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