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