.
3030                  */
3031                 mutex_unlock(&mapping->i_mmap_mutex);
3032                 if (!test_and_clear_bit(AS_MM_ALL_LOCKS,
3033                                         &mapping->flags))
3034                         BUG();
3035         }
3036 }
3037
3038 /*
3039  * The mmap_sem cannot be released by the caller until
3040  * mm_drop_all_locks() returns.
3041  */
3042 void mm_drop_all_locks(struct mm_struct *mm)
3043 {
3044         struct vm_area_struct *vma;
3045         struct anon_vma_chain *avc;
3046
3047         BUG_ON(down_read_trylock(&mm->mmap_sem));
3048         BUG_ON(!mutex_is_locked(&mm_all_locks_mutex));
3049
3050         for (vma = mm->mmap; vma; vma = vma->vm_next) {
3051                 if (vma->anon_vma)
3052                         list_for_each_entry(avc, &vma->anon_vma_chain, same_vma)
3053                                 vm_unlock_anon_vma(avc->anon_vma);
3054                 if (vma->vm_file && vma->vm_file->f_mapping)
3055                         vm_unlock_mapping(vma->vm_file->f_mapping);
3056         }
3057
3058         mutex_unlock(&mm_all_locks_mutex);
3059 }
3060
3061 /*
3062  * initialise the VMA slab
3063  */
3064 void __init mmap_init(void)
3065 {
3066         int ret;
3067
3068         ret = percpu_counter_init(&vm_committed_as, 0);
3069         VM_BUG_ON(ret);
3070 }